Unsupervised Approaches to Sequence Tagging, Morphology Induction, and

Lexical Resource Acquisition

Reza Bosaghzadeh & Nathan Schneider

LS2 ~ 1 December 2008

Unsupervised methods– Sequence labeling (Part-of-Speech tagging)

– Morphology Induction

– Lexical Resource acquisition

.

She ran to the station quickly

pronoun verb preposition det noun adverb

un-supervise-d learn-ing

Part of Speech (POS) Tagging

• Prototype-driven model– Cluster based on a few examples – (Haghighi and Klein 2006)

• Contrastive estimation – Discount positive examples at cost of

implicit negative example– (Smith and Eisner 2005)

Prototype Driven taggingOverview

+

PrototypesTarget Label

Unlabeled Data

Prototype List

Annotated Data

(Haghighi and Klein 2006)

Prototypes

Newly remodeled 2 Bdrms/1 Bath, spacious upper unit, located in Hilltop Mall area. Walking distance to shopping, public transportation, schools and park. Paid water and garbage. No dogs allowed.

FEATURE kitchen, laundry

LOCATION near, closeTERMS paid, utilitiesSIZE large, feetRESTRICT cat, smoking

Information Extraction: Classified Ads

FeaturesLocationTermsRestrictSize

Prototype List

Newly remodeled 2 Bdrms/1 Bath, spacious upper unit, located in Hilltop Mall area. Walking distance to shopping, public transportation, schools and park. Paid water and garbage. No dogs allowed.

(Haghighi and Klein 2006)

Prototypes

Newly remodeled 2 Bdrms/1 Bath, spacious upper unit, located in Hilltop Mall area. Walking distance to shopping, public transportation, schools and park. Paid water and garbage. No dogs allowed.

Newly remodeled 2 Bdrms/1 Bath, spacious upper unit, located in Hilltop Mall area. Walking distance to shopping, public transportation, schools and park. Paid water and garbage. No dogs allowed.

Prototype List

NN VBN CC JJ CD PUNCIN NNS IN NNP RB DET

NN president IN of

VBD said NNS shares

CC and TO to

NNP Mr. PUNC .

JJ new CD million

DET the VBP are

English POS

(Haghighi and Klein 2006)

Where to get Prototypes

• 3 prototypes per tag

• Automatically extracted by frequency

(Haghighi and Klein 2006)

Prototypes

Tie each word to its most similar prototype

Features are same as (Smith & Eisner 05)

Trigram taggerWord type, suffixes up to length 3, contains hyphen, contains digit, initial capitalization

Similarity (Schuetze 93)• SVD dimensionality reduction• cos() between context vectors

(Haghighi and Klein 2006)

Prototypes – Language Variation

Pros• Can get prototypes for popular

languages• Doesn’t require tagging dictionary

Cons• Needs tag set• Needs prototypes for each tag

Contrastive Estimation• Already discussed in class• Key Idea:

– Mutating training examples often gives ungrammatical (negative) sentences

– During training, shift probability mass from generated negative examples to given positive examples

(Smith and Eisner 2005)

Unsupervised POS taggingCurrent results

Best supervised results (CRFs): 99.5% !

Unsupervised methods– Sequence labeling (Part-of-Speech tagging)

– Morphology Induction

– Lexical Resource acquisition

.

She ran to the station quickly

pronoun verb preposition det noun adverb

un-supervise-d learn-ing

Unsupervised Approaches to Morphology

• Morphology refers to the internal structure of words– A morpheme is a minimal meaningful

linguistic unit– Morpheme segmentation is the

process of dividing words into their component morphemes

unsupervised learning

.

Unsupervised Approaches to Morphology

• Morphology refers to the internal structure of words– A morpheme is a minimal meaningful

linguistic unit– Morpheme segmentation is the process of

dividing words into their component morphemes

un-supervise-d learn-ing– Word segmentation is the process of

finding word boundaries in a stream of speech or text

unsupervisedlearningofnaturallanguage

Unsupervised Approaches to Morphology

• Morphology refers to the internal structure of words– A morpheme is a minimal meaningful

linguistic unit– Morpheme segmentation is the process of

dividing words into their component morphemes

un-supervise-d learn-ing– Word segmentation is the process of

finding word boundaries in a stream of speech or textunsupervised learning of natural language

ParaMor: Monson et al. (2007, 2008)

• Learns inflectional paradigms from raw text– Requires only the vocabulary of a corpus– Looks at word counts of substrings, and

proposes (stem, suffix) pairings based on type frequency

• 3-stage algorithm– Stage 1: Candidate paradigms based on

frequencies– Stages 2-3: Refinement of paradigm set via

merging and filtering• Paradigms can be used for morpheme

segmentation or stemming

ParaMor: Monson et al. (2007, 2008)

speak dancehablar bailarhablo bailohablamos bailamoshablan bailan… …• A sampling of Spanish verb

conjugations

ParaMor: Monson et al. (2007, 2008)

speak dancehablar bailarhablo bailohablamos bailamoshablan bailan… …• A proposed paradigm with stems {habl,

bail} and suffixes {-ar, -o, -amos, -an}

ParaMor: Monson et al. (2007, 2008)

speak dance buyhablar bailar comprarhablo bailo comprohablamos bailamos compramoshablan bailan compran… … …• Same paradigm from previous slide,

but with stems {habl, bail, compr}

ParaMor: Monson et al. (2007, 2008)

speak dancehablar bailarhablo bailohablamos bailamoshablan bailan… …• From just this list, other paradigm analyses

(which happen to be incorrect) are possible

ParaMor: Monson et al. (2007, 2008)

speak dancehablar bailarhablo bailohablamos bailamoshablan bailan… …• Another possibility: stems {hab, bai},

suffixes {-lar, -lo, -lamos, -lan}

ParaMor: Monson et al. (2007, 2008)

speak dance buyhablar bailar comprarhablo bailo comprohablamos bailamos compramoshablan bailan compran… … …• Spurious segmentations—this paradigm

doesn’t generalize to comprar (or most verbs)

ParaMor: Monson et al. (2007, 2008)

• What if not all conjugations were in the corpus?

speak dance buyhablar bailar comprar

bailo comprohablamos bailamos compramoshablan… … …

ParaMor: Monson et al. (2007, 2008)

• We have two similar paradigms that we want to merge

speak dance buyhablar bailar comprar

bailo comprohablamos bailamos compramoshablan… … …

ParaMor: Monson et al. (2007, 2008)

speak dance buyhablar bailar comprarhablo bailo comprohablamos bailamos compramoshablan bailan compran… … …• This amounts to smoothing, or

“hallucinating” out-of-vocabulary items

ParaMor: Monson et al. (2007, 2008)

• Heuristic-based, deterministic algorithm can learn inflectional paradigms from raw text

• Paradigms can be used straightforwardly to predict segmentations– Combining the outputs of ParaMor and

Morfessor (another system) won the segmentation task at MorphoChallenge 2008 for every language: English, Arabic, Turkish, German, and Finnish

• Currently, ParaMor assumes suffix-based morphology

• Word segmentation results – comparison

• See Narges & Andreas’s presentation for more on this model

Goldwater Unigram DPGoldwater Bigram HDP

Goldwater et al. (2006; in submission)

Multilingual morpheme segmentation Snyder & Barzilay (2008)

speak ES speak FRhablar parlerhablo parlehablamos parlonshablan parlent… …• Abstract morphemes cross

languages: (ar, er), (o, e), (amos, ons), (an, ent), (habl, parl)

• Considers parallel phrases and tries to find morpheme correspondences

• Stray morphemes don’t correspond across languages

Morphology papers: inputs & outputs

Unsupervised methods– Sequence labeling (Part-of-Speech

tagging)

– Morphology Induction

– Lexical Resource acquisition

.

She ran to the station quickly

pronoun verb preposition det noun adverb

un-supervise-d learn-ing

Lexical Resource Acquisition• Learning Bilingual Lexicons from

Monolingual Corpora (Haghighi et al. 2008)

• Narrative Event Chain Acquisition (Chambers and Jurafsky 2008)

• Labeling semantic roles for verbs (Grenager and Manning 2006)

Learning Bilingual Lexicons from Monolingual Corpora

SourceText

TargetText

Matching

mstate

world

name

Source Words

s

nation

estado

política

Target Words

t

mundo

nombre

(Haghighi et al. 2008)

Bilingual Lexicons: Data Representation

state

Orthographic Features 1.0

1.0

1.0

#sttatte#

5.0

20.0

10.0

Context Featuresworldpoliticssociety

SourceText

estado

Orthographic Features 1.0

1.0

1.0

#esstado#

10.0

17.0

6.0

Context Featuresmundopoliticasociedad

TargetText

(Haghighi et al. 2008)

Use CCA (Canonical Correlation Analysis) Trained with Hard EM to find best matching

Feature Experiments

Series10

25

50

75

100

61.1

80.1 80.289.0

Edit Dist Ortho Context MCCA

Prec

isio

n• MCCA: Orthographic and context

features

4k EN-ES Wikipedia Articles (Haghighi et al. 2008)

Narrative events: Chambers & Jurafsky (2008)

• Given a corpus, identifies related events that constitute a “narrative” and (when possible) predict their typical temporal ordering– E.g.: CRIMINAL PROSECUTION narrative, with

verbs: arrest, accuse, plead, testify, acquit/convict

• Key insight: related events tend to share a participant in a document– The common participant may fill different

syntactic/semantic roles with respect to verbs: arrest.OBJECT, accuse.OBJECT, plead.SUBJECT

Narrative events: Chambers & Jurafsky (2008)

• A temporal classifier can reconstruct pairwise canonical event orderings, producing a directed graph for each narrative

Narrative events: Grenager & Manning (2006)

• From dependency parses, a generative model predicts semantic roles corresponding to each verb’s arguments, as well as their syntactic realizations– PropBank-style: ARG0, ARG1, etc. per verb (do

not necessarily correspond across verbs)– Learned syntactic patterns of the form:

(subj=give.ARG0, verb=give, np#1=give.ARG2, np#2=give.ARG1) or (subj=give.ARG0, verb=give, np#2=give.ARG1, pp_to=give.ARG2)

• Used for semantic role labeling

“Semanticity”: Our proposed scale of semantic richness

• text < POS < syntax/morphology/alignments < coreference/semantic roles/temporal ordering < translations/narrative event sequences

• We score each model’s inputs and outputs on this scale, and call the input-to-output increase “semantic gain”– Haghighi et al.’s bilingual lexicon induction wins

in this respect, going from raw text to lexical translations

Robustness to language variation• About half of the papers we examined

had English-only evaluations• We considered which techniques were

most adaptable to other (esp. resource-poor) languages. Two main factors:– Reliance on existing tools/resources for

preprocessing (parsers, coreference resolvers, …)

– Any linguistic specificity in the model (e.g. suffix-based morphology)

SummaryWe examined three areas of unsupervised NLP:

1. Sequence tagging: How can we predict POS (or topic) tags for words in sequence?

2. Morphology: How are words put together from morphemes (and how can we break them apart)?

3. Lexical resources: How can we identify lexical translations, semantic roles and argument frames, or narrative event sequences from text?

In eight recent papers we found a variety of approaches, including heuristic algorithms, Bayesian methods, and EM-style techniques.

Thanks to Noah and Kevin for their feedback on the paper; Andreas and Narges for their collaboration on the presentations; and all of you for giving us your attention!

Questions?

un-supervise-d learn-inghablar bailarhablo bailohablamos

bailamos

hablan bailan

subj=give.ARG0 verb=give np#1=give.ARG2 np#2=give.ARG1