CSE 435/535

Information Retrieval
Fall 2019

Chapter 2: Tokenization, Stemming,

Lemmatization

* Starred slides are additional content not from IR textbook

Introduction to Information Retrieval

Recall the basic indexing pipeline

Documents to Friends, Romans, countrymen.
be indexed

Tokenizer
Token stream Friends Romans Countrymen
Linguistic
modules
Modified tokens friend roman countryman

Indexer friend 2 4
roman 1 2
Inverted index
countryman 13 16
Introduction to Information Retrieval Sec. 2.1

Parsing a document
§ What format is it in?
§ pdf/word/excel/html?
§ What language is it in?
§ What character set is in use?
§ (CP1252, UTF-8, …)

Each of these is a classification problem,

which we will study later in the course.

But these tasks are often done heuristically …

Introduction to Information Retrieval Sec. 2.1

Complications: Format/language
§ Documents being indexed can include docs from
many different languages
§ A single index may contain terms from many languages.
§ Sometimes a document or its components can
contain multiple languages/formats
§ French email with a German pdf attachment.
§ French email quote clauses from an English-language
contract

§ There are commercial and open source libraries that

can handle a lot of this stuff
Introduction to Information Retrieval Sec. 2.1

Complications: What is a document?

We return from our query documents but there are
often interesting questions of grain size:

What is a unit document?

§ A file?
§ An email? (Perhaps one of many in a single mbox file)
§ What about an email with 5 attachments?
§ A group of files (e.g., PPT or LaTeX split over HTML pages)
Introduction to Information Retrieval

Introduction to
Information Retrieval
Tokens
Introduction to Information Retrieval Sec. 2.2.1

Tokenization
§ Input: Friends, Romans and Countrymen
§ Output: Tokens
§ Friends
§ Romans
§ Countrymen
§ A token is an instance of a sequence of characters
§ Each such token is now a candidate for an index
entry, after further processing
§ Described below
§ But what are valid tokens to emit?
Introduction to Information Retrieval Sec. 2.2.1

Tokenization
§ Issues in tokenization:
§ Finland s capital ®
Finland AND s? Finlands? Finland s?
§ Hewlett-Packard ® Hewlett and Packard as two
tokens?
§ state-of-the-art: break up hyphenated sequence.
§ co-education
§ lowercase, lower-case, lower case ?
§ It can be effective to get the user to put in possible hyphens

§ San Francisco: one token or two?

§ How do you decide it is one token?
Introduction to Information Retrieval Sec. 2.2.1

Numbers
§ 3/20/91 Mar. 12, 1991 20/3/91
§ 55 B.C.
§ B-52
§ My PGP key is 324a3df234cb23e
§ (800) 234-2333
§ Often have embedded spaces
§ Older IR systems may not index numbers
§ But often very useful: think about things like looking up error
codes/stacktraces on the web
§ (One answer is using n-grams: IIR ch. 3)
§ Will often index meta-data separately
§ Creation date, format, etc.
Introduction to Information Retrieval Sec. 2.2.1

Tokenization: language issues

§ French
§ L'ensemble ® one token or two?
§ L ? L’ ? Le ?
§ Want l’ensemble to match with un ensemble
§ Until at least 2003, it didn’t on Google
§ Internationalization!

§ German noun compounds are not segmented

§ Lebensversicherungsgesellschaftsangestellter
§ ‘life insurance company employee
§ German retrieval systems benefit greatly from a compound splitter
module
§ Can give a 15% performance boost for German
Introduction to Information Retrieval Sec. 2.2.1

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
500 $500K( 6,000 )

Katakana Hiragana Kanji Romaji

End-user can express query entirely in hiragana!

Introduction to Information Retrieval Sec. 2.2.1

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

§ ← → ←→ ← 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
Introduction to Information Retrieval

Equivalence Classes*
§ Handle synonyms and homonyms
§ Equivalence Classes: Mapping rules that remove
characters like hyphens: terms that happen to
become identical are the equivalence classes
§ Hand-constructed equivalence classes (implicit)
§ e.g., car = automobile
§ your and you re
§ Soundex: Chebyshev, Tchebycheff
§ 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
Introduction to Information Retrieval

Introduction to
Information Retrieval
Terms
The things indexed in an IR system
Introduction to Information Retrieval Sec. 2.2.2

Stop words
§ With a stop list, you exclude from the dictionary
entirely the commonest words. Intuition:
§ They have little semantic content: the, a, and, to, be
§ There are a lot of them: ~30% of postings for top 30 words
§ But the trend is away from doing this:
§ Good compression techniques (IIR 5) means the space for including
stop words in a system is very small
§ Good query optimization techniques (IIR 7) 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
OAC Stopword list (~ 275 words)*
The Complete OAC-Search Stopword List
a did has nobody somewhere usually

about do have noone soon very

all does having nor still was

almost doing here not such way

along done how nothing than ways

already during however now that we

also each i of the well

although either if once their went

always enough ii one theirs were

am etc in one's then what

among even including only there when

an ever indeed or therefore whenever

and every instead other these where

any everyone into others they wherever

anyone everything is our thing whether

anything for it ours things which

Stopword list contd.*
anywhere from its out this while
are further itself over those who

as gave just own though whoever

at get like perhaps through why
away getting likely rather thus will
because give may really to with

between given maybe same together within

beyond giving me shall too without
both go might should toward would

but going mine simply until yes

by gone much since upon yet
can good must so us you
cannot got neither some use your

come great never someone used yours

could had no something uses yourself
sometimes using
Introduction to Information Retrieval Sec. 2.2.3

Normalization to terms
§ We may need to normalize words in indexed text
as well as query words into the same form
§ We want to match U.S.A. and USA
§ Result is terms: a term is a (normalized) word type,
which is an entry in our IR system dictionary
§ We most commonly implicitly define equivalence
classes of terms by, e.g.,
§ deleting periods to form a term
§ U.S.A., USA è USA
§ deleting hyphens to form a term
§ anti-discriminatory, antidiscriminatory è antidiscriminatory
Introduction to Information Retrieval Sec. 2.2.3

Normalization: other languages

§ Accents: e.g., French résumé vs. resume.
§ Umlauts: e.g., German: Tuebingen vs. Tübingen
§ Should be equivalent
§ 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
§ Often best to normalize to a de-accented term
§ Tuebingen, Tübingen, Tubingen è Tubingen
Introduction to Information Retrieval Sec. 2.2.3

Normalization: other languages

§ Normalization of things like date forms
§ 7 30 vs. 7/30
§ Japanese use of kana vs. Chinese characters

§ Tokenization and normalization may depend on the

language and so is intertwined with language
detection Is this
Morgen will ich in MIT … German mit ?

§ Crucial: Need to normalize indexed text as well as

query terms identically
Introduction to Information Retrieval Sec. 2.2.3

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…

§ Longstanding Google example: [fixed in 2011…]

§ Query C.A.T.
§ #1 result is for cats (well, Lolcats) not Caterpillar Inc.
Introduction to Information Retrieval Sec. 2.2.3

Normalization to terms

§ An alternative to equivalence classing is to do

asymmetric expansion
§ An example of where this may be useful
§ Enter: window Search: window, windows
§ Enter: windows Search: Windows, windows, window
§ Enter: Windows Search: Windows
§ Potentially more powerful, but less efficient
Introduction to Information Retrieval

Thesauri and soundex

§ Do we handle synonyms and homonyms?
§ E.g., by hand-constructed equivalence classes
§ car = automobile color = colour
§ We can rewrite to form equivalence-class terms
§ When the document contains automobile, index it under car-
automobile (and vice-versa)
§ Or we can expand a query
§ When the query contains automobile, look under car as well
§ What about spelling mistakes?
§ One approach is Soundex, which forms equivalence classes
of words based on phonetic heuristics
§ More in IIR 3 and IIR 9
Introduction to Information Retrieval

Introduction to
Information Retrieval
Stemming and Lemmatization
Introduction to Information Retrieval Sec. 2.2.4

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
Introduction to Information Retrieval Sec. 2.2.4

Stemming
§ Reduce terms to their roots before indexing
§ Stemming suggests 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.
Introduction to Information Retrieval Sec. 2.2.4

Porter s algorithm
§ Commonest algorithm for stemming English
§ Results suggest it s 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.
Introduction to Information Retrieval Sec. 2.2.4

Typical rules in Porter

§ sses ® ss
§ ies ® i
§ ational ® ate
§ tional ® tion

§ Weight of word sensitive rules

§ (m>1) EMENT →
§ replacement → replac
§ cement → cement
Introduction to Information Retrieval Sec. 2.2.4

Other stemmers
§ Other stemmers exist:
§ Lovins stemmer
§ http://www.comp.lancs.ac.uk/computing/research/stemming/general/lovins.htm
§ Single-pass, longest suffix removal (about 250 rules)
§ Paice/Husk stemmer
§ Snowball

§ Full morphological analysis (lemmatization)

§ At most modest benefits for retrieval
Stemming variants*
Porter Stemmer : Representing Strings *
A consonant in a word is a letter other than A, E, I, O or U, and other than Y preceded by a
consonant.
So in TOY the consonants are T and Y, and in SYZYGY they are S, Z and G.

A consonant will be denoted by c, a vowel by v. A list ccc... of length greater than 0 will be
denoted by C,
and a list vvv... of length greater than 0 will be denoted by V. Any word, or part of a word,
therefore has one of the four forms:
CVCV ... C
CVCV ... V
VCVC ... C
VCVC ... V
These may all be represented by the single form
[C]VCVC ... [V]
Using (VC)m to denote VC repeated m times, this may again be written as
[C](VC)m[V]

m will be called the measure of any word or word part. m = 0 covers the null word.
m=0 TR, EE, TREE, Y, BY.
m=1 TROUBLE, OATS, TREES, IVY.

m=2 TROUBLES, PRIVATE, OATEN, ORRERY.

 Porter Stemmer - Types of Rules. *
The rules for removing a suffix will be given in the form
(condition) S1 -> S2
This means that if a word ends with the suffix S1, and the stem before S1 satisfies the given condition, S1 is replaced by S2..
(m > 1) EMENT ->
Here S1 is EMENT and S2 is null. This would map REPLACEMENT to REPLAC, since REPLAC is a word part for which m = 2.

The condition part may also contain the following:

*S - the stem ends with S (and similarly for the other letters).

*v* - the stem contains a vowel.

*d - the stem ends with a double consonant (e.g. -TT, -SS).

*o - the stem ends cvc, where the second c is not W, X or Y (e.g. -WIL, -HOP).

And the condition part may also contain expressions with and, or and not, so that
(m>1 and (*S or *T))
tests for a stem with m>1 ending in S or T, while
(*d and not (*L or *S or *Z))
tests for a stem ending with a double consonant other than L, S or Z. Elaborate conditions like this are required only rarely.

In a set of rules written beneath each other, only one is obeyed, and this will be the one with the longest matching S1 for the given word. For ex
SSES -> SS

IES -> I

SS -> SS

S ->

(here the conditions are all null) CARESSES maps to CARESS since SSES is the longest match for S1.

Equally CARESS maps to CARESS (S1= SS ) and CARES to CARE (S1= S ).

 Porter Stemmer - Step 1a. *
In the rules below, examples of their application, successful or otherwise, are given
on the right in lower case. Only one rule from each step can fire. The algorithm now follow

Step 1a
SSES -> SS caresses -> caress
IES -> I ponies -> poni

ties -> ti
SS -> SS caress -> caress
S -> cats -> cat
Porter Stemmer- Step 1b
Step 1b

(m>0) EED -> EE feed -> feed

agreed -> agree

(*v*) ED -> plastered -> plaster

bled -> bled

(*v*) ING -> motoring -> motor

sing -> sing

If the second or third of the rules in Step 1b is successful, the following is done:

AT -> ATE conflat(ed) -> conflate

BL -> BLE troubl(ed) -> trouble

IZ -> IZE siz(ed) -> size

(*d and not (*L or *S or *Z)) -> single letter hopp(ing) -> hop

tann(ed) -> tan

fall(ing) -> fall

hiss(ing) -> hiss

fizz(ed) -> fizz

(m=1 and *o) -> E fail(ing) -> fail

fil(ing) -> file

The rule to map to a single letter causes the removal of one of the double letter pair. The -E is put back on -AT, -BL and -IZ, so that the suffixe

-ATE, -BLE and -IZE can be recognised later. This E may be removed in step 4.
Porter Stemmer- Step 1c. *
Step 1c
(*v*) Y -> I happy -> happ
sky -> sky
Step 1 deals with plurals and past participles. The subseque
much more straightforward.
 Porter Stemmer - Step 2. *
(m>0) ATIONAL -> ATE relational -> relate

(m>0) TIONAL -> TION conditional -> condition

rational -> rational

(m>0) ENCI -> ENCE valenci -> valence

(m>0) ANCI -> ANCE hesitanci -> hesitance

(m>0) IZER -> IZE digitizer -> digitize

(m>0) ABLI -> ABLE conformabli -> conformable

(m>0) ALLI -> AL radicalli -> radical

(m>0) ENTLI -> ENT differentli -> different

(m>0) ELI -> E vileli -> vile

(m>0) OUSLI -> OUS analogousli -> analogous

(m>0) IZATION -> IZE vietnamization -> vietnamize

(m>0) ATION -> ATE predication -> predicate

(m>0) ATOR -> ATE operator -> operate

(m>0) ALISM -> AL feudalism -> feudal

(m>0) IVENESS -> IVE decisiveness -> decisive

(m>0) FULNESS -> FUL hopefulness -> hopeful

(m>0) OUSNESS -> OUS callousness -> callous

(m>0) ALITI -> AL formaliti -> formal

(m>0) IVITI -> IVE sensitiviti -> sensitive

(m>0) BILITI -> BLE sensibiliti -> sensible

The test for the string S1 can be made fast by doing a program switch on the penultimate letter of the word being tested. This gives a fairly even breakdown

of the possible values of the string S1. It will be seen in fact that the S1-strings in step 2 are presented here in the alphabetical order of their penultimate let

Similar techniques may be applied in the other steps.

Porter Stemmer - Step 3. *
(m>0) ICATE -> IC triplicate -> triplic
(m>0) ATIVE -> formative -> form
(m>0) ALIZE -> AL formalize -> formal
(m>0) ICITI -> IC electriciti -> electric
(m>0) ICAL -> IC electrical -> electric
(m>0) FUL -> hopeful -> hope

(m>0) NESS -> goodness -> good

Porter Stemmer- Step 4
(m>1) AL

(m>1) ANCE
->

->
revival

allowance
*->

->
reviv

allow

(m>1) ENCE -> inference -> infer

(m>1) ER -> airliner -> airlin

(m>1) IC -> gyroscopic -> gyroscop

(m>1) ABLE -> adjustable -> adjust

(m>1) IBLE -> defensible -> defens

(m>1) ANT -> irritant -> irrit

(m>1) EMENT -> replacement -> replac

(m>1) MENT -> adjustment -> adjust

(m>1) ENT -> dependent -> depend

(m>1 and (*S or *T)) ION -> adoption -> adopt

(m>1) OU -> homologou -> homolog

(m>1) ISM -> communism -> commun

(m>1) ATE -> activate -> activ

(m>1) ITI -> angulariti -> angular

(m>1) OUS -> homologous -> homolog

(m>1) IVE -> effective -> effect

(m>1) IZE -> bowdlerize -> bowdler

The suffixes are now removed. All that remains is a little tidying up.
Porter Stemmer - Step 5 *
Step 5a
(m>1) E -> probate -> probat
rate -> rate
(m=1 and not *o) E -> cease -> ceas

Step 5b
(m > 1 and *d and *L) -> single letter controll -> control
roll -> roll
Introduction to Information Retrieval Sec. 2.2.4

Language-specificity
§ The above methods 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 are
available for handling these
Introduction to Information Retrieval Sec. 2.2.4

Does stemming help?

§ English: very mixed results. Helps recall for some
queries but harms precision on others
§ E.g., operative (dentistry) ⇒ oper
§ Definitely useful for Spanish, German, Finnish, …
§ 30% performance gains for Finnish!
 Character N-Grams*
• alternative to complex language-specific tokenization
• treat overlapping sequences of n characters as tokens
• e.g. If n=5, orienteering is split into
_orie, orien, rient, iente, entee, nteer, teeri, eerin, ering, ring_
• at query time, split query into n-grams for retrieval
• in principle, need not account for punctuation, capitalization, white space or other
language characteristics
• optimal n varies from language to language
• n-grams can substitute for a stemmer
• e.g. orienteers matches more n-grams from orienteering than oriental
Introduction to Information Retrieval

Introduction to
Information Retrieval
Faster postings merges:
Skip pointers/Skip lists
Introduction to Information Retrieval Sec. 2.3

Recall basic merge

§ Walk through the two postings simultaneously, in
time linear in the total number of postings entries

2 4 8 41 48 64 128 Brutus
2 8
1 2 3 8 11 17 21 31 Caesar

If the list lengths are m and n, the merge takes O(m+n)

operations.

Can we do better?
Yes (if the index isn t changing too fast).
Introduction to Information Retrieval Sec. 2.3

Augment postings with skip pointers

(at indexing time)
41 128
2 4 8 41 48 64 128

11 31
1 2 3 8 11 17 21 31

§ Why?
§ To skip postings that will not figure in the search
results.
§ How?
§ Where do we place skip pointers?
Introduction to Information Retrieval Sec. 2.3

Query processing with skip pointers

41 128
2 4 8 41 48 64 128

11 31
1 2 3 8 11 17 21 31

Suppose we ve stepped through the lists until we

process 8 on each list. We match it and advance.

We then have 41 and 11 on the lower. 11 is smaller.

But the skip successor of 11 on the lower list is 31, so

we can skip ahead past the intervening postings.
Introduction to Information Retrieval Sec. 2.3

Where do we place skips?

§ Tradeoff:
§ More skips ® shorter skip spans Þ more likely to skip.
But lots of comparisons to skip pointers.
§ Fewer skips ® few pointer comparison, but then long skip
spans Þ few successful skips.
Introduction to Information Retrieval Sec. 2.3

Placing skips
§ Simple heuristic: for postings of length L, use ÖL
evenly-spaced skip pointers [Moffat and Zobel 1996]
§ This ignores the distribution of query terms.
§ Easy if the index is relatively static; harder if L keeps
changing because of updates.

§ This definitely used to help; with modern hardware it

may not unless you re memory-based [Bahle et al. 2002]
§ The I/O cost of loading a bigger postings list can outweigh
the gains from quicker in memory merging!