Detecting Anaphoricity and Antecedenthood for Coreference Resolution
description
Transcript of Detecting Anaphoricity and Antecedenthood for Coreference Resolution
Detecting Anaphoricity and Antecedenthood for Coreference
Resolution
Olga Uryupina ([email protected])
Institute of Linguistics, RAS 13.11.08
Overview
• Anaphoricity and Antecedenthood• Experiments• Incorporating A&A detectors into a
CR system• Conclusion
A&A: example
Shares in Loral Space will be distributed to Loral shareholders. The new company will start life with no debt and $700 million in cash. Globalstar still needs to raise $600 million, and Schwartz said that the company would try to raise the money in the debt market.
A&A: example
Shares in Loral Space will be distributed to Loral shareholders. The new company will start life with no debt and $700 million in cash. Globalstar still needs to raise $600 million, and Schwartz said that the company would try to raise the money in the debt market.
Anaphoricity
Likely anaphors:- pronouns, definite descriptions
Unlikely anaphors:- indefinites
Unknown:- proper names
Poesio&Vieira: more than 50% of definite descriptions in a newswire text are not anaphoric!
A&A: example
Shares in Loral Space will be distributed to Loral shareholders. The new company will start life with no debt and $700 million in cash. Globalstar still needs to raise $600 million, and Schwartz said that the company would try to raise the money in the debt market.
A&A: example
Shares in Loral Space will be distributed to Loral shareholders. The new company will start life with no debt and $700 million in cash. Globalstar still needs to raise $600 million, and Schwartz said that the company would try to raise the money in the debt market.
Antecedenthood
Related to referentiality (Karttunen, 1976):
„no debt“ etc
Antecedenthood vs. Referentiality: corpus-based decision
Experiments
• Can we learn anaphoricity/antecedenthood classifiers?
• Do they help for coreference resolution?
Methodology
• MUC-7 dataset • Anaphoricity/antecedenthood
induced from the MUC annotations• Ripper, SVM
Features
• Surface form (12)• Syntax (20)• Semantics (3)• Salience (10)• „same-head“ (2)• From Karttunen, 1976 (7)
49 features – 123 boolean/continuous
Results: anaphoricity
Feature groups R P F
Baseline 100 66.5 79.9
All 93.5 82.3 87.6
Surface 100 66.5 79.9
Syntax 97.4 72.0 82.8
Semantics 98.5 68.9 81.1
Salience 91.2 69.3 78.7
Same-head 84.5 81.1 82.8
Karttunen‘s 91.6 71.1 80.1
Synt+SH 90.0 83.5 86.6
Results: antecedenthood
Feature groups R P F
Baseline 100 66.5 79.9
All 95.7 69.2 80.4
Surface 94.6 68.5 79.5
Syntax 95.7 69.2 80.3
Semantics 94.9 69.4 80.2
Salience 98.9 67.0 79.9
Same-head 100 66.5 79.9
Karttunen‘s 99.3 67.3 80.2
Integrating A&A into a CR system
Apply an A&A prefiltering before CR starts:
- Saves time- Improves precision
Problem: we can filter out good candidates..:
- Will loose some recall
Oracle-based A&A prefiltering
Take MUC-based A&A classifier („gold standard“
CR system: Soon et al. (2001) with SVMs
MUC-7 validation set (3 „training“ documents)
Oracle-based A&A prefiltering
R P F
No prefilteing 54.5 56.9 55.7
±ana 49.6 73.6 59.3
±ante 54.2 69.4 60.9
±ana & ±ante 52.9 81.9 64.3
Automatically induced classifiers
Precision more crucial than Recall
Learn Ripper classifiers with different Ls (Loss Ratio)
Anaphoricity prefiltering
Antecedenthood prefiltering
Conclusion
Automatically induced detectors:• Reliable for anaphoricity• Much less reliable for antecedenthood(a corpus, explicitly annotated for
referentiality could help)A&A prefiltering:• Ideally, should help• In practice – substantial optimization
required
Thank You!