Text Processing &

Characteristics
Kron
 Text
• Text parsing
– Tokenization, terms
– A bit of linguistics
• Text characteristics
– Zipfs law
 Query Engine Index

Interface

Text processing
Indexer
Users

Crawler

Web
A Typical Web Search Engine
 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
• .,:;-()_
 
• 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
 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
Documents to Friends, Romans, countrymen.
be indexed.

Tokenizer
Token stream. Friends Romans Countrymen
Linguistic
modules
Modified tokens (terms). friend roman countryman

Indexer friend 2 4
roman 1 2
Inverted index.
countryman 13 16
 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
• ‫ عاما من‬132 ‫ بعد‬1962 ‫استقلت الجزائر في سنة‬
‫االحتالل 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
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 for exampl compress and
and compression are both compress ar both accept
accepted as equivalent to as equival to compress
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
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

7 月 30 日 vs. 7/30
• Character-level alphabet detection and
conversion
– Tokenization not separable from this.
– Sometimes ambiguous: Is this
Morgen will ich in MIT … German “mit”?
 Stop Lists
•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?")
 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
 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.
 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
Example of analyzing a document
 Other Analyzers
• Also available
– GermanAnalyzer
– RussianAnalyzer
– (Lucene Sandbox)
• BrazilianAnaylzer
• ChineseAnalyzer (UTF-8)
• CzechAnalyzer
• DutchAnalyzer
• FrenchAnalyzer
• GreekAnalyzer
• KoreanAnalyzer
• JapaneseAnalyzer
 Indexing Subsystem
documents
Documents assign document IDs

text document
break into tokens
numbers
tokens stop list* and *field
numbers
non-stoplist stemming*
tokens
*Indicates
optional stemmed term weighting*
operation. terms

terms with Index

weights database
 Search Subsystem
query parse query
query tokens
ranked
document set stop list* non-stoplist
tokens
ranking*
stemming*
stemmed
terms
Boolean
*Indicates retrieved operations*
optional document set
Index
operation.
relevant database
document set
 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
Government documents, 157734 tokens, 32259 unique
8164 the 969 on 1 ABC
4771 of 915 FT 1 ABFT
4005 to 883 Mr 1 ABOUT
2834 a 860 was 1 ACFT
2827 and 855 be 1 ACI
2802 in 849 Pounds 1 ACQUI
1592 The 798 TEXT 1 ACQUISITIONS
1370 for 798 PUB 1 ACSIS
1326 is 798 PROFILE 1 ADFT
1324 s 798 PAGE 1 ADVISERS
1194 that 798 HEADLINE 1 AE
973 by 798 DOCNO
 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.
 Most and Least Frequent Terms
Rank Freq Term
1 37 system 150 2 enhanc
2 32 knowledg 151 2 energi
3 24 base 152 2 emphasi
4 20 problem 153 2 detect
5 18 abstract 154 2 desir
6 15 model
7 15 languag 155 2 date
8 15 implem 156 2 critic
9 13 reason 157 2 content
10 13 inform 158 2 consider
11 11 expert 159 2 concern
12 11 analysi 160 2 compon
13 10 rule 161 2 compar
14 10 program 162 2 commerci
15 10 oper
16 10 evalu 163 2 clause
17 10 comput 164 2 aspect
18 10 case 165 2 area
19 9 gener 166 2 aim
20 9 form 167 2 affect
 The Corresponding Zipf Curve
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
Zoom in on the Knee of the Curve
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

f C 1 / r
C  N / 10
• 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)
 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
p(k) = Ck-

• 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
zk  z
p(k) P(k;z)  e
k!
– highly concentrated around the mean
– the probability of very high degree nodes is exponentially small
 Power-law signature
• Power-law distribution gives a line in the log-log plot

log p(k) = - logk + logC

frequency log frequency α

degree log degree

  : power-law exponent (typically 2 ≤  ≤ 3)

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
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