Morphology from a computational point of view

March 2001

Today

Minimal Edit Distance, and Viterbi more generally;

Letter to Sound What is morphology? Finite-state automata Finite-state phonological rules

1. What is morphology?

Study of the internal structure of words: morph-ology word-s jump-ingWhy?1. For some purposes, we need to know

what the internal pieces are.2. Knowledge of the words of a language

can’t be summarized in a finite list: we need to know the principles of word-formation

Some resources

Richard Sproat: Morphology and Computation (MIT Press, 1992)

Excellent overview of computational morphology and phonoloy by Harald Trost at

http://www.ai.univie.ac.at/~harald/handbook.html

2. What applications need knowledge of words?

Any high-level linguistic analysis: syntactic parser

machine translation

speech recognition, text-to-speech (TTS)

information retrieval (IR)

dictionary, spell-checker

3. A list is not enoughAn empirical fact:

AP newswire: mid-Feb – Dec 30 1988

Nearly 300,000 words.

“New” words that appeared on Dec 31 1988:

compounds: prenatal-care, publicly-funded, channel-switching, owner-president, logic-loving, part-Vulcan, signal-emitting, landsite, government-aligned, armhole, signal-emitting...

...new words...

dumbbells, groveled, fuzzier, oxidized

ex-presidency, puppetry, boulderlike, over-emphasized, hydrosulfite, outclassing, non-passengers, racialist, counterprograms, antiprejudice, re-unification, traumatological, refinancings, instrumenting, ex-critters, mega-lizard

ex-presidency: prefix ex- boulder-like: suffix –like over-emphasized: prefix over- antiprejudice: prefix anti

This is often called the OOV problem (“out of vocabulary”).

If we work out the principles of word-formation, we will simultaneously:

1. compress the size of our internalized list of words;

2. become able to deal with new words on the fly.

4. Overview of applications that need knowledge of words Speech generation: text to speech

(TTS) Speech recognition

Text to speech

Problem: take text, in standard spelling, and produce a sequence of phonemes which can be synthesized by the “backend”.

Severe problems: Proper names (persons, places), OOV words

boathouse B OW1 T H AU2 S

Speech recognition

Take a sound file (e.g., *.wav) and produce a list of words in standard orthography.

Bill Clinton is a recent ex-president.

If someone says it, we need to figure out what the word was.

Do we know what a word is?

This is actually not an easy question! – especially if we turn to Asian languages, without a tradition of putting in “white space” between “words”, as we do in the West.

German writes more compounds without white space than English does.

Basic principles of morphology

For some purposes, we need to think about phonemes, while for others it’s more convenient to talk about letters.

For our purposes, I’ll talk about letters whenever we don’t need to specifically focus on phonemes.

Morpheme

It is convenient to be able to talk about the pieces into which words may be broken, and linguists call these pieces morphemes: the smallest parts of a language that can be regularly assigned a meaning.

Morphemes

Uncontroversial morphemes:

door, dog, jump, -ing, -s, to

More controversial morphemes

sing/sang: s-ng + i/a

cut/cut: cut + PAST

Classic distinctions in morphology:

Analytic (isolating) languages:– no morphology of derivational or

inflectional sort.

Synthetic (inflecting) languages:– Agglutinative: 1 function per morpheme– Fusional: > 1 function per morpheme

Agglutinative:Finnish Nominal Declension

talo 'the-house' kaup-pa 'the-shop'

talo-ni 'my house' kaup-pa-ni 'my shop'

talo-ssa 'in the-house' kaup-a-ssa 'in the-shop'

talo-ssa-ni 'in my house’ kaup-a-ssa-ni 'in my shop'

talo-i-ssa 'in the-houses’ kaup-o-i-ssa 'in the-shops'

talo-i-ssa-ni 'in my houses’ kaup-o-i-ssa-ni 'in my shops'

Courtesy of Bucknell Univ. web page

Fusional: LatinLatin Declension of hortus 'garden'

Singular Plural

Nominative (Subject) hort-us hort-i

Genitive (of) hort-i hort-rum

Dative (for/to) hort-o hort-is

Accusative (Direct Obj) hort-um hort-us

Vocative (Call) hort-e hort-i

Ablative (from/with) hort-o hort-is

Morphemes vs. morphs

Some analysts distinguish between “morphemes” and “morphs”.

Morphemes are motivated by an analysis, and include “plural” and “past”

Morphs are strings of letters or phones that “realize” or “manifest” a morpheme.

Free and bound morphemes

Free morphemes can form (free-standing) words; bound morphemes are only found in combination with other morphemes.

Examples?

Functions of morphology

Derivational morphology: creates one lexeme from another

compute > computer > computerize > computerization

Inflectional morphology: creates the form of a lexeme that’s right for a sentence:

the nominative singular form of a noun; or the past 3rd person singular form of a verb.

Word: an identifiable string of letters (or phonemes) sing

Word-form: a word with a specific set of syntactic and morphological features. The sing in I sing is 1st person sg, and is a distinct word-from from the sing in you sing.

Lexeme: a complete set of inflectionally related word-forms, including sing, sings, and sang

Lemma: a complete set of morphologically related lexemes: sing, sings, song, sang.

A lexeme’s stem

In many languages (unlike English), constellations of word-forms forming a lexeme demand the recognition of a basic stem which does not stand freely as a word:

Italian ragazzo, ragazzi (boy, girl)ragazzi, ragazze (boys, girls)

ragazz-

Compounds

Compounds are composed of 2 (or more) words or stems

Compounds: hot dog, White House, bookstore, cherry-covered

Languages vary in the amount of morphology they have and use

English has a lot of derivational morphology and relatively little inflectional morphology

English verb’s inflectional forms:

bare stem, -s, -ed, -ing

European languages

Not uncommon for a verb to have 30 to 50+ forms:

marking tense, person and number of the subject

Derivation Derivational morphology usually consists of

adding a prefix or suffix to a base (= stem).

The base has a lexical category (it is a noun, verb, adjective), and the suffix typically assigns a different category to the whole word.

sad ness

Adj

Noun -ness: suffix that takes an adjective, & makes a noun.

un interest ing

Adj

Adj

VerbVb

Adj

Adj

Distinct from contractions…

English (and some other languages) permit the collapsing together of common words. In some extremely rare cases, only the collapsed form exists (English possessive ’s).

He will arrive tonight > he’ll arrive…

The [King of England]’s children

Some basics of English morphologyInflectional morphologyNouns: -NULL, -s, -’sVerbs: -NULL, s, -ed, -ing (so-called weak verbs)Strong verbs: 3 major groupsa. Internal verb change (sing/sang,

drive/drove/driven, dive/dove)b. –t suffix, typically with vowel-shortening

dream/dreamt, sleep/sleptc. –aught replacement: catch, teach,

seek,

Derivational morphology in complex

This morphology creates new words, by adding prefixes or suffixes.

It is helpful to divide them into two groups, depending on whether they leave the pronunciation of the base unchanged or not.

There are, as always, some fuzzy cases.

Level 1

ize, ization, al, ity, al, ic, al, ity, ion, y (nominaliz-ing), al, ate, ous, ive, ation

Can attach to non-word stems (fratern-al, paternal; parent-al)

Typically change stress and vowel quality of stem

Level 2

Never precede Level 1 suffixes

Never change stress pattern or vowel quality

Almost always attach to words that already exist

hood, ness, ly, s, ing, ish, ful, ly, ize, less, y (adj.)

Combinations of Class 1,2

Class 1 + Class 1: histor-ic-al, illumina-at-tion, indetermin-at-y;

Class 1 + Class 2: frantern-al-ly, transform-ate-ion-less;

Class 2 + Class 2: weight-less-ness ?? Class 2 + Class 1: *weight-less-ity,

fatal-ism-al

Signature

Set of suffixes (or prefixes) that occurs in a corpus with a set of stems.

NULL.ed.ing.s

look interest add claim mark extend demand remain want succeed record offer represent cover return end explain follow help belong attempt talk fear happen assault account point award appeal train contract result request staff view fail kick visit confront attack comment sponsor

NULL.s paper retain improvement missile song truth doctor

indictment window conductor dick misunderstanding struggle stake tank belief cafeteria material mind operator bassi lot movement chain notion marriage dancer scholarship reservoir sweet right battalion hold mr shot cardinal athletic revenue duel confrontation solo talent guest shoe russian commitment average monk election street roger rifle worker area plane pinch-hitter dozen browning conclusion teacher narcotic appearance alternative dealer producer mile stock shrine sometime bag successor career mistake ankle weapon model front spotlight rhode pace debate payment requirement fairway consultation chip dollar employer thank mustang rocket-bomb hat string precinct robert employee action detective pressure measure spirit forbid hitter breast yankee partner floor member

NULL.d.s increase tie hole associate reserve price fire receive

challenge rate purchase propose feature celebrate decide suite single change sculpture combine privilege pledge issue frame indicate believe damage include use aide graduate surprise intervene practice trouble serve oppose promise charge note schedule continue raise decline cause operate emphasize relieve hope share judge birdie produce exchange

NULL.ed.er.ing.s report turn walk park pick flowNULL.d.ment enforce announce engage arrange replace

improve encourageNULL.n.s rose low take law drive rise undertakeNULL.al intern profession logic fat tradition extern margin

jurisdiction historic education promotion constitution addition sensation roy ration origin classic convention

NULL.man sand news police states gross sun fresh sports boss sales 3- patrol bonds

ed.er.ing slugg manag crush publish robbNULL.ity.s major senior moral hospitalNULL.ry hung mason ave summit scene surge rival

forestNULL.a.s indian kind american

Finite state morphology

FSA: finite-state automataConsists of

1. a set of states

2. a starting state

3. a set of final or accepting states

4. a finite set of symbols

5. a set of transitions: each is defined by a from-state, a to-state, and a symbol

It’s natural to think of this as describing an annotated directed graph.

q0 q1 q2 q3

b a a

q3

!

aa a !

An FSA can be thought of as judging (accepting) a string, or as generating one.

How does it judge? Find a start to finish path that matches the string.

How does it generate? Walk through any start-to-finish path.

Deterministic and Non-deterministic FSAsJust a little difference: Deterministic case: For every state,

there is a maximum of one transition associated with any given symbol. You can say that there’s a function from {states}X{symbols} {states}

Nondeterministic case: There is no such restriction; hence, given a state and a symbol, it is not necessarily certain which transition is to be taken.

q0 q1 q2 q3

b a a

q3

!

a

deterministic…

q0 q1 q2 q3

b a a

q3

!

a

non-deterministicThe best things in lifeare non-deterministic.

Figure 3.4 p. 68

q0 q1 q2 q3

un- adj-root -er –est -ly

)()(

ly

est

er

rootsadjun

Alternate notation

Yet a third way: rows in an array(to-column can consist of pointers)

From To Output

0 1 un

0 1 NULL

1 2 adj-root-list

2 3 er;est;ly

Stop states: 2,3

Figure 3.5 p. 69

q0q1 q2

q5

un-adj-root-1

-er –est -ly

q4 -er –est

q3adj-root-2

)()( 1

ly

est

er

rootsadjun )(2

est

errootsadj

adj-root-1

Yet a third way: rows in an array(to-column can consist of pointers)

From To Output

0 1 un

0 3 NULL

1 2 adj-root-list-1

2 5 er;est;ly

3 2 adj-root-list-1

3 4 adj-root-list-2

4 5 er;est

Stop states: 2,4,5

0 1 noun 11 2 ize2 3 ation2 4 er2 5 able0 5 -al “adj”0 1 adj-al5 6 ity;ness0 7 verbs17 8 ive “adj”8 6 ness8 9 ly0 10 verb210 8 ative0 11 noun211 8 ful

Figure 3.6, p. 70

ation ize

noun

adj

verb

adverb

er

nouns

ativeive

able

ly

ly

ity, ness

ful

verbs

adjectives

adverbs

We can simplify greatly (generating a bit more….)

Finite-State Transducers (FST)The symbols of the FST are complex:

they’re really pairs of symbols, one for each of two “tapes” or levels.

Recognizer: decides if a given pair of representations fits together “OK”

Generator: generates pairs of representations that fit together

Translator: takes a representation on one level and produces the appropriate representation on the other level

Finite state transducers

can be inverted, or composed and you get another FST.

Complex symbols Usually of the form a:b, which means a

can appear on the upper tape when b appears on the lower tape.

So a:b means that’s a permissible pairing up of symbols.

“a” along means a:a, etc. epsilon means null character. Remember, “other” means “any feasible

pair that is not in this transducer” (p. 78)

Using FSTs for orthographic rules

#__/ s

z

s

x

e

#

q0 q1 q2 q3 q4

q5:̂#

other

otherZ! = Z, s, x

Z! Z!

Z!

S

#, other

:e

#, other z,x

^:

^:

s

Using FSTs for orthographic rules

fox^s#…we get to q1 with ‘x’

#

q0 q1 q2 q3 q4

q5:̂#

other

otherZ! = Z, s, x

Z! Z!

Z!

S

#, other

:e

#, other z,x

^:

^:

s

Using FSTs for orthographic rules

#

q0 q1 q2 q3 q4

q5:̂#

other

otherZ! = Z, s, x

Z! Z!

Z!

S

#, other

:e

#, other z,x

^:

^:

s

fox^s#…we get to q2 with ‘^’

Using FSTs for orthographic rules

#

q0 q1 q2 q3 q4

q5:̂#

other

otherZ! = Z, s, x

Z! Z!

Z!

S

#, other

:e

#, other z,x

^:

^:

s

fox^s#…we can get to q3 with ‘NULL’

Using FSTs for orthographic rules

#

q0 q1 q2 q3 q4

q5:̂#

other

otherZ! = Z, s, x

Z! Z!

Z!

S

#, other

:e

#, other z,x

^:

^:

s

fox^s#…we also get to q5 with ‘s’but we don’t want to!

#

q0 q1 q2 q3 q4

q5:̂#

other

otherZ! = Z, s, x

Z! Z!

Z!

S

#, other

:e

#, other z,x

^:

^:

s

fox^s#…we also get to q5 with ‘s’but we don’t want to!

So why is this transition there??friend^ship, ?fox^s^s (= foxes’s)

#

q0 q1 q2 q3 q4

q5:̂#

other

otherZ! = Z, s, x

Z! Z!

Z!

S

#, other

:e

#, other z,x

^:

^:

s

fox^s#…q4 with s

#

q0 q1 q2 q3 q4

q5:̂#

other

otherZ! = Z, s, x

Z! Z!

Z!

S

#, other

:e

#, other z,x

^:

^:

s

fox^s#…q0 with # (accepting state)

#

q0 q1 q2 q3 q4

q5:̂#

other

otherZ! = Z, s, x

Z! Z!

Z!

S

#, other

:e

#, other z,x

^:

^:

s

arizona: we leave q0 but return

Other transitions…

#

q0 q1 q2 q3 q4

q5:̂#

other

otherZ! = Z, s, x

Z! Z!

Z!

S

#, other

:e

#, other z,x

^:

^:

s

m i s s ^ s

Other transitions…