Text Processing & Characteristics Kron Text •Text parsing –Tokenization, terms –A bit of linguistics •Text characteristics –Zipfs law Interface Query Engine Indexer Index Crawler Users Web A Typical Web Search Engine Text processing Focus on documents •Decide what is an individual document –Can vary depending on problem •Documents are basic units consisting of a sequence of tokens or terms and are to be indexed. •Terms (derived from tokens) are words or roots of words, semantic units or phrases which are the atoms of indexing •Repositories (databases) and corpora are collections of documents. •Query is a request for documents on a query-related topic. Building an index •Collect documents to be indexed –Create your corpora •Tokenize the text •Linguistic processing •Build the inverted index from terms What is a Document? •A document is a digital object –Indexable •Can be queried and potentially retrieved. •Many types of documents –Text –Image –Audio –Video –Data –Email –Others? • What is Text? •Text is so common that we often ignore its importance •What is text? –Strings of characters (alphabets, ideograms, ascii, unicode, etc.) •Words •. , : ; - ( ) _ •Symbols •1 2 3, 3.1415, 1010 •f = ma, H20 •Tables •Figures – –Anything that is not an image, etc. –Why is text important? •Text is language capture –an instantiation of language, culture, science, etc. Collection of text •Corpora: collection of texts –especially if complete and self contained; the corpus of Anglo-Saxon verse –Special collection •In linguistics and lexicography, a body of texts, utterances or other specimens considered more or less representative of a language and usually stored as an electronic database (The Oxford Companion to the English Language 1992) •A collection of naturally occurring language text chosen to characterize a state or variety of a language (John Sinclair Corpus Concordance Collocation OUP 1991) • •Types: –Written vs Spoken –General vs Specialized –Monolingual vs Multilingual •e.g. Parallel, Comparable –Synchronic (at a particular pt in time) vs Diachronic (over time) –Annotated vs Unannotated –Indexed vs unindexed –Static vs dynamic • Written corpora B r o w n L O B T i m e o f c o m p i l a t i o n 1 9 6 0 s 1 9 7 0 s C o m p i l e d a t B r o w n U n i v e r s i t y ( U S ) L a n c a s t e r , O s l o , B e r g e n L a n g u a g e v a r i e t y W r i t t e n A m e r i c a n E n g l i s h W r i t t e n B r i t i s h E n g l i s h S i z e 1 m i l l i o n w o r d s ( 5 0 0 t e x t s o f 2 0 0 0 w o r d s e a c h ) D e s i g n B a l a n c e d c o r p o r a ; 1 5 g e n r e s o f t e x t , i n c l . p r e s s r e p o r t a g e , e d i t o r i a l s , r e v i e w s , r e l i g i o n , g o v e r n m e n t d o c u m e n t s , r e p o r t s , b i o g r a p h i e s , s c i e n t i f i c w r i t i n g , f i c t i o n Text Processing •Standard Steps: –Recognize document structure •titles, sections, paragraphs, etc. –Break into tokens – type of markup •Tokens are delimited text –Hello, how are you. –_hello_,_how_are_you_._ •usually space and punctuation delineated •special issues with Asian languages –Stemming/morphological analysis –What is left are terms –Store in inverted index •Lexical analysis is the process of converting a sequence of characters into a sequence of tokens. –A program or function which performs lexical analysis is called a lexical analyzer, lexer or scanner. – Basic indexing pipeline Tokenizer Token stream. Friends Romans Countrymen Linguistic modules Modified tokens (terms). friend roman countryman Indexer Inverted index. friend roman countryman 2 4 2 13 16 1 Documents to be indexed. Friends, Romans, countrymen. Parsing a document (lexical analysis) •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 which can be solved using heuristics or Machine Learning methods. But there are complications … Format/language stripping •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 Portuguese pdf attachment. •What is a unit document? –An email? –With attachments? –An email with a zip containing documents? Document preprocessing •Convert byte sequences into a linear sequence of characters •Trivial with ascii, but not so with Unicode or others –Use ML classifiers or heuristics. – •Crucial problem for commercial system! • Tokenization •Parsing (chopping up) the document into basic units that are candidates for later indexing –What parts of text to use and what not •Issues with –Punctuation –Numbers –Special characters –Equations –Formula –Languages –Normalization (often by stemming) – 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? 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 ? –San Francisco: one token or two? How do you decide it is one token? Numbers •3/12/91 •Mar. 12, 1991 •55 B.C. •B-52 •My PGP key is 324a3df234cb23e •100.2.86.144 –Generally, don’t index as text. –Will often index “meta-data” separately •Creation date, format, etc. Tokenization: Language issues •L'ensemble ® one token or two? –L ? L’ ? Le ? –Want ensemble to match with un ensemble – •German noun compounds are not segmented –Lebensversicherungsgesellschaftsangestellter –‘life insurance company employee’ Tokenization: language issues •Chinese and Japanese have no spaces between words: –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) Katakana Hiragana Kanji “Romaji” End-user can express query entirely in hiragana! 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 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 limited expansion: –Enter: window Search: window, windows –Enter: windows Search: Windows, windows –Enter: Windows Search: Windows •Potentially more powerful, but less efficient – 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 Normalizing Punctuation •Ne’er vs. never: use language-specific, handcrafted “locale” to normalize. –Which language? –Most common: detect/apply language at a pre-determined granularity: doc/paragraph. •U.S.A. vs. USA – remove all periods or use locale. •a.out 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 Soundex •Traditional class of heuristics to expand a query into phonetic equivalents –Language specific – mainly for names –E.g., chebyshev ® tchebycheff Stemming and Morphological Analysis •Goal: “normalize” similar words •Morphology (“form” of words) –Inflectional Morphology •E.g,. inflect verb endings and noun number •Never change grammatical class –dog, dogs –Derivational Morphology •Derive one word from another, •Often change grammatical class –build, building; health, healthy • • 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 Stemming Morphological variants of a word (morphemes). Similar terms derived from a common stem: engineer, engineered, engineering use, user, users, used, using Stemming in Information Retrieval. Grouping words with a common stem together. For example, a search on reads, also finds read, reading, and readable Stemming consists of removing suffixes and conflating the resulting morphemes. Occasionally, prefixes are also removed. 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 example compressed and compression are both accepted as equivalent to compress. for exampl compress and compress ar both accept as equival to compress 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. • Typical rules in Porter •sses ® ss •ies ® i •ational ® ate •tional ® tion • • Weight of word sensitive rules • (m>1) EMENT → •replacement → replac •cement → cement • 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 Automated Methods are the norm •Powerful multilingual tools exist for morphological analysis –PCKimmo, Xerox Lexical technology –Require a grammar and dictionary –Use “two-level” automata •Stemmers: –Very dumb rules work well (for English) –Porter Stemmer: Iteratively remove suffixes –Improvement: pass results through a lexicon Porter’s algorithm •Commonest algorithm for stemming English •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. •Typical rules –sses ® ss –ies ® i –ational ® ate –tional ® tion • Categories of Stemmer The following diagram illustrate the various categories of stemmer. Porter's algorithm is shown by the red path. Conflation methods Manual Automatic (stemmers) Affix Successor Table n-gram removal variety lookup Longest Simple match removal Comparison of stemmers Stemming in Practice Evaluation studies have found that stemming can affect retrieval performance, usually for the better, but the results are mixed. • Effectiveness is dependent on the vocabulary. Fine distinctions may be lost through stemming. • Automatic stemming is as effective as manual conflation. • Performance of various algorithms is similar. Porter's Algorithm is entirely empirical, but has proved to be an effective algorithm for stemming English text with trained users. 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 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 standardly have accents, users often may not type them • •German: Tuebingen vs. Tübingen –Should be equivalent 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: 7月30日 vs. 7/30 Morgen will ich in MIT … Is this German “mit”? Dictionary entries – first cut ensemble.french 時間.japanese MIT.english mit.german guaranteed.english entries.english sometimes.english tokenization.english These may be grouped by language. More on this in ranking/query processing. Text Documents A text digital document consists of a sequence of words and other symbols, e.g., punctuation. The individual words and other symbols are known as tokens or terms. A textual document can be: • Free text, also known as unstructured text, which is a continuous sequence of tokens. • Fielded text, also known as structured text, in which the text is broken into sections that are distinguished by tags or other markup. Example? Why the focus on text? •Language is the most powerful query model •Language can be treated as text •Others? Text Based Information Retrieval Most matching methods are based on Boolean operators. Most ranking methods are based on the vector space model. Web search methods combine vector space model with ranking based on importance of documents. Many practical systems combine features of several approaches. In the basic form, all approaches treat words as separate tokens with minimal attempt to interpret them linguistically. Interface Query Engine Indexer Index Crawler Users Web A Typical Web Search Engine Content Analysis of Text •Automated Transformation of raw text into a form that represent some aspect(s) of its meaning •Including, but not limited to: –Token creation –Matrices and Vectorization –Phrase Detection –Categorization –Clustering –Summarization Techniques for Content Analysis •Statistical / vector –Single Document –Full Collection •Linguistic –Syntactic –Semantic –Pragmatic •Knowledge-Based (Artificial Intelligence) •Hybrid (Combinations) •Very common words, such as of, and, the, are rarely of use in information retrieval. •A stop list is a list of such words that are removed during lexical analysis. •A long stop list saves space in indexes, speeds processing, and eliminates many false hits. •However, common words are sometimes significant in information retrieval, which is an argument for a short stop list. (Consider the query, "To be or not to be?") Stop Lists Suggestions for Including Words in a Stop List •• Include the most common words in the English language (perhaps 50 to 250 words). •• Do not include words that might be important for retrieval (Among the 200 most frequently occurring words in general literature in English are time, war, home, life, water, and world). •• In addition, include words that are very common in context (e.g., computer, information, system in a set of computing documents). Example: the WAIS stop list (first 84 of 363 multi-letter words) about above according across actually adj after afterwards again against all almost alone along already also although always among amongst an another any anyhow anyone anything anywhere are aren't around at be became because become becomes becoming been before beforehand begin beginning behind being below beside besides between beyond billion both but by can can't cannot caption co could couldn't did didn't do does doesn't don't down during each eg eight eighty either else elsewhere end ending enough etc even ever every everyone everything Stop list policies How many words should be in the stop list? • Long list lowers recall Which words should be in list? • Some common words may have retrieval importance: -- war, home, life, water, world • In certain domains, some words are very common: -- computer, program, source, machine, language There is very little systematic evidence to use in selecting a stop list. Stop Lists in Practice The modern tendency is: (a)have very short stop lists for broad-ranging or multi-lingual document collections, especially when the users are not trained. (b)have longer stop lists for document collections in well-defined fields, especially when the users are trained professional. Token generation - stemming •What are tokens for documents? –Words (things between spaces) •Some words equivalent •Stemming finds equivalences among words •Removal of grammatical suffixes – Stemming •Reduce terms to their roots before indexing –language dependent –e.g., automate(s), automatic, automation all reduced to automat. for example compressed and compression are both accepted as equivalent to compress. for exampl compres and compres are both accept as equival to compres. Selection of tokens, weights, stop lists and stemming Special purpose collections (e.g., law, medicine, monographs) Best results are obtained by tuning the search engine for the characteristics of the collections and the expected queries. It is valuable to use a training set of queries, with lists of relevant documents, to tune the system for each application. General purpose collections (e.g., web search) The modern practice is to use a basic weighting scheme (e.g., tf.idf), a simple definition of token, a short stop list and no stemming except for plurals, with minimal conflation. Web searching combine similarity ranking with ranking based on document importance. Analyser for Lucene •Tokenization: Create an Analyser –Options •WhitespaceAnalyzer –divides text at whitespace •SimpleAnalyzer –divides text at non-letters –convert to lower case •StopAnalyzer –SimpleAnalyzer –removes stop words •StandardAnalyzer –good for most European Languages –removes stop words –convert to lower case • Picture 2 Example of analyzing a document Other Analyzers •Also available –GermanAnalyzer –RussianAnalyzer –(Lucene Sandbox) •BrazilianAnaylzer •ChineseAnalyzer (UTF-8) •CzechAnalyzer •DutchAnalyzer •FrenchAnalyzer •GreekAnalyzer •KoreanAnalyzer •JapaneseAnalyzer Summary of Text •Text is reduced to tokens •Stop words can be removed •Stemmers widely used for token generation –Porter stemmer most common Indexing Subsystem Documents break into tokens stop list* stemming* term weighting* Index database text non-stoplist tokens tokens stemmed terms terms with weights *Indicates optional operation. assign document IDs documents document numbers and *field numbers Search Subsystem Index database query parse query stemming* stemmed terms stop list* non-stoplist tokens query tokens Boolean operations* ranking* relevant document set ranked document set retrieved document set *Indicates optional operation. Statistical Properties of Text •Token occurrences in text are not uniformly distributed •They are also not normally distributed •They do exhibit a Zipf distribution A More Standard Collection 8164 the 4771 of 4005 to 2834 a 2827 and 2802 in 1592 The 1370 for 1326 is 1324 s 1194 that 973 by 969 on 915 FT 883 Mr 860 was 855 be 849 Pounds 798 TEXT 798 PUB 798 PROFILE 798 PAGE 798 HEADLINE 798 DOCNO 1 ABC 1 ABFT 1 ABOUT 1 ACFT 1 ACI 1 ACQUI 1 ACQUISITIONS 1 ACSIS 1 ADFT 1 ADVISERS 1 AE Government documents, 157734 tokens, 32259 unique Plotting Word Frequency by Rank •Main idea: count –How many times tokens occur in the text •Over all texts in the collection •Now rank these according to how often they occur. This is called the rank. Rank Freq Term 1 37 system 2 32 knowledg 3 24 base 4 20 problem 5 18 abstract 6 15 model 7 15 languag 8 15 implem 9 13 reason 10 13 inform 11 11 expert 12 11 analysi 13 10 rule 14 10 program 15 10 oper 16 10 evalu 17 10 comput 18 10 case 19 9 gener 20 9 form 150 2 enhanc 151 2 energi 152 2 emphasi 153 2 detect 154 2 desir 155 2 date 156 2 critic 157 2 content 158 2 consider 159 2 concern 160 2 compon 161 2 compar 162 2 commerci 163 2 clause 164 2 aspect 165 2 area 166 2 aim 167 2 affect Most and Least Frequent Terms short-zipf Rank Freq 1 37 system 2 32 knowledg 3 24 base 4 20 problem 5 18 abstract 6 15 model 7 15 languag 8 15 implem 9 13 reason 10 13 inform 11 11 expert 12 11 analysi 13 10 rule 14 10 program 15 10 oper 16 10 evalu 17 10 comput 18 10 case 19 9 gener 20 9 form The Corresponding Zipf Curve Zoom in on the Knee of the Curve short-zipf-zoom 43 6 approach 44 5 work 45 5 variabl 46 5 theori 47 5 specif 48 5 softwar 49 5 requir 50 5 potenti 51 5 method 52 5 mean 53 5 inher 54 5 data 55 5 commit 56 5 applic 57 4 tool 58 4 technolog 59 4 techniqu Zipf Distribution – •The Important Points: –a few elements occur very frequently –a medium number of elements have medium frequency –many elements occur very infrequently –Self similarity •Same shape for large and small frequency words –Long tail –Not necessarily obeys central limit theorem Zipf Distribution •The product of the frequency of words (f) and their rank (r) is approximately constant –Rank = order of words’ frequency of occurrence – – – • • •Another way to state this is with an approximately correct rule of thumb: –Say the most common term occurs C times –The second most common occurs C/2 times –The third most common occurs C/3 times –… • Zipf Distribution (linear and log scale) zipf_linear zipf_log What Kinds of Data Exhibit a Zipf Distribution? •Words in a text collection –Virtually any language usage •Library book checkout patterns •Incoming Web Page Requests (Nielsen) •Outgoing Web Page Requests (Cunha & Crovella) •Document Size on Web (Cunha & Crovella) •Many sales with certain retailers Power Laws Power Law Statistics - problems with means Power-law distributions •The degree distributions of most real-life networks follow a power law • • •Right-skewed/Heavy-tail distribution –there is a non-negligible fraction of nodes that has very high degree (hubs) –scale-free: no characteristic scale, average is not informative • •In stark contrast with the random graph model! –Poisson degree distribution, z=np – – – –highly concentrated around the mean –the probability of very high degree nodes is exponentially small p(k) = Ck-a Power-law signature •Power-law distribution gives a line in the log-log plot • • • • • • • • • • • • • •a : power-law exponent (typically 2 ≤ a ≤ 3) degree frequency log degree log frequency α log p(k) = -a logk + logC Examples of degree distribution for power laws Taken from [Newman 2003] Power Law Statistics - long tails Power of the long tail: The phrase The Long Tail, as a proper noun, was first coined by Chris Anderson. The concept drew in part from an influential February 2003 essay by Clay Shirky, "Power Laws, Weblogs and Inequality" that noted that a relative handful of weblogs have many links going into them but "the long tail" of millions of weblogs may have only a handful of links going into them. Beginning in a series of speeches in early 2004 and culminating with the publication of a Wired magazine article in October 2004, Anderson described the effects of the long tail on current and future business models. Anderson later extended it into the book The Long Tail: Why the Future of Business is Selling Less of More (2006). Anderson argued that products that are in low demand or have low sales volume can collectively make up a market share that rivals or exceeds the relatively few current bestsellers and blockbusters, if the store or distribution channel is large enough. Examples of such mega-stores include the online retailer Amazon.com and the online video rental service Netflix. The Long Tail is a potential market and, as the examples illustrate, the distribution and sales channel opportunities created by the Internet often enable businesses to tap into that market successfully. Word Frequency vs. Resolving Power zipf-resolving The most frequent words are not the most descriptive. van Rijsbergen 79 Consequences of Zipf for IR •There are always a few very frequent tokens that are not good discriminators. –Called “stop words” in IR –Usually correspond to linguistic notion of “closed-class” words •English examples: to, from, on, and, the, ... •Grammatical classes that don’t take on new members. •There are always a large number of tokens that occur once and can mess up algorithms. •Medium frequency words most descriptive Text •Perform lexical analysis - processing text into tokens –Many issues: normalization, lemmatization •Stemming reduces the number of tokens –Porter stemmer most common •Stop words removed to improve performance •What remains are terms to be indexed •Text has power law distribution –Words with resolving power in the middle and tail of the distribution