9

Click here to load reader

Anaphora Resolution

Embed Size (px)

Citation preview

Page 1: Anaphora Resolution

ANAPHORA RESOLUTIONFINDWISE

Page 2: Anaphora Resolution

Anaphoric Pronoun Resolution

Finding links• Pronoun to antecedent

Enriching text• Input: preprocessed document• Output: All found anaphoric pronoun

references to words/phrases

Page 3: Anaphora Resolution

Areas of use

Document summarization• Improving sentence comparisons• Enriching results

Entity level sentiment analysis• Adding more information to indata.

Ontology enrichment• Populating with more

data.Question answering• Extracting more RDF-

tripples

Page 4: Anaphora Resolution

Preprocessing

Required• Sentence splitting• Tokenization• Part of Speech-tagging• Named Entity Reconition• Gender Detection

Additional• Dependency parsing

Page 5: Anaphora Resolution

Model representation

Anaphora pairs• Pronoun• Antecedent

- Entities- Nouns, cardinals, foreign words

Candidate selection/ranking• Find pronoun• Pair with antecedent candidates• Filter out improbable pairs (rules)• Rank candidate pairs• Select the most probable

candidate (if any)

Page 6: Anaphora Resolution

Feature representation

Distance Features• Sentence distance• Hobbs distance

Antecedent Features• PoS-tag• Gender• Animacity• Number• Entity tag• ...

Overlap Features/Filters• Gender• Animacity• Number• Entity

Pronoun Features• Word string• Gender• Animacity• ...

Page 7: Anaphora Resolution

Machine learning models

Models• Condidtional Random Fields (CRF)

- Mallet• Logistic Regression

- LiblinearTraining the models• OntoNotes Conll 2012• English• 1667 documents• Various domains

Running the models• Control confiedence

threshold - Precision/Recall trade off

Page 8: Anaphora Resolution

Further Work/Ideas for Improvement

Full coreference/anaphora resolution• Change model representations

- Clusters- Chains

• Generalize comparisons (not only pronoun - antecedent)Non referential/cataphora detection• Training separate models• Rule based

Improved Features• Improved gender detection• Improved animacity detection• Additional overlap features

Multi pass approach• First pass(es) rule based• Harder classifications with

machine learning models

Page 9: Anaphora Resolution

Demonstration