TL;DR
This paper tackles the challenging task of event anaphora resolution, specifically identifying which event verb a pronoun refers to. They introduce a novel "twin-candidate" approach that learns preferences between potential antecedents and a composite kernel that combines various linguistic features, achieving improved performance over existing methods.
What Problem Does This Paper Address?
Resolving pronouns to their antecedents (anaphora resolution) is crucial for understanding text. While object anaphora (e.g., "John ate an apple. It was red.") has been extensively studied, event anaphora (e.g., "The company was acquired. It happened last week.") refers to events, facts, or propositions and is more complex. Previous work on event anaphora has been limited, often lacking systematic feature exploration or computational systems. This paper aims to provide a comprehensive study of event pronoun resolution, focusing on identifying event verbs as antecedents, and proposes a robust system to tackle this challenge.
Methods
The authors propose a learning-based framework for event pronoun resolution, building upon common approaches for object anaphora. Key methodological contributions include:
- Feature Engineering: They explore a rich set of features, including positional, lexical, and syntactic features, noting that many standard anaphora features are less relevant for event antecedents (verbs) compared to noun phrases.
- Composite Kernel: To effectively integrate diverse information, they develop a composite kernel. This combines a standard "flat" feature kernel (for positional, lexical, and POS tag information) with a novel "tree kernel" that captures structural syntactic information from parse trees.
- Structural Syntactic Features: They experiment with three types of tree structures derived from parse trees: Minimum Expansion Tree, Simple Expansion Tree, and Full Expansion Tree, each capturing progressively more syntactic context around the pronoun and its candidate antecedent.
- Twin-Candidate Preference Learning: A core innovation is the "twin-candidate" approach, inspired by ranking SVMs. Instead of classifying a single candidate as correct, it learns preferences between pairs of candidates, assuming the antecedent is the one preferred over all others.
- Negative Instance Handling: They incorporate negative instances from non-event anaphoric pronouns (pronouns whose antecedents are not verbs) to improve the classifier's discriminative power. To address the imbalance this creates, they employ random down-sampling.
- Hyper-plane Adjustment: They propose adjusting the SVM hyper-plane using a development dataset to find a more accurate decision boundary, leading to improved performance.
Datasets
The experiments are conducted on the OntoNotes Release 2.0 English corpus. This corpus contains a mix of newswire and broadcast news data. The authors specifically focus on resolving 502 event pronouns identified within this corpus. They note that event anaphoric pronouns constitute about 14.7% of all pronouns in OntoNotes.
Key Insights and Ideas
- Event Anaphora is Different: Features effective for object anaphora (like gender agreement) are largely irrelevant for event anaphora, highlighting the need for specialized feature sets.
- Syntax Matters, But How? While syntactic information is crucial, the way it's represented matters. The "Simple Expansion Tree" strikes a good balance between capturing relevant syntactic context and avoiding noise, leading to the best F-score when combined with flat features.
- Preferences are Powerful: Learning pairwise preferences between candidates (the "twin-candidate" approach) significantly improves recall, suggesting that understanding relative suitability is more effective than absolute classification for some candidates.
- Negative Examples are Useful (with caution): Including negative instances from non-event anaphoric pronouns boosts precision, but careful handling (like balanced sampling) is essential to prevent overwhelming the model.
- Development Data is Key: Fine-tuning the SVM hyper-plane using a dedicated development set leads to substantial performance gains, demonstrating the value of data-driven hyper-parameter optimization.
Key Results
- The baseline system using only flat features achieved an F-score of 40.6%.
- Combining flat features with the "Simple Expansion Tree" structure, and using the adjusted hyper-plane, yielded the best performance among single-candidate approaches, achieving an F-score of 0.561.
- The "twin-candidate" approach, using the Simple Expansion Tree and hyper-plane adjustment, further improved the F-score to 0.579, with a significant increase in recall.
- Incorporating negative instances from non-event anaphoric pronouns, using balanced sampling, improved the F-score to 0.549 (compared to 0.472 without them).
Limitations and Future Work
- Corpus Specificity: The authors note that the OntoNotes corpus, particularly broadcast news, can have "nasty and noisy" sentence structures, which might affect the performance of tree-based features.
- Future Work: The authors plan to incorporate more semantic information, such as semantic role labels and verb frames, to further enhance the system's understanding of events.
Summary updated