Transcript Finite-state transducers, Morphology and Phonology

Finite-state automata and Morphology
Outline
Morphology
•
What is it?
•
Why do we need it?
•
How do we model it?
Computational Model: Finite-state transducer
Structure of Words
What are words?
•
Orthographic tokens separated by white space.
In some languages the distinction between words and sentences is
less clear.
•
Chinese, Japanese: no white space between words
–
•
nowhitespace  no white space/no whites pace/now hit esp ace
Turkish: words could represent a complete “sentence”
–
Eg: uygarlastiramadiklarimizdanmissinizcasina
•
“(behaving) as if you are among those whom we could not civilize”
Morphology: the structure of words
•
Basic elements: morphemes
•
Morphological Rules: how to combine morphemes.
Syntax: the structure of sentences
•
Rules for ordering words in a sentence
Morphology and Syntax
Interplay between syntax and morphology
•
How much information does a language allow to be packed
in a word, and how easy is it to unpack.
•
More information  less rigid syntax  more free word
order
•
Eg: Hindi: John likes Mary – all six orders are possible, due
to rich morphological information.
Why Study Morphology?
Morphology provides
•
systematic rules for forming new words in a language.
•
can be used to verify if a word is legitimate in a language.
•
efficient storage methods.
•
improving lexical coverage of a system.
•
group words into classes.
Applications
•
Improving recall in search applications
–
•
Text-to-speech synthesis
–
•
Try “fish” as a query in a search engine
category of a word determines its pronunciation
Parsing
–
Morphological information eliminates spurious parses for a sentence
Structure of Words
Morphology: The study of how words are composed from smaller, meaningbearing units (morphemes)
• Stems: children, undoubtedly,
•
Affixes (prefixes, suffixes, circumfixes, infixes)
– Immaterial, Trying, Gesagt, Absobl**dylutely
Different ways of creating words
• Concatenative
– Adding strings to the stems (examples above)
• Non-concatenative
– Infixation (e.g. Tagalog: hingi (order)  humingi)
– Templatic (e.g. Arabic root-and-pattern) morphological systems
•
•
Triconsonant CCC expands into words using templates that represent more semantic
information.
Hebrew: lmd (study)  lamad (he studied), limed (he taught),
lumad (he was taught)
Stacking of affixes in English is limited (three or four); but in Turkish nine or
ten affixes can be stacked.
• Agglutinative languages
Different Classifications
Sets of words display (almost) regular patterns.
Open class words: Noun, Verb, Adverb, Adjectives
Closed class words: Pronoun, Preposition, Conjunction, Determiners.
Kinds of morphology
Inflectional
Word stem + grammatical morpheme results in a word of the same class.
Morpheme serves a syntactic function: grammatical agreement, case marker
-s for plural on nouns, -ed for past tense on verbs
Only Nouns and Verbs inflect in English for grammatical reasons
In French, adjectives agree (match on gender and number) with nouns they modify.
Derivational
Word stem + grammatical morpheme results in a different class of word.
Nominalization: -ation suffix makes verbs into nouns (e.g. Computerize  computerization
-ly suffix makes adjectives into adverbs (e.g. beautiful  beautifully)
Inflectional Morphology
Word stem + grammatical morpheme
Morpheme serves a syntactic function: grammatical agreement, case marker
• -s for plural on nouns, -ed for past tense on verbs
Results in word of the same class as the stem
• (bat  bats; man  man’s; jump  jumps, jumped, jumping)
Noun morphology:
• Regular nouns: bat  bats, fish  fishes
•
Irregular nouns: spelling changes (mouse  mice; ox  oxen)
Verb morphology:
• Regular verbs: jump  jumps (-s form), jumped (past/ -ed participle), jumping (ing participle)
•
Irregular verbs: eat  eats (-s form), ate (past), eaten (-ed participle), eating (ing participle)
Irregular forms have to be stored.
• Language learners typically make errors on irregular forms.
•
hit  hitted
Inflectional Morphology
Nominal morphology
• Plural forms
–
–
–
•
s or es
Irregular forms (goose/geese)
Mass vs. count nouns (fish/fish,email or emails?)
Possessives (cat’s, cats’)
Verbal inflection
• Main verbs (sleep, like, fear) verbs relatively regular
–
–
–
•
Primary (be, have, do) and modal verbs (can, will, must) often irregular and not
productive
–
•
Be: am/is/are/were/was/been/being
Irregular verbs few (~250) but frequently occurring
–
•
-s, ing, ed
And productive: Emailed, instant-messaged, faxed, homered
But some are not regular: eat/ate/eaten, catch/caught/caught
Irregularity a consequence of frequency?
So….English inflectional morphology is fairly easy to model….with some special
cases...
(English) Derivational Morphology
Word stem + grammatical morpheme
•
Usually produces word of different class
•
More complicated than inflectional
E.g. verbs --> nouns
•
-ize verbs  -ation nouns
•
generalize, realize  generalization, realization
E.g.: verbs, nouns  adjectives
•
embrace, pity embraceable, pitiable
•
care, wit  careless, witless
E.g.: adjective  adverb
•
happy  happily
•
But “rules” have many exceptions
•
Less productive: *evidence-less, *concern-less, *go-able, *sleep-able
•
Meanings of derived terms harder to predict by rule
–
clueless, careless, nerveless
How do humans represent words?
Hypotheses:
• Full listing hypothesis: words listed
•
Minimum redundancy hypothesis: morphemes listed
– Words are formed using rules.
Experimental evidence:
• Priming experiments (Does seeing/hearing one word facilitate
recognition of another?) suggest neither
•
Regularly inflected forms prime stem but not derived forms
•
But spoken derived words can prime stems if they are semantically
close (e.g. government/govern but not department/depart)
Speech errors suggest affixes must be represented separately in the
mental lexicon
• easy enoughly
Finite-State Morphological Parsing
Morphological Parsing:
•
takes : root=take, category=V, person=3, number=sg
•
cities : root=city, category=N, number=plural
•
ships: a. root=ship, category=N, number=plural
b. root=ship, category=V, person=3, number=sg
What does a parser need?
•
Lexicon: root forms of the words in a language
•
Morphological Rules: morpheme ordering rules
•
Orthographic Rules: spelling change rules
We will represent each of these components as finite-state automata.
Parsing versus Recognition
•
Recognition: yes/no
•
Parsing: derivation structure
–
indecipherable  [in [[de [cipher]] able]]
Lexicon as an FSA
Words can be represented as paths in a FSA with characters as
transition symbols.
•
Also called a “trie” structure
List of words: bat, cat, mat, sat
c
b
m
s
a
t
Morphological Recognition
Check if a word is a valid word
in the language.
reg-n
For nouns: cats, geese, goose
q0
Adj-root1: clear, happy,
real (clearly)
•
Adj-root2: big, red (~bigly)
q1
q2
irreg-pl-n
For adjectives: Derivational
morphology
•
plural (-s)
irreg-sg-n
un-
adj-root1 -er, -ly, -est
q1
q2
adj-root1
q0

q3
q4
adj-root2
q5
-er, -est
Morphotactic Rules
An automaton that encodes valid English word forming rules.
Koskenniemi’s two-level morphology
•
Lexical string: stem and morphemes
•
Surface string: spelling of the derived word
•
Pair up lexical string with the surface string.
•
(N,bat,pl; bats)
•
(V,bat,3sg; bats)
•
Align the strings at the character level
N: b:b a:a t:t pl:s
– V: b:b a:a t:t 3sg:s
–
Relation that maps characters to other characters.
Finite-State Transducers
Finite State Acceptors represent regular sets.
Finite State Transducers represent regular relations.
Relation
•
If A and B are two regular sets; relation R ⊆ A x B
•
Example: {(x,y) | x ∊ a*, y ∊ b*}
FSTs can be considered as
•
Translators (Hello:Ciao)
•
Parser/generators (Hello:How may I help you?)
•
As well as Kimmo-style morphological parsing
Finite State Transducers – formally speaking
FST is a 5-tuple consisting of
•
Q: set of states {q0,q1,q2,q3,q4}
•
: an alphabet of complex symbols, each an i:o pair s.t. i  I (an input
alphabet) and o  O (an output alphabet) and  ⊆ I x O
•
q0: a start state
•
F: a set of final states in Q {q4}
•
(q,i:o): a transition function mapping Q x  to Q
a:o
b:m
q0
a:o
q1
a:o
q2
!:?
q3
q4
More FST examples
FSTs are functions,
• Compose (L1,L2) = {(x,y) | (x,z) ∊ L1 and (z,y) ∊ L2}
–
3((q1,q2),(x,y)) = (q3,q4) if ∃z 1(q1,(x,z)) = q3 and 2(q2,(z,y)) = q4
•
Inverse(L) = {(x,y) | (y,x) ∊ L}
•
FirstProjection(L) = {x | (x,y) ∊ L}
•
SecondProjection(L) = {y | (x,y) ∊ L}
Speech Recognition as FST composition
•
A: Acoustic  phones
•
Lex: phones  Words
•
Gram: Words  Words
•
Speech Recognition: A ● Lex ● Gram
Machine Translation as FST composition
•
LexTrans: SourceWords  TargetWords
•
LexReorder: UnOrderedTargetWords  OrderedTargetWords
Weighted FSTs allow for ranking of the generated outputs.
FST for a 2-level Lexicon
E.g.
q0
c:c
q1
q4
g:g
a:a
q5
q6
e:o
e:o
t:t
q2
q7
q3
e:e
s:s
Nominal Morphology
reg-n
q0
+N:
q4
+PL:^s#
+SG:#
q2
irreg-n-sg
irreg-n-pl
q5
q6
q7
+SG:#
+PL:#
Orthographic Rules
Define additional FSTs to implement spelling rules
•
Gemination: consonant doubling (beg  begging),
•
Elision: ‘e’ deletion (make  making),
•
Epenthesis: ‘e’ insertion (fox  foxes), etc.
•
Y replacement: ‘y’ changes to ‘ie’ before ‘ed’ (trytries)
•
I spelling: ‘I’ changes to ‘y’ before a vowel (lielying)
Rewrite rule notation: a  b / p _ 
Examples: Rule: a  b/a _ b
•
aaabaab  aabbabb (Sequential; there are other control strategies:
parallel, feedback).
Spelling rules:   e / x ^ _ s (eg. fox^s  foxes)
Multiple spelling rules might apply for a given word.
•
Eg: spy^s  spies (yi, e)
All spelling rules execute in parallel.
Rewrite rules can be compiled into FSTs.
FST-based Morphological Parsing
Remember FSTs can be composed.
Lexicon : List of stems
Morphotactics: Rules for adding morphemes
Orthographic Rules: Spelling Rules
Morphology =
Lexicon ● Morphotactics ● Orhtographics
All morphological analyses generated for a given word,
…but in a context a preference ordering on the analysis is needed.
•
Weighted FSTs to the rescue.
Porter Stemmer
Lexicon free stemmer; heuristic rules for deriving the stem.
Rules to rewrite the suffix of a word.
•
“ational”  “ate” (eg. relational  relate)
•
“-ing”   (eg. motoring  motor)
Details of the rules in Appendix B in the book.
Purported to improve recall in IR engines.
Role of Morphology in Machine Translation
Every MT system contains a bilingual lexicon
Bilingual lexicon: a table mapping the source language token to target
language token(s).
Two options:
1. Full-form lexicon
–
–
every word form of the source token is paired with the target token
large table if the vocabulary is large for morphologically rich languages
2. Root-form lexicon
–
–
–
–
–
–
pairing of stems from the two languages
Reduces the size of the lexicon
requires morphological analysis for source language
bats  (bat, V, 3sg) (bat, N, pl)
morphological generation for target language
(bat, V, 3sg)  bats
Unknown words: words not covered in the bilingual lexicon
- with morphology, one can guess the syntactic function
Compounding of words:
English: simple juxtaposition (“car seat”), some times (“seaweed”)
German: fusion is more common
(“Dampfschiffahrtsgesellschaft”  steamship company
Readings
For this class
HS: pages 15-16, 81-85
JM: chapter on Morphology
For next class
Machine Translation divergences (Dorr 94)
Example based Machine Translation (Somers 99)