Morphology & FSTs Shallow Processing Techniques for NLP Ling570 October 17, 2011.

Morphology & FSTsShallow Processing Techniques for NLP

Ling570October 17, 2011

RoadmapTwo-level morphology summary

Unsupervised morphology

Combining FST Lexicon & Rules

Two-level morphological system: ‘Cascade’Transducer from Lexicon to IntermediateRule transducers from Intermediate to Surface

Integrating the LexiconReplace classes with stems

Using the E-insertion FST

(fox,fox): q0, q0,q0,q1, accept(fox#,fox#): q0.q0.q0.q1,q0, accept (fox^s#,foxes#): q0,q0,q0,q1,q2,q3,q4,q0,accept(fox^s,foxs): q0,q0,q0,q1 ,q2,q5,reject(fox^z#,foxz#) ?

IssuesWhat do you think of creating all the rules for a

languages – by hand?Time-consuming, complicated



Proposed approach: Unsupervised morphology induction



Proposed approach: Unsupervised morphology induction

Potentially useful for many applications IR, MT

Unsupervised MorphologyStart from tokenized text (or word frequencies)

talk 60talked 120walked 40walk 30

Unsupervised MorphologyStart from tokenized text (or word frequencies)

talk 60talked 120walked 40walk 30

Treat as coding/compression problemFind most compact representation of lexicon

Popular model MDL (Minimum Description Length) Smallest total encoding:

Weighted combination of lexicon size & ‘rules’

ApproachGenerate initial model:

Base set of words, compute MDL length



Iterate:Generate a new set of words + some model to

create a smaller description size



Iterate:Generate a new set of words + some model to

create a smaller description size

E.g. for talk, talked, walk, walked4 words



Iterate:Generate a new set of words + some model to create

a smaller description size

E.g. for talk, talked, walk, walked4 words2 words (talk, walk) + 1 affix (-ed) + combination info2 words (t,w) + 2 affixes (alk,-ed) + combination info

Successful ApplicationsInducing word classes (e.g. N,V) by affix patterns

Unsupervised morphological analysis for MT

Word segmentation in CJK

Word text/sound segmentation in English

Unit #1 Summary

Formal Languages Formal Languages and Grammars

Chomsky hierarchy Languages and the grammars that

accept/generate

Formal Languages Formal Languages and Grammars

Chomsky hierarchy Languages and the grammars that

accept/generate

EquivalencesRegular languagesRegular grammarsRegular expressionsFinite State Automata

Finite-State Automata & Transducers

Finite-State Automata:Deterministic & non-deterministic automata

Equivalence and conversionProbabilistic & weighted FSAs




Packages and operations: Carmel




Packages and operations: Carmel

FSTs & regular relationsClosures and equivalencesComposition, inversion

FSA/FST ApplicationsRange of applications:

ParsingTranslationTokenization…

FSA/FST ApplicationsRange of applications:

ParsingTranslationTokenization…

Morphology:Lexicon: cat: N, +Sg; -s: PlMorphotactics: N+PLOrthographic rules: fox + s foxesParsing & Generation

ImplementationTokenizers

FSA acceptors

FST acceptors/translators

Orthographic rule as FST

Language Modeling

RoadmapMotivation:

LM applications

N-grams

Training and Testing

Evaluation: Perplexity

Predicting WordsGiven a sequence of words, the next word is

(somewhat) predictable: I’d like to place a collect …..



Ngram models: Predict next word given previous N

Language models (LMs):Statistical models of word sequences



Ngram models: Predict next word given previous N

Language models (LMs):Statistical models of word sequences

Approach: Build model of word sequences from corpusGiven alternative sequences, select the most

probable

N-gram LM ApplicationsUsed in

Speech recognition

Spelling correction

Augmentative communication

Part-of-speech tagging

Machine translation

Information retrieval

TerminologyCorpus (pl. corpora):

Online collection of text of speechE.g. Brown corpus: 1M word, balanced text collectionE.g. Switchboard: 240 hrs of speech; ~3M words



Wordform: Full inflected or derived form of word: cats,

glottalized




glottalized

Word types: # of distinct words in corpus




glottalized

Word types: # of distinct words in corpus

Word tokens: total # of words in corpus

Corpus CountsEstimate probabilities by counts in large

collections of text/speech

Should we count:Wordform vs lemma ?

Case? Punctuation? Disfluency?

Type vs Token ?

Words, Counts and Prediction

They picnicked by the pool, then lay back on the grass and looked at the stars.


They picnicked by the pool, then lay back on the grass and looked at the stars.Word types (excluding punct):


They picnicked by the pool, then lay back on the grass and looked at the stars.Word types (excluding punct): 14Word tokens (“ ):


They picnicked by the pool, then lay back on the grass and looked at the stars.Word types (excluding punct): 14Word tokens (“ ): 16.

I do uh main- mainly business data processingUtterance (spoken “sentence” equivalent)



I do uh main- mainly business data processingUtterance (spoken “sentence” equivalent)What about:

Disfluenciesmain-: fragmentuh: filler (aka filled pause)



I do uh main- mainly business data processingUtterance (spoken “sentence” equivalent)What about:

Disfluenciesmain-: fragmentuh: filler (aka filled pause)

Keep, depending on app.: can help prediction; uh vs um

LM TaskTraining:

Given a corpus of text, learn probabilities of word sequences

LM TaskTraining:


Testing:Given trained LM and new text, determine

sequence probabilities, orSelect most probable sequence among

alternatives

LM TaskTraining:


Testing:Given trained LM and new text, determine

sequence probabilities, orSelect most probable sequence among

alternatives

LM types:Basic, Class-based, Structured

Word PredictionGoal:

Given some history, what is probability of some next word?

Formally, P(w|h)e.g. P(call|I’d like to place a collect)




How can we compute?




How can we compute?Relative frequency in a corpus

C(I’d like to place a collect call)/C(I’d like to place a collect)

Issues?


Given some history, what is probability of some next word? Formally, P(w|h)

e.g. P(call|I’d like to place a collect)

How can we compute? Relative frequency in a corpus

C(I’d like to place a collect call)/C(I’d like to place a collect)

Issues? Zero counts: language is productive! Joint word sequence probability of length N:

Count of all sequences of length N & count of that sequence

Word Sequence ProbabilityNotation:

P(Xi=the) written as P(the)

P(w1w2w3…wn) =



P(w1w2w3…wn) =

Compute probability of word sequence by chain rule Links to word prediction by history



P(w1w2w3…wn) =


Issues?



P(w1w2w3…wn) =


Issues? Potentially infinite history

Word Sequence Probability Notation:


P(w1w2w3…wn) =


Issues? Potentially infinite history Language infinitely productive

Markov AssumptionsExact computation requires too much data


Approximate probability given all prior wordsAssume finite history


Approximate probability given all prior wordsAssume finite historyUnigram: Probability of word in isolation (0th order)Bigram: Probability of word given 1 previous

First-order Markov

Trigram: Probability of word given 2 previous


Approximate probability given all prior words Assume finite history Unigram: Probability of word in isolation (0th order) Bigram: Probability of word given 1 previous

First-order Markov

Trigram: Probability of word given 2 previous

N-gram approximation

)|()|( 11

11

nNnn

nn wwPwwP

)|()( 11

1 k

n

kk

n wwPwPBigram sequence

Unigram ModelsP(w1w2…w3)~

Unigram ModelsP(w1w2…w3) ~ P(w1)*P(w2)*…*P(wn)

Training: Estimate P(w) given corpus


Training: Estimate P(w) given corpusRelative frequency:


Training: Estimate P(w) given corpusRelative frequency: P(w) = C(w)/N, N=# tokens in

corpusHow many parameters?


Training: Estimate P(w) given corpusRelative frequency: P(w) = C(w)/N, N=# tokens in

corpusHow many parameters?

Testing: For sentence s, compute P(s)

Model with PFA: Input symbols? Probabilities on arcs? States?

Bigram ModelsP(w1w2…w3) = P(BOS w1w2….wnEOS)


~ P(BOS)*P(w1|BOS)*P(w2|w1)*…*P(wn|wn-1)*P(EOS|wn)

Training: Relative frequency:



Training: Relative frequency: P(wi|wi-1) = C(wi-1wi)/C(wi-1)

How many parameters?



Training: Relative frequency: P(wi|wi-1) = C(wi-1wi)/C(wi-1)


Testing: For sentence s, compute P(s)

Model with PFA: Input symbols? Probabilities on arcs? States?

Trigram ModelsP(w1w2…w3) = P(BOS w1w2….wnEOS)


~ P(BOS)*P(w1|BOS)*P(w2|BOS,w1)*… *P(wn|wn-2,wn-

1)*P(EOS|wn-1,wn)

Training: P(wi|wi-2,wi-1)



1)*P(EOS|wn-1,wn)

Training: P(wi|wi-2,wi-1) = C(wi-2 wi-1wi)/C(wi-2wi-1)




1)*P(EOS|wn-1,wn)

Training: P(wi|wi-2,wi-1) = C(wi-2 wi-1wi)/C(wi-2wi-1)


How many states?

Speech and Language Processing - Jurafsky and Martin

An Example<s> I am Sam </s>

<s> Sam I am </s>

<s> I do not like green eggs and ham </s>

RecapNgrams:

# FSA states:

RecapNgrams:

# FSA states: |V|n-1

# Model parameters:

RecapNgrams:


# Model parameters: |V|n

Issues:

RecapNgrams:


# Model parameters: |V|n

Issues:Data sparseness, Out-of-vocabulary elements

(OOV) Smoothing

Mismatches between training & test dataOther Language Models

Morphology & FSTs Shallow Processing Techniques for NLP Ling570 October 17, 2011.

Documents

complicated slide

summary slide

acceptgenerate slide

mt slide

carmel slide

english slide

surface slide

compute mdl length slide