Abstract
Event Anaphora Resolution is a critical task for advanced natural language processing, particularly in template extraction and information retrieval. While previous research has primarily focused on event pronoun resolution, this paper presents the first systematic study of resolving event noun phrases (NPs) to their verbal mentions over long distances. Our research identifies that traditional coreference features are largely insufficient for this task. Instead, we propose a model utilizing diverse lexical, syntactic, and positional features. By combining flat features with syntactic structural information via composite kernels and a twin-candidate ranking model, we achieve a significant performance improvement, reaching a 61.36% F-score on the OntoNotes corpus.
The Challenge of Event Coreference
In standard coreference resolution, the goal is to link mentions of the same physical object, such as connecting "President Obama" to the pronoun "he." However, human language frequently uses noun phrases to refer back to events previously described by verbs. For instance, in the sentence "U.S. warplanes hit targets... These air strikes lasted for hours," the noun phrase "air strikes" refers back to the verbal action "hit."
This task is significantly more complex than object coreference for several reasons. First, the triggers involve different parts of speech (verb vs. noun), rendering traditional features like gender, number, or person agreement ineffective. Second, event mentions often span much larger distances in text than object pronouns.
Methodology
To bridge the gap between verbal and nominal mentions, we developed a multi-layered approach:
1. Flat Feature Set
We utilized a suite of lexical and contextual features designed to capture the relationship between verbs and nouns. This included morphology and synonymy analysis using WordNet (e.g., linking "confess" to "confession") and contextual similarity measures to compare the surrounding environments of the noun phrase and the candidate verb.
2. Structural Syntactic Features
Recognizing that grammatical position is a strong indicator of coreference, we employed Convolution Parse Tree Kernels. We evaluated three levels of tree complexity:
- Minimum-Expansion Tree (MET): Focuses strictly on the path connecting the two mentions.
- Simple Expansion Tree (SET): Includes the MET plus immediate child nodes.
- Full Expansion Tree (FET): Encompasses the entire tree structure between mentions.
3. Twin-Candidate Framework
Rather than evaluating candidate verbs in isolation, we implemented a "twin-candidate" model. This framework compares two candidates simultaneously to determine which is a better match for the noun phrase, providing a more robust ranking mechanism than binary classification.
Dataset and Insights
Our study utilized the OntoNotes Release 2.0 English corpus, which provides gold-standard annotations for both object and event coreference. Analysis of this dataset revealed that nearly 20% of entity chains contain at least one verb mention, underscoring the necessity of event-specific resolution systems.
Key findings from our experiments include:
- Syntax is Essential: Integrating syntactic tree information boosted the F-score by over 15% compared to flat features alone.
- Noise in Complexity: The Minimum-Expansion Tree (MET) outperformed the Full Expansion Tree (FET), suggesting that excessive structural detail can introduce noise that hinders model performance.
- Positional Cues: The average distance between an event NP and its verbal antecedent is approximately three sentences. Consequently, positional features are among the most powerful indicators in our model.
Results and Performance
The progression of our model's performance demonstrates the value of each added layer:
- Baseline (Flat Features): 43.35% F-score
- Syntactic Integration (Flat + MET): 59.01% F-score
- Final Model (Twin-Candidate + MET): 61.36% F-score
Limitations and Future Directions
While our results are encouraging, several challenges remain. Distinguishing between noun phrases that refer to events and those that refer to physical objects (e.g., "the break" as an event vs. a physical gap) remains difficult. Furthermore, our learning curve analysis indicates that performance begins to plateau after utilizing 80% of the training data, suggesting that further progress will require more sophisticated semantic features, such as Semantic Role Labels (SRL) or verb frames, rather than simply more data.
Summary updated