Understanding how children learn the components of their mother tongue and the meanings of each word has long fascinated linguists and cognitive scientists. Equally, robots face a similar challenge in understanding language and perception to allow for a natural and effortless human-robot interaction. Acquiring such knowledge is a challenging task, unless this knowledge is preprogrammed, which is no easy task either, nor does it solve the problem of language difference between individuals or learning the meaning of new words. In this thesis, the problem of bootstrapping knowledge in language and vision for autonomous robots is addressed through novel techniques in grammar induction and word grounding to the perceptual world. The learning is achieved in a cognitively plausible loosely-supervised manner from raw linguistic and visual data. The visual data is collected using different robotic platforms deployed in real-world and simulated environments and equipped with different sensing modalities, while the linguistic data is collected using online crowdsourcing tools and volunteers. The presented framework does not rely on any particular robot or any specific sensors; rather it is flexible to what the modalities of the robot can support.
The learning framework is divided into three processes. First, the perceptual raw data is clustered into a number of Gaussian components to learn the ‘visual concepts’. Second, frequent co-occurrence of words and visual concepts are used to learn the language grounding, and finally, the learned language grounding and visual concepts are used to induce probabilistic grammar rules to model the language structure.
In this thesis, the visual concepts refer to: (i) people’s faces and the appearance of their garments; (ii) objects and their perceptual properties; (iii) pairwise spatial relations; (iv) the robot actions; and (v) human activities. The visual concepts are learned by first processing the raw visual data to find people and objects in the scene using state-of-the-art techniques in human pose estimation, object segmentation and tracking, and activity analysis. Once found, the concepts are learned incrementally using a combination of techniques: Incremental Gaussian Mixture Models and a Bayesian Information Criterion to learn simple visual concepts such as object colours and shapes; spatio-temporal graphs and topic models to learn more complex visual concepts, such as human activities and robot actions.
Language grounding is enabled by seeking frequent co-occurrence between words and learned visual concepts. Finding the correct language grounding is formulated as an integer programming problem to find the best many-to-many matches between words and concepts. Grammar induction refers to the process of learning a formal grammar (usually as a collection of re-write rules or productions) from a set of observations. In this thesis, Probabilistic Context Free Grammar rules are generated to model the language by mapping natural language sentences to learned visual concepts, as opposed to traditional supervised grammar induction techniques where the learning is only made possible by using manually annotated training examples on large datasets.
The learning framework attains its cognitive plausibility from a number of sources. First, the learning is achieved by providing the robot with pairs of raw linguistic and visual inputs in a “show-and-tell” procedure akin to how human children learn about their environment. Second, no prior knowledge is assumed about the meaning of words or the structure of the language, except that there are different classes of words (corresponding to observable actions, spatial relations, and objects and their observable properties). Third, the knowledge in both language and vision is obtained in an incremental manner where the gained knowledge can evolve to adapt to new observations without the need to revisit previously seen ones (previous observations). Fourth, the robot learns about the visual world first, then it learns about how it maps to language, which aligns with the findings of cognitive studies on language acquisition in human infants that suggest children come to develop considerable cognitive understanding about their environment in the pre-linguistic period of their lives. It should be noted that this work does not claim to be modelling how humans learn about objects in their environments, but rather it is inspired by it. For validation, four different datasets are used which contain temporally aligned video clips of people or robots performing activities, and sentences describing these video clips. The video clips are collected using four robotic platforms, three robot arms in simple block-world scenarios and a mobile robot deployed in a challenging real-world office environment observing different people performing complex activities. The linguistic descriptions for these datasets are obtained using Amazon Mechanical Turk and volunteers. The analysis performed on these datasets suggest that the learning framework is suitable to learn from complex real-world scenarios. The experimental results show that the learning framework enables (i) acquiring correct visual concepts from visual data; (ii) learning the word grounding for each of the extracted visual concepts; (iii) inducing correct grammar rules to model the language structure; (iv) using the gained knowledge to understand previously unseen linguistic commands; and (v) using the gained knowledge to generate well-formed natural language descriptions of novel scenes.
