スライド 1 - Stanford University

Download Report

Transcript スライド 1 - Stanford University

Anaphora Resolution

Sanghoon Kwak Takahiro Aoyama

Anaphora Resolution

• Moussa said the presidents discussed the situation in Iraq and would return to it Monday. However, Syria and Egypt were planning to send planes to Baghdad. he would not comment on whether • Anaphora resolution is the process of determining whether a noun phrase refers to an antecedent noun phrase.

Approach

• Extracted 12 features for each noun phrase pair from a news article corpora • Constructed a training set using these 12 features. Used 70 news articles from NY Times and Broadcast News. • Built a decision tree with this training data using the C4.5 decision tree builder

Features

•      Some of 12 features include… Distance Distance by the number of sentences Pronoun Whether a noun phrase is a pronoun or not String match Whether two phrases have the same character sequence Semantic Class Whether two phrases are in the same semantic class (Person, Organization, etc.) Number Whether two phrases are both plural or both singular

Result

• Training and Test Result Phrase Pairs Incorrectly Classified Error Rate Training 406,400 24,762 6.1% Testing 212,865 14,213 6.7% • Rules based on decision tree   DIST <= 0, JPRO = True, NUM = True, SEMCLASS = True Coreferent   IPRO = True, STR MATCH = true, DIST > 0 Not Coreferent

Future Work

• Train and test data from different domains (Blog posts, E-Mail messages, etc.) • Further subdivide decision tree classes