Introduction to Information Retrieval

Introduction to
Information Retrieval
CS276: Information Retrieval and Web Search
Christopher Manning and Pandu Nayak

Lecture 4: Index Compression

Introduction to Information Retrieval

Last lecture – index construction

 Sort-based indexing
 Naïve in-memory inversion
 Blocked Sort-Based Indexing (BSBI)
 Merge sort is effective for hard disk–based sorting (avoid seeks!)
 Single-Pass In-Memory Indexing (SPIMI)
 No global dictionary
 Generate separate dictionary for each block
 Don’t sort postings
 Accumulate postings in postings lists as they occur
 Distributed indexing using MapReduce
 Dynamic indexing: Multiple indices, logarithmic merge
2
Introduction to Information Retrieval Ch. 5

Today

 Collection statistics in more detail (with RCV1)

 How big will the dictionary and postings be?
 Dictionary compression
 Postings compression

3
Introduction to Information Retrieval Ch. 5

Why compression (in general)?

 Use less disk space
 Save a little money; give users more space
 Keep more stuff in memory
 Increases speed
 Increase speed of data transfer from disk to memory
 [read compressed data | decompress] is faster than
[read uncompressed data]

 Premise: Decompression algorithms are fast

 True of the decompression algorithms we use

4
Introduction to Information Retrieval Ch. 5

Why compression for inverted indexes?

 Dictionary
 Make it small enough to keep in main memory
 Make it so small that you can keep some postings lists in
main memory too
 Postings file(s)
 Reduce disk space needed
 Decrease time needed to read postings lists from disk
 Large search engines keep a significant part of the postings
in memory.
 Compression lets you keep more in memory
 We will devise various IR-specific compression schemes
5
Introduction to Information Retrieval Sec. 5.1

Recall Reuters RCV1

 symbol statistic
value
 N documents
800,000
 L avg. # tokens per doc
200
 M terms (= word types)
~400,000
 avg. # bytes per token 6
(incl. spaces/punct.)
 avg. # bytes per token 4.5 6
Introduction to Information Retrieval Sec. 5.1

Lossless vs. lossy compression

 Lossless compression: All information is preserved.
 What we mostly do in IR.
 Lossy compression: Discard some information
 Several of the preprocessing steps can be viewed as
lossy compression: case folding, stop words,
stemming, number elimination.
 Chapter 7: Prune postings entries that are unlikely to
turn up in the top k list for any query.
 Almost no loss of quality in top k list.

7
Introduction to Information Retrieval Sec. 5.1

Vocabulary size vs. collection size

 How big is the term vocabulary?
 That is, how many distinct words are there?
 Can we assume an upper bound?
 Not really: At least 7020 = 1037 different words of length 20
 In practice, the vocabulary will keep growing with the
collection size
 Especially with Unicode 

8
Introduction to Information Retrieval Ch. 5

Compression
 Now, we will consider compressing the space
for the dictionary and postings. We’ll do:
 Basic Boolean index only
 No study of positional indexes, etc.
 But these ideas can be extended

 We will consider compression schemes

9
Introduction to Information Retrieval Sec. 5.2

DICTIONARY COMPRESSION

10
Introduction to Information Retrieval Sec. 5.2

Why compress the dictionary?

 Search begins with the dictionary
 We want to keep it in memory
 Memory footprint competition with other
applications
 Embedded/mobile devices may have very little
memory
 Even if the dictionary isn’t in memory, we want it to
be small for a fast search startup time
 So, compressing the dictionary is important

11
Introduction to Information Retrieval Sec. 5.2

Dictionary storage – naïve version

 Array of fixed-width entries
 ~400,000 terms; 28 bytes/term = 11.2 MB.

Terms Freq. Postings ptr.

a 656,265
aachen 65
…. ….
zulu 221

Dictionary search 20 bytes 4 bytes each

structure 12
Introduction to Information Retrieval Sec. 5.2

Fixed-width terms are wasteful

 Most of the bytes in the Term column are wasted –
we allot 20 bytes for 1 letter terms.
 And we still can’t handle supercalifragilisticexpialidocious or
hydrochlorofluorocarbons.
 Written English averages ~4.5 characters/word.
 Exercise: Why is/isn’t this the number to use for estimating
the dictionary size?
 Ave. dictionary word in English: ~8 characters
 How do we use ~8 characters per dictionary term?
 Short words dominate token counts but not type
average.
13
Introduction to Information Retrieval Sec. 5.2

Compressing the term list:

Dictionary-as-a-String
 Store dictionary as a (long) string of characters:

Pointer to next word shows end of current word

Hope to save up to 60% of dictionary space
….systilesyzygeticsyzygialsyzygyszaibelyiteszczecinszomo….

Freq. Postings ptr. Term ptr.

Total string length =
33
400K x 8B = 3.2MB
29
44
Pointers resolve 3.2M
126 positions: log23.2M =
22bits = 3bytes

14
Introduction to Information Retrieval Sec. 5.2

Space for dictionary as a string

 4 bytes per term for Freq.
 Now avg. 11
 4 bytes per term for pointer to Postings.  bytes/term,
 3 bytes per term pointer  not 20.
 Avg. 8 bytes per term in term string
 400K terms x 19  7.6 MB (against 11.2MB for fixed
width)

15
Introduction to Information Retrieval Sec. 5.2

Blocking
 Store pointers to every kth term string.
 Example below: k=4.
 Need to store term lengths (1 extra byte)
….7systile9syzygetic8syzygial6syzygy11szaibelyite8szczecin9szomo….

Freq. Postings ptr. Term ptr.

33
29
 Save 9 bytes Lose 4 bytes on
44
 on 3 term lengths.
126  pointers.
7
16
Introduction to Information Retrieval Sec. 5.2

Blocking Net Gains

 Example for block size k = 4
 Where we used 3 bytes/pointer without blocking
 3 x 4 = 12 bytes,
now we use 3 + 4 = 7 bytes.

Shaved another ~0.5MB. This reduces the size of the

dictionary from 7.6 MB to 7.1 MB.
We can save more with larger k.

Question: Why not go with larger k?

17
Introduction to Information Retrieval Sec. 5.2

Dictionary search without blocking

 Assuming each
dictionary term equally
likely in query (not really
so in practice!), average
number of comparisons
= (1+2∙2+4∙3+4)/8 ~2.6

Exercise: what if the frequencies

of query terms were non-uniform
but known, how would you
structure the dictionary search
tree?
18
Introduction to Information Retrieval Sec. 5.2

Dictionary search with blocking

 Binary search down to 4-term block;

 Then linear search through terms in block.
 Blocks of 4 (binary tree), avg. =
(1+2∙2+2∙3+2∙4+5)/8 = 3 compares
19
Introduction to Information Retrieval Sec. 5.2

Exercises
 Estimate the space usage (and savings compared to
7.6 MB) with blocking, for block sizes of k = 4, 8 and
16.

 Estimate the impact on search performance (and

slowdown compared to k=1) with blocking, for block
sizes of k = 4, 8 and 16.

20
Introduction to Information Retrieval Sec. 5.2

Front coding
 Front-coding:
 Sorted words commonly have long common prefix – store
differences only
 (for last k-1 in a block of k)
8automata8automate9automatic10automation

8automat*a1e2ic3ion

Encodes prefix automat Extra length

beyond automat.
Begins to resemble general string compression. 21
Introduction to Information Retrieval Sec. 5.2

RCV1 dictionary compression summary

Technique Size in MB

Fixed width 11.2

Dictionary-as-String with pointers to every term 7.6

+ blocking, k = 4 7.1

+ blocking + front coding 5.9

22
Introduction to Information Retrieval Sec. 5.3

POSTINGS COMPRESSION

23
Introduction to Information Retrieval Sec. 5.3

Postings compression
 The postings file is much larger than the dictionary,
factor of at least 10, often over 100 times larger
 Key desideratum: store each posting compactly.
 A posting for our purposes is a docID.
 For Reuters (800,000 documents), we would use 32
bits per docID when using 4-byte integers.
 Alternatively, we can use log2 800,000 ≈ 20 bits per
docID.
 Our goal: use far fewer than 20 bits per docID.

24
Introduction to Information Retrieval Sec. 5.3

Postings: two conflicting forces

 A term like arachnocentric occurs in maybe one doc
out of a million – we would like to store this posting
using log2 1M ≈ 20 bits.
 A term like the occurs in virtually every doc, so 20
bits/posting ≈ 2MB is too expensive.
 Prefer 0/1 bitmap vector in this case (≈100K)

25
Introduction to Information Retrieval Sec. 5.3

Gap encoding of postings file entries

 We store the list of docs containing a term in
increasing order of docID.
 computer: 33,47,154,159,202 …
 Consequence: it suffices to store gaps.
 33,14,107,5,43 …
 Hope: most gaps can be encoded/stored with far
fewer than 20 bits.
 Especially for common words

26
Introduction to Information Retrieval Sec. 5.3

Three postings entries

27
Introduction to Information Retrieval Sec. 5.3

Variable length encoding

 Aim:
 For arachnocentric, we will use ~20 bits/gap entry.
 For the, we will use ~1 bit/gap entry.
 If the average gap for a term is G, we want to use
~log2G bits/gap entry.
 Key challenge: encode every integer (gap) with about
as few bits as needed for that integer.
 This requires a variable length encoding
 Variable length codes achieve this by using short
codes for small numbers
28
Introduction to Information Retrieval

Unary code
 Represent n as n 1s with a final 0.
 Unary code for 3 is 1110.
 Unary code for 40 is
11111111111111111111111111111111111111110 .
 Unary code for 80 is:
11111111111111111111111111111111111111111111
1111111111111111111111111111111111110
 This doesn’t look promising, but….
 Optimal if P(n) = 2–n
 We can use it as part of our solution
29
Introduction to Information Retrieval Sec. 5.3

Gamma codes
 We can compress better with bit-level codes
 The Gamma code is the best known of these.
 Represent a gap G as a pair length and offset
 offset is G in binary, with the leading bit cut off
 For example 13 → 1101 → 101
 length is the length of offset
 For 13 (offset 101), this is 3.
 We encode length with unary code: 1110.
 Gamma code of 13 is the concatenation of length
and offset: 1110101
30
Introduction to Information Retrieval Sec. 5.3

Gamma code examples

number length offset g-code
0 none
1 0 0
2 10 0 10,0
3 10 1 10,1
4 110 00 110,00
9 1110 001 1110,001
13 1110 101 1110,101
24 11110 1000 11110,1000
511 111111110 11111111 111111110,11111111
1025 11111111110 0000000001 11111111110,0000000001

31
Introduction to Information Retrieval

Reminder: bitwise operations

 For compression, you need to use bitwise operators

 Python (and most everything else):

 & bitwise and; | bitwise or; ^ bitwise xor; ~ ones
complement
 << left shift bits, >> right shift; LACKS >>> zero fill right shift
 Recipes:
 Extract 7 bits: a & 0x7f00 >> 8 ; if take high-order bit add: & 0x7f
 Combine 3 5-bit numbers: a | (b << 5) | (c << 10)
32
 Lookup tables rather than decoding can be faster, yet still small
Introduction to Information Retrieval Sec. 5.3

Gamma code properties

 G is encoded using 2 log G + 1 bits
 Length of offset is log G bits
 Length of length is log G + 1 bits
 All gamma codes have an odd number of bits
 Almost within a factor of 2 of best possible, log2 G

 Gamma code is uniquely prefix-decodable, like VB

 Gamma code can be used for any distribution
 Optimal for P(n)  1/(2n2)
 Gamma code is parameter-free
33
Introduction to Information Retrieval Sec. 5.3

Gamma seldom used in practice

 Machines have word boundaries – 8, 16, 32, 64 bits
 Operations that cross word boundaries are slower
 Compressing and manipulating at the granularity of
bits can be too slow

 All modern practice is to use byte or word aligned

codes
 Variable byte encoding is a faster, conceptually simpler
compression scheme, with decent compression

34
Introduction to Information Retrieval Sec. 5.3

Variable Byte (VB) codes

 For a gap value G, we want to use close to the fewest
bytes needed to hold log2 G bits
 Begin with one byte to store G and dedicate 1 bit in it
to be a continuation bit c
 If G ≤127, binary-encode it in the 7 available bits and
set c =1
 Else encode G’s lower-order 7 bits and then use
additional bytes to encode the higher order bits
using the same algorithm
 At the end set the continuation bit of the last byte to
1 (c =1) – and for the other bytes c = 0. 35
Introduction to Information Retrieval Sec. 5.3

Example
docIDs 824 829 215406
gaps 5 214577
VB code 00000110 10000101 00001101
10111000 00001100
10110001

Postings stored as the byte concatenation

000001101011100010000101000011010000110010110001

Key property: VB-encoded postings are

uniquely prefix-decodable.

For a small gap (5), VB

uses a whole byte. 36
Introduction to Information Retrieval Sec. 5.3

RCV1 compression
Data structure Size in MB
dictionary, fixed-width 11.2
dictionary, term pointers into string 7.6
with blocking, k = 4 7.1
with blocking & front coding 5.9
collection (text, xml markup etc) 3,600.0
collection (text) 960.0
Term-doc incidence matrix 40,000.0
postings, uncompressed (32-bit words) 400.0
postings, uncompressed (20 bits) 250.0
postings, variable byte encoded 116.0
postings, g-encoded 101.0

37
Introduction to Information Retrieval Sec. 5.3

Other variable unit codes

 Variable byte codes are used by many real systems
 Good low-tech blend of variable-length coding and
sensitivity to computer memory alignment matches
 Byte alignment wastes space if you have many small
gaps – as gap encoding often makes
 More modern work mainly uses the ideas:
 Be word aligned (32 or 64 bits; even faster)
 Encode several gaps at the same time
 Often assume a maximum gap size, perhaps with an
escape

38
Introduction to Information Retrieval

Group Variable Integer code

 Used by Google around turn of millennium….
 Jeff Dean, keynote at WSDM 2009 and presentations at CS276
 Encodes 4 integers in blocks of size 5–17 bytes
 First byte: four 2-bit binary length fields
 L1 L2 L3 L4 , Lj{1,2,3,4}
 Then, L1+L2+L3+L4 bytes (between 4–16) hold 4 numbers
 Each number can use 8/16/24/32 bits. Max gap length ~4 billion
 It was suggested that this was about twice as fast as VB
encoding
 Decoding gaps is much simpler – no bit masking
 First byte can be decoded with lookup table or switch

39
Introduction to Information Retrieval

Simple-9 [Anh & Moffat, 2004]

A word-aligned, multiple number encoding scheme
How can we store several numbers in 32 bits with a format selector?

28 1-bit numbers

14 2-bit numbers

9 3-bit numbers

7 4-bit numbers
“selectors”
(9 total ways)
Introduction to Information Retrieval

Simple9 Encoding Scheme [Anh & Moffat, 2004]

 Encoding block: 4 bytes (32 bits)
 Most significant nibble (4 bits) describe the layout of the 28
other bits as follows: n numbers of b bits each
Layout

 0: a single 28-bit number (4 bits) n * b  28

 1: two 14-bit numbers

 2: three 9-bit numbers (and one spare bit)
 3: four 7-bit numbers
 4: five 5-bit numbers (and three spare bits)
 5: seven 4-bit numbers
 6: nine 3-bit numbers (and one spare bit)
 7: fourteen two-bit numbers
 8: twenty-eight one-bit numbers
 Simple16 is a variant with 5 additional (uneven) configurations
 Efficiently decoded with hand-coded decoder, using bit masks
 Extended Simple Family – idea applies to 64-bit words, etc.
41
Introduction to Information Retrieval Sec. 5.3

Index compression summary

 We can now create an index for highly efficient
Boolean retrieval that is very space efficient
 Only 4% of the total size of the collection
 Only 10-15% of the total size of the text in the
collection

 We’ve ignored positional information

 Hence, space savings are less for indexes used in
practice
 But techniques substantially the same
42
Introduction to Information Retrieval Ch. 5

Resources for today’s lecture

 IIR 5
 MG 3.3, 3.4.
 F. Scholer, H.E. Williams and J. Zobel. 2002.
Compression of Inverted Indexes For Fast Query
Evaluation. Proc. ACM-SIGIR 2002.
 Variable byte codes
 V. N. Anh and A. Moffat. 2005. Inverted Index
Compression Using Word-Aligned Binary Codes.
Information Retrieval 8: 151–166.
 Word aligned codes

43