Information Retrieval: Parsing, Tokenization & Storage

1. INF 2914Information Retrieval and Web Search Lecture 3: Parsing/Tokenization/Storage These slides are adapted from Stanford’s class CS276 / LING 286 Information Retrieval and Web Mining

2. (Offline) Search Engine Data Flow Parse & Tokenize Global Analysis Index Build Crawler - Scan tokenized web pages, anchor text, etc- Generate text index web page - Parse- Tokenize- Per page analysis - Dup detection- Static rank comp- Anchor text - … 1 2 3 4 in background duptable tokenizedweb pages anchortext ranktable invertedtext index

3. Brutus Calpurnia Caesar Dictionary Postings lists Inverted index Posting 2 4 8 16 32 64 128 1 2 3 5 8 13 21 34 13 16 Sorted by docID (more later on why).

4. Tokenizer Friends Romans Countrymen Token stream. Linguistic modules friend friend roman countryman Modified tokens. roman Indexer 2 4 countryman 1 2 Inverted index. 16 13 Inverted index construction Documents to be indexed. Friends, Romans, countrymen.

5. Plan for this lecture • The Dictionary • Parsing • Tokenization • What terms do we put in the index? • Storage • Log structured file systems • XML Introduction

6. Parsing a document • What format is it in? • pdf/word/excel/html? • What language is it in? • What character set is in use? Each of these is a classification problem. But these tasks are often done heuristically …

7. Complications: Format/language • Documents being indexed can include docs from many different languages • A single index may have to contain terms of several languages. • Sometimes a document or its components can contain multiple languages/formats • French email with a German pdf attachment. • What is a unit document? • A file? • An email? (Perhaps one of many in an mbox.) • An email with 5 attachments? • A group of files (PPT or LaTeX in HTML)

8. Tokenization

9. Tokenization • Input: “Friends, Romans and Countrymen” • Output: Tokens • Friends • Romans • Countrymen • Each such token is now a candidate for an index entry, after further processing • Described below • But what are valid tokens to emit?

10. Tokenization • Issues in tokenization: • Finland’s capital  Finland? Finlands? Finland’s? • Hewlett-Packard  Hewlett and Packard as two tokens? • State-of-the-art: break up hyphenated sequence. • co-education ? • the hold-him-back-and-drag-him-away-maneuver ? • It’s effective to get the user to put in possible hyphens • San Francisco: one token or two? How do you decide it is one token?

11. Numbers • 3/12/91 Mar. 12, 1991 • 55 B.C. • B-52 • My PGP key is 324a3df234cb23e • 100.2.86.144 • Often, don’t index as text. • But often very useful: think about things like looking up error codes/stacktraces on the web • (One answer is using n-grams, lectures 6 and 7) • Will often index “meta-data” separately • Creation date, format, etc.

12. Tokenization: Language issues • L'ensemble one token or two? • L ? L’ ? Le ? • Want l’ensemble to match with un ensemble • German noun compounds are not segmented • Lebensversicherungsgesellschaftsangestellter • ‘life insurance company employee’

13. Katakana Hiragana Kanji Romaji Tokenization: language issues • Chinese and Japanese have no spaces between words: • 莎拉波娃现在居住在美国东南部的佛罗里达。 • Not always guaranteed a unique tokenization • Further complicated in Japanese, with multiple alphabets intermingled • Dates/amounts in multiple formats フォーチュン500社は情報不足のため時間あた$500K(約6,000万円) End-user can express query entirely in hiragana!

14. Tokenization: language issues • Arabic (or Hebrew) is basically written right to left, but with certain items like numbers written left to right • Words are separated, but letter forms within a word form complex ligatures • استقلت الجزائر في سنة 1962 بعد 132 عاما من الاحتلال الفرنسي. • ← → ← → ← start • ‘Algeria achieved its independence in 1962 after 132 years of French occupation.’ • With Unicode, the surface presentation is complex, but the stored form is straightforward

15. Normalization • Need to “normalize” terms in indexed text as well as query terms into the same form • We want to match U.S.A. and USA • We most commonly implicitly define equivalence classes of terms • e.g., by deleting periods in a term • Alternative is to do asymmetric expansion: • Enter: window Search: window, windows • Enter: windows Search: windows • Potentially more powerful, but less efficient • Execute queries in parallel or do a second pass over the index

16. Normalization: other languages • Accents: résumé vs. resume. • Most important criterion: • How are your users like to write their queries for these words? • Even in languages that have accents, users often may not type them • German: Tuebingen vs. Tübingen • Should be equivalent

17. 7月30日 vs. 7/30 Is this German “mit”? Morgen will ich in MIT … Normalization: other languages • Need to “normalize” indexed text as well as query terms into the same form • Character-level alphabet detection and conversion • Tokenization not separable from this. • Sometimes ambiguous:

18. Case folding • Reduce all letters to lower case • exception: upper case (in mid-sentence?) • e.g., General Motors • Fed vs. fed • SAIL vs. sail • Often best to lower case everything, since users will use lowercase regardless of ‘correct’ capitalization…

19. Stop words • With a stop list, you exclude from dictionary entirely the commonest words. Intuition: • They have little semantic content: the, a, and, to, be • They take a lot of space: ~30% of postings for top 30 • But the trend is away from doing this: • Good compression techniques means the space for including stopwords in a system is very small • Good query optimization techniques mean you pay little at query time for including stop words. • You need them for: • Phrase queries: “King of Denmark” • Various song titles, etc.: “Let it be”, “To be or not to be” • “Relational” queries: “flights to London”

20. Thesauri and soundex • Handle synonyms and homonyms • Hand-constructed equivalence classes • e.g., car = automobile • color = colour • Rewrite to form equivalence classes • Index such equivalences • When the document contains automobile, index it under car as well (usually, also vice-versa) • Or expand query? • When the query contains automobile, look under car as well

21. Soundex • Traditional class of heuristics to expand a query into phonetic equivalents • Language specific – mainly for names • E.g., chebyshev tchebycheff

22. Lemmatization • Reduce inflectional/variant forms to base form • E.g., • am, are,is be • car, cars, car's, cars'car • the boy's cars are different colorsthe boy car be different color • Lemmatization implies doing “proper” reduction to dictionary headword form

23. Stemming • Reduce terms to their “roots” before indexing • “Stemming” suggest crude affix chopping • language dependent • e.g., automate(s), automatic, automation all reduced to automat. for exampl compress and compress ar both accept as equival to compress for example compressed and compression are both accepted as equivalent to compress.

24. Porter’s algorithm • Commonest algorithm for stemming English • Results suggest at least as good as other stemming options • Conventions + 5 phases of reductions • phases applied sequentially • each phase consists of a set of commands • sample convention: Of the rules in a compound command, select the one that applies to the longest suffix.

25. Typical rules in Porter • sses ss • ies i • ational ate • tional tion

26. Other stemmers • Other stemmers exist, e.g., Lovins stemmer http://www.comp.lancs.ac.uk/computing/research/stemming/general/lovins.htm • Single-pass, longest suffix removal (about 250 rules) • Motivated by linguistics as well as IR • Full morphological analysis – at most modest benefits for retrieval • Do stemming and other normalizations help? • Often very mixed results: really help recall for some queries but harm precision on others

27. Language-specificity • Many of the above features embody transformations that are • Language-specific and • Often, application-specific • These are “plug-in” addenda to the indexing process • Both open source and commercial plug-ins available for handling these

28. Per document analysis Global analysis Indexing Search indexes Index Build Flow - Overview Crawled documents

29. Per document analysis • Multi-format parsing • Handles different files types (HTML, PDF, PowerPoint, etc) • Multi-language tokenization, stemming, synonyms, user-defined annotations, etc. • Per document analysis is tipically the bottleneck of the index build process • 50 times slower than I/O • Indexing can be done at I/O speed

30. Per document analysis Incorporating per document analysis • Per document analysis is much slower than indexing • Store tokenized documents in a scalable document store Document store Crawled documents

31. Storage

32. Document store • Log-structured file system • Only the most recent version of each document is accessible • No in place updates • Documents are grouped into bundles to optimize I/O • Typically built over the file system • 3 basic operation modes • Document insertion (during per-document analysis) • Sequential access for index build • Random access during query processing

33. Header attributes # docs hash(URL) timestamp attributes offset # docs hash(URL) timestamp attributes offset Doc 2 # fields field ID length offset # fields field ID length offset field ID length offset data data # fields data Doc 1 # fields field ID length offset data Store design (1/5) • Bundle disk layout • Fixed bundle size (for instance 8MB) • All fields are 64-bit aligned • All fields are binary • Store does not know how to interpret fields • Compression • Fields are tokens, anchor text, URL, shingle, statistics, etc.

34. Store design (2/5) • Document insertion uses a double buffering algorithm and asynchronous I/O • Try to fit as many documents as it can in a bundle • Schedule write for bundle • Start writing the next bundle

35. Store design (3/5) • Store is sequentially scanned during index build and global analysis • A double buffering algorithm with asynchronous I/O is also used here • Return only the newest version of each document • Store is accessed in reverse order • LFS semantics currentBuffer nextBuffer Bundle# 1053 Bundle# 1052

36. probe copy set Store design (4/5) • Store cleanup algorithm • “Smarter” algorithms can be used if we are not I/O bound • Avoid seeks New documents bundle Bloom filter Storei+1 D1’ 1 D5’ bundle 1 D6 0 D1’ bundle 1 D5’ 0 D3 Storei D6 1 D4 0 bundle D2 1 0 * D1 bundle D3 D5 * D4 D2 * garbage collected

37. Bloom Filters (1/2) • Compact data structures for a probabilistic representation of a set • Appropriate to answer membership queries • False positives!

38. Bloom Filters (2/2) Query for b: check the bits at positions H1(b), H2(b), ..., H4(b).

39. Store design (5/5) • During runtime the summarizer uses the store to fetch the tokens (random access) • Store provides an API call for retrieving a set of documents (e.g. 20) given its bundle number and offset in the file • Internally the store uses a buffer pool for documents • Asynchronous I/O is used for exploiting parallelism from the storage subsystem • Summarizer releases the documents after it is done

40. Storage Issues • Performance • Fault tolerance • Distribution • Redundancy • Field compression Google File System tries to address these issues

41. XML Introduction

42. Preliminaries: XML x0 <conference> <name> PODS </name> <speaker> <name> Josifovski </name> <paper_cnt> 1 </paper_cnt> </speaker> <speaker> <name> Fagin </name> <paper_cnt> 3 </paper_cnt> </speaker> </conference> root conference x1 x2 name x6 PODS x3 speaker speaker x8 x4 x5 x7 paper_cnt name paper_cnt name Josifovski 3 1 Fagin

43. Preliminaries: XPath 1.0 /conference[name =PODS]/speaker[paper_cnt >1]/name Query Document root x0 root conference conference x1 x2 name name speaker x6 = PODS PODS x3 speaker speaker x8 paper_cnt x4 x5 > 1 x7 name paper_cnt name paper_cnt name Josifovski 3 1 Fagin Result: { x7 }

44. //article//section[ //title contains(‘Query Processing’) AND //figure//caption contains(‘XML’)] In an index-based method, 8 tags and text elements need to be verified to process this query (lessons 6 and 7) article section title figure caption “XML” XML Indexing “Query Processing”

45. (0,7,0) R (1,5,1) B3 A1 (6,7,1) (3,5,2) (7,7,2) B2 B1 C2 (2,2,2) (4,4,3) C1 D1 (5,5,3) Position Encoding • Scheme #1: Begin/End/Level • Begin: preorder position of tag/text • End: preorder position of last descendent • Level: depth • Containment: X contains Y iff X.begin < Y.begin <= X.end (assuming well-formed)

46. (1) R (1.1) B3 A1 (1.2) (1.1.2) (1.2.1) B2 B1 C2 (1.1.1) C1 D1 (1.1.2.1) (1.1.2.2) Position Encoding • Scheme #2: Dewey • Position of element E = {position of parent}.n, where E is the nth child of its parent • Containment: X contains Y iff X is a prefix of Y

47. Position Encoding • Begin/End/Level • Typically more compact • Fewer implementation issues • Dewey • Encodes positions of all ancestors

48. R B3 A1 B1 B2 C2 C1 D1 Path Index Path ID /R 1 /R/A 2 /R/A/B 3 /R/A/B/C 4 /R/A/B/D 5 /R/B 6 /R/B/C 7 Path Pattern -> Set of matching path IDs /R/B -> {6} //R//C -> {4, 7}

49. Basic Access Path • Inverted posting lists • Posting: <Token, Location> • Token = <Term/Tag> • Location = <DocumentID, Position> • Exercise: Create the posting list representation for the following XML document R B3 A1 B1 B2 C2 C1 D1

50. Brutus Calpurnia Caesar Dictionary Postings lists Inverted index Posting 2 4 8 16 32 64 128 1 2 3 5 8 13 21 34 13 16 Sorted by docID (Why on lessons 6/7).