Information Fusion 115 (2025) 102755

Contents lists available at ScienceDirect

Information Fusion
journal homepage: www.elsevier.com/locate/inffus

Natural language processing in finance: A survey

Kelvin Du a , Yazhi Zhao b , Rui Mao a , Frank Xing c , Erik Cambria a ,∗
a
College of Computing and Data Science, Nanyang Technological University, 50 Nanyang Avenue, 639798, Singapore
b
Visa Inc, 71 Robinson Road, Singapore, 068895, Singapore
c
School of Computing, National University of Singapore, 15 Computing Drive, Singapore, 117418, Singapore

ARTICLE INFO ABSTRACT

Keywords: This survey presents an in-depth review of the transformative role of Natural Language Processing (NLP)
Natural language processing in finance, highlighting its impact on ten major financial applications: (1) financial sentiment analysis, (2)
Finance financial narrative processing, (3) financial forecasting, (4) portfolio management, (5) question answering,
Financial sentiment analysis
virtual assistant and chatbot, (6) risk management, (7) regulatory compliance monitoring, (8) Environmental,
Financial narrative processing
Social, Governance (ESG) and sustainable finance, (9) explainable artificial intelligence (XAI) in finance and
Financial forecasting
Portfolio management
(10) NLP for digital assets. With the integration of vast amounts of unstructured financial data and advanced
Question answering NLP techniques, the study explores how NLP enables data-driven decision-making and innovation in the
Risk management financial sector, alongside the limitations and challenges. By providing a comprehensive analysis of NLP
Regulatory compliance applications combining both academic and industrial perspectives, this study postulates the future trends
ESG and sustainable finance and evolution of financial services. It introduces a unique review framework to understand the interaction of
Explainable AI financial data and NLP technologies systematically and outlines the key drivers, transformations, and emerging
Digital assets areas in this field. This survey targets researchers, practitioners, and professionals, aiming to close their
knowledge gap by highlighting the significance and future direction of NLP in enhancing financial services.

1. Introduction While prior reviews have often leaned towards NLP or finance,
our survey seeks to provide an all-encompassing review that connects
In the rapidly evolving landscape of finance, the integration of Nat- recent research across both dimensions. Crucially, our study involves
ural Language Processing (NLP) technologies has emerged as a trans- a detailed investigation of the latest advancements in NLP within the
formative force, unlocking new dimensions of data-driven decision- finance sector, providing insights from both the perspectives of NLP
making and innovation. This survey aims to explore the multifaceted methodologies and financial applications. Moreover, we systematically
applications of NLP techniques within the financial sector, highlighting analyze NLP in finance by considering financial textual data, NLP
its profound impact on areas such as financial sentiment analysis, techniques, and financial applications, along with emerging trends in
natural language-based financial forecasting, portfolio management, both NLP techniques, such as large language models (LLMs), as well as
financial narrative processing, question answering, virtual assistant and financial applications, including NLP for ESG and sustainable finance.
chatbot, risk management, regulatory compliance, ESG and sustainable This survey begins with an overview of NLP’s foundational concepts
finance, explainable AI in finance and NLP for digital assets from both and techniques, providing a solid groundwork for understanding its ap-
academic and industrial research perspective. As financial institutions plication in finance. It then delves into specific applications where NLP
increasingly rely on vast amounts of unstructured data, from corporate
has been successfully implemented, together with the financial data
releases to news articles and reports to social media and beyond,
adopted, illustrating the technology’s versatility and power. Further, it
the capability to efficiently process, understand, and act upon this
addresses the challenges and limitations of applying NLP in the finance
information has become a competitive necessity. In the field of NLP
sector, including issues related to data quality, privacy concerns, and
in finance, Fisher et al. [1] executed a thorough exploration, mainly
the need for highly specialized models. It also explores the future
concentrating on tasks related to classification and prediction. Chen
trajectory of NLP technologies in finance, considering advances in
et al. [2] presented a broad review of financial application areas,
AI and computational linguistics that promise to further enhance the
including three categories: Know Your Customer (KYC), Know Your
Product (KYP), and Satisfy Your Customer (SYC). sophistication and effectiveness of financial services.

∗ Corresponding author.
E-mail addresses: zidong001@e.ntu.edu.sg (K. Du), yazzhao@visa.com (Y. Zhao), rui.mao@ntu.edu.sg (R. Mao), xing@nus.edu.sg (F. Xing),
cambria@ntu.edu.sg (E. Cambria).

https://doi.org/10.1016/j.inffus.2024.102755
Received 27 June 2024; Received in revised form 9 October 2024; Accepted 18 October 2024
Available online 28 October 2024
1566-2535/© 2024 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
K. Du et al. Information Fusion 115 (2025) 102755

Fig. 1. NLP in finance: Financial textual data, NLP techniques, and financial applications.

Through this comprehensive exploration, we aim to provide aca- transformative effects on services, products, and operations. This
demics, industry professionals, and technology enthusiasts with a thor- includes insights into practical challenges and the transformative
ough understanding of the current state and future potential of NLP potential of these technologies.
in the financial domain. By showcasing the technology’s significant
The remainder of this survey is structured as follows: Section 2 de-
contributions and pondering its future directions, the survey aims to
scribes the literature review framework; Section 3 conducts an in-depth
foster a deeper appreciation for the role of NLP in shaping the next gen-
review of studies on NLP in finance by financial application areas;
eration of the financial sector. Specifically, this study aims to answer
Section 4 demonstrates the research findings of this review includ-
the following three groups of research questions:
ing trends, challenges, opportunities, and future directions; Section 5
1. What are the key drivers behind the increasing importance of provides concluding remarks.
NLP in the finance sector, and how do these drivers address the
2. Literature review framework
unique challenges and opportunities within this industry?
2. What emerging areas of application for NLP in finance are being
Fig. 1 presents our literature review framework, focusing specifi-
explored, and what potential do these areas hold for the future
cally on the financial textual data, NLP techniques, key players and
of financial services?
financial applications. We started with a review of a study on AI
3. What are the prevailing trends in NLP research within the fi-
in finance conducted by the International Monetary Fund (IMF) [3].
nance sector, and how do these trends reflect the evolving
We then zoomed into each potential application area, examining the
challenges and opportunities in the field?
relevant literature. We also summarized the data sources available
Our contributions can be outlined in the following three aspects: in finance. The financial data can be broadly categorized into three
types: corporate data, public news and announcements, and social
1. We presented a detailed literature review on NLP techniques media data. Corporate data includes data produced by corporations
in finance, bridging a significant gap by providing a definitive such as financial transactions, financial statements, customer feedback,
reference for researchers and practitioners. It encapsulates the disclosures, annual reports, filings, press releases, earnings calls, and
evolution and diverse applications of NLP across both academic conference calls etc. The public news and announcements category
research and industrial implementations. encompasses financial and economic news, government policies, and
2. We developed a comprehensive review framework for NLP in analyst reports, among others. Social media data consists of information
finance, which maps the relationship between data, NLP tech- from social media platforms, e.g., X, StockTwits, and various online
niques, and financial applications. We identified ten key finan- forums. Subsequently, we defined the scope of our study to include
cial application areas that are potentially revolutionized by NLP, retail banking, investment banking, investment management, govern-
and our framework provides a systematic overview to compre- ment institutions, exchanges and clearing houses, payment processors,
hending the intricacies of financial data, NLP techniques, and and insurance providers. We proposed key financial application areas
their real-world applications in the finance industry. significantly enhanced by adopting various NLP techniques in the
3. We analyzed key research contributions that have significantly finance sector. This includes financial sentiment analysis, financial fore-
influenced NLP’s application in the finance sector, highlighting casting, portfolio management, financial narrative processing, question

2
K. Du et al. Information Fusion 115 (2025) 102755

Table 1
Key words for literature search.
Research topic Key words
Financial sentiment analysis financial sentiment analysis
Financial narrative processing finance AND (summarization OR information extraction OR information retrieval OR financial narrative processing)
Financial forecasting (nlp OR text mining OR news OR tweet) AND (market prediction OR financial forecasting)
Portfolio management (nlp OR text mining OR news OR tweet) AND portfolio AND (selection OR optimization OR management)
Question answering, virtual assistant and chatbot (question answering OR virtual assistant OR chatbot) AND (finance OR financial services)
Risk management (nlp OR text mining OR news OR tweet) AND (risk management OR risk prediction OR credit risk OR fraud detection)
Regulatory compliance monitoring (nlp OR text mining) AND (regulatory compliance OR anti-money laundering)
ESG and sustainable finance (nlp OR text mining) AND (esg OR (sustainable finance) OR (environmental AND social AND governance))
Explainable artificial intelligence (XAI) in finance (nlp OR text mining) AND (explainable artificial intelligence OR explainable AI OR explainability) AND finance
NLP for digital assets (nlp OR text mining OR news OR tweet) AND (digital asset OR cryptocurrency OR non-fungible token OR blockchain)

answering, virtual assistant and chatbot, risk management, regulatory 3.1. Financial sentiment analysis
compliance monitoring, explainable AI in finance, ESG and sustainable
finance and NLP for digital assets. Additionally, we examine how these Financial Sentiment Analysis (FSA) is one of the widely adopted
applications are adopted across businesses in the finance sector in NLP techniques in the finance domain due to the advancements of
subsequent sections. sentiment analysis in the general domain [4–7]. Du et al. [8] has
Following PRISMA guidelines, we designed a three-step process, conducted a comprehensive survey on FSA which categorizes FSA into
namely identification, screening and inclusion, to select the literature two principal research tracks. The FSA techniques concentrate on iden-
for review. Firstly, leveraging industry experience, we defined key- tifying specific tasks such as document-level [9], paragraph-level [10],
words, as illustrated in Table 1, which closely align with the topics sentence-level [11] and aspect-level [12] sentiment analysis and aim
from the International Monetary Fund’s (IMF) study on AI in finance. to propose methodologies to improve the performance of different
Subsequently, we used these keywords to perform targeted searches on FSA tasks by using human-annotated datasets. The FSA applications
Web of Science and Google Scholar, prioritizing relevance as defined emphasize the application of financial sentiments, whether direct or
by the database. Our aim is to ensure the coverage of the range of indirect, in subsequent applications within financial markets, attracting
NLP tasks applicable to this field. We assessed each article by screening greater research attention.
through the abstracts to ascertain their relevance and contribution, Presented in Table 2, we explore the advancement of research in
which determined their inclusion in the review. We balanced the FSA techniques, tracing a path from the lexicon-based method, through
breadth of coverage to prevent overwhelming the review given the traditional machine learning models, to the deep learning and pre-
broad topic. Our focus extended to the coverage of tasks, the recency of trained language models. Notably, a majority of FSA research has
the research, and its impact within the field. This methodical approach converged on the established benchmarks of PhraseBank [13], SemEval
not only ensured the thoroughness of our review but also presented 2017 Task 5 [14], and FiQA Task 1 [15] for evaluation purposes. The
the status and achievements of AI applications in specific fields, laying construction of financial lexicons remains a focal point for researchers,
a foundation for future research directions. and lexicons can be employed either independently or in conjunction
The number of publications across the ten research topics are pre- with learning-based methods for FSA [16]. A notable shift in lexicon
sented in Fig. 2. The fields are ranked by the volume of publications, development moves from simplistic single-word expression to complex,
in descending order, as follows: risk management, question answering, multi-word, and context-aware phrases, a critical adaptation for the
virtual assistant and chatbot, financial sentiment analysis, financial financial sector where the implication of terms can dramatically change
narrative processing, financial forecasting, NLP for digital assets, port- with context. At the same time, there is a trend towards automated
folio management, ESG and sustainable finance, regulatory compliance approaches, highlighted by studies such as Oliveira et al. [17] and Du
monitoring and explainable AI in finance. All fields have demonstrated et al. [18], which contrasts with the traditional manual compilation
an increasing trend in publication volume over the past decade, high- of lexicons, demanding significant expert input. In the context of ma-
lighting a growing interest in NLP applications within the finance chine learning models, the process of feature engineering stands out
sector. Notably, emerging fields such as ESG and sustainable finance, as a pivotal stage, with the categorization into lexicon-based, linguis-
regulatory compliance monitoring and explainable AI in finance are tic, domain-specific, and word embedding features, etc. Among these,
also drawing significant scholarly attention. domain-specific attributes, especially numerical data or combinations
of keywords with numerical values, though not extensively examined,
3. An in-depth review of NLP in finance have shown their potential in FSA. An example is the term ‘‘revenue’’
coupled with a positive sign and a numerical value, indicating a posi-
NLP in finance refers to the broader application of NLP techniques tive financial forecast. Furthermore, the forefront of FSA includes deep
to the financial sector, which generally includes corporate data, public learning approaches, notably through the use of Convolutional Neural
financial news and announcements, and social media data such as Networks (CNN) and Long Short-Term Memory (LSTM) networks, as
StockTwits. It encompasses a wide array of tasks aimed at understand- well as pre-trained language models, which have significantly improved
ing, interpreting, and extracting information from textual data in the FSA performance. Specifically, a BERT variant tailored for the financial
financial domain. This includes processing corporate data, public news sector, known as FinBERT [19,20], has been developed through the
and announcement and social media data, and any other form of textual incorporation of a variety of data sources including Reuters Corpora,
data that can impact financial markets and decision-making. NLP in Yahoo Finance, and financial reports, marking a significant leap in FSA
finance leverages techniques such as sentiment analysis, information research methodologies. The latest study by Du et al. [16] achieved
extraction, named entity recognition, language modeling, LLMs and unparalleled results on FiQA Task 1 and SemEval 2017 Task 5 by
more, to analyze financial texts for various purposes like financial integrating the knowledge into the process of fine-tuning language
sentiment analysis, financial forecasting, portfolio management, finan- models such as RoBERTa, thus pushing the performance boundaries of
cial narrative processing, question answering, virtual assistant and FSA technique research further.
chatbot, risk management, financial regulatory compliance monitoring, The development of domain-specific transformer models, with no-
explainable AI in finance, ESG and sustainable finance and NLP for table examples, including FinBERT [19,20], has significantly enhanced
digital assets. FSA. However, the potential and applicability of autoregressive decoder

3
K. Du et al. Information Fusion 115 (2025) 102755

Fig. 2. Number of publications in the field of NLP in finance.

architectures such as Generative Pre-trained Transformer (GPT) [21] classification, question answering, among others. Du et al. [29] pro-
in the context of FSA have not been thoroughly investigated yet [22]. posed a prompting framework to evaluate the reasoning capabilities
In recent studies, Fatouros et al. [23] adopted zero-shot prompting to of LLMs for FSA by identifying six specific financial attributes that
examine multiple ChatGPT prompts on a curated dataset comprised potentially influence financial sentiment. Xing [30] designed special-
of forex-related news headlines, measuring performance using a range ized LLM agents for FSA, based on guiding knowledge. Mao et al.
of metrics such as precision, recall, F1-score, which outperforms Fin- [31] analyzed the market sentiment perception from the perspective
BERT. Zhang et al. [24] presented a framework that integrates a of cognition, and they employed MetaPro [32,33] to parse concept
retrieval-augmented mechanism with LLMs specifically for FSA. The mappings from metaphorical expressions, and then detected the distinct
framework consists of two key modules which are instruction-finetuned cognitive patterns during different market conditions. Manro et al. [34]
LLMs and a retrieval-augmented module. The performance metrics, examined the influence of CEOs’ cognitive states on stock price trends
specifically accuracy and F1 score, are significantly improved, under- by analyzing the opinions expressed in their letters to shareholders
scoring the efficacy of the framework in FSA. Fatemi and Hu [22] using MetaPro.
conducted a thorough comparative analysis to examine the effective-
ness of fine-tuning LLMs versus employing few-shot learning techniques 3.2. Financial forecasting
in the context of FSA. In particular, the in-context learning is adopted
on GPT-3.5-turbo and the fine-tuning is performed on Flan-T5. The The primary application of FSA is to investigate the causality and
study highlights the remarkable capabilities of LLMs, even smaller mod- correlation between financial sentiment and financial markets, and
els, in both in-context learning and fine-tuning for the FSA task. Zhang to perform financial forecasting [8]. Natural language-based financial
et al. [25] introduced a novel approach named Instruct-FinGPT, which forecasting was brought up by Xing et al. [50] with a focus on the
transforms a limited subset of supervised FSA data into instruction stock market. Subsequently, Du et al. [8] broadened the scope of review
data to fine-tune a general-purpose LLM. The Instruct-FinGPT approach to include the FOREX and cryptocurrency markets. Additionally, Du
has demonstrated significant advancements in its ability to perform et al. [8] brought attention to the interconnected relationship between
zero-shot generalization across other financial datasets. Hu et al. [26] financial sentiment, investor sentiment, and market sentiment, arguing
conducted a study to evaluate the sentiment of management’s dis- that financial forecasting is an essential application of FSA. This is
cussion and analysis disclosures using three distinct methods: GPT-3, because financial sentiment, whether used implicitly or explicitly, is
FinBERT, and the dictionary-based method. The findings reveal that crucial in predicting financial market movements. Natural language-
GPT-3 surpasses the dictionary-based method in its effectiveness for based financial forecasting serves to augment algorithmic trading by
financial sentiment classification tasks. However, it was observed that leveraging NLP to analyze news articles and social media, thereby
GPT-3 slightly underperforms when compared to FinBERT. Deng et al. informing automated trading decisions, enabling algorithms to parse
[27] implemented a semi-supervised learning approach using LLMs to and interpret vast amounts of textual data in real time and enhancing
generate weak sentiment labels for Reddit posts and then used that data decision-making processes and market responsiveness.
to train a small model which is to be served in production. Additionally, Predicting market movements in financial markets is notably chal-
Bloomberg has introduced BloombergGPT [28], an LLM tailored for lenging, attributed to their volatile and unpredictable nature. The
financial contexts, which has demonstrated superior performance in foundation of earlier financial market prediction research rests on his-
financial NLP tasks, including FSA, named entity recognition, news torical trading data, technical indicators, and macroeconomic variables,

4
K. Du et al. Information Fusion 115 (2025) 102755

Table 2
FSA techniques.
Literature Dataset Task Method Performance
Park et al. [35] PhraseBank Sentence-level FSA Lexicon Precision: 0.8238, Recall: 0.8128
F1-Score: 0.8105
Du et al. [18] PhraseBank, SEntFiN, etc. Sentence-level FSA Lexicon Accuracy: 0.7619, F1-Score: 0.7216
Krishnamoorthy [36] PhraseBank Sentence-level FSA Machine learning Precision:
Pos: 0.83, Neg: 0.93, Neut: 0.86
Recall:
Pos: 0.82, Neg: 0.93, Neut: 0.81
F1-Score:
Pos: 0.83, Neg: 0.93, Neut: 0.83
Araci [37] PhraseBank, FiQA Task 1 Sentence-level FSA Pre-trained language model PhraseBank
(Accuracy: 0.86, F1-Score: 0.84)
FiQA Task 1 (MSE: 0.07)
Liu et al. [20] PhraseBank Sentence-level FSA Pre-trained language model Accuracy: 0.94, F1-Score: 0.93
Zhao et al. [38] CCF BDCI, CCKS Sentence-level FSA Pre-trained language model Accuracy: 96.774%
Wu et al. [28] PhraseBank, FiQA Task 1 Sentence-level FSA LLMs PhraseBank: 0.5107
FiQA Task 1: 0.7507
Du et al. [29] PhraseBank, Fin. News Sentence-level FSA LLMs PhraseBank: (Accuracy: 0.9639)
Fin. News: (Accuracy: 0.8234)
Xing [30] PhraseBank, Fin. News, etc. Sentence-level FSA LLMs PhraseBank:
Accuracy: 0.8048, F1-Score: 0.8141
FiQA:
Accuracy: 0.9391, F1-Score: 0.9522
SEntFiN:
Accuracy: 0.7745, F1-Score: 0.7693
Jiang et al. [39] SemEval 2017 Task 5 Targeted, Sentence-level FSA Machine learning WCS (News: 0.7779, Microblogs: 0.7107)
Schouten et al. [40] SemEval 2017 Task 5 Targeted, Sentence-level FSA Machine learning WCS (News and Microblogs: 0.7050)
Dridi et al. [41] SemEval 2017 Task 5 Targeted, Sentence-level FSA Machine learning WCS (News: 0.655, Microblogs: 0.726)
Akhtar et al. [42] SemEval 2017 Task 5 Targeted, Sentence-level FSA Hybrid approach WCS (News: 0.786, Microblogs: 0.797)
Ghosal et al. [43] SemEval 2017 Task 5 Targeted, Sentence-level FSA Hybrid approach WCS (News: 0.697, Microblogs: 0.751)
Mao et al. [44] SemEval 2017 Task 5 Targeted, Sentence-level FSA Metaphor Paraphrasing Accuracy: 0.696, F1-Score: 0.696
Xiang et al. [45] SemEval 2017 Task 5 Targeted, Sentence-level FSA Pre-trained language model WCS (News: 0.8441, Microblogs: 0.8333)
Sinha et al. [46] SemEval 2017 Task 5 Targeted, Sentence-level FSA Pre-trained language model Accuracy: 0.9429, F1-Score: 0.9327
Du et al. [16] SemEval 2017 Task 5 Targeted, Sentence-level FSA Pre-trained language model WCS (News: 0.8483, Microblogs: 0.8122)
de França Costa and FiQA Task 1 Targeted, Aspect-level FSA Machine learning Aspect Extraction:
da Silva [47] F1-Score (News: 0.4240, Post: 0.5775)
Sentiment Analysis:
MSE (News: 0.1052, Microblogs: 0.1281)
Piao and Breslin [48] FiQA Task 1 Targeted, Aspect-level FSA Deep learning Aspect Extraction: F1-Score (0.6530)
Sentiment Analysis: MSE (0.0926)
Xiang et al. [45] FiQA Task 1 Targeted, Aspect-level FSA Pre-trained language model MSE (0.0717), R2 (0.4878)
Du et al. [16] FiQA Task 1 Targeted, Aspect-level FSA Pre-trained language model MSE: 0.0490
Luo et al. [49] Financial documents Document-level FSA Hierarchical Query-driven Attention Accuracy: 0.9446, F1-Score: 0.9449
Mao et al. [31] Financial news Perception-level FSA MetaPro and Statistical Analysis
Manro et al. [34] Letters to shareholders Perception-level FSA MetaPro and Statistical Analysis

with recent advancements incorporating financial textual data using tuple features with trading data [58], and integrating sentiment in-
NLP techniques for enhanced financial forecasting [50–53]. Established dicators with technical analysis for enhanced price prediction [59].
financial news providers, including Bloomberg and Thomson Reuters, Novel strategies also involve parsing earnings calls, which are shown to
have been at the forefront of offering commercial FSA service [52, influence short-term investor sentiment and predict stock returns [60–
54]. Mudinas et al. [52] highlighted that financial institutions e.g., D. 63]. Further advancements include the application of Graph Convo-
E. Shaw and Two Sigma have integrated sentiment signals from fi- lutional Neural Networks to incorporate company relationships [64],
nancial texts with traditional structured transaction data, to enhance utilizing social media texts [65], and employing transformers and
their machine learning models for algorithmic trading. Convention- self-supervised learning for dynamic market prediction and improved
ally, fundamental analysis and technical analysis are the two principal stock forecasting accuracy [53,66,67]. Additionally, combining various
approaches in stock market analysis [55]. Financial forecasting us-
sentiment sources through Principal Component Analysis and Kalman
ing natural language can be classified as part of technical analysis,
Filter has proven effective for generating robust daily sentiment indica-
since it does not alter the intrinsic value of assets. Tables 3 and 4
tors, crucial for accurate market predictions [54]. Event-driven trading
showcase studies in financial forecasting in stock and FOREX markets,
strategies have also been developed to capitalize on corporate events
respectively, demonstrating the effectiveness of incorporating financial
detected in news articles, offering substantial returns [68]. Investor
textual information. The financial forecasting for the cryptocurrency
market is reviewed in Section 3.10 NLP for Digital Assets. sentiment not only influences time-series returns but also affects the
Investor sentiment, or the deviation of investors’ beliefs about future cross-section of stock returns, with certain stocks being more suscep-
firm valuations from fundamental information, significantly impacts tible to sentiment-driven fluctuations [69]. Recently, Du et al. [70]
stock prices and market activities [56]. Research in stock market pre- proposed a dual-graph neural network, which dynamically generates
diction spans various methodologies, including time series models, and integrates price relationships and semantic relationships between
machine learning, deep learning, and reinforcement learning, focusing companies for stock price movement prediction.
on aspects like stock indexes, prices, movements, return rates, and The focus on predicting stock market trends has been significant,
volatility. Notable approaches include the use of neural tensor net- with the FOREX market receiving less attention. Early research ex-
works for event-driven predictions [57], combining news and event plored the link between macroeconomic fundamentals and short-term

5
K. Du et al. Information Fusion 115 (2025) 102755

Table 3
Financial forecasting in stock market.
Literature Data source Method Task Performance
Ding et al. [57] Thomson Reuters, CNN Price prediction Accuracy: 0.6548, MCC: 0.41
Bloomberg S&P500 prediction Accuracy: 0.6421, MCC: 0.40
Nguyen and Shirai [71] Yahoo Finance SVM Price movement prediction
Message Board
Oliveira et al. [54] Twitter MR, NN, SVM, RF, Ensemble Daily return NMAE: NDQ, 7.58
Daily trading volume NMAE: DJIA, 5.84
Daily volatility NMAE: DJIA, 2.79
Liu et al. [58] Thomson Reuters SVM, LSTM Price movement prediction Accuracy: 0.5544, F1-Score: 0.7133
Xu and Cohen [72] Twitter GRU, VAEs Price movement prediction Accuracy: 0.5823, MCC: 0.080796
Wu et al. [73] Twitter Cross-modal attention-based Price movement prediction Accuracy: 0.5915
Hybrid RNN
Guo and Li [74] Twitter Linear Regression Price prediction Accuracy: 0.6722
Keith and Stent [63] Earning calls Ridge/Logistic Regression Change in target price Regression (MSE: 0.00137, 𝑅2 : 0.1718)
LSTM, Ensemble Classification (Accuracy: 0.482, F1-Score: 0.475)
Jin et al. [75] Stocktwits Empirical Mode Price prediction Accuracy: 0.7056, MAPE: 1.65%, MAE: 2.39
Decomposition-based LSTM
Sawhney et al. [65] Twitter Graph Attention Network Price movement prediction F1-Score: 0.605, Accuracy: 0.608, MCC: 0.195
Zhou et al. [68] PRNewswire Multi-class Classification Price movement prediction Trade at End Strategy: Average Return: 1.74%
Businesswire with MLM Loss Trade at Best Strategy: Average Return: 9.11%
Wu et al. [76] EastMoney.com CNN, LSTM Price prediction MAE: 2.38, MSE: 7.27, RMSE: 2.69
Soun et al. [67] Twitter Attentive LSTM Price movement prediction Accuracy (BIGDATA22: 54.81%, ACL18: 58.72%,
CIKM18: 55.86%)
MCC (BIGDATA22: 0.0952, ACL18: 0.2065,
CIKM18: 0.0899)
Ma et al. [53] Hundsun Electronics Pre-training, BiLSTM Price movement prediction Accuracy: 0.6726, MCC: 0.3452
and GCN
Du et al. [70] Twitter Attentive LSTM, GATs Price movement prediction Accuracy (BIGDATA22: 0.5833, CIKM18: 0.5730)
MCC (BIGDATA22: 0.1556, CIKM18: 0.0389)

Table 4
Financial forecasting in FOREX market.
Literature Data source Method Task Performance
Jin et al. [77] Bloomberg Linear Regression Change in currency value and Precision (Argentina: 0.18, Brazil: 0.28, Chile:
generate warning messages 0.33, Colombia: 0.25)
Recall (Argentina: 0.60, Brazil: 0.63, Chile: 1,
Colombia: 1)
Nassirtoussi et al. [55] MarketWatch.com SVM FOREX price movement Accuracy: 0.8333, Precision (Pos: 0.6667, Neg:
Google RSS prediction 0.8889)
Recall (Pos: 0.6667, Neg: 0.8889)
Seifollahi and Shajari [78] MarketWatch.com SVM FOREX price movement Accuracy: 0.5926, Precision: 0.5710, Recall: 0.5735
prediction
Semiromi et al. [79] Forexfactory.com, SVM, RF, XGB FOREX price movement Accuracy (EUR/USD: 0.638, GBP/USD: 0.663)
prediction
First Word FX News Accuracy (USD/CHF: 0.631, USD/JPY: 0.641)
Xing et al. [80] Dow Jones Newswire SVM FOREX price movement Accuracy: 0.503, F1-Score: 0.538
prediction

exchange rate volatility using the Flexible Fourier Form regression, that news sentiment alone offers predictive value, albeit limited, for
focusing on the impact of macroeconomic announcements (e.g., interest FOREX price movements [80]. Lastly, researchers and practitioners are
rates, GDP, consumer confidence indexes) on USD/EUR volatility [81]. expanding their focus to include other financial forecasting areas, like
Findings from data segmented into 5-min intervals over a few months option pricing and IPO pricing [83], and firm value [84] etc.
indicated that such news significantly affects exchange rate volatility
immediately post-announcement, with the impact varying by news 3.3. Portfolio management
type and country. Subsequent studies confirmed that macroeconomic
news influences currency markets over time, with immediate absorp- Portfolio optimization is an important element of portfolio manage-
tion of the direct price impact but delayed total reflection [82]. The ment, which involves strategies to identify the optimal investment mix
Forex-foreteller (FF) model was introduced to forecast currency market to maximize expected returns for a given risk level, or to minimize
movements using news articles, language models, topic clustering, and risk for a targeted return. The use of NLP in this field is innovative,
sentiment analysis in conjunction with historical data, proving effective focusing on enhancing decision-making by analyzing qualitative data
for generating predictive signals [77]. Other approaches have leveraged (see Table 5). In portfolio optimization, NLP helps by capturing se-
text mining of news headlines, such as TF–IDF weighted by sentiment mantic relations of investment targets [85] or incorporating opinions
scores using SentiWordNet, to predict FOREX market movement [55]. from investor posts for better stock market investments [86]. This shift
Further advancements include sentiment analysis of news headlines towards utilizing financial sentiment in portfolio management has led
for predicting EUR/USD directional movement with enhanced accu- to various approaches, including Follow-the-Loser strategies, based on
racy [78], incorporating news story events from the economic calendar stock microblogs [87], and models predicting the quality of investment
to forecast currency pair movements with SVM, RF, and XGBoost opinions for portfolio creation [88].
algorithms [79], and evaluating high-frequency news sentiment with- Moreover, techniques like Gaussian inverse reinforcement learning,
out other semantic features, where a FinBERT-based model showed model market dynamics and investor sentiment for trading systems that

6
K. Du et al. Information Fusion 115 (2025) 102755

Table 5
Portfolio management.
Literature Data source Method Performance
Tu et al. [88] StockTwits Linear regression for opinion quality score Cumulative return: 30%
Koyano and Ikeda [87] Yahoo! JAPAN textream MLP, Follow-the-Winner/Loser strategy Cumulative return: 1.151
Sun et al. [98] Sina Guba, Eastmoney Guba, RESSET Probabilistic Neural Network Accuracy: 86.3%
Yang et al. [89] Thomson Reuters Gaussian inverse reinforcement learning Return: 17.39%, Sharpe ratio: 0.85
Sterling Ratio: 0.76
Malandri et al. [86] Financial and sentiment data EW, LSTM, MLP, RF Wealth of the portfolio (LSTM + Sentiment)
Five portfolios: 2.23, 2.72, 2.30, 2.81 and 1.65
Xing et al. [85] Company profiles Doc2vec, covariance matrix correction CAGR: 0.2368, Sharpe ratio: 1.01
Ye et al. [91] Financial news State augmented reinforcement learning Portfolio value improved (Bitcoin: 140.9%,
HighTech: 15.7%)
Sharpe ratio (Bitcoin: 14.78, HighTech: 7.73)
Du and Tanaka-Ishii [92] Wall Street Journal (WSJ), Reuters, Deep learning Average realized annual gain (WSJ: 17.2%, R&B:
Bloomberg 35.5%)
Chen et al. [94] RESSET database DT, LR, SVM, RF Accuracy: 79.6%, Holding Period yield: 5.41
Sawhney et al. [95] Twitter, Wind.com.cn Time-aware LSTM (t-LSTM) US S&P 500: Return ratio: 1.34, Sharpe ratio: 0.96
China and Hong Kong: Return ratio: 1.44, Sharpe
ratio:1.19
Koratamaddi et al. [93] Wharton Research Data Services Deep reinforcement learning Sharpe ratio: 2.07, Annualized return: 22%
Ghosh et al. [99] Bloomberg RF, LSTM Daily return (LSTM: 0.64%, RF: 0.54%)
Hung et al. [96] Reuters, CNBC GRU, Black–Litterman Annualized return: 46.6262%, Sharpe ratio:
13.02%
Sortino ratio: 17.96%, MDD: 4.45%, Turnover:
68.35%
Ma et al. [97] RESSET, Qichacha, Hundsun Electronics Multitask learning, pre-training IRR: 0.3923, Sharpe ratio: 2.0334, MDD: 0.0754

filter out noisy signals [89]. Strategies e.g., PROFIT (Policy for Return FNP workshop since 2018.1 Moreover, the automation of financial
Optimization using FInancial news and online Text) leverages financial reporting [101] and financial statement analysis [102] using NLP,
news and social media for trading decision optimization, showing sig- particularly with LLMs, has also emerged as a promising area of re-
nificant performance improvements over benchmarks [90]. Similarly, search. In our review, we focus on the financial text summarization
State Augmented Reinforcement Learning (SARL) integrates diverse and information extraction task.
data sources for learning portfolio management strategies, achieving
3.4.1. Financial text summarization
superior results [91]. Stock embedding, representing stocks as neu-
Companies generate a multitude of reports incorporating both tex-
ral network-generated vectors from news and price history, enhances
tual narratives and numerical data throughout their financial year, such
portfolio optimization beyond price prediction [92]. Including on-
as financial annual reports, earnings announcements, and conference
line public mood in models has proven to consistently outperform
calls. The information encapsulated within these financial reports holds
equal-weighted strategies, highlighting the performance benefits of significant importance for both companies and investors. It serves to
incorporating financial sentiment [86]. Sentiment-aware models for showcase company accomplishments, garner support from shareholders
daily portfolio allocation demonstrate the robustness and higher re- in the stock market, as well as unveil risks and opportunities crucial for
turns compared to benchmarks [93]. Additionally, sentiment-based investor decision-making. Moreover, it plays a pivotal role in the due
models adapted to market volatility, like those for Chinese stocks, diligence procedures during mergers, acquisitions, and audit processes.
yield superior returns [94]. Time-aware LSTM models, which capture However, the vast volume of financial data poses challenges in terms of
market trends and rank stocks, based on predicted returns, outperform navigation, management, and monitoring. Consequently, there arises a
state-of-the-art methods in both intra-day situations and risk-adjusted critical need for automated financial text summarization.
returns [95]. Hung et al. [96] constructed a portfolio using news Financial text summarization endeavors to generate succinct, infor-
sentiment measured by BERT and demonstrated that Black–Litterman mative, and non-repetitive summaries from either single or multiple
portfolio model with GRU model to predict stock price yields the input texts. This objective can be pursued through two distinct method-
highest performance. Ma et al. [97] developed a multi-task learning ologies: the extractive approach and the abstractive approach. The
framework, learning risk and return factors, simultaneously, for port- extractive method involves the identification and ranking of pertinent
subsets within the input text [103]. Conversely, the abstractive ap-
folio optimization. The framework also leveraged data from multiple
proach generates summaries from scratch, based on the contents of the
sources, e.g., news, numerical features, and company relationships.
input document [104].
Table 6 showcases a variety of techniques and methods ranging
3.4. Financial narrative processing from traditional machine learning techniques to high-performing deep
learning models and word embeddings. The majority of techniques
Financial Narrative Processing (FNP) [100], is a subset of NLP in have converged on the established benchmark dataset from FNS Shared
finance with a specific focus on analyzing narrative content within Tasks since 2020 [105–107], with a focus on producing concise sum-
maries for annual financial reports of firms listed on the stock ex-
financial documents. While NLP in finance has a broader application
changes of the UK, Greece, and Spain. Extractive summarizer continues
range, encompassing any textual analysis that can inform financial de-
to be dominant on either sentence-level or section-level text summa-
cisions and strategies, FNP is more specifically focused on the narrative
rization within the FNS shared Task dataset. El-Haj and Ogden [108]
and qualitative aspects of financial texts. It involves understanding
proposed to use the TF–IDF representations of the sentence and K-
the stories and qualitative disclosures that companies and financial means clustering model to rank the importance of sentences, and com-
analysts provide, such as management’s discussion and analysis sec- bine them into a final summary. As deep learning models have demon-
tion in annual reports, earnings call transcripts, analyst reports, and strated effectiveness in generating pertinent text features, the approach
financial news. The key objective is to extract insights from narrative
texts that are not readily quantifiable but have significant implications
1
for financial decision-making. El-Haj et al. [100] has organized the http://wp.lancs.ac.uk/cfie.

7
K. Du et al. Information Fusion 115 (2025) 102755

Table 6
Financial text summarization.
Literature Dataset Type of summarization Method Performance
El-Haj and Ogden [108] FNS Shared Task Extractive at sentence level TF–IDF + Clustering Rouge-1F (English: 0.317,
Spanish: 0.448, Greek: 0.334)
Gokhan et al. [109] FNS Shared Task Extractive at sentence level SentenceBERT-based Clustering Rouge-1F: 0.48
La Quatra and Cagliero [110] FNS Shared Task Extractive at sentence level BERT Rouge-1F: 0.424
Abdaljalil and Bouamor [111] FNS Shared Task Extractive at sentence level BERT Rouge-1F: 0.38
Orzhenovskii [112] FNS Shared Task Extractive at sentence level T5 Rouge-1F: 0.54
Shukla et al. [114] FNS Shared Task Extractive at section level LR, k-maximal word Rouge-1F: 0.508
allocation, top-k summarizer
Shukla et al. [115] FNS Shared Task Extractive at section level GPT3.5 Rouge-1F (English: 0.45,
Spanish: 0.41, Greek: 0.12)
Li et al. [116] Financial News Abstractive Transformer-BiLSTM encoder, Rouge-1F (LCFNS: 0.3814,
Graph attention decoder Fin-LCSTS, 0.3669, CNN/Daily Mail: 0.4164)
Liu et al. [117] FINDSum Abstractive BART-large Rouge-1F (FINDSum-ROO: 0.5390,
FINDSum-Liquidity: 0.5412)
Zmandar et al. [118] FNS Shared Task Hybrid Reinforcement learning Rouge-1F: 0.52
Singh [119] FNS Shared Task Hybrid Pointer Network + T5 Rouge-1F: 0.466
Mukherjee et al. [120] ECTSum Hybrid FinBERT + T5 Rouge-1F: 0.467

of fine-tuning language models has been employed to produce sentence- summaries for short input and output sequences in the abstractive
level embeddings customized for specific financial contexts. Gokhan summarization task. Building upon these techniques, Li et al. [116]
et al. [109] introduced a sentence-BERT methodology, leveraging BERT introduced an enhanced Seq2Seq model named TLGA. TLGA utilizes a
to generate sentence-level embeddings and subsequently clustering hierarchical Transformer-BiLSTM encoder to capture long-range inter-
these embeddings using the K-means model. Additionally, both BERT actions and sequential semantics, while employing a Graph Attention-
and T5 have been utilized through fine-tuning processes to enhance based decoder to leverage historical information of decoded tokens and
performance in ranking the importance of sentences [110–112]. capture key causal relations. This approach captures both qualitative
Financial reports, comprising financial statements and extensive and quantitative information as well as latent causal relationships from
information, often suffer from repetitiveness, posing a challenge when financial news.
summarizing them at the sentence level. Rather than summarizing the Several other studies [123,124] have suggested integrating extrac-
entire report, a strategy has been adopted to identify key narrative tive and abstractive methodologies to enhance performance. Zmandar
sections within the report and generate summaries, based on these iden- et al. [118] proposed a reinforcement learning model utilizing the
tified sections. As outlined by Litvak et al. [113], financial documents standard policy gradient method to construct an end-to-end train-
typically feature 13 predefined narrative section titles. Abdaljalil and able computation graph that combines both extractor and abstractor
Bouamor [111] employed a similar approach for section extraction. It agents. Singh [119] implemented pointer networks to extract important
compares BERT embeddings of reference sections to extracted ones, narrative sentences, subsequently using T5 to paraphrase extracted
naming the extracted sections after the closest reference sections by sentences into a concise yet informative sentence. Shukla et al. [115]
cosine similarity. A frequency counter is then utilized to determine the introduced a novel approach to enhance the summarization of financial
weight of each section in reference summaries. Following a clustering reports, the first instance of applying LLMs to the financial narrative
process to identify the top three sections for each document, the section summarization task. The emergence of various formats of financial
with the highest frequency weight is retained, and the top 1000 words text data has ignited research interest in techniques for financial text
of that section are included in the final summary. Shukla et al. [114] summarization. Mukherjee et al. [120] presented ECTSum,2 a novel
built an annotated dataset to define sections as ‘‘narrative’’ or ‘‘non- benchmark dataset for telegram-style bullet-point summarization of
long transcripts of earning calls (ECT). Additionally, ECT-BPS is pro-
narrative’’. Tables of content within the report are parsed to extract
posed, a hybrid approach that fine-tunes FinBERT to identify the most
section features: section name, page number, and length. Utilizing these
relevant sentences from the input ECT documents, followed by fine-
features, a classification model is employed to categorize each section
tuning T5 [125] to paraphrase these sentences to the telegram-style
as narrative (‘‘true’’) or non-narrative (‘‘false’’), based on the annotated
format of target summary sentence. Liu et al. [117] proposed FINDSum,
dataset. The weight of a section is defined as the probability of it being
a large-scale dataset encompassing both long text and multiple tables,
narrative, as assigned by the model. Subsequently, it adopts the ‘‘K-
enabling extractive summaries to incorporate quantitative descriptions
Maximal Word Allocation’’ approach, which optimally distributes the
of critical metrics in tables. The model processes long text by adopting
required number of words among sections, based on their weights and
the Maximum Marginal Recall Gain (MMRG) method to select salient
the number of available words in each section. The final summary
text segments as inputs, while transforming each table cell into a tuple
of the report is then generated, based on a set of pairs consisting of
and treating tuple selection as a binary classification task for tabular
narrative sections and the number of words to be generated. Guided by
content. Three types of methods are evaluated on FINDSum dataset.3
DiMSum framework [114], Shukla et al. [115] showcased the signifi-
1. combine-and-generate (CG) concatenates selected text segments and
cant enhancement in financial report summarization quality through
tuples as two types of input to a sequence-to-sequence summarizer 2.
the utilization of LLMs. It employs Retrieval Augmented Generation
generate-and-combine (GC) parallelly generates summary on selected
(RAG) to generate few-shot examples from the DiMSum annotated text segments and tuples, and concatenate two summaries as one. 3.
dataset, facilitating LLM-based few-shot classification to identify nar- generate-combine-and-generate (GCG) adopts a tuple-to-text generator
rative sections and determine their weights. Subsequently, the LLM to produce text descriptions of input tuples, concatenate tuples’ text
summarizer generates summaries, based on section name, content, and descriptions with selected text segments, and generate a final summary.
word count. Overall, CG and GCG methods demonstrate superior performance over
In scenarios where gold summaries have been removed from orig- text-only baselines on FINDSum’s two subsets.
inal annual reports, abstractive models offer the potential for greater
conciseness by generating summaries from scratch. Sequence-to-
Sequence model [121] and RNN-based encoder–decoder technique 2
https://github.com/rajdeep345/ECTSum.
3
[122] have demonstrated effective performance in generating succinct https://github.com/StevenLau6/FINDSum.

8
K. Du et al. Information Fusion 115 (2025) 102755

Table 7
Information extraction.
Literature Type of extraction Method Performance
Xiao et al. [126] Corporate event FLAN-T5, FLAN-UL2, FLAN-Alpaca F1-Score (Acquisition: 0.70, SEO: 0.79, CBI: 0.61)
Jacobs and Hoste [127] Corporate event BERT, RoBERTa, XLM, and XLNet Precision: 0.737, Recall: 0.564, F1-Score: 0.591
Jacobs et al. [128] Corporate event SVM, RNN-LSTM Precision: 0.80, Recall: 0.71, F1-Score: 0.73
Ein-Dor et al. [129] Corporate event BERT Precision: 0.93, Recall: 0.95, F1-Score: 0.94
Carta et al. [130] Corporate event Hierarchical clustering Silhouette Coefficient: 0.3, Dunn Index: 0.46
Guo et al. [131] Corporate event GNN + Transformer Precision: 0.844, Recall: 0.791, F1-Score: 0.817
Cheng et al. [132] Corporate event and relation OpenIE, Bidirectional LSTM Macro-F1: 0.7686, Micro-F1: 0.7515,
Weighted-F1: 0.7527
Zhou et al. [68] Corporate event Bi-level event detector Win Rate: 0.545 Big Win Rate: 0.342
Balashankar et al. [133] Causal relationship Predictive causal graph RMSE: 0.020
Izumi and Sakaji [134] Causal relationship Causal-chain search algorithm F1-Score (Financial Summaries: 0.71, Newspaper: 0.83)
Takayanagi et al. [135] Causal relationship ChatGPT FinCausal (Precision: 0.646, Recall: 0.795, F1-Score: 0.713)
Rajpoot and Parikh [136] Relation extraction LLMs F1-Score: 0.718
Hillebrand et al. [137] Relation extraction KPI-BERT F1-Score: 0.7032
Zhang and Zhang [138] Entity extraction FinBERT-MRC F1-Score: (ChFinAnn: 0.9278, AdminPunish: 0.9680)
Oral et al. [139] Entity extraction BiLSTM-CRF Macro-F1: 0.8925, Micro-F1: 0.9223

3.4.2. Information extraction token, while the high-level takes the predicted results and integrates
Information Extraction (IE) focuses on identifying specific types them with the global contextual information from the input article
of entities, relationships, events, and other factual information from to predict the probabilities of each event’s existence. The multi-task
unstructured textual sources [140]. IE is extensively used in the finance learning approach is also explored in identifying financial events. For
sector, particularly for the identification and detection of entities, instance, Li and Zhang [141] introduced a unified model for financial
corporate events, relations, causal relationships, financial document event classification, detection, and summarization. Specifically, the
structure, etc (see Table 7). Event extraction is a fundamentally chal- model leveraged pre-trained BERT to encode financial documents, with
lenging task in NLP, extensively applied to manage vast and rapidly the encoded information shared across the event type classification,
expanding collections of financial corpus which often contain multiple detection, and summarization components. A transformer structure
events with their elements scattered and mixed across the documents, serves as the decoder for the event summarization task. Alongside the
complicating the extraction process [131]. In terms of event extrac- input document encoded by BERT, the decoder incorporates predicted
tion from financial texts, Jacobs et al. [128] demonstrated that a event types and cluster information, enabling it to focus on specific
feature-engineered linear kernel SVM approach obtains better perfor- aspects of the event type during the summary generation.
mance than RNN-LSTM word-vector in the identification of corporate Extraction of relations and causality [142,143] represents a promi-
events from news articles. Jacobs and Hoste [127] proposed SEN- nent research domain within information extraction. Relation extrac-
TiVENT, an annotated English economic news corpus of fine-grained tion stands as a pivotal task within NLP, focusing on identifying and
company-specific events at the sentence level, and performed text categorizing connections between entities referenced in text. Rajpoot
classification using pre-trained BERT, RoBERTa, XLM, and XLNet mod- and Parikh [136] proposed GPT-FinRE, which explored the poten-
els. Xiao et al. [126] fine-tuned the FLAN series of models to evaluate tial of GPT with two retrieval strategies, KNN and Efficient Prompt
the effectiveness of LLMs in the context of corporate event prediction. A Retrieval (EPR), on the financial relation extraction task on REFinD
weakly supervised approach has also been adopted for event extraction. dataset [144]. Hillebrand et al. [137] introduced a novel model that
includes three main building blockers, a BERT-based sentence en-
For example, Ein-Dor et al. [129] leveraged information contained
coder, a named entity recognition decoder, and a relation extrac-
in Wikipedia to generate a weakly labeled sentence dataset, demon-
tion decoder, to extract and link key performance indicators from
strating its effectiveness for company-specific events in news articles.
financial reports. Causality detection, meanwhile, seeks to unveil the
In terms of unsupervised approaches, Carta et al. [130] proposed a
causal relationships between entities, with applications ranging from
hierarchical clustering-based approach that integrates news articles and
stock movement prediction [133] and supporting financial informa-
StockTwits messages to detect corporate events.
tion services [134]. For example, Balashankar et al. [133] proposed
The construction of knowledge graphs and their integration into
the Predictive Causal Graph by discovering the influence between
event extraction has garnered significant attention from researchers.
words, based on temporal co-variance and demonstrated that PCG
Guo et al. [131] enhanced the document-level event extraction by
outperforms other graph-based methods in stock market prediction.
integrating a knowledge graph capturing entity relations and their at-
To assess GPT’s performance in causal reasoning, Takayanagi et al.
tributes. Specifically, the proposed model employed a graph encoder to
[135] conducted a comprehensive evaluation of ChatGPT’s causal text
embed knowledge graph information for entities, applied named entity
mining capabilities and observed that causal hallucination is evident
recognition to identify potentially relevant entities, and incorporated a in the low precision on the FinCausal dataset when implementing
transformer and linear layer to integrate the graph embeddings of the ChatGPT to detect causality relationship. Sharma et al. [145] released
extracted entities into the event extraction model. Cheng et al. [132] FinRED,4 a relation extraction dataset curated from financial news and
introduced an event embedding framework, based on a knowledge earning call transcripts containing relations from the finance domain.
graph, which not only extracts structured events from unstructured Lastly, Becquin [146] formulates the causality extraction into a span
texts, but also builds the knowledge graph using the entities and re- extraction and sequence labeling task. Span extraction aims to extract
lations mentioned therein, and then incorporates the knowledge graph text spans, such as words or phrases, from plain texts [147]. Specif-
information into the objective function of an event embedding learning ically, features are generated from an input text using transformers
model using a joint model. The learned representations serve as inputs with high-performance tokenizers, and question-answering state-of-the-
for subsequent quantitative trading strategies. Meanwhile, Zhou et al. art architecture is implemented with two spans extracted (cause, effect)
[68] presented a bi-level event detection model to detect corporate pairs. The architecture not only can extract cause and effect but also
events from news articles which are subsequently used to predict predict the number of cause/effect relationships in a document.
stock movement for the event-driven trading strategy. The model in-
cludes low-level and high-level event detectors. The low-level detector
4
identifies specific event-describing sub-sequences by classifying each https://github.com/soummyaah/FinRED.

9
K. Du et al. Information Fusion 115 (2025) 102755

Financial named entity recognition (FinNER), which refers to the dataset, FINQA, for numerical reasoning over financial reports. Other
process of identifying and categorizing specific entities, such as compa- areas of focus include open-domain question answering [160], which
nies, currencies, dates, and numerical values, within financial texts, is involves extracting answers from large, unstructured datasets using so-
a challenging task in financial text information extraction [138]. Zhang phisticated retrieval and reading strategies. The quest for sophisticated
and Zhang [138] proposed a new model termed FinBERT-MRC by QA systems in finance also extends to the ability to handle hybrid data,
formulating FinNER task as a machine reading comprehension (MRC) combining both tabular and textual content. For example, Zhu et al.
problem, and achieved competitive performance. Oral et al. [139] [161] introduced TAT-QA, a benchmark leveraging both tabular and
extracted named entities with a BiLSTM-CRF architecture from text- textual content, demonstrating the importance of context in finance-
intensive and visually rich scanned documents and predicted binary re- related queries, and further proposed a novel QA model termed TAGOP,
lations for each entity pair using a graph-based MLP architecture. Shah which is capable of reasoning over both tables and text. Deng et al.
et al. [148] proposed a novel domain-specific model named FLANG (Fi- [162] also presented the PACIFIC dataset to facilitate conversational
nancial LANGuage) which enhances masking using financial keywords question answering (CQA) over hybrid contexts in finance. Further, the
and phrases, incorporating span boundary objective and in-filing ob- knowledge-enhanced QA system, which enables information retrieval
jective. Together, an open-source comprehensive suite of benchmarks from an external knowledge base or knowledge graph, has improved
for the financial domain, named Financial Language Understanding the performance on QA tasks [163,164]. Lastly, the evaluation of LLM’s
Evaluation (FLUE), is released. It includes new benchmarks across NLP capability including mathematical reasoning for QA over financial
tasks which are sentiment analysis, FinNER, financial structure bound- documents has also attracted researchers’ attention significantly [165,
ary detection, and financial question answering. The source dataset for 166].
FinNER is from [149].
Lastly, researchers also explored financial information extraction 3.5.2. Virtual assistant and chatbot
tasks such as financial document structure extraction [150], which
Unlike research in QA, which primarily leverages golden datasets,
extracts table-of-contents (TOC) from financial documents by detecting
virtual assistant and chatbot are more application-oriented, spanning
the document titles and organizing them hierarchically into a TOC,
various functions in finance such as customer service [167], personal
and structure boundary detection [151], an extension of boundary
finance advice [168], cryptocurrency [169], and anti-fraud [170] etc.
detection, which extracts the boundaries of sentences, lists and list
Virtual assistants are designed to perform tasks, based on user re-
items, including structure elements like header, footer and tables.
quests. These tasks can range from finding information online, like
the best local restaurants, to more complex activities such as schedul-
3.5. Question answering, virtual assistant and chatbot
ing appointments. The effectiveness of a virtual assistant hinges on
its ability to accurately parse the user’s intent and to retrieve and
Question answering systems, virtual assistants, and chatbots lever-
interact with the relevant information effectively. Research in this
aging NLP techniques serve distinct yet overlapping roles in the realm
domain often focuses on enhancing the assistants’ understanding of
of dialogue interactions. Question answering systems are optimized for
user intents through improved dialogue management systems [171,
precision and are adept at fetching exact answers from structured and
172] and by refining the query processing capabilities that help in
unstructured data, primarily used in settings requiring high accuracy
better understanding and actioning user requests. The development
such as academic research or customer support [152]. Virtual assistants
of LLMs has significantly influenced the evolution of chatbots. These
are crafted for a wide range of functionalities, from managing daily
activities to handling enterprise processes. These systems prioritize user models, trained on vast text collections, are capable of generating
interaction and strive to simplify and enhance the user experience. responses that mimic human conversation and handle a wide range
Chatbots, particularly those integrated with LLMs, focus on simulating of linguistic tasks. The quality of these responses is heavily dependent
conversational human interactions to enhance customer engagement on the context provided. Major tech companies like OpenAI, Meta,
and service with responses that are context-aware and increasingly per- and Google primarily oversee these sophisticated models. Factors such
sonalized. Despite their differences, all these technologies interpret and as high computational costs, extensive data collection requirements,
respond to human language, tailoring their functionalities to specific specialized expertise and labor, infrastructure and maintenance needs,
user needs and contexts. licensing and intellectual property costs, and substantial energy con-
sumption make it prohibitively expensive for corporations to develop
3.5.1. Question answering and train these models from scratch. Consequently, practical appli-
Question Answering (QA) involves designing a system that can cations of LLMs often involve methods like fine-tuning or advanced
answer questions posed by humans in natural language. It requires context management strategies. Recent innovations in this field include
the integration of several fundamental components, such as tokeniza- RAG techniques [173], which bolster the domain-specific abilities of
tion, entity extraction, intent detection, and more. Key challenges chatbots, and developments like MEM-GPT [174], which approach
include understanding the intent behind the query and retrieving the LLMs as operating systems capable of both long and short-term memory
most relevant information from a set of documents [153]. In addi- management to efficiently update and manage context. Furthermore,
tion to general QA which answers natural language questions [15], the infinite attention mechanism [175] employs transformers to dynam-
the finance domain requires complex numerical reasoning and un- ically manage context length, addressing the inherent token limitations
derstanding of heterogeneous representations [154]. Recent research of LLMs. Additional research has concentrated on model fine-tuning
in QA has delved into various aspects, including the development of strategies such as parameter-efficient fine-tuning (PEFT) by Xu et al.
datasets [155] that challenge the reading comprehension abilities of [176] and LOw Rank Adoption (LORA) by Hu et al. [177], which
models, like SQuAD [156] and TriviaQA [157] and the enhancement train LLMs through low-rank parameterization or in computationally
of reasoning capabilities necessary for answering questions more ac- efficient ways. Recent alignment efforts in LLMs include techniques
curately (see Table 8). Numerical reasoning capability is another area like Reinforcement Learning with Human Feedback (RLHF) by Kauf-
of focus for QA in finance. Chen et al. [158] propose ConvFinQA, a mann et al. [178] and Supervised Fine-Tuning (SFT) by Mecklenburg
new large-scale dataset aiming to study the chain of numerical rea- et al. [179]. Reinforcement Learning with Generative Adversarial Feed-
soning in conversational question answering. Li et al. [159] introduced back [180] provides an alternative alignment method that overcomes
FinMath, which injects a tree-structured neural model to perform multi- the challenges posed by RLHF, which is constrained by the expertise
step numerical reasoning, aiming to improve the numerical reasoning and productivity limitations of human evaluators, and SFT, which relies
capacity of the model. Chen et al. [154] released a new large-scale on additional, carefully selected expert demonstrations.

10
K. Du et al. Information Fusion 115 (2025) 102755

Table 8
Question answering, virtual assistant and chatbot.
Literature Data sources Method Performance
Li et al. [159] TAT-QA RoBERTa-large, Auto-regressive sequence-to-tree model Exact Match: 0.586, numeracy-focused F1: 0.641
Zhu et al. [161] TAT-QA RoBERTa, Sequence tagging, aggregation operator, scale prediction Exact Match: 0.501, numeracy-focused F1: 0.580
Chen et al. [158] CONVFINQA RoBERTa-large, FinQANet Execution accuracy: 0.6890, program accuracy: 0.6824
Chen et al. [154] FinQA RoBERTa-large, FinQANet Execution accuracy: 0.6124, program accuracy: 0.5886

Lastly, the trends and characteristics of chatbot development in the subsequent day’s market volatility, by analyzing the relationship
finance include personalization [181], decentralization [182], and between financial news, tweets, and the FTSE100 index. The findings
ontology-based approaches [183], among others. As chatbots become also highlight that the accuracy of the model depends on the selected
more integral to finance operations, ensuring they operate reliably and number of topics.
ethically is crucial. This involves designing systems that are not only
accurate but also trustworthy and free from biases. Current research 3.6.2. Fraud detection
[184] is exploring how to implement safety features and ethical guide- Fraud detection in finance is a broad field that encompasses various
lines to prevent the generation of harmful or misleading responses. strategies and methodologies to identify unusual patterns that may
These guardrails are essential for maintaining user trust and ensuring indicate fraudulent activities. In the finance sector, fraudulent activities
that interactions are both helpful and safe. can be categorized into transaction fraud, mortgage fraud, corporate
and financial statement fraud, money laundering fraud, securities and
3.6. Risk management commodities fraud, insurance fraud, and cryptocurrency fraud [193].
The advent of NLP techniques marks a significant transformation in
3.6.1. Financial risk prediction detecting and preventing fraudulent activities. By leveraging NLP’s
Research in financial risk prediction leverages a variety of data power to sift through vast amounts of unstructured data, ranging
sources including financial reports [185–188], financial news [187, from customer correspondences and transaction narratives to financial
188] and earning calls [189] (see Table 9). Wang et al. [186] explored statements and digital social interactions, financial institutions are now
the impact of sentiment words on financial risk, by establishing a better equipped to identify suspicious patterns and inconsistencies that
regression model to predict future real-value risk, based on sentiment may indicate fraud.
analysis and a ranking model to assess risk levels, using financial re- The primary research focus of NLP in fraud detection centers on
ports i.e., management’s discussion and analysis of financial conditions identifying fraudulent transactions as well as detecting fraud within
and results of operations in the 10-K Form. The findings reveal that corporate and financial statements (see Table 10). Rodríguez et al.
models developed using sentiment words outperform those trained on [194] presented a novel approach for fraud transaction detection, based
original texts, underscoring the significance of financial sentiment lex- on the transformer model. Boulieris et al. [195] proposed FraudNLP,
icons in predicting financial risk. Furthermore, the results also indicate the first anonymized dataset accessible to the public for detecting
a strong correlation between financial risk and financial sentiment online fraud, and benchmarked both machine learning and deep learn-
words. Wang and Hua [189] defined financial risk as stock price ing techniques with multiple evaluation measures. Additionally, they
volatility in the subsequent week and investigated its relationship argue that online behaviors adhere to patterns similar to natural lan-
with earnings calls. It utilized uni-grams, bi-grams, part-of-speech tags, guage, suggesting that natural language processing techniques can
named entities, and probabilistic frame-semantic features to construct effectively address these phenomena. Yang et al. [196] introduced
a Gaussian copula model, evaluated by Kendall’s tau and Spearman’s FinChain-BERT, demonstrating a successful application of BERT models
correlation. The research examined three datasets consisting of tran- in detecting financial fraud and showing effectiveness in handling
scribed quarterly earnings calls from the U.S. stock market during the complex financial text information. In terms of corporate and finan-
period of the Great Recession. The datasets were categorized into three cial statement fraud, Chen et al. [197] employs a fusion of NLP,
periods: pre-2009 (2006–2008), which marks the onset of the economic Queen Genetic Algorithm (QGA), and SVM techniques to create a novel
downturn; 2009, capturing the global spread of the financial crisis; and approach for identifying fraudulent content within the narratives of
post-2009 (2010–2013), during the global economic recovery. It has annual reports. Dong et al. [198] automatically captures various signals
demonstrated the potential of using quarterly earnings calls to predict including sentiment and emotional attributes, topic characteristics,
short-term stock volatility. Another study by Rekabsaz et al. [190] lexical features, and social network dynamics from financial social
explored volatility forecasting using SVM with Radial Basis Function media data, which are subsequently integrated into machine learning
(RBF) by employing textual features extracted from financial disclo- classifiers to detect corporate fraud. Craja et al. [199] detects fraud
sures, such as Term Count (TC), Term Frequency (TF), TF–IDF, and from financial statements using a hierarchical attention network (HAN)
BM25, combined with a dimension reduction strategy using Principal to extract text features from annual reports including the management’s
Component Analysis (PCA). Additionally, the study integrated market discussion and analysis section. Achakzai and Peng [200] proposed
features like current volatility and sector specifics from factual market a Dynamic Ensemble Section (DES) algorithm to detect fraudulent
data. This sentiment analysis approach notably surpassed existing lead- firms. A novel approach involves integrating a knowledge graph for
ing methods, demonstrating that data from 10-K reports are valuable fraud detection. Specifically, Mao et al. [201] developed a related-
for predicting volatility. Xing et al. [191] introduced the Sentiment- party transactions (RPTs) knowledge graph, by using techniques such
Aware Volatility Forecasting (SAVING) model, combining symbolic and as information extraction, entity extraction, relation extraction, and
sub-symbolic AI approaches by infusing grounded knowledge into neu- attribute extraction, to detect financial fraud.
ral networks. This model uses market sentiment to enhance predictions
of stock return fluctuations. It surpasses traditional statistical models 3.6.3. Credit scoring
like GARCH and its variants, the Gaussian-process volatility model, and This section explores how NLP can enhance the qualification and
the latest neural stochastic volatility model and variational recurrent quantification of an individual’s creditworthiness. Traditional credit
neural networks. In a separate study, Deveikyte et al. [192] developed scoring methods rely heavily on factors such as credit history, income,
a classifier for market volatility using Latent Dirichlet Allocation for and outstanding debts. However, NLP techniques offer an alternative
topic modeling. The study demonstrates a significant negative corre- approach by analyzing an individual’s digital footprint, including so-
lation between positive social media sentiments, such as tweets, and cial media messages, website interactions, and online browsing and

11
K. Du et al. Information Fusion 115 (2025) 102755

Table 9
Financial risk.
Literature Data source Method Performance
Wang et al. [186] 10-K reports SVR Regression (MSE: 0.14894),
Ranking (Kendall’s Tau: 0.60458, Spearman’s Rho: 0.63403)
Wang and Hua [189] Earning calls Linear Regression, Linear SVM, Spearman (Pre-2009: 0.425, 2009: 0.422, Post-2009: 0.375)
Gaussian SVM, Gaussian Copula Models Kendall: (Pre-2009: 0.315, 2009: 0.310, Post-2009: 0.282)
Oliveira et al. [54] Twitter MR, NN, SVM, RF, Ensemble Daily return (Lowest NMAE: NDQ, 7.58)
Daily trading volume (Lowest NMAE: DJIA, 5.84)
Daily volatility (Lowest NMAE: DJIA, 2.79)
Rekabsaz et al. [190] 10-K reports GARCH, SVM MSE: 0.111, 𝑅2 : 0.527
Xing et al. [191] StockTwits SAVING, (E)GARCH, TARCH, GJR, Negative Log-Likelihood (NLL): −3.0642
GP-vol, VRNN, NSVM, LSTM, s+LSTM
Deveikyte et al. [192] RavenPack, Twitter, Logistic Regression Accuracy (Headlines: 65%, Tweets: 65%, Stories: 67%)
Thomson Reuters F1-Score (Headlines: 64%, Tweets: 64%, Stories: 63%)

Table 10
Fraud detection.
Literature Task Method Performance
Rodríguez et al. [194] Fraudulent transactions RF, IF, SVM, KNN, LSTM, transformer F1-Score: 0.830, MCC: 0.801, AUC: 0.871
Yang et al. [196] Fraud detection from financial texts FinChain-BERT Precision: 0.97, Recall: 0.96, Accuracy: 0.97
Boulieris et al. [195] Fraudulent transactions LR, RF, KNN, SVM, LSTM, CNN, TCN AUPRC: 0.404, F1 Score: 0.467
Craja et al. [199] Financial statement fraud LR, RF, SVM, XGB, ANN, HAN, GPT-2 w/ Attn AUC: 0.9264, F1-score: 0.7500
Chen et al. [197] Financial statement fraud QGA-SVM Accuracy: 0.852482
Achakzai and Peng [200] Financial statement fraud Dynamic ensemble selection algorithm AUC: 0.763, MCC: 0.354, AP: 0.443
Seemakurthi et al. [202] Financial statement fraud LR, SVM, NN, LDA, Ensemble Accuracy: 0.887, Specificity: 0.58, AUC: 0.72
Dong et al. [198] Corporate fraud detection SVM, NN, DT, LR Accuracy: 0.8000, Recall: 0.8304,
F1-Score: 0.7908, AUC: 0.8503
Mao et al. [201] Corporate fraud detection LR, DT, RF, XGB Accuracy: 0.8502, AUC: 0.7508

purchasing behaviors, to assess their ability to repay debts. NLP-based (LDA), Logistic Regression (LR), Support Vector Machine (SVM), Ran-
credit scoring has the potential to provide financial solutions to un- dom Forest (RF) and XGBoost (XGB), and deep learning models, such
banked communities that may not have access to traditional credit as Average Embedding Neural Networks (AE), CNN, RNN, LSTM, BERT
evaluation methods. By leveraging alternative data sources, NLP tech- and RoBERTa, are widely adopted.
niques can enhance the assessment of an individual’s creditworthiness
without the need for intrusive data collection methods, thereby offering 3.7. Financial regulatory compliance monitoring
a more accessible and inclusive approach (see Table 11). In the realm
of business credit scoring, the extensive corpus of texts issued by or The potential of AI in enhancing financial regulatory compliance
related to companies, including management reports, news articles, and monitoring is significantly bolstered by NLP technologies [216]. The
social media content, offers significant opportunities for gaining deeper integration of NLP into financial regulatory compliance monitoring
insights into a company’s financial health [203]. has become increasingly transformative, potentially reshaping the land-
Markov et al. [204] reviewed recent trends and considerations, scape of regulatory compliance management [217]. Despite these ad-
noting a sustained increase in research focused on credit scoring. The vancements, research in specific application areas remains sparse. NLP
research on NLP application in credit scoring can be broadly cate- is applied to analyze communications within financial institutions to
gorized into consumer credit scoring [205–207] such as default risk ensure compliance with regulatory requirements. It examines emails,
and business credit scoring [208–210] such as default risk, corporate chats, and documents for any suspicious content or language that
distress and bankruptcy prediction [211]. Within the consumer credit might suggest manipulative actions or insider trading, boosting the
scoring category, the focus is predominantly on peer-to-peer (P2P) surveillance effectiveness and proactively preventing potential breaches
lending, Wang et al. [205] proposed a novel consumer credit scoring of compliance [218]. Additionally, Al-Shabandar et al. [219] empha-
approach using attention-based LSTM by incorporating the sequence sized the importance of real-time monitoring of transaction data, the
of events in the online user operation behavior data, outperforming identification of anomalies, and the application of NLP to align reg-
models without the operation behavior data, suggesting the value of ulatory rules with institutional data. Furthermore, Abualhaija et al.
user actions for improved credit risk assessment. Kriebel and Stitz [206] [220] introduced an automated approach for monitoring and analyzing
adopted machine learning and deep learning methods to predict the changes in financial regulations.
risk of default using user-generated text such as the brief description of
themselves and the reason for their loan request and the results indicate 3.8. Explainable AI in finance
that the text provided by users is valuable for predictions. Netzer et al.
[207] implemented text mining and machine learning techniques to The notion of explainability holds tremendous importance in highly
automatically process and analyze raw text in loan requests, and the regulated domains such as finance where decisions can have significant
findings indicated that incorporating textual information from the loan consequences [221,222]. Yeo et al. [223] has categorized the expla-
into the predictive model significantly enhances the prediction of loan nation procedure into feature relevance, simplification, by example,
defaults. In terms of business credit scoring, Stevenson et al. [208] pro- visual, and textual. Most studies in NLP in finance, such as credit scor-
posed a deep learning approach using BERT to demonstrate the value of ing, financial forecasting, and FSA, emphasize explainability through
text for small business default prediction. Matin et al. [209] predicted visual, feature relevance, and simplification procedures. Specifically, in
corporate distress using deep learning of text segments in annual re- visual explanation, Kumar et al. [224] introduced a method called Class
ports and Mai et al. [210] predicted corporate bankruptcy forecasting Enhanced Attentive Response (CLEAR), which utilizes deconvolution
using textual disclosures in annual reports. The credit scoring task is on the penultimate layer before the output to create a visual attention
formulated into a classification problem. The machine learning models map. This method graphically highlights the periods with the highest
such as Multilayer Perceptron (MLP), Linear Discrimination Analysis degree of attention by the stock-picking agent, along with a separate

12
K. Du et al. Information Fusion 115 (2025) 102755

Table 11
Credit scoring.
Literature Data source Method Performance
Wang et al. [205] Online operation behavior data of borrowers Attentive LSTM AUC: 0.672, KS: 0.28
for peer-to-peer (P2P) lending
Kriebel and Stitz [206] Text such as brief description of themselves and RF, XGB, CNN, RNN, AE, BERT, AUC (Incl. text: 0.712,
the reason for their loan request provided by users Convolutional RNN, and RoBERTa Excl. text: 0.702)
Yang et al. [212] Personality mining from social media LDA, NB, SVM, MLP, and LR Precision: 0.8325, Recall: 0.9437
information of borrowers F1-Score: 0.8830, Accuracy: 0.8757
Stevenson et al. [208] Short description of the applicant’s business LR, RF, BERT AUC (New: 0.86, Existing: 0.88)
context, the requirement of the loan etc.
Netzer et al. [207] Text in loan requests LR, RF AUC: 0.7260
Gao et al. [213] Written description of the borrower’s financial Hypothesis testing Marginal effects are reported to
situation and purpose of the loan show statistical significance
Matin et al. [209] Corporate distress prediction using annual CNN-LSTM, XGB, LR AUC: 0.844, Log score: 0.1064
reports and financial
Mai et al. [210] Corporate bankruptcy prediction using textual CNN, LR, SVM, RF Accuracy: 0.712, AUC: 0.856
disclosures in annual reports
Ahmadi et al. [203] Corporate bankruptcy prediction using business Dependency Sensitive Cohen’s Kappa: 0.6544,
management reports Convolutional Neural Networks Accuracy: 0.8414, F1-Score: 0.8760
El-Qadi et al. [214] Comments including company description, LR, SVM, RF, XGB, LightGBM, F1-Score: LR-0.209, Accuracy: MLP-0.923
business context, and activities MLP Precision: RF-0.340
Nguyen and Huynh [215] Assess manager’s reflections towards company LR Pseudo R2 : 0.2008
performance in financial reports

plot corresponding to the sentiment class of the stock. Deng et al. [225], streams, which have attracted researchers from both academia [236]
on the other hand, applied knowledge graphs to visually link event and industry [237–240]. The first focuses on ESG disclosure, measure-
entities that are extracted from stock news, offering a graphical repre- ment, and governance including the identification of ESG topics and
sentation of the relationships between features and their corresponding issues [236,241–245], addressing how NLP can enhance transparency,
predictions. In terms of feature relevance, Carta et al. [226] explored accuracy, and effectiveness in reporting and managing ESG criteria.
the explainability of feature relevance by adopting various configura- The second stream explores the integration of ESG factors into broader
tions of a permutation importance technique to prune less significant financial applications, seeking innovative ways to incorporate ESG con-
technical indicators, followed by the implementation of decision tree siderations into sentiment analysis [246], portfolio management [247],
techniques for predicting stock market trends. The proposed method and risk management [248–251] etc. This distinction emphasizes the
was found to be more reliable than LIME. Similarly, Ong et al. [227] role of NLP in enhancing the ESG reporting framework and its potential
used aspect-based sentiment analysis to investigate the correlation to weave ESG considerations into the broader landscape of financial
between stock price movements and key aspects identified in tweets, decision-making and product development.
determining the sentiment of each aspect through a SenticNet-based Table 12 presents the research on NLP in ESG and sustainable
graph convolutional network (GCN) [228]. This approach mirrors the finance. From the standpoint of NLP techniques, the fine-tuning of
feature relevance technique, with its focus on identifying key con- transformer-based models for ESG criteria has emerged as a prominent
tributing aspects and their associated polarity values, emphasizing area of research [239,244,251]. Specifically, Raman et al. [240] used
the interplay among financial variables rather than direct financial pre-trained BERT, XLNet, and RoBERTa to analyze the transcripts of
forecasting. For simplification procedure, Bandi et al. [229] combined corporate earning calls and detect historical trends of ESG discus-
sentiment analysis and technical analysis to develop a random forest sions. Mehra et al. [251] proposed ESGBERT for classification tasks
model for stock forecasting, explained through LIME. Moreover, Gite related to corporate ESG practices. Chen et al. [236] adopted BERT-
et al. [230] employed LIME together with the LSTM-CNN model to like language models with data augmentation for multi-lingual ESG
effectively identify keywords that align with the target sentiment. In issue identification. The same ESG issue identification task is per-
related work, Yuan and Zhang [231] adopted GPT-2 [232] for generat- formed by Glenn et al. [237] using synthetic data and transfer learning
ing text explanations, incorporating specific keywords in the generated and Wang et al. [238] by contrastive learning with BERT.
text. The presented method, named SC-DBA (soft-constrained dynamic
beam allocation), uses a distinct network designed for analyzing news 3.10. NLP for digital assets
titles to extract keywords associated with different tiers of anticipated
market volatility. The efficacy of the proposed method is evaluated, Another promising application of NLP in finance is within the realm
based on the fluency and practical relevance of the generated expla- of digital assets. NLP for digital assets is a distinct area of application,
nation. Specifically for NLP, Danilevsky et al. [233] categorized XAI reflecting our belief in NLP’s potential to significantly influence the
for NLP into local versus global and self-explaining versus post-hoc, landscape of digital assets. Over the past decade, there has been an
with explainability techniques including feature importance, surro- explosion of digital assets and digitalization of financial services [258].
gate model, example-driven and provenance-based, and declarative Cryptocurrencies and Non-Fungible Tokens (NFTs) represent two of
induction. the most significant types of digital assets. Cryptocurrencies, or digital
currencies, leverage cryptographic techniques to offer secure and de-
3.9. ESG and sustainable finance centralized financial transactions. Bitcoin (BTC) and Ethereum (ETH)
are the most well-known cryptocurrencies. On the other hand, NFTs,
Lim [234] conducted an extensive review of ESG and AI in the share the digital token format with cryptocurrencies but differ signif-
finance sector. However, our primary interest lies in the application of icantly in their nature and usage. Unlike cryptocurrencies, which are
NLP techniques within this domain. While research like that of Capelli fungible and can be exchanged on a one-to-one basis like traditional
et al. [235], which integrates structured ESG data into financial fore- money, NFTs are inherently unique and cannot be exchanged on a like-
casting, falls within the broader category of ESG and AI in finance, for-like basis. This uniqueness allows NFTs to represent ownership of
it does not align with our specific focus on NLP methodologies. The specific, often one-of-a-kind digital items such as artwork, collectibles,
NLP research in ESG and sustainable finance encompasses two primary and even real estate in virtual worlds. Both cryptocurrencies and NFTs

13
K. Du et al. Information Fusion 115 (2025) 102755

Table 12
NLP in ESG and sustainable finance.
Literature Task Method Performance
Guo et al. [248] Predict stock volatility BERT, Bayesian Model Ensemble RMSE (MSCI-US: 0.289, AC-EU: 0.281)
MAE (MSCI-US: 0.249, AC-EU: 0.241)
Apel et al. [249] Approximate changes in transition risk BERT and multivariate regression Significance is observed for pure-play
from climate-related news events and decarbonized indices
Jan [250] Detect information asymmetry for RNN, LSTM Accuracy: 0.9488, F1-Score: 0.9231,
sustainable development AUC: 0.9586
Mehra et al. [251] Classify environmental score change BERT Accuracy for Change/No Change: 0.6709
Accuracy for Positive/Negative: 0.793
Sandwidi and Mukkolakal [246] ESG aspect-oriented sentiment analysis RoBERTa, ESG fine-tuning, Accuracy: 0.9130, F1-Score: 0.9020
Attention Augmentation
Koloski et al. [241] Classify financial texts as sustainable or Knowledge-based latent heterogeneous Precision: 0.90, Recall: 0.89, F1-Score:
unsustainable representation 0.89
Chen et al. [236] Identify multi-lingual ESG issue BERT-like models with data F1-Score (English: 0.69,
augmentation
French: 0.78, Chinese: 0.392)
Haase and Sassen [252] Uncover lobbying strategies in Pre-trained hierarchical topic model Metadata helps to identify stakeholders
sustainable finance disclosure regulations and lobbying strategies
Kouloukoui et al. [243] Analyze tweets and identify the network Structural topic model The popularity of Twitter as a platform
of ESG topics and their trend to discuss CSR and ESG issues, clean
and cleaner production has increased
significantly
Lee et al. [253] Classify E, S, or G using ESG documents BERT, RoBERTa, ALBERT Accuracy: 0.8030, Recall: 0.79,
and calculate ESG rating Precision: 0.79, F1-Score: 0.79
Schimanski et al. [254] Classify Environment, Social, Governance FinBERT-ESG Accuracy: (E: 0.957, S: 0.934, G: 0.897)
Sokolov et al. [255] ESG categorization BERT AUC: 0.88 to 0.97, AUPR: 0.19 to 0.76
Pasch and Ehnes [239] Classify ESG behavior Fine-tuning transformer models for ESG Accuracy: 0.79, F1-Score: 0.78
Raman et al. [240] Detect historical trends of ESG discussion BERT, XLNet and RoBERTa Accuracy: 0.782, F1-Score: 0.784
Nugent et al. [244] Detect ESG topics BERT, Data Augmentation Precision: 0.84, Recall: 0.84, F1-Score:
0.84
Huang et al. [256] Classify ESG-related discussion FinBERT Accuracy: 0.895, Precision: 0.90
Recall: 0.895, F1-Score: 0.896
Fan and Wu [257] Study effect of environmental regulations Multivariate regression Stricter EPA regulations reduce pollution
on firm valuation and policies and increase firm innovation, especially
green innovation

operate on blockchain technology, ensuring transparency, security, and Weighted Cosine Similarity (WCS), Mean Squared Error (MSE), Mean
permanence in transactions and ownership records. Absolute Percentage Error (MAPE), and coefficient of determination or
Research in NLP for digital assets primarily focuses on predict- R-squared (𝑅2 ).
∑𝑛
ing trends in the cryptocurrency market and providing insights into |𝑃 | 𝑖=1 (𝐺𝑖 × 𝑃𝑖 )
WCS = ×√ √∑ (1)
the NFTs market. Cryptocurrencies have seen unprecedented value |𝐺| ∑𝑛 2) × 𝑛 2)
growth, surpassing significant historical bubbles [259]. Recent studies 𝑖=1 (𝐺 𝑖 (𝑃
𝑖=1 𝑖
have deep-dived into cryptocurrency market dynamics, as shown in where P is the vector of scores predicted by the model and G is the
Table 13, focusing on causality and correlation to understand behav- vector of ground truth scores.
ior over time [260–262]. It has been shown that investor sentiment
1∑
𝑛
plays a crucial role in predicting returns of major cryptocurrencies, MSE = (𝑦 − 𝑦̂𝑖 )2 (2)
𝑛 𝑖=1 𝑖
with nonlinear effects [263]. A hidden Markov Model was developed
1 ∑ 𝑦𝑖 − 𝑦̂𝑖
𝑛
to predict market trends, based on sentiment, trading volume, and
MAPE = | | (3)
price, finding that markets react differently to sentiments depend- 𝑛 𝑖=1 𝑦𝑖
ing on the overall market condition; however, this research focused
where 𝑦𝑖 is the gold standard score and 𝑦̂𝑖 is the score predicted by the
solely on Bitcoin [264]. Another study employed an LSTM-based RNN
model.
model to analyze Chinese social media sentiment’s impact on cryptocur- ∑𝑛
(𝑦𝑖 − 𝑦̂𝑖 )
rency prices, showing improved precision and recall over traditional 𝑅2 = 1 − ∑𝑛𝑖=1 (4)
autoregressive models [265]. 𝑖=1 𝑖 − 𝑦̄𝑖 )
(𝑦
In terms of market insights for NFTs, Leitter and Cambria [266] where 𝑦𝑖 is the gold standard score and 𝑦̂𝑖 is the score predicted by the
analyzed 200,000 tweets about NFTs using state-of-the-art neurosym- model.
bolic AI tools, which aims to decode the drivers of online conversations The second type of metrics measures the categorical accuracy be-
and sentiments surrounding NFTs, thereby uncovering factors that tween predicted value and ground truth in the context of classification
contribute to their perceived value. Additionally, Meyns and Dalipi tasks. The popular metrics include Accuracy, Matthews Correlation
[267] examined the human perceptions of, or attitudes towards, NFTs, Coefficient (MCC), Area Under the Curve (AUC), F1-Score, and the
specifically aiming to identify concerns expressed by social media users more generalized 𝐹𝛽 -Score where 𝛽 can be 1 and 2 for instance.
engaged with NFTs on Twitter. 𝑇𝑃 + 𝑇𝑁
Accuracy = (5)
𝑇𝑃 + 𝐹𝑃 + 𝑇𝑁 + 𝐹𝑁
3.11. Benchmarks and evaluation metrics 𝑇𝑃 × 𝑇𝑁 − 𝐹𝑃 × 𝐹𝑁
MCC = √ (6)
(𝑇 𝑃 + 𝐹 𝑃 ) × (𝑇 𝑃 + 𝐹 𝑁) × (𝑇 𝑁 + 𝐹 𝑃 ) × (𝑇 𝑁 + 𝐹 𝑁)
We summarize the common benchmark datasets and evaluation
TP
metrics for various applications in Table 14. The first type of met- TPR = (7)
TP + FN
rics measures the closeness between the predicted value and ground
FP
truth in the context of a regression task. The popular metrics include FPR = (8)
FP + TN

14
K. Du et al. Information Fusion 115 (2025) 102755

Table 13
NLP for cryptocurrency market.
Literature Data source Method Task Performance
Shen et al. [260] Twitter Granger causality test Volume, return and realized volatility Significant for Bitcoin RV and volume
Vector Autoregression
Kraaijeveld and De Smedt [261] Twitter Granger causality test Cryptocurrency price return Significant predictive power (p < 0.05) on
Bitcoin, Bitcoin Cash and Litecoin
Gao et al. [268] Twitter Regression Return and volatility Sentiment from Twitter can predict Bitcoin
returns and volatility, based on significance level
Kim et al. [264] Google Trends Hidden Markov Model Market movement prediction Accuracy: 54%, AUC: 52%
Twitter
Huang et al. [265] Sina-Weibo Autoregression, LSTM Market movement prediction Precision: 87%, Recall: 92.5%
Saha et al. [269] Twitter LSTM Financial sentiment, Sentiment: Accuracy: 0.8352, F1-Score: 0.8515
BTC volume correlation BTC Volume: Pearson’s R: 0.1584
Oikonomopoulos et al. [262] Twitter Granger causality test Price MAPE: 0.0038
Vector Autoregression

Table 14
Benchmarks and evaluation metrics.
Research topic Benchmark dataset Common evaluation metrics
Financial sentiment analysis PhraseBank, SemEval 2017 Task 5, Accuracy, MCC, F1-Score, WCS, MSE, 𝑅2
FiQA Task 1, StockSen, SentiEcon GS-1000,
SEntFiN, FinLin, Twitter Financial News
Financial forecasting BIGDATA22, ACL18, CIKM18, EDT Accuracy, MCC, F1-Score, WCS, MSE, MAE, 𝑅2
Portfolio management No public benchmark dataset has been identified MAE, Accuracy, MCC, Return, Sharpe ratio, Sterling Ratio, Sortino Ratio
Financial narrative processing FNP, FINDSum, ECTSum, EDT, ROUGE, F1-Score, Precision, Recall, Silhouette Coefficient, Dunn Index
CGRAPH, REFinD, Chinese FinNER
Question answering, virtual assistant ConvFinQA, FinQA, TAT-QA, PACIFIC Accuracy, Exact Match, Numeracy-focused F1-Score, ROUGE
and chatbot
Risk management No public benchmark dataset has been identified Accuracy, F1-Score, Precision, Recall, AUC, AUPRC, Kendall’s Tau,
Spearman’s Rho, 𝑅2 , MSE, NMAE, NLL, Cohen’s Kappa
Regulatory compliance monitoring No public benchmark dataset has been identified No common metrics has been identified
ESG and sustainable finance FinSim-ESG Accuracy, AUC, F1-Score, Precision, Recall, RMSE, MAE
Explainable AI in finance No public benchmark dataset has been identified No common metrics has been identified
NLP for digital assets No public benchmark dataset has been identified Accuracy, AUC, F-Score, MAPE, Pearson’s Correlation Coefficient

1 𝐿_𝑐 𝑟
AUC = TPR 𝑑(FPR) (9) Precision𝑅𝑂𝑈 𝐺𝐸−𝐿 = (18)
∫0 𝑁_𝑐
TP Precision𝑅𝑂𝑈 𝐺𝐸−𝐿 × Recall𝑅𝑂𝑈 𝐺𝐸−𝐿
Recall = (10) F1-Score𝑅𝑂𝑈 𝐺𝐸−𝐿 =2× (19)
TP + FN Precision𝑅𝑂𝑈 𝐺𝐸−𝐿 + 𝑅𝑒𝑐 𝑎𝑙𝑙𝑅𝑂𝑈 𝐺𝐸−𝐿
TP
Precision = (11) Exact-match accuracy (EM) and numeracy-focused F1 are intro-
TP + FP
duced into the evaluation of question answering. EM is set to 1 when
Precision × Recall
F1-Score = 2 × (12) the characters of the prediction exactly match the characters of (one of)
Precision + Recall
the true answer(s), otherwise will be 0. Numeracy-focused F1 defines
Precision × Recall
𝐹𝛽 -Score = (1 + 𝛽 2 ) × (13) F1 to be 0 when there is a number mismatch between the gold and
(𝛽 2 × Precision) + Recall
predicted answers, regardless of other word overlap. In addition to
execution accuracy, Chen et al. [154] also proposed program accuracy,
In addition, Recall-Oriented Understudy for Gisting Evaluation
by replacing all the arguments in a program with symbols, and then
(ROUGE) is widely employed in assessing the quality of automatic
evaluating if two symbolic programs are mathematically equivalent.
summarization, machine translation and question answering. ROUGE-
For example, the following two programs are equivalent programs:
1 evaluates the overlap of individual words between the generated
text (i.e., candidate) and a reference. ROUGE-2, on the other hand, add(𝑎1 , 𝑎2 ), add(𝑎3 , 𝑎4 ), subtract(#0, #1)
measures the overlap of bigrams. ROUGE-N generalizes this approach add(𝑎4 , 𝑎3 ), add(𝑎1 , 𝑎2 ), subtract(#1, #0)
to N-grams, which can include both unigrams (one word) and bigrams
(two words). Lastly, ROUGE-L focuses on the Longest Common Subse-
quence, assessing the longest sequence of words that appears in both For ranking tasks, Kendall’s Tau (𝜏) and Spearman’s Rho (𝜌) are
the generated text and the reference in the same order. Given candidate commonly used non-parametric metrics to measure the strength and
𝐶 and reference 𝑅, 𝑂_𝑐 𝑟 represents the number of overlapping N-grams, direction of the association between two ranked variables.
𝑛 − 𝑛𝑑
𝐿_𝑐 𝑟 represents the longest common subsequence, 𝑁_𝑐 is the number of 𝜏= 𝑐 (20)
1
N-grams in candidate and 𝑁_𝑟 is the number of N-grams in reference, 2
𝑛(𝑛 − 1)
ROUGE-N and ROUGE-L metrics can be calculated as follows:
where 𝑛 represents the sample size, 𝑛𝑐 is the number of concordant
𝑂_𝑐 𝑟
Recall𝑅𝑂𝑈 𝐺𝐸−𝑁 = (14) pairs, and 𝑛𝑑 is the number of discordant pairs.
𝑁_𝑟 ∑
6 𝑛𝑖=1 𝑑𝑖2
𝑂_𝑐 𝑟 𝜌=1− (21)
Precision𝑅𝑂𝑈 𝐺𝐸−𝑁 = (15) 𝑛(𝑛2 − 1)
𝑁_𝑐
Precision𝑅𝑂𝑈 𝐺𝐸−𝑁 × Recall𝑅𝑂𝑈 𝐺𝐸−𝑁 where 𝑑𝑖 is the difference in rankings for each object 𝑖, 𝑖 ∈ {1, 2, … , 𝑛}.
F1-Score𝑅𝑂𝑈 𝐺𝐸−𝑁 = 2 × (16) Portfolio management involves comparing the returns of an in-
Precision𝑅𝑂𝑈 𝐺𝐸−𝑁 + 𝑅𝑒𝑐 𝑎𝑙𝑙𝑅𝑂𝑈 𝐺𝐸−𝑁
vestment against its associated risks. Commonly used metrics for this
𝐿_𝑐 𝑟
Recall𝑅𝑂𝑈 𝐺𝐸−𝐿 = (17) purpose include the Sharpe ratio, Sterling Ratio, and Sortino Ratio,
𝑁_𝑟

15
K. Du et al. Information Fusion 115 (2025) 102755

which are employed to evaluate the effectiveness of the proposed NLP in XAI in finance, portfolio management, ESG and sustainable
portfolio selection strategy. Given 𝑅𝑝 the return of a portfolio, 𝑅𝑓 the finance, and financial forecasting are at the initial awareness stage, and
risk-free rate, 𝜎𝑝 the standard deviation of the portfolio’s excess return, have garnered significant enthusiasm for their potential to drive tech-
and 𝜎𝑑 the standard deviation of the downside. nological breakthroughs, although their practical impact may not yet be
R𝑝 − R𝑓 fully realized. Generative AI, particularly promising in wealth manage-
Sharpe Ratio = (22)
𝜎𝑝 ment, has the capability to revolutionize the industry by synthesizing
information and creating personalized advisory services and investment
The Sterling Ratio evaluates the risk-adjusted returns of an invest- strategies. Meanwhile, NLP applications in financial risk prediction,
ment by focusing on drawdowns rather than volatility. financial regulatory compliance monitoring, fraud detection, and credit
R𝑝 − R𝑓 scoring are now undergoing a more realistic assessment of their chal-
Sterling Ratio = (23)
Average Annual Max Drawdown lenges and limitations during operational integration, striving to meet
the expectations for accuracy and reliability of early adopters. NLP
The Sortino ratio, a modified version of the Sharpe ratio, dis- technologies, such as financial sentiment analysis, information extrac-
tinguishes detrimental volatility from overall volatility by using the tion, financial text summarization, QA systems, and virtual assistants,
standard deviation of negative returns in a portfolio. and chatbots, are beginning to demonstrate clear benefits for enter-
R𝑝 − R𝑓
Sortino Ratio = (24) prises and are becoming more widely understood, analyzing financial
𝜎𝑑 documents and markets, supporting decision-making processes across
the finance sector.

4. Main findings
4.3. Trends, challenges, and opportunities
The application of NLP in the financial sector has been transfor-
mative across various facets, from enhancing operational efficiencies
To address our third research question: What are the prevailing
to enabling more personalized customer engagements and improving
trends in NLP research within the finance sector, and how do these
strategic decision-making processes. We summarize our main findings
trends reflect the evolving challenges and opportunities in the field? We
from the following aspects.
summarize the trends, challenges, and opportunities from the following
4.1. Growing importance of NLP in finance key perspectives.

To address our first research question: What are the key drivers
4.3.1. Increasing reasoning capabilities
behind the increasing importance of NLP in the finance sector, and how
The ability to perform complex numerical reasoning and understand
do these drivers address the unique challenges and opportunities within
this industry? The growing importance of NLP in the finance sector is heterogeneous representations is particularly important in the finance
driven by several key factors. First, the exponential growth in data, domain. Financial data encompasses a broad spectrum of formats and
including unstructured data from financial documents, news, reports, content types, from numerical tables to varied textual documents. As
social media, etc, necessitates sophisticated tools like NLP for efficient such, NLP models operating within the finance domain must not only
data processing and insight generation. Second, the advancement of parse and interpret the data but also reason across. Applications such
NLP technologies, particularly from pre-trained language models to as QA demand the ability to infer answers from fragmented or implicit
large language models, plays a crucial role in propelling the use of NLP data, while FSA requires models to assess the sentiment and potential
in finance. NLP enables enhanced financial forecasting and sentiment market impacts of nuanced textual information. Similarly, tasks like
analysis, allowing firms to make more informed decisions. Additionally, financial textual summarization and information extraction call for the
the complexity and volume of financial regulations demand advanced capacity to distill and synthesize essential information from extensive
solutions for compliance monitoring, where NLP can automate and textual and numerical data efficiently. Hence, increasing the reasoning
streamline the extraction and management of pertinent information. capabilities of NLP models is crucial, as it enhances their ability to
Moreover, as financial services increasingly move towards digital in- make informed decisions and predictions by effectively navigating and
terfaces, NLP-powered virtual assistants and chatbots are essential for integrating the complex, multi-faceted landscape of financial data.
improving customer interaction and service efficiency. Lastly, the need
for transparency and better risk management prompts the integration
of NLP in areas like fraud detection and financial risk prediction, help- 4.3.2. Understanding tabular and textual data
ing firms identify potential threats faster and with greater accuracy. The application of NLP in finance demands adept handling of both
These drivers collectively address the unique challenges of managing tabular and textual data, reflecting the complex nature of financial
vast amounts of diverse data, ensuring compliance, enhancing user
information which encompasses both structured numerical data and
engagement, and bolstering security in the finance industry.
unstructured textual content. Tabular data, which includes financial
ratios, stock prices, and transaction histories, provides essential quan-
4.2. Stages of maturity and adoption of NLP in finance
titative insights fundamental to financial analysis. On the other hand,
textual data from news articles, financial reports, and regulatory fil-
To address our second research question: How has NLP transformed
financial services and products from the past to the present, and what ings introduces qualitative information that can profoundly influence
are the anticipated future developments in this area? What emerging market movements and investment strategies. This combination of
areas of application for NLP in finance are being explored, and what capabilities enables NLP to not only extract meaningful insights from
potential do these areas hold for the future of financial services? We the vast and varied data streams but also enhance more nuanced
assess the stages of maturity and adoption of various NLP applications and context-aware financial decision-making. Current research in fields
in finance as shown in Fig. 3. We aim to evaluate their potential to such as QA, FSA, and FNP focus on advancing these capabilities,
address real-world business challenges and to identify new opportuni- improving how models comprehend and integrate both forms of data
ties for innovation and growth within the financial sector. Currently, in financial contexts.

16
K. Du et al. Information Fusion 115 (2025) 102755

Fig. 3. Stages of maturity and adoption of NLP in finance.

4.3.3. Leveraging large language models and product development, thereby promoting a more sustainable and
Generative AI, particularly LLMs, is revolutionizing the financial transparent financial landscape.
sector by automating complex decision-making processes, personalizing
customer experiences, synthesizing information, generating financial 4.3.5. Transforming portfolio management
insights, and providing advisory services [270]. By leveraging vast NLP is potentially revolutionizing the field of portfolio management
structured and unstructured datasets, they not only enhance opera- by enhancing the speed and intelligence of decision-making processes.
tional efficiency but also drive the creation of innovative financial It enables real-time data processing and analysis, offering predictive
products and services, transforming banking into a more agile and insights that refine investment strategies using current market data and
customer-centric industry. LLMs are reshaping the landscape of NLP trends. An essential application of NLP is in portfolio optimization,
in finance [271–273], influencing areas from FSA, FNP to QA, Virtual where it analyzes quantitative and qualitative data to identify the best
Assistant and Chatbot. Advancements in prompt engineering, retrieval- investment mix to maximize returns or minimize risk. Techniques such
augmented generation (RAG), and multi-agent systems are demon- as Gaussian inverse reinforcement learning are employed to model
strating significant potential for use in industry. Additionally, ongoing investor sentiments and market dynamics, enabling the development
research is addressing challenges faced by LLMs, such as hallucination, of trading systems that effectively filter out market noise. Additionally,
enhancing the ability of NLP systems to generate contextually relevant strategies, e.g., PROFIT and SARL utilize financial news and social me-
and highly precise responses in complex financial scenarios. dia to optimize trading decisions, showing marked improvements over
traditional benchmarks. Furthermore, NLP models incorporate senti-
4.3.4. Capturing the sustainability wave ments from investor posts and public moods online, enhancing stock
Leveraging NLP to sift through the massive volumes of ESG data market investments and portfolio allocations. This sentiment-driven
facilitates informed decision-making for sustainability investments and approach not only aligns with dynamic market conditions but also con-
ensures compliance with evolving regulatory frameworks. This critical sistently outperforms equal-weighted strategies. Advanced models, e.g.,
application of NLP is pivotal in aligning financial strategies with global time-aware LSTM and those using BERT for sentiment analysis, have
sustainability objectives. By focusing primarily on the utilization of NLP proven their efficacy by significantly surpassing conventional methods
methodologies within the finance sector, we identify two main research in predicting stock returns and managing portfolio risks. The integra-
streams in the domain of ESG and sustainable finance. The first stream tion of these innovative NLP strategies into portfolio management is
emphasizes ESG disclosure, measurement, and governance, spotlighting setting new standards for achieving superior financial outcomes.
the role of NLP in enhancing transparency, accuracy, and effectiveness
in reporting and managing ESG criteria. The second stream delves into 4.3.6. Improving risk management and regulatory compliance monitoring
the integration of ESG factors into broader financial applications such NLP is transforming the landscape of financial risk management and
as sentiment analysis, portfolio management, and risk management. regulatory compliance monitoring. By leveraging the rich information
This highlights the transformative potential of NLP to embed ESG contained in various financial documents, such as earnings calls, finan-
considerations seamlessly into the financial decision-making process cial reports, and social media data, NLP facilitates more precise and

17
K. Du et al. Information Fusion 115 (2025) 102755

predictive risk assessment models. In financial risk prediction, studies In the field of information extraction, a key trend is leveraging
have leveraged sentiment analysis and textual features from financial pre-trained language models like BERT, RoBERTa, and GPT for tasks
reports, news, and earnings calls to forecast stock volatility and fi- such as event extraction and financial relation identification, which are
nancial risk with notable accuracy. For instance, models that utilize demonstrating promising results in improving the accuracy and depth
sentiment from financial statements have shown strong correlations of extracted information, with knowledge graphs further enhancing
with financial risk indicators, thereby improving prediction capabilities the models’ contextual understanding. The challenges include handling
over traditional models. In fraud detection, NLP techniques analyze the complex, dispersed information typical of financial texts and the
vast arrays of unstructured data, such as transaction narratives and resource-intensive nature of creating well-annotated datasets necessary
corporate financial statements, to identify suspicious patterns indicative for effective model training. However, there is potential for expanding
of fraudulent activities. Innovative NLP applications like transformer the use of weakly supervised and unsupervised learning methods to
models and knowledge graphs have demonstrated enhanced effec- reduce dependency on extensive labeled datasets. Exploring multi-
tiveness in detecting complex financial fraud scenarios. Furthermore, task learning frameworks that simultaneously perform classification,
NLP has introduced new methodologies in credit scoring by assessing detection, and summarization of financial events can lead to more
individuals’ digital footprints, providing inclusive financial solutions efficient processing of financial documents. Moreover, the development
especially beneficial to unbanked populations. Additionally, in the of methods for detecting causality, such as those using predictive causal
realm of regulatory compliance, NLP has been instrumental in moni- graphs, offers new avenues for understanding the intricate dynamics of
toring internal communications within financial institutions, ensuring financial markets, providing insights into cause-and-effect relationships
adherence to regulatory norms, and proactively preventing compliance crucial for market prediction. Lastly, automated financial reporting and
breaches. Through these applications, NLP not only enhances the pre- financial statement analysis using NLP, particularly with LLMs, has also
cision and efficiency of risk and compliance management processes but emerged as a promising area of research.
also supports more informed decision-making in the finance sector.
4.3.9. Robustness and trustworthiness
4.3.7. Augmenting algorithmic trading NLP technologies have undergone significant advancements, be-
Natural language-based financial forecasting augments the land- coming increasingly sophisticated to address the specialized linguistic
scape of algorithmic trading by enabling the rapid processing and challenges inherent in the financial sector. This progression enhances
integration of financial documents, news, and social media content, precision and resilience, leading to reduced operational risks associ-
thereby improving model performance. This allows trading systems to ated with errors and biases. As NLP systems evolve, their ability to
quickly capitalize on market movements driven by real-time events, understand and process complex financial jargon and semantics im-
significantly enhancing the speed and effectiveness of trading strate- proves, thereby enhancing decision-making processes and compliance
gies. Natural language-based financial forecasting in this area includes monitoring in finance. Trustworthiness forms a crucial pillar in evalu-
forecasting in stock, foreign exchange, cryptocurrency, options pricing, ating AI systems within the financial industry. It measures the extent
firm valuation, etc. A significant trend involves incorporating com- to which AI models and algorithms can be relied upon to function
pany relationships into financial forecasting models, refining methods as expected, make decisions that are both accurate and ethical, and
for preprocessing textual data, and enhancing data representation. By mitigate potential adverse impacts. This concept is closely associated
analyzing textual data from diverse sources, NLP helps in identifying with ‘‘Intention Awareness’’, which pertains to the AI’s capability to
sentiment and key information that directly impacts financial markets, align its operations with intended goals. However, trustworthiness also
and predict market signals, enabling algorithmic traders to execute encompasses explainability and interpretability, key aspects of what
more timely and informed trades, based on the latest news and social is often referred to as XAI. In finance, XAI facilitates transparency
media trends, leading to potentially higher returns and more robust by enabling stakeholders to understand and trust the decision-making
trading strategies. From this perspective, NLP serves as a powerful tool processes of AI systems, thus bridging the gap between complex AI
in the arsenal of modern algorithmic trading, bridging the gap between operations and user-friendly outputs. These advancements signify a
unstructured data and actionable trading insights. pivotal shift towards more reliable and accountable NLP applications
in the financial sector, paving the way for broader acceptance and
4.3.8. Enhancing financial narrative processing integration of AI technologies in high-stakes environments. Lastly, the
Financial text summarization and information extraction are two presence of algorithmic bias, which inherits biases from the training
primary application areas of FNP, each significantly enhancing oper- data, can lead to decisions that disproportionately impact certain indi-
ational efficiency. A prominent trend in financial text summarization viduals and groups such as credit scoring. Consequently, the pursuit of
is the increasing reliance on both extractive and abstractive tech- fairness and equity in models is crucial within financial services. Lastly,
niques, with extractive methods particularly dominant in sentence the robustness and trustworthiness also extend to considerations of data
or section-level summarization using pre-trained language models. A privacy and security, especially when managing sensitive compliance-
key challenge in this field stems from the repetitiveness and density related data, which requires robust measures to ensure data privacy and
of financial texts, which complicate the summarization process. To prevent unauthorized access, safeguarding integrity and confidentiality.
address this, researchers have proposed models that focus on extracting
and summarizing key narrative sections rather than entire reports, thus 4.3.10. NLP system implementation
improving the relevance and utility of summaries. The integration of This section zooms into the practical aspects of NLP system imple-
extractive and abstractive methods, capitalizing on the strengths of mentation, focusing on data and platform integration, process integra-
both, has shown considerable potential. Innovative approaches such as tion, risk management, regulatory compliance, data privacy, scalability
reinforcement learning models and pointer networks, which combine and performance, ethics and fairness, and ROI estimation. Firstly, fi-
the extraction of key sentences with advanced paraphrasing techniques, nancial institutions are expected to seamlessly integrate both structured
are showing promise. Additionally, the use of LLMs for narrative sec- and unstructured data sources into their enterprise data platforms. This
tion identification and summarization is emerging as a powerful tool, requires the development of robust data ingestion and preprocessing
enhancing the quality of financial report summaries. There is also a pipelines that cleanse, normalize, and prepare data for NLP systems.
growing interest in developing specialized datasets and methods for Integrating these platforms necessitates compatibility with existing IT
summarizing diverse financial document formats, such as earnings call infrastructures, which may require upgrading legacy systems or deploy-
transcripts, which could further refine the precision and applicability ing new middleware solutions. Next, integrating NLP systems into the
of financial text summarization technologies. routine workflows of financial institutions demands forward planning.

18
K. Du et al. Information Fusion 115 (2025) 102755

It is crucial to synchronize NLP systems with established business corporate performance, or economic indicators, thus providing more
processes. Adopting change management strategies is essential to facili- nuanced sentiment insights that are crucial for investment decisions.
tate smooth transitions and foster user acceptance, ensuring the benefits Lastly, in the realm of compliance, there is a growing need to demon-
of NLP technologies are fully realized. Additionally, implementing NLP strate the effectiveness of NLP applications. Explainable AI models
systems introduces new risks associated with data accuracy, algorithmic in particular are crucial in finance, where stakeholders require clear,
bias, and system reliability. Comprehensive risk management protocols understandable explanations for automated decisions to ensure compli-
must be established to mitigate these issues. This includes continuous ance with regulatory standards and to build trust in AI-driven processes.
model monitoring, periodic updates to models to reflect evolving data, Overall, the application of NLP in finance is transitioning from its
and implementing fallback mechanisms to prevent system failures. initial, task-driven early stages to a more in-depth exploration and
Another key consideration is regulatory compliance, which requires adoption of innovative algorithms. This evolution enhances the sys-
NLP systems to be developed with an awareness of the regulatory tem’s ability to effectively handle complex data, thereby improving
frameworks governing financial institutions. Compliance with finan-
functionality and precision in financial contexts.
cial regulations, data protection laws, and international standards is
paramount. In terms of data privacy, financial institutions must pri-
oritize the safeguarding of sensitive information. NLP systems should 5. Conclusion
be designed to anonymize and secure personal data, ensuring com-
pliance with stringent data protection laws. Integrating robust data This survey conducted a comprehensive review of research in NLP in
privacy measures into system design and operations is essential, un- finance, targeting researchers, practitioners, and industry professionals.
derpinned by clear policies and procedures to prevent data breaches We began with the basic structure of an IMF industry report and
and unauthorized access. Scalability and performance are also crucial expanded to include a detailed review of research literature. This
for NLP systems. As financial institutions manage ever-larger datasets
review serves as a reference for both researchers and practitioners,
and require faster processing for timely decision-making, scalability
highlighting the evolution and diverse applications of NLP across both
and performance are paramount. NLP systems must be designed to
academic research and industrial implementations. Our review frame-
efficiently handle scalability, ensuring they can accommodate increas-
work systematically maps the relationships between financial data,
ing data volumes without compromising performance. This involves
NLP techniques, key players, and financial applications, identifying ten
leveraging scalable cloud infrastructures, employing efficient data stor-
principal areas where NLP has potentially revolutionized the finance
age solutions, and utilizing advanced algorithms capable of high-speed
data processing. Moreover, the deployment of NLP systems must also sector. Additionally, we analyzed key research contributions that have
consider ethical implications and strive for fairness in automated de- significantly influenced the application of NLP in finance, emphasizing
cisions. Financial institutions should develop frameworks that ensure its transformative effects on services, products, and operations. This
NLP models are devoid of biases that may lead to discriminatory survey not only addresses practical challenges but also highlights the
treatment, based on demographic factors. Regular audits and updates transformative potential of NLP technologies in the finance industry.
of models to identify and rectify biases, coupled with transparency Lastly, we summarized the current trends, challenges, and opportunities
in decision-making processes, are essential for upholding fairness and from nine perspectives, and provided eight promising directions for
ethical standards in automated systems. Lastly, the ROI estimation and future research.
economic value of NLP systems need to be measured, potentially from
perspectives of revenue impact and productivity gains.
6. Limitations

4.3.11. Future directions

The landscape of NLP in finance is poised for transformative ad- This review is designed to explore broad themes within the field of
vancements with several emerging areas ripe for exploration. The NLP in finance. It does not include detailed data extraction, data quality
adoption of time series transformers (TST) in financial forecasting assessment, and risk of bias evaluation due to the expansive nature of
signifies a significant advancement in the modeling of financial time NLP’s application in finance. This is intended as part of the review’s
series data. TST models, known for their ability to capture complex tem- design which allows for a more extensive discussion across a broader
poral relationships, promise to greatly enhance forecasting accuracy, array of topics, which was deemed more beneficial for our intended
especially in volatile financial environments. Reinforcement learning audience of researchers, practitioners, and professionals in the finance
is emerging as a potent tool for portfolio management, leveraging dy- sector. By delineating these boundaries, we aim to clarify this article’s
namic decision-making to optimize investment strategies in real time. intended scope and encourage readers to interpret the findings within
This approach could revolutionize portfolio management by continu- the context of an exploratory and thematic analysis.
ously learning and adapting to new market conditions. Neuro-symbolic
AI presents a compelling frontier, integrating the interpretability of
CRediT authorship contribution statement
symbolic AI with the learning capabilities of neural networks. This
hybrid approach is particularly promising for complex financial tasks
that require both deep learning for pattern recognition and symbolic Kelvin Du: Writing – original draft, Methodology, Investigation,
reasoning for rule-based decision-making. In the realm of QA systems, Conceptualization. Yazhi Zhao: Writing – original draft, Investigation,
enhancing numerical reasoning capabilities to interpret and answer Conceptualization. Rui Mao: Writing – review & editing, Methodology,
queries involving textual and tabular data is crucial. This will im- Formal analysis. Frank Xing: Writing – review & editing, Formal
prove decision-making processes by providing precise and contextually analysis, Conceptualization. Erik Cambria: Writing – review & editing,
relevant answers to financially oriented questions. FNP is another Supervision, Project administration.
promising area, where the accuracy of information extraction and
summarization from financial texts can be greatly enhanced. The in-
Declaration of competing interest
tegration of knowledge graphs into NLP models could enable more so-
phisticated processing of financial narratives, identifying relationships
and insights that are not readily apparent. The authors declare that they have no known competing finan-
Fine-grained FSA within financial documents can be targeted more cial interests or personal relationships that could have appeared to
precisely with NLP, focusing on specific aspects such as market trends, influence the work reported in this paper.

19
K. Du et al. Information Fusion 115 (2025) 102755

Acknowledgment [17] N. Oliveira, P. Cortez, N. Areal, Stock market sentiment lexicon acquisition
using microblogging data and statistical measures, Decis. Support Syst. 85
(2016) 62–73.
The authors would like to extend their thanks to Jiaxi He for her [18] K. Du, F. Xing, R. Mao, E. Cambria, FinSenticNet: A concept-level lexicon for
valuable comments on the financial applications of question-answering, financial sentiment analysis, in: 2023 IEEE Symposium Series on Computational
virtual assistants, and chatbots. This research/project is supported by Intelligence (SSCI), 2023, pp. 109–114.
the Ministry of Education, Singapore under its MOE Academic Research [19] D. Araci, FinBERT: Financial sentiment analysis with pre-trained language
models, 2019, arXiv:1908.10063.
Fund Tier 2 (STEM RIE2025 Award MOE-T2EP20123-0005) and by the [20] Z. Liu, D. Huang, K. Huang, Z. Li, J. Zhao, FinBERT: A pre-trained financial
RIE2025 Industry Alignment Fund – Industry Collaboration Projects language representation model for financial text mining, in: IJCAI, 2020, pp.
(IAF-ICP) (Award I2301E0026), administered by A*STAR, as well as 4513–4519.
supported by Alibaba Group and NTU Singapore. [21] A. Radford, K. Narasimhan, T. Salimans, I. Sutskever, et al., Improving Language
Understanding by Generative Pre-Training, OpenAI, 2018, pp. 1–12.
[22] S. Fatemi, Y. Hu, A comparative analysis of fine-tuned LLMs and few-shot
Data availability learning of LLMs for financial sentiment analysis, 2023, arXiv:2312.08725.
[23] G. Fatouros, J. Soldatos, K. Kouroumali, G. Makridis, D. Kyriazis, Transforming
sentiment analysis in the financial domain with ChatGPT, 2023, arXiv:2308.
No data was used for the research described in the article. 07935.
[24] B. Zhang, H. Yang, T. Zhou, M. Ali Babar, X.-Y. Liu, Enhancing financial senti-
ment analysis via retrieval augmented large language models, in: Proceedings of
References the Fourth ACM International Conference on AI in Finance, 2023, pp. 349–356.
[25] B. Zhang, H. Yang, X.-Y. Liu, Instruct-FinGPT: Financial sentiment analysis
by instruction tuning of general-purpose large language models, 2023, arXiv:
[1] I.E. Fisher, M.R. Garnsey, M.E. Hughes, Natural language processing in account-
2306.12659.
ing, auditing and finance: A synthesis of the literature with a roadmap for future
[26] N. Hu, P. Liang, X. Yang, Whetting All Your Appetites for Financial Tasks
research, Intell. Syst. Account. Finance Manage. 23 (3) (2016) 157–214.
with One Meal from GPT? A Comparison of GPT, FinBERT, and Dictionaries
[2] C.-C. Chen, H.-H. Huang, H.-H. Chen, NLP in FinTech applications: past, present
in Evaluating Sentiment Analysis, Technical Report, SSRN: , 2023, https://ssrn.
and future, 2020, arXiv preprint arXiv:2005.01320.
com/abstract=4426455.
[3] E.B. Boukherouaa, M.G. Shabsigh, K. AlAjmi, J. Deodoro, A. Farias, E.S.
[27] X. Deng, V. Bashlovkina, F. Han, S. Baumgartner, M. Bendersky, What do
Iskender, M.A.T. Mirestean, R. Ravikumar, Powering the Digital Economy:
llms know about financial markets? a case study on reddit market sentiment
Opportunities and Risks of Artificial Intelligence in Finance, International
analysis, in: Companion Proceedings of the ACM Web Conference 2023, 2023,
Monetary Fund, 2021.
pp. 107–110.
[4] R. Mao, X. Li, Bridging towers of multi-task learning with a gating mechanism [28] S. Wu, O. Irsoy, S. Lu, V. Dabravolski, M. Dredze, S. Gehrmann, P. Kambadur,
for aspect-based sentiment analysis and sequential metaphor identification, in: D. Rosenberg, G. Mann, BloombergGPT: A large language model for finance,
Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35, 2021, 2023, arXiv:2303.17564.
pp. 13534–13542. [29] K. Du, F. Xing, R. Mao, E. Cambria, An evaluation of reasoning capabilities of
[5] A. Valdivia, V. Luzón, E. Cambria, F. Herrera, Consensus vote models for large language models in financial sentiment analysis, in: IEEE Conference on
detecting and filtering neutrality in sentiment analysis, Inf. Fusion 44 (2018) Artificial Intelligence, IEEE CAI, Singapore, 2024, pp. 189–194.
126–135. [30] F. Xing, Designing heterogeneous LLM agents for financial sentiment analysis,
[6] E. Cambria, X. Zhang, R. Mao, M. Chen, K. Kwok, SenticNet 8: Fusing emotion ACM Trans. Manag. Inf. Syst. (2024) http://dx.doi.org/10.1145/3688399.
AI and commonsense AI for interpretable, trustworthy, and explainable affective [31] R. Mao, K. Du, Y. Ma, L. Zhu, E. Cambria, Discovering the cognition behind
computing, in: Proceedings of International Conference on Human-Computer language: Financial metaphor analysis with MetaPro, in: 2023 IEEE Interna-
Interaction, HCII, Washington DC, USA, 2024. tional Conference on Data Mining, ICDM, IEEE, Shanghai, China, 2023, pp.
[7] L. Zhu, R. Mao, E. Cambria, B.J. Jansen, Neurosymbolic AI for person- 1211–1216.
alized sentiment analysis, in: Proceedings of International Conference on [32] R. Mao, X. Li, K. He, M. Ge, E. Cambria, MetaPro Online: A computa-
Human-Computer Interaction, HCII, Washington DC, USA, 2024. tional metaphor processing online system, in: Proceedings of the 61st Annual
[8] K. Du, F. Xing, R. Mao, E. Cambria, Financial sentiment analysis: Techniques Meeting of the Association for Computational Linguistics (Volume 3: System
and applications, ACM Comput. Surv. 56 (9) (2024) http://dx.doi.org/10.1145/ Demonstrations), Vol. 3, 2023, pp. 127–135.
3649451. [33] R. Mao, K. He, C.B. Ong, Q. Liu, E. Cambria, MetaPro 2.0: Computational
[9] R. Moraes, J.F. Valiati, W.P.G. Neto, Document-level sentiment classification: metaphor processing on the effectiveness of anomalous language modeling, in:
An empirical comparison between SVM and ANN, Expert Syst. Appl. 40 (2) Findings of the Association for Computational Linguistics: ACL, Association for
(2013) 621–633. Computational Linguistics, Bangkok, Thailand, 2024, pp. 9891–9908.
[10] P. Ferguson, N. O’Hare, M. Davy, A. Bermingham, P. Sheridan, C. Gurrin, [34] R. Manro, R. Mao, L. Dahiya, Y. Ma, E. Cambria, A cognitive analysis of
A.F. Smeaton, Exploring the Use of Paragraph-Level Annotations for Sentiment CEO speeches and their effects on stock markets, in: Proceedings of the 5th
Analysis of Financial Blogs, Technical Report, Dublin City University, 2009. International Conference on Financial Technology, ICFT, Singapore, 2024.
[11] P. Zhang, Z. He, Using data-driven feature enrichment of text representation [35] J. Park, H.J. Lee, S. Cho, Automatic construction of context-aware sentiment
and ensemble technique for sentence-level polarity classification, J. Inf. Sci. 41 lexicon in the financial domain using direction-dependent words, 2021, arXiv
(4) (2015) 531–549. preprint arXiv:2106.05723.
[36] S. Krishnamoorthy, Sentiment analysis of financial news articles using
[12] M. Pontiki, D. Galanis, H. Papageorgiou, I. Androutsopoulos, S. Manandhar,
performance indicators, Knowl. Inf. Syst. 56 (2) (2018) 373–394.
M. AL-Smadi, M. Al-Ayyoub, Y. Zhao, B. Qin, O. De Clercq, V. Hoste, M.
[37] D. Araci, Finbert: Financial sentiment analysis with pre-trained language
Apidianaki, X. Tannier, N. Loukachevitch, E. Kotelnikov, N. Bel, S.M. Jiménez-
models, 2019, arXiv preprint arXiv:1908.10063.
Zafra, G. Eryiğit, SemEval-2016 task 5: Aspect based sentiment analysis, in: S.
[38] L. Zhao, L. Li, X. Zheng, A BERT based sentiment analysis and key entity
Bethard, M. Carpuat, D. Cer, D. Jurgens, P. Nakov, T. Zesch (Eds.), Proceedings
detection approach for online financial texts, in: 2021 IEEE CSCWD, 2021, pp.
of the 10th International Workshop on Semantic Evaluation, SemEVal-2016,
1233–1238.
Association for Computational Linguistics, San Diego, California, 2016, pp.
[39] M. Jiang, M. Lan, Y. Wu, Ecnu at semeval-2017 task 5: An ensemble of
19–30.
regression algorithms with effective features for fine-grained sentiment analysis
[13] P. Malo, A. Sinha, P. Korhonen, J. Wallenius, P. Takala, Good debt or bad debt:
in financial domain, in: Proceedings of SemEval-2017, 2017, pp. 888–893.
Detecting semantic orientations in economic texts, J. Assoc. Inf. Sci. Technol.
[40] K. Schouten, F. Frasincar, F. de Jong, Commit at semeval-2017 task 5: Ontology-
65 (4) (2014) 782–796.
based method for sentiment analysis of financial headlines, in: Proceedings of
[14] K. Cortis, A. Freitas, T. Daudert, M. Huerlimann, M. Zarrouk, S. Handschuh, SemEval-2017, 2017, pp. 883–887.
B. Davis, Semeval-2017 task 5: Fine-grained sentiment analysis on financial [41] A. Dridi, M. Atzeni, D.R. Recupero, FineNews: fine-grained semantic sentiment
microblogs and news, in: Proceedings of the 11th International Workshop on analysis on financial microblogs and news, Int. J. Mach. Learn. Cybern. (2019)
Semantic Evaluation, SemEval-2017, 2017, pp. 519–535. 1–9.
[15] M. Maia, S. Handschuh, A. Freitas, B. Davis, R. McDermott, M. Zarrouk, [42] M.S. Akhtar, A. Kumar, D. Ghosal, A. Ekbal, P. Bhattacharyya, A multilayer per-
A. Balahur, WWW’18 open challenge: financial opinion mining and question ceptron based ensemble technique for fine-grained financial sentiment analysis,
answering, in: Companion Proceedings of the the Web Conference 2018, 2018, in: Proceedings of EMNLP, 2017, pp. 540–546.
pp. 1941–1942. [43] D. Ghosal, S. Bhatnagar, M.S. Akhtar, A. Ekbal, P. Bhattacharyya, IITP at
[16] K. Du, F. Xing, E. Cambria, Incorporating multiple knowledge sources for SemEval-2017 task 5: an ensemble of deep learning and feature based models
targeted aspect-based financial sentiment analysis, ACM Trans. Manage. Inf. for financial sentiment analysis, in: Proceedings of SemEval-2017, 2017, pp.
Syst. 14 (3) (2023) 23, http://dx.doi.org/10.1145/3580480. 899–903.

20
K. Du et al. Information Fusion 115 (2025) 102755

[44] R. Mao, X. Li, M. Ge, E. Cambria, MetaPro: A computational metaphor [72] Y. Xu, S.B. Cohen, Stock movement prediction from tweets and historical prices,
processing model for text pre-processing, Inf. Fusion (ISSN: 1566-2535) 86–87 in: Proceedings of ACL, 2018, pp. 1970–1979.
(2022) 30–43. [73] H. Wu, W. Zhang, W. Shen, J. Wang, Hybrid deep sequential modeling for social
[45] C. Xiang, J. Zhang, F. Li, H. Fei, D. Ji, A semantic and syntactic enhanced neural text-driven stock prediction, in: Proceedings of the 27th ACM International
model for financial sentiment analysis, Inf. Process. Manage. 59 (4) (2022) Conference on Information and Knowledge Management, CIKM, 2018, pp.
102943. 1627–1630.
[46] A. Sinha, S. Kedas, R. Kumar, P. Malo, SEntFiN 1.0: Entity-aware sentiment [74] X. Guo, J. Li, A novel Twitter sentiment analysis model with baseline correlation
analysis for financial news, J. Assoc. Inf. Sci. Technol. (2022). for financial market prediction with improved efficiency, in: 2019 Sixth Inter-
[47] D. de França Costa, N.F.F. da Silva, INF-ufg at FiQA 2018 task 1: predicting national Conference on Social Networks Analysis, Management and Security,
sentiments and aspects on financial tweets and news headlines, in: Companion SNAMS, IEEE, 2019, pp. 472–477.
Proceedings of the the Web Conference 2018, 2018, pp. 1967–1971. [75] Z. Jin, Y. Yang, Y. Liu, Stock closing price prediction based on sentiment
[48] G. Piao, J.G. Breslin, Financial aspect and sentiment predictions with deep analysis and LSTM, Neural Comput. Appl. 32 (2020).
neural networks: an ensemble approach, in: Companion Proceedings of the the [76] S. Wu, Y. Liu, Z. Zou, T.-H. Weng, S_I_LSTM: stock price prediction based on
Web Conference 2018, 2018, pp. 1973–1977. multiple data sources and sentiment analysis, Connect. Sci. 34 (1) (2022) 44–62.
[49] L. Luo, X. Ao, F. Pan, J. Wang, T. Zhao, N. Yu, Q. He, Beyond polarity: Inter- [77] F. Jin, N. Self, P. Saraf, P. Butler, W. Wang, N. Ramakrishnan, Forex-foreteller:
pretable financial sentiment analysis with hierarchical query-driven attention, Currency trend modeling using news articles, in: Proceedings of ACM SIGKDD,
in: IJCAI, 2018, pp. 4244–4250. 2013, pp. 1470–1473.
[50] F.Z. Xing, E. Cambria, R.E. Welsch, Natural language based financial [78] S. Seifollahi, M. Shajari, Word sense disambiguation application in sentiment
forecasting: a survey, Artif. Intell. Rev. 50 (1) (2018) 49–73. analysis of news headlines: an applied approach to FOREX market prediction,
[51] F. Xing, E. Cambria, R. Welsch, Intelligent asset allocation via market sentiment J. Intell. Inf. Syst. 52 (1) (2019) 57–83.
views, IEEE Comput. Intell. Mag. 13 (4) (2018) 25–34. [79] H.N. Semiromi, S. Lessmann, W. Peters, News will tell: Forecasting foreign
[52] A. Mudinas, D. Zhang, M. Levene, Market trend prediction using sentiment exchange rates based on news story events in the economy calendar, North
analysis: lessons learned and paths forward, 2019, arXiv preprint arXiv:1903. Am. J. Econom. Finance 52 (2020) 101181.
05440. [80] F. Xing, D.H. Hoang, D.-V. Vo, High-frequency news sentiment and its ap-
[53] Y. Ma, R. Mao, Q. Lin, P. Wu, E. Cambria, Multi-source aggregated classification plication to forex market prediction, in: Proceedings of HICSS, 2020, pp.
for stock price movement prediction, Inf. Fusion 91 (2023) 515–528. 1583–1592.
[54] N. Oliveira, P. Cortez, N. Areal, The impact of microblogging data for stock [81] H. Laakkonen, The Impact of Macroeconomic News on Exchange Rate Volatility,
market prediction: Using Twitter to predict returns, volatility, trading volume Bank of Finland Discussion Paper 24, 2004.
and survey sentiment indices, Expert Syst. Appl. 73 (2017) 125–144. [82] M.D. Evans, R.K. Lyons, Do currency markets absorb news quickly? J. Int.
[55] A.K. Nassirtoussi, S. Aghabozorgi, T.Y. Wah, D.C.L. Ngo, Text mining of news- Money Finance 24 (2) (2005) 197–217.
[83] E. Fedorova, S. Druchok, P. Drogovoz, Impact of news sentiment and topics
headlines for FOREX market prediction: A multi-layer dimension reduction
on IPO underpricing: US evidence, Int. J. Account. Inf. Manage. 30 (1) (2022)
algorithm with semantics and sentiment, Expert Syst. Appl. 42 (1) (2015)
73–94.
306–324.
[84] S.J. Benjamin, P.K. Biswas, M.S. M Marathamuthu, M. Arunachalam, Social
[56] G.M. Mian, S. Sankaraguruswamy, Investor sentiment and stock market response
media sentiments and firm value, Appl. Econ. 54 (26) (2022) 2983–2997.
to earnings news, Account. Rev. 87 (4) (2012) 1357–1384.
[85] F. Xing, E. Cambria, R. Welsch, Growing semantic vines for robust asset
[57] X. Ding, Y. Zhang, T. Liu, J. Duan, Deep learning for event-driven stock
allocation, Knowl.-Based Syst. 165 (2019) 297–305, http://dx.doi.org/10.1016/
prediction, in: Proceedings of the 24th International Conference on Artificial
j.knosys.2018.11.035.
Intelligence, 2015, pp. 2327–2333.
[86] L. Malandri, F.Z. Xing, C. Orsenigo, C. Vercellis, E. Cambria, Public mood–
[58] Y. Liu, Q. Zeng, H. Yang, A. Carrio, Stock price movement prediction from
driven asset allocation: The importance of financial sentiment in portfolio
financial news with deep learning and knowledge graph embedding, in: Pacific
management, Cogn. Comput. 10 (6) (2018) 1167–1176.
Rim Knowledge Acquisition Workshop, Springer, 2018, pp. 102–113.
[87] S. Koyano, K. Ikeda, Online portfolio selection based on the posts of winners and
[59] S. Deng, T. Mitsubuchi, K. Shioda, T. Shimada, A. Sakurai, Combining technical
losers in stock microblogs, in: 2017 IEEE Symposium Series on Computational
analysis with sentiment analysis for stock price prediction, in: IEEE DASC, 2011,
Intelligence, SSCI, IEEE, 2017, pp. 1–4.
pp. 800–807.
[88] W. Tu, D.W. Cheung, N. Mamoulis, M. Yang, Z. Lu, Investment recommendation
[60] R. Frankel, M. Johnson, D.J. Skinner, An empirical examination of conference
using investor opinions in social media, in: Proceedings of ACM SIGIR, 2016,
calls as a voluntary disclosure medium, J. Account. Res. 37 (1) (1999) 133–150.
pp. 881–884.
[61] R.M. Bowen, A.K. Davis, D.A. Matsumoto, Do conference calls affect analysts’
[89] S.Y. Yang, Y. Yu, S. Almahdi, An investor sentiment reward-based trading
forecasts? Account. Rev. 77 (2) (2002) 285–316.
system using Gaussian inverse reinforcement learning algorithm, Expert Syst.
[62] L. Cohen, D. Lou, C.J. Malloy, Casting conference calls, Manage. Sci. 66 (11)
Appl. 114 (2018) 388–401.
(2020) 5015–5039. [90] R. Sawhney, A. Wadhwa, S. Agarwal, R. Shah, Quantitative day trading from
[63] K. Keith, A. Stent, Modeling financial analysts’ decision making via the natural language using reinforcement learning, in: Proceedings of NAACL-HLT,
pragmatics and semantics of earnings calls, in: Proceedings of ACL, 2019, pp. 2021, pp. 4018–4030.
493–503. [91] Y. Ye, H. Pei, B. Wang, P.-Y. Chen, Y. Zhu, J. Xiao, B. Li, Reinforcement-learning
[64] Y. Chen, Z. Wei, X. Huang, Incorporating corporation relationship via graph based portfolio management with augmented asset movement prediction states,
convolutional neural networks for stock price prediction, in: Proceedings in: Proceedings of AAAI, Vol. 34, 2020, pp. 1112–1119.
of the 27th ACM International Conference on Information and Knowledge [92] X. Du, K. Tanaka-Ishii, Stock embeddings acquired from news articles and price
Management, 2018, pp. 1655–1658. history, and an application to portfolio optimization, in: Proceedings of ACL,
[65] R. Sawhney, S. Agarwal, A. Wadhwa, R. Shah, Deep attentive learning for 2020, pp. 3353–3363.
stock movement prediction from social media text and company correlations, in: [93] P. Koratamaddi, K. Wadhwani, M. Gupta, S.G. Sanjeevi, Market sentiment-aware
Proceedings of the 2020 Conference on Empirical Methods in Natural Language deep reinforcement learning approach for stock portfolio allocation, Eng. Sci.
Processing, EMNLP, 2020, pp. 8415–8426. Technol. Int. J. (2021).
[66] J. Yoo, Y. Soun, Y.-c. Park, U. Kang, Accurate multivariate stock movement [94] M. Chen, Z. Zhang, J. Shen, Z. Deng, J. He, S. Huang, A quantitative investment
prediction via data-axis transformer with multi-level contexts, in: Proceedings model based on random forest and sentiment analysis, J. Phys. Conf. Ser. 1575
of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, (1) (2020) 012083, http://dx.doi.org/10.1088/1742-6596/1575/1/012083.
2021, pp. 2037–2045. [95] R. Sawhney, A. Wadhwa, S. Agarwal, R.R. Shah, FAST: Financial news and
[67] Y. Soun, J. Yoo, M. Cho, J. Jeon, U. Kang, Accurate stock movement prediction tweet based time aware network for stock trading, in: Proceedings of EACL,
with self-supervised learning from sparse noisy tweets, in: IEEE International 2021, pp. 2164–2175.
Conference on Big Data, 2022, pp. 1691–1700. [96] M.-C. Hung, P.-H. Hsia, X.-J. Kuang, S.-K. Lin, Intelligent portfolio construction
[68] Z. Zhou, L. Ma, H. Liu, Trade the event: Corporate events detection for news- via news sentiment analysis, Int. Rev. Econ. Finance 89 (2024) 605–617.
based event-driven trading, in: C. Zong, F. Xia, W. Li, R. Navigli (Eds.), Findings [97] Y. Ma, R. Mao, Q. Lin, P. Wu, E. Cambria, Quantitative stock portfolio
of the Association for Computational Linguistics: ACL-IJCNLP 2021, Association optimization by multi-task learning risk and return, Inf. Fusion 104 (2024)
for Computational Linguistics, Online, 2021, pp. 2114–2124. 102165.
[69] M. Baker, J. Wurgler, Investor sentiment and the cross-section of stock returns, [98] Y. Sun, M. Fang, X. Wang, A novel stock recommendation system using guba
J. Finance 61 (4) (2006) 1645–1680. sentiment analysis, Pers. Ubiquitous Comput. 22 (2018) 575–587.
[70] K. Du, R. Mao, F. Xing, E. Cambria, A dynamic dual-graph neural network for [99] P. Ghosh, A. Neufeld, J.K. Sahoo, Forecasting directional movements of stock
stock price movement prediction, in: 2024 International Joint Conference on prices for intraday trading using LSTM and random forests, Finance Res. Lett.
Neural Networks, IJCNN, Yokohama, Japan, 2024. (ISSN: 1544-6123) 46 (2022) 102280.
[71] T.H. Nguyen, K. Shirai, Topic modeling based sentiment analysis on social [100] M. El-Haj, P. Rayson, A. Moore, The first financial narrative processing
media for stock market prediction, in: Proceedings of ACL-IJCNLP, 2015, pp. workshop (FNP 2018), in: Proceedings of the LREC 2018 Workshop, 2018, pp.
1354–1364. 1–65.

21
K. Du et al. Information Fusion 115 (2025) 102755

[101] A.T. Oyewole, O.B. Adeoye, W.A. Addy, C.C. Okoye, O.C. Ofodile, C.E. [126] Z. Xiao, Z. Mai, Z. Xu, Y. Cui, J. Li, Corporate event predictions using large
Ugochukwu, Automating financial reporting with natural language processing: language models, in: 2023 10th International Conference on Soft Computing &
A review and case analysis, World J. Adv. Res. Rev. 21 (3) (2024) 575–589. Machine Intelligence, ISCMI, IEEE, 2023, pp. 193–197.
[102] A. Kim, M. Muhn, V.V. Nikolaev, Financial Statement Analysis with Large [127] G. Jacobs, V. Hoste, SENTiVENT: enabling supervised information extraction of
Language Models, Chicago Booth Research Paper Forthcoming, Fama-Miller company-specific events in economic and financial news, Lang. Resour. Eval.
Working Paper, 2024. 56 (1) (2022) 225–257.
[103] V. Gupta, G.S. Lehal, A survey of text summarization extractive techniques, J. [128] G. Jacobs, E. Lefever, V. Hoste, Economic event detection in company-specific
Emerg. Technol. Web Intell. 2 (3) (2010) 258–268. news text, in: 1st Workshop on Economics and Natural Language Processing
[104] N. Moratanch, S. Chitrakala, A survey on abstractive text summarization, in: (ECONLP) At Meeting of the Association-for-Computational-Linguistics, ACL,
2016 International Conference on Circuit, Power and Computing Technologies, Association for Computational Linguistics (ACL), 2018, pp. 1–10.
ICCPCT, IEEE, 2016, pp. 1–7. [129] L. Ein-Dor, A. Gera, O. Toledo-Ronen, A. Halfon, B. Sznajder, L. Dankin, Y.
[105] M. El-Haj, M. Litvak, N. Pittaras, G. Giannakopoulos, et al., The financial Bilu, Y. Katz, N. Slonim, Financial event extraction using wikipedia-based weak
narrative summarisation shared task (FNS 2020), in: Proceedings of the 1st supervision, in: U. Hahn, V. Hoste, Z. Zhang (Eds.), Proceedings of the Second
Joint Workshop on Financial Narrative Processing and MultiLing Financial Workshop on Economics and Natural Language Processing, Association for
Summarisation, 2020, pp. 1–12. Computational Linguistics, Hong Kong, 2019, pp. 10–15.
[106] N. Zmandar, M. El-Haj, P. Rayson, M. Litvak, G. Giannakopoulos, N. Pittaras, et [130] S. Carta, S. Consoli, L. Piras, A.S. Podda, D.R. Recupero, Event detection in
al., The financial narrative summarisation shared task fns 2021, in: Proceedings finance using hierarchical clustering algorithms on news and tweets, PeerJ
of the 3rd Financial Narrative Processing Workshop, 2021, pp. 120–125. Comput. Sci. 7 (2021) e438.
[107] E. Zavitsanos, A. Kosmopoulos, G. Giannakopoulos, M. Litvak, B. Carbajo- [131] K. Guo, T. Jiang, H. Zhang, Knowledge graph enhanced event extraction in
Coronado, A. Moreno-Sandoval, M. El-Haj, The financial narrative summari- financial documents, in: 2020 IEEE International Conference on Big Data, Big
sation shared task (FNS 2023), in: 2023 IEEE International Conference on Big Data, IEEE, 2020, pp. 1322–1329.
Data, BigData, IEEE, 2023, pp. 2890–2896. [132] D. Cheng, F. Yang, X. Wang, Y. Zhang, L. Zhang, Knowledge graph-based event
[108] M. El-Haj, A. Ogden, Financial narrative summarisation using a hybrid tf-idf embedding framework for financial quantitative investments, in: Proceedings of
and clustering summariser: Ao-lancs system at fns 2022, in: Proceedings of the the 43rd International ACM SIGIR Conference on Research and Development
4th Financial Narrative Processing Workshop@ LREC2022, 2022, pp. 79–82. in Information Retrieval, 2020, pp. 2221–2230.
[109] T. Gokhan, P. Smith, M. Lee, Extractive financial narrative summarisation using [133] A. Balashankar, S. Chakraborty, S. Fraiberger, L. Subramanian, Identifying
sentencebert based clustering, in: Proceedings of the 3rd Financial Narrative predictive causal factors from news streams, in: Proceedings of the 2019
Processing Workshop, 2021, pp. 94–98. Conference on Empirical Methods in Natural Language Processing and the 9th
[110] M. La Quatra, L. Cagliero, End-to-end training for financial report summa- International Joint Conference on Natural Language Processing, EMNLP-IJCNLP,
rization, in: Proceedings of the 1st Joint Workshop on Financial Narrative 2019, pp. 2338–2348.
Processing and MultiLing Financial Summarisation, 2020, pp. 118–123. [134] K. Izumi, H. Sakaji, Economic causal-chain search using text mining technology,
[111] S. Abdaljalil, H. Bouamor, An exploration of automatic text summarization in: Artificial Intelligence. IJCAI 2019 International Workshops: Macao, China,
of financial reports, in: Proceedings of the Third Workshop on Financial August 10–12, 2019, Revised Selected Best Papers 28, Springer, 2020, pp.
Technology and Natural Language Processing, 2021, pp. 1–7. 23–35.
[112] M. Orzhenovskii, T5-LONG-EXTRACT at FNS-2021 shared task, in: Proceedings [135] T. Takayanagi, M. Suzuki, R. Kobayashi, H. Sakaji, K. Izumi, Is ChatGPT the
of the 3rd Financial Narrative Processing Workshop, 2021, pp. 67–69. future of causal text mining? A comprehensive evaluation and analysis, 2024,
[113] M. Litvak, N. Vanetik, Z. Puchinsky, Hierarchical summarization of financial arXiv preprint arXiv:2402.14484.
reports with RUNNER, in: Proceedings of the 1st Joint Workshop on Finan- [136] P. Rajpoot, A. Parikh, GPT-finRE: In-context learning for financial relation ex-
cial Narrative Processing and MultiLing Financial Summarisation, 2020, pp. traction using large language models, in: C.-C. Chen, H.-H. Huang, H. Takamura,
213–225. H.-H. Chen, H. Sakaji, K. Izumi (Eds.), Proceedings of the Sixth Workshop
[114] N. Shukla, A. Vaid, R. Katikeri, S. Keeriyadath, M. Raja, Dimsum: Distributed on Financial Technology and Natural Language Processing, Association for
and multilingual summarization of financial narratives, in: Proceedings of the Computational Linguistics, Bali, Indonesia, 2023, pp. 42–45.
4th Financial Narrative Processing Workshop@ LREC2022, 2022, pp. 65–72. [137] L. Hillebrand, T. Deuß er, T. Dilmaghani, B. Kliem, R. Loitz, C. Bauckhage,
[115] N.K. Shukla, R. Katikeri, M. Raja, G. Sivam, S. Yadav, A. Vaid, S. Prabhakararao, R. Sifa, Kpi-bert: A joint named entity recognition and relation extraction
Generative AI approach to distributed summarization of financial narratives, model for financial reports, in: 2022 26th International Conference on Pattern
in: 2023 IEEE International Conference on Big Data, BigData, IEEE, 2023, pp. Recognition, ICPR, IEEE, 2022, pp. 606–612.
2872–2876. [138] Y. Zhang, H. Zhang, FinBERT–MRC: financial named entity recognition using
[116] H. Li, Q. Peng, X. Mou, Y. Wang, Z. Zeng, M.F. Bashir, Abstractive financial BERT under the machine reading comprehension paradigm, Neural Process.
news summarization via transformer-bilstm encoder and graph attention-based Lett. 55 (6) (2023) 7393–7413.
decoder, IEEE/ACM Trans. Audio Speech Lang. Process. (2023). [139] B. Oral, E. Emekligil, S. Arslan, G. Eryiǧit, Information extraction from text
[117] S. Liu, J. Cao, R. Yang, Z. Wen, Long text and multi-table summarization: intensive and visually rich banking documents, Inf. Process. Manage. 57 (6)
Dataset and method, in: Y. Goldberg, Z. Kozareva, Y. Zhang (Eds.), Findings (2020) 102361.
of the Association for Computational Linguistics: EMNLP 2022, Association [140] Y. Yang, Z. Wu, Y. Yang, S. Lian, F. Guo, Z. Wang, A survey of information
for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022, pp. extraction based on deep learning, Appl. Sci. 12 (19) (2022) 9691.
1995–2010. [141] Q. Li, Q. Zhang, A unified model for financial event classification, detection
[118] N. Zmandar, A. Singh, M. El-Haj, P. Rayson, Joint abstractive and extractive and summarization, in: Proceedings of the Twenty-Ninth International Confer-
method for long financial document summarization, in: Proceedings of the 3rd ence on International Joint Conferences on Artificial Intelligence, 2021, pp.
Financial Narrative Processing Workshop, 2021, pp. 99–105. 4668–4674.
[119] A. Singh, Point-5: Pointer network and t-5 based financial narrativesummarisa- [142] D. Mariko, H. Abi Akl, K. Trottier, M. El-Haj, The financial causality extraction
tion, 2020, arXiv preprint arXiv:2010.04191. shared task (FinCausal 2022), in: Proceedings of the 4th Financial Narrative
[120] R. Mukherjee, A. Bohra, A. Banerjee, S. Sharma, M. Hegde, A. Shaikh, S. Processing Workshop@ LREC2022, 2022, pp. 105–107.
Shrivastava, K. Dasgupta, N. Ganguly, S. Ghosh, P. Goyal, ECTSum: A new [143] A. Moreno-Sandoval, J. Porta-Zamorano, B. Carbajo-Coronado, D. Samy, D.
benchmark dataset for bullet point summarization of long earnings call tran- Mariko, M. El-Haj, The financial document causality detection shared task
scripts, in: Y. Goldberg, Z. Kozareva, Y. Zhang (Eds.), Proceedings of the 2022 (FinCausal 2023), in: 2023 IEEE International Conference on Big Data, BigData,
Conference on Empirical Methods in Natural Language Processing, Association IEEE, 2023, pp. 2855–2860.
for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022, pp. [144] S. Kaur, C. Smiley, A. Gupta, J. Sain, D. Wang, S. Siddagangappa, T. Aguda,
10893–10906. S. Shah, REFinD: Relation extraction financial dataset, in: Proceedings of the
[121] A. See, P.J. Liu, C.D. Manning, Get to the point: Summarization with 46th International ACM SIGIR Conference on Research and Development in
pointer-generator networks, 2017, arXiv preprint arXiv:1704.04368. Information Retrieval, 2023, pp. 3054–3063.
[122] R. Paulus, C. Xiong, R. Socher, A deep reinforced model for abstractive [145] S. Sharma, T. Nayak, A. Bose, A.K. Meena, K. Dasgupta, N. Ganguly, P. Goyal,
summarization, 2017, arXiv preprint arXiv:1705.04304. FinRED: A dataset for relation extraction in financial domain, in: Companion
[123] A.P. Widyassari, S. Rustad, G.F. Shidik, E. Noersasongko, A. Syukur, A. Affandy, Proceedings of the Web Conference 2022, 2022, pp. 595–597.
et al., Review of automatic text summarization techniques & methods, J. King [146] G. Becquin, GBe at FinCausal 2020, task 2: Span-based causality extraction for
Saud Univ.-Comput. Inf. Sci. 34 (4) (2022) 1029–1046. financial documents, in: Proceedings of the 1st Joint Workshop on Financial
[124] V. Tretyak, D. Stepanov, Combination of abstractive and extractive approaches Narrative Processing and MultiLing Financial Summarisation, 2020, pp. 40–44.
for summarization of long scientific texts, 2020, arXiv preprint arXiv:2006. [147] P. Yang, X. Cong, Z. Sun, X. Liu, Enhanced language representation with label
05354. knowledge for span extraction, in: M.-F. Moens, X. Huang, L. Specia, S.W.-
[125] C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. t. Yih (Eds.), Proceedings of the 2021 Conference on Empirical Methods in
Li, P.J. Liu, Exploring the limits of transfer learning with a unified text-to-text Natural Language Processing, Association for Computational Linguistics, Online
transformer, J. Mach. Learn. Res. 21 (140) (2020) 1–67. and Punta Cana, Dominican Republic, 2021, pp. 4623–4635.

22
K. Du et al. Information Fusion 115 (2025) 102755

[148] R. Shah, K. Chawla, D. Eidnani, A. Shah, W. Du, S. Chava, N. Raman, C. [166] P. Srivastava, M. Malik, T. Ganu, Assessing LLMs’ mathematical reasoning in
Smiley, J. Chen, D. Yang, When FLUE meets FLANG: Benchmarks and large financial document question answering, 2024, arXiv preprint arXiv:2402.11194.
pretrained language model for financial domain, in: Y. Goldberg, Z. Kozareva, [167] T. Okuda, S. Shoda, AI-based chatbot service for financial industry, Fujitsu Sci.
Y. Zhang (Eds.), Proceedings of the 2022 Conference on Empirical Methods in Tech. J. 54 (2) (2018) 4–8.
Natural Language Processing, Association for Computational Linguistics, Abu [168] V. Kobets, K. Kozlovskyi, Application of chat bots for personalized financial
Dhabi, United Arab Emirates, 2022, pp. 2322–2335. advice, Herald Adv. Inf. Technol. 3 (5) (2022) 229–242.
[149] J.C.S. Alvarado, K. Verspoor, T. Baldwin, Domain adaption of named entity [169] Q. Xie, D. Tan, T. Zhu, Q. Zhang, S. Xiao, J. Wang, B. Li, L. Sun, P. Yi, Chatbot
recognition to support credit risk assessment, in: Proceedings of the Australasian application on cryptocurrency, in: 2019 IEEE Conference on Computational
Language Technology Association Workshop 2015, 2015, pp. 84–90. Intelligence for Financial Engineering & Economics, CIFEr, IEEE, 2019, pp. 1–8.
[150] J. Kang, A. Ait Azzi, S. Bellato, B. Carbajo Coronado, M. El-Haj, I. El Maarouf, [170] J.-W. Chang, N. Yen, J.C. Hung, Design of a NLP-empowered finance fraud
M. Gan, A. Gisbert, A. Moreno Sandoval, The financial document structure ex- awareness model: the anti-fraud chatbot for fraud detection and fraud classi-
traction shared task (finTOC 2022), in: M. El-Haj, P. Rayson, N. Zmandar (Eds.), fication as an instance, J. Ambient Intell. Humaniz. Comput. 13 (10) (2022)
Proceedings of the 4th Financial Narrative Processing Workshop @LREC2022, 4663–4679.
Marseille, France, 2022, pp. 83–88. [171] S.E. Finch, J.D. Choi, Towards unified dialogue system evaluation: A compre-
[151] W. Au, A. Ait-Azzi, J. Kang, FinSBD-2021: The 3rd shared task on structure hensive analysis of current evaluation protocols, in: O. Pietquin, S. Muresan,
boundary detection in unstructured text in the financial domain, in: Companion V. Chen, C. Kennington, D. Vandyke, N. Dethlefs, K. Inoue, E. Ekstedt, S. Ultes
Proceedings of the Web Conference 2021, 2021. (Eds.), Proceedings of the 21th Annual Meeting of the Special Interest Group on
[152] K.S.D. Ishwari, A.K.R.R. Aneeze, S. Sudheesan, H.J.D.A. Karunaratne, A. Discourse and Dialogue, Association for Computational Linguistics, 1st virtual
Nugaliyadde, Y. Mallawarrachchi, Advances in natural language question meeting, 2020, pp. 236–245.
answering: A review, 2019, arXiv:1904.05276. [172] F. Zhu, W. Lei, C. Wang, J. Zheng, S. Poria, T.-S. Chua, Retrieving and
[153] E. Cambria, Understanding Natural Language Understanding, Springer, ISBN: reading: A comprehensive survey on open-domain question answering, 2021,
978-3-031-73973-6, 2024. arXiv:2101.00774.
[154] Z. Chen, W. Chen, C. Smiley, S. Shah, I. Borova, D. Langdon, R. Moussa, M. [173] A. Balaguer, V. Benara, R.L.d. Cunha, R. de M. Estevão Filho, T. Hendry, D.
Beane, T.-H. Huang, B. Routledge, W.Y. Wang, FinQA: A dataset of numerical Holstein, J. Marsman, N. Mecklenburg, S. Malvar, L.O. Nunes, R. Padilha, M.
reasoning over financial data, in: M.-F. Moens, X. Huang, L. Specia, S.W.- Sharp, B. Silva, S. Sharma, V. Aski, R. Chandra, RAG vs fine-tuning: Pipelines,
t. Yih (Eds.), Proceedings of the 2021 Conference on Empirical Methods in tradeoffs, and a case study on agriculture, 2024, arXiv:2401.08406.
Natural Language Processing, Association for Computational Linguistics, Online [174] C. Packer, S. Wooders, K. Lin, V. Fang, S.G. Patil, I. Stoica, J.E. Gonzalez,
and Punta Cana, Dominican Republic, 2021, pp. 3697–3711. MemGPT: Towards LLMs as operating systems, 2024, arXiv:2310.08560.
[155] Z. Wang, Modern question answering datasets and benchmarks: A survey, 2022, [175] T. Munkhdalai, M. Faruqui, S. Gopal, Leave no context behind: Efficient infinite
arXiv:2206.15030. context transformers with infini-attention, 2024, arXiv:2404.07143.
[156] P. Rajpurkar, J. Zhang, K. Lopyrev, P. Liang, SQuAD: 100,000+ questions [176] L. Xu, H. Xie, S.-Z.J. Qin, X. Tao, F.L. Wang, Parameter-efficient fine-tuning
for machine comprehension of text, in: J. Su, K. Duh, X. Carreras (Eds.), methods for pretrained language models: A critical review and assessment,
Proceedings of the 2016 Conference on Empirical Methods in Natural Language 2023, arXiv:2312.12148.
[177] E.J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen,
Processing, Association for Computational Linguistics, Austin, Texas, 2016, pp.
LoRA: Low-rank adaptation of large language models, 2021, arXiv:2106.09685.
2383–2392.
[178] T. Kaufmann, P. Weng, V. Bengs, E. Hüllermeier, A survey of reinforcement
[157] M. Joshi, E. Choi, D. Weld, L. Zettlemoyer, TriviaQA: A large scale distantly
learning from human feedback, 2024, arXiv:2312.14925.
supervised challenge dataset for reading comprehension, in: R. Barzilay, M.-Y.
[179] N. Mecklenburg, Y. Lin, X. Li, D. Holstein, L. Nunes, S. Malvar, B. Silva, R.
Kan (Eds.), Proceedings of the 55th Annual Meeting of the Association for Com-
Chandra, V. Aski, P.K.R. Yannam, T. Aktas, T. Hendry, Injecting new knowledge
putational Linguistics (Volume 1: Long Papers), Association for Computational
into large language models via supervised fine-tuning, 2024, arXiv:2404.00213.
Linguistics, Vancouver, Canada, 2017, pp. 1601–1611.
[180] Z.Z. Yu, L.J. Jaw, Z. Hui, B.K.H. Low, Fine-tuning language models with
[158] Z. Chen, S. Li, C. Smiley, Z. Ma, S. Shah, W.Y. Wang, ConvFinQA: Exploring
generative adversarial reward modelling, 2024, arXiv:2305.06176.
the chain of numerical reasoning in conversational finance question answering,
[181] S. Khan, M.R. Rabbani, Chatbot as islamic finance expert (CaIFE) when finance
in: Y. Goldberg, Z. Kozareva, Y. Zhang (Eds.), Proceedings of the 2022
meets artificial intelligence, in: Proceedings of the 2020 4th International
Conference on Empirical Methods in Natural Language Processing, Association
Symposium on Computer Science and Intelligent Control, 2020, pp. 1–5.
for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022, pp.
[182] M.S.I. Bhuiyan, A. Razzak, M.S. Ferdous, M.J.M. Chowdhury, M.A. Hoque, S.
6279–6292.
Tarkoma, BONIK: A blockchain empowered chatbot for financial transactions,
[159] C. Li, W. Ye, Y. Zhao, Finmath: Injecting a tree-structured solver for question
in: 2020 IEEE 19th International Conference on Trust, Security and Privacy in
answering over financial reports, in: Proceedings of the Thirteenth Language
Computing and Communications, TrustCom, IEEE, 2020, pp. 1079–1088.
Resources and Evaluation Conference, 2022, pp. 6147–6152.
[183] S.F. Suhel, V.K. Shukla, S. Vyas, V.P. Mishra, Conversation to automation in
[160] Q. Zhang, S. Chen, D. Xu, Q. Cao, X. Chen, T. Cohn, M. Fang, A survey
banking through chatbot using artificial machine intelligence language, in:
for efficient open domain question answering, in: A. Rogers, J. Boyd-Graber,
2020 8th International Conference on Reliability, Infocom Technologies and
N. Okazaki (Eds.), Proceedings of the 61st Annual Meeting of the Associa-
Optimization (Trends and Future Directions), IEEE, 2020, pp. 611–618.
tion for Computational Linguistics (Volume 1: Long Papers), Association for [184] H.A. Pandya, B.S. Bhatt, Question answering survey: Directions, challenges,
Computational Linguistics, Toronto, Canada, 2023, pp. 14447–14465. datasets, evaluation matrices, 2021, arXiv:2112.03572.
[161] F. Zhu, W. Lei, Y. Huang, C. Wang, S. Zhang, J. Lv, F. Feng, T.-S. Chua, TAT-QA: [185] S. Kogan, D. Levin, B.R. Routledge, J.S. Sagi, N.A. Smith, Predicting risk from
A question answering benchmark on a hybrid of tabular and textual content financial reports with regression, in: Proceedings of NAACL-HLT, 2009, pp.
in finance, in: C. Zong, F. Xia, W. Li, R. Navigli (Eds.), Proceedings of the 272–280.
59th Annual Meeting of the Association for Computational Linguistics and the [186] C.-J. Wang, M.-F. Tsai, T. Liu, C.-T. Chang, Financial sentiment analysis for risk
11th International Joint Conference on Natural Language Processing (Volume prediction, in: Proceedings of IJCNLP, 2013, pp. 802–808.
1: Long Papers), Association for Computational Linguistics, Online, 2021, pp. [187] M.-F. Tsai, C.-J. Wang, Financial keyword expansion via continuous word vector
3277–3287. representations, in: Proceedings of EMNLP, 2014, pp. 1453–1458.
[162] Y. Deng, W. Lei, W. Zhang, W. Lam, T.-S. Chua, PACIFIC: Towards proactive [188] C. Nopp, A. Hanbury, Detecting risks in the banking system by sentiment
conversational question answering over tabular and textual data in finance, analysis, in: Proceedings of EMNLP, 2015, pp. 591–600.
in: Y. Goldberg, Z. Kozareva, Y. Zhang (Eds.), Proceedings of the 2022 [189] W.Y. Wang, Z. Hua, A semiparametric Gaussian copula regression model for
Conference on Empirical Methods in Natural Language Processing, Association predicting financial risks from earnings calls, in: Proceedings of ACL, 2014, pp.
for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022, pp. 1155–1165.
6970–6984. [190] N. Rekabsaz, M. Lupu, A. Baklanov, A. Hanbury, A. Dür, L. Anderson, Volatility
[163] M. Bulla, L. Hillebrand, M. Lübbering, R. Sifa, Knowledge graph based question prediction using financial disclosures sentiments with word embedding-based
answering system for financial securities, in: KI 2021: Advances in Artificial In- IR models, in: Proceedings of the 55th Annual Meeting of the Association for
telligence: 44th German Conference on AI, Virtual Event, September 27–October Computational Linguistics (Volume 1: Long Papers), 2017.
1, 2021, Proceedings 44, Springer, 2021, pp. 44–50. [191] F.Z. Xing, E. Cambria, Y. Zhang, Sentiment-aware volatility forecasting,
[164] R. Nararatwong, N. Kertkeidkachorn, R. Ichise, KIQA: Knowledge-infused ques- Knowl.-Based Syst. 176 (2019) 68–76.
tion answering model for financial table-text data, in: E. Agirre, M. Apidianaki, [192] J. Deveikyte, H. Geman, C. Piccari, A. Provetti, A sentiment analysis approach
I. Vulić (Eds.), Proceedings of Deep Learning Inside Out (DeeLIO 2022): The to the prediction of market volatility, 2020, arXiv arXiv:2012.05906.
3rd Workshop on Knowledge Extraction and Integration for Deep Learning [193] K.G. Al-Hashedi, P. Magalingam, Financial fraud detection applying data mining
Architectures, Association for Computational Linguistics, Dublin, Ireland and techniques: A comprehensive review from 2009 to 2019, Comp. Sci. Rev. 40
Online, 2022, pp. 53–61. (2021) 100402.
[165] S. Panwar, A. Bansal, F. Zareen, Comparative analysis of large language models [194] J.F. Rodríguez, M. Papale, M. Carminati, S. Zanero, et al., A natural lan-
for question answering from financial documents, in: International Conference guage processing approach for financial fraud detection, in: CEUR workshop
on Communication and Intelligent Systems, Springer, 2023, pp. 297–308. proceedings, Vol. 3260, CEUR-WS. org, 2022, pp. 135–149.

23
K. Du et al. Information Fusion 115 (2025) 102755

[195] P. Boulieris, J. Pavlopoulos, A. Xenos, V. Vassalos, Fraud detection with natural [223] W.J. Yeo, W. van der Heever, R. Mao, E. Cambria, R. Satapathy, G. Mengaldo,
language processing, Mach. Learn. (2023) 1–22. A comprehensive review on financial explainable AI, 2023, arXiv preprint
[196] X. Yang, C. Zhang, Y. Sun, K. Pang, L. Jing, S. Wa, C. Lv, FinChain-BERT: arXiv:2309.11960.
A high-accuracy automatic fraud detection model based on NLP methods for [224] D. Kumar, G.W. Taylor, A. Wong, Opening the black box of financial ai with
financial scenarios, Information 14 (9) (2023) 499. clear-trade: A class-enhanced attentive response approach for explaining and
[197] Y.-J. Chen, C.-H. Wu, Y.-M. Chen, H.-Y. Li, H.-K. Chen, Enhancement of fraud visualizing deep learning-driven stock market prediction, 2017, arXiv preprint
detection for narratives in annual reports, Int. J. Account. Inf. Syst. 26 (2017) arXiv:1709.01574.
32–45. [225] S. Deng, N. Zhang, W. Zhang, J. Chen, J.Z. Pan, H. Chen, Knowledge-driven
[198] W. Dong, S. Liao, Z. Zhang, Leveraging financial social media data for corporate stock trend prediction and explanation via temporal convolutional network, in:
fraud detection, J. Manage. Inf. Syst. 35 (2) (2018) 461–487. Companion Proceedings of the 2019 World Wide Web Conference, 2019, pp.
[199] P. Craja, A. Kim, S. Lessmann, Deep learning for detecting financial statement 678–685.
fraud, Decis. Support Syst. 139 (2020) 113421. [226] S. Carta, A.S. Podda, D. Reforgiato Recupero, M.M. Stanciu, Explainable AI
[200] M.A.K. Achakzai, J. Peng, Detecting financial statement fraud using dynamic for financial forecasting, in: International Conference on Machine Learning,
ensemble machine learning, Int. Rev. Financ. Anal. 89 (2023) 102827. Optimization, and Data Science, Springer, 2021, pp. 51–69.
[201] X. Mao, H. Sun, X. Zhu, J. Li, Financial fraud detection using the related-party [227] K. Ong, W. van der Heever, R. Satapathy, G. Mengaldo, E. Cambria, FinX-
transaction knowledge graph, Procedia Comput. Sci. 199 (2022) 733–740. ABSA: Explainable finance through aspect-based sentiment analysis, in: 2023
[202] P. Seemakurthi, S. Zhang, Y. Qi, Detection of fraudulent financial reports with IEEE International Conference on Data Mining Workshops, ICDMW, 2023, pp.
machine learning techniques, in: 2015 Systems and Information Engineering 773–782.
Design Symposium, IEEE, 2015, pp. 358–361. [228] B. Liang, H. Su, L. Gui, E. Cambria, R. Xu, Aspect-based sentiment analysis
[203] Z. Ahmadi, P. Martens, C. Koch, T. Gottron, S. Kramer, Towards bankruptcy via affective knowledge enhanced graph convolutional networks, Knowl.-Based
prediction: Deep sentiment mining to detect financial distress from business Syst. 235 (2022) 107643.
management reports, in: 2018 IEEE 5th International Conference on Data [229] H. Bandi, S. Joshi, S. Bhagat, D. Ambawade, Integrated technical and sentiment
Science and Advanced Analytics, DSAA, IEEE, 2018, pp. 293–302. analysis tool for market index movement prediction, comprehensible using xai,
[204] A. Markov, Z. Seleznyova, V. Lapshin, Credit scoring methods: Latest trends in: 2021 ICCICT, IEEE, 2021, pp. 1–8.
and points to consider, J. Finance Data Sci. 8 (2022) 180–201. [230] S. Gite, H. Khatavkar, K. Kotecha, S. Srivastava, P. Maheshwari, N. Pandey,
[205] C. Wang, D. Han, Q. Liu, S. Luo, A deep learning approach for credit scoring Explainable stock prices prediction from financial news articles using sentiment
of peer-to-peer lending using attention mechanism LSTM, IEEE Access 7 (2018) analysis, PeerJ Comput. Sci. 7 (2021) e340.
2161–2168. [231] J. Yuan, Z. Zhang, Connecting the dots: forecasting and explaining short-term
market volatility, in: Proceedings of ACM ICAIF, 2020, pp. 1–8.
[206] J. Kriebel, L. Stitz, Credit default prediction from user-generated text in peer-
[232] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, Language models
to-peer lending using deep learning, European J. Oper. Res. 302 (1) (2022)
are unsupervised multitask learners, OpenAI 1 (8) (2019) 9.
309–323.
[233] M. Danilevsky, K. Qian, R. Aharonov, Y. Katsis, B. Kawas, P. Sen, A survey of
[207] O. Netzer, A. Lemaire, M. Herzenstein, When words sweat: Identifying signals
the state of explainable AI for natural language processing, in: K.-F. Wong,
for loan default in the text of loan applications, J. Mar. Res. 56 (6) (2019)
K. Knight, H. Wu (Eds.), Proceedings of the 1st Conference of the Asia-
960–980.
Pacific Chapter of the Association for Computational Linguistics and the 10th
[208] M. Stevenson, C. Mues, C. Bravo, The value of text for small business default
International Joint Conference on Natural Language Processing, Association for
prediction: A deep learning approach, European J. Oper. Res. 295 (2) (2021)
Computational Linguistics, Suzhou, China, 2020, pp. 447–459.
758–771.
[234] T. Lim, Environmental, social, and governance (ESG) and artificial intelligence
[209] R. Matin, C. Hansen, C. Hansen, P. Mø lgaard, Predicting distresses using deep
in finance: State-of-the-art and research takeaways, Artif. Intell. Rev. 57 (4)
learning of text segments in annual reports, Expert Syst. Appl. 132 (2019)
(2024) 1–64.
199–208.
[235] P. Capelli, F. Ielasi, A. Russo, Forecasting volatility by integrating financial risk
[210] F. Mai, S. Tian, C. Lee, L. Ma, Deep learning models for bankruptcy prediction
with environmental, social, and governance risk, Corp. Soc. Responsib. Environ.
using textual disclosures, European J. Oper. Res. 274 (2) (2019) 743–758.
Manage. 28 (5) (2021) 1483–1495.
[211] T. Nieß ner, D.H. Gross, M. Schumann, Evidential strategies in financial
[236] C.-C. Chen, Y.-M. Tseng, J. Kang, A. Lhuissier, M.-Y. Day, T.-T. Tu, H.-H. Chen,
statement analysis: a corpus linguistic text mining approach to bankruptcy
Multi-lingual ESG issue identification, in: Proceedings of the Fifth Workshop
prediction, J. Risk Financ. Manage. 15 (10) (2022) 459.
on Financial Technology and Natural Language Processing and the Second
[212] K. Yang, H. Yuan, R.Y. Lau, PsyCredit: an interpretable deep learning-
Multimodal AI for Financial Forecasting, 2023, pp. 111–115.
based credit assessment approach facilitated by psychometric natural language
[237] P. Glenn, A. Gon, N. Kohli, S. Zha, P.P. Dakle, P. Raghavan, Jetsons at the
processing, Expert Syst. Appl. 198 (2022) 116847.
finnlp-2023: Using synthetic data and transfer learning for multilingual esg issue
[213] Q. Gao, M. Lin, R. Sias, Words matter: The role of readability, tone, and classification, in: Proceedings of the Fifth Workshop on Financial Technology
deception cues in online credit markets, J. Financ. Quant. Anal. 58 (1) (2023) and Natural Language Processing and the Second Multimodal AI for Financial
1–28. Forecasting, 2023, pp. 133–139.
[214] A. El-Qadi, M. Trocan, P. Conde-Cespedes, T. Frossard, N. Díaz-Rodríguez, [238] W. Wang, W. Wei, Q. Song, Y. Wang, Leveraging contrastive learning with bert
Credit risk scoring using a data fusion approach, in: International Conference for esg issue identification, in: Proceedings of the Fifth Workshop on Financial
on Computational Collective Intelligence, Springer, 2023, pp. 769–781. Technology and Natural Language Processing and the Second Multimodal AI
[215] B.-H. Nguyen, V.-N. Huynh, Textual analysis and corporate bankruptcy: A for Financial Forecasting, 2023, pp. 116–120.
financial dictionary-based sentiment approach, J. Oper. Res. Soc. 73 (1) (2022) [239] S. Pasch, D. Ehnes, Nlp for responsible finance: Fine-tuning transformer-based
102–121. models for esg, in: 2022 IEEE International Conference on Big Data, Big Data,
[216] M. Hassan, L.A.-R. Aziz, Y. Andriansyah, The role artificial intelligence in IEEE, 2022, pp. 3532–3536.
modern banking: an exploration of AI-driven approaches for enhanced fraud [240] N. Raman, G. Bang, A. Nourbakhsh, Mapping ESG trends by distant supervision
prevention, risk management, and regulatory compliance, Rev. Contemp. Bus. of neural language models, Mach. Learn. Knowl. Extract. 2 (4) (2020) 453–468.
Anal. 6 (1) (2023) 110–132. [241] B. Koloski, S. Montariol, M. Purver, S. Pollak, Knowledge informed sustainabil-
[217] T. Butler, L. O’Brien, Artificial intelligence for regulatory compliance: Are we ity detection from short financial texts, in: Proceedings of the Fourth Workshop
there yet? J. Financ. Compliance 3 (1) (2019) 44–59. on Financial Technology and Natural Language Processing, FinNLP, 2022, pp.
[218] V. Jain, A. Balakrishnan, D. Beeram, M. Najana, P. Chintale, Leveraging 228–234.
artificial intelligence for enhancing regulatory compliance in the financial [242] N. Smeuninx, B. De Clerck, W. Aerts, Measuring the readability of sustainability
sector, Int. J. Comput. Trends Technol. 72 (5) (2024) 124–140. reports: A corpus-based analysis through standard formulae and NLP, Int. J. Bus.
[219] R. Al-Shabandar, G. Lightbody, F. Browne, J. Liu, H. Wang, H. Zheng, The Commun. 57 (1) (2020) 52–85.
application of artificial intelligence in financial compliance management, in: [243] D. Kouloukoui, N. de Marcellis-Warin, S.M. da Silva Gomes, T. Warin, Mapping
Proceedings of the 2019 International Conference on Artificial Intelligence and global conversations on twitter about environmental, social, and governance
Advanced Manufacturing, 2019, pp. 1–6. topics through natural language processing, J. Clean. Prod. 414 (2023) 137369.
[220] S. Abualhaija, M. Ceci, N. Sannier, D. Bianculli, D.A. ZETZSCHE, M. Bodellini, [244] T. Nugent, N. Stelea, J.L. Leidner, Detecting environmental, social and
Toward automated change impact analysis of financial regulations, in: 2024 governance (ESG) topics using domain-specific language models and data
Workshop on Software Engineering Challenges in Financial Firms, FinanSE’24, augmentation, in: Flexible Query Answering Systems: 14th International Con-
ACM, 2024, pp. 31–32. ference, FQAS 2021, Bratislava, Slovakia, September 19–24, 2021, Proceedings
[221] E. Cambria, R. Mao, M. Chen, Z. Wang, S.-B. Ho, Seven pillars for the future 14, Springer, 2021, pp. 157–169.
of artificial intelligence, IEEE Intell. Syst. 38 (6) (2023) 62–69. [245] S. Ghosh, S.K. Naskar, Ranking environment, social and governance related
[222] P. Weber, K.V. Carl, O. Hinz, Applications of explainable artificial intelligence concepts and assessing sustainability aspect of financial texts, in: Proceedings of
in finance—a systematic review of finance, information systems, and computer the Fourth Workshop on Financial Technology and Natural Language Processing
science literature, Manage. Rev. Q. (2023) 1–41. (FinNLP), 2022, pp. 243–249.

24
K. Du et al. Information Fusion 115 (2025) 102755

[246] B. Sandwidi, S.P. Mukkolakal, Transformers-based approach for a sustainability [261] O. Kraaijeveld, J. De Smedt, The predictive power of public Twitter sentiment
term-based sentiment analysis (STBSA), in: Proceedings of the Second Workshop for forecasting cryptocurrency prices, J. Int. Financ. Mark. Inst. Money 65
on NLP for Positive Impact, NLP4PI, 2022, pp. 157–170. (2020) 101188.
[247] A. Sokolov, K. Caverly, J. Mostovoy, T. Fahoum, L. Seco, Weak supervision and [262] S. Oikonomopoulos, K. Tzafilkou, D. Karapiperis, V. Verykios, Cryptocurrency
black-litterman for automated ESG portfolio construction, J. Financ. Data Sci. price prediction using social media sentiment analysis, in: 2022 13th Interna-
(2021). tional Conference on Information, Intelligence, Systems & Applications, IISA,
[248] T. Guo, N. Jamet, V. Betrix, L.-A. Piquet, E. Hauptmann, ESG2risk: A deep IEEE, 2022, pp. 1–8.
learning framework from ESG news to stock volatility prediction, ERN: Stock [263] M.A. Naeem, I. Mbarki, S.J.H. Shahzad, Predictive role of online investor
Market Risk (Topic) (2020). sentiment for cryptocurrency market: Evidence from happiness and fears, Int.
[249] M. Apel, A. Betzer, B. Scherer, Real-time transition risk, Finance Res. Lett. 53 Rev. Econ. Finance 73 (2021) 496–514.
(2023) 103600. [264] K. Kim, S.-Y. Lee, S. Assar, The dynamics of cryptocurrency market behavior:
[250] C.-L. Jan, Detection of financial statement fraud using deep learning for sentiment analysis using Markov chains, Ind. Manage. Data Syst. ahead-of-print
sustainable development of capital markets under information asymmetry, (2021).
Sustainability 13 (17) (2021) 9879. [265] X. Huang, W. Zhang, X. Tang, M. Zhang, J. Surbiryala, V. Iosifidis, Z. Liu,
[251] S. Mehra, R. Louka, Y. Zhang, Esgbert: Language model to help with classifica- J. Zhang, LSTM based sentiment analysis for cryptocurrency prediction, in:
tion tasks related to companies environmental, social, and governance practices, Database Systems for Advanced Applications: 26th International Conference,
2022, arXiv preprint arXiv:2203.16788. Springer, 2021, pp. 617–621.
[252] E. Haase, R. Sassen, Uncovering lobbying strategies in sustainable finance [266] Z. Leitter, E. Cambria, Non-fungible tokens: What makes them valuable? in:
disclosure regulations using machine learning, J. Environ. Manag. 356 (2024) 2023 IEEE International Conference on Data Mining Workshops, ICDMW, 2023,
120562. pp. 750–756.
[253] H. Lee, S.H. Lee, H. Park, J.H. Kim, H.S. Jung, ESG2PreEM: Automated [267] S.C. Meyns, F. Dalipi, What users tweet on NFTs: mining Twitter to understand
ESG grade assessment framework using pre-trained ensemble models, Heliyon NFT-related concerns using a topic modeling approach, IEEE Access 10 (2022)
(2024). 117658–117680.
[254] T. Schimanski, A. Reding, N. Reding, J. Bingler, M. Kraus, M. Leippold, Bridging
[268] X. Gao, W. Huang, H. Wang, Financial Twitter sentiment on bitcoin return and
the gap in ESG measurement: Using NLP to quantify environmental, social, and
high-frequency volatility, Virtual Econom. 4 (1) (2021) 7–18.
governance communication, Finance Res. Lett. 61 (2024) 104979.
[269] J. Saha, S. Patel, F. Xing, E. Cambria, Does social media sentiment predict
[255] A. Sokolov, J. Mostovoy, J. Ding, L. Seco, Building machine learning systems
bitcoin trading volume? in: Proceedings of the 43rd International Conference
for automated ESG scoring, J. Impact ESG Invest. 1 (3) (2021) 39–50.
[256] A.H. Huang, H. Wang, Y. Yang, FinBERT: A large language model for extracting on Information Systems, ICIS, 2022, pp. 1–9.
information from financial text, Contemp. Account. Res. 40 (2) (2023) 806–841. [270] R. Mao, G. Chen, X. Zhang, F. Guerin, E. Cambria, GPTEval: A survey
[257] G. Fan, X. Wu, Going Green: The Governance Role of Environmental Regulations on assessments of ChatGPT and GPT-4, in: Proceedings of the 2024 Joint
on Firm Innovation and Value, Singapore Management University School of International Conference on Computational Linguistics, Language Resources
Accountancy Research Paper 2023–163, 2022. and Evaluation, LREC-COLING 2024, ELRA and ICCL, Torino, Italia, 2024, pp.
[258] K. Król, D. Zdonek, Digital assets in the eyes of generation z: Perceptions, 7844–7866.
outlooks, concerns, J. Risk Financ. Manage. 16 (1) (2022) 22. [271] B. Chen, Z. Wu, R. Zhao, From fiction to fact: the growing role of generative
[259] K.H. Al-Yahyaee, M.U. Rehman, W. Mensi, I.M.W. Al-Jarrah, Can uncertainty AI in business and finance, J. Chin. Econ. Bus. Stud. 21 (4) (2023) 471–496.
indices predict bitcoin prices? A revisited analysis using partial and multivariate [272] N. Rane, Role and challenges of ChatGPT and similar generative artificial
wavelet approaches, North Am. J. Econom. Finance 49 (2019) 47–56. intelligence in finance and accounting, 2023, Available at SSRN 4603206.
[260] D. Shen, A. Urquhart, P. Wang, Does twitter predict bitcoin? Econom. Lett. 174 [273] G. Shabsigh, E.B. Boukherouaa, Generative artificial intelligence in finance,
(2019) 118–122. FinTech Not. 2023 (006) (2023).

25