Abstract
Events and their coreference offer useful semantic and discourse resources. Traditional approaches address event extraction and event coreference resolution either separately or sequentially, limiting their potential interactions. This paper proposes a document-level structured learning model that simultaneously identifies event triggers and resolves event coreference. Using a biology-domain corpus, we demonstrate that this joint model significantly outperforms pipelined models, specifically by 6.9 BLANC F1 and 1.8 CoNLL F1 points in coreference resolution.
Introduction
In Natural Language Processing (NLP), "events" are typically handled in stages. First, a system identifies event triggers—the specific words or phrases that signal an event (e.g., "acquired" in "The company acquired its rival"). Later, a separate system handles coreference resolution, determining if two mentions refer to the same event (e.g., linking "the acquisition" in a later sentence back to the earlier "acquired").
Traditional "pipelined" approaches are fundamentally limited because these two tasks are interdependent. A vague word like "the combination" might be difficult to identify as an event trigger in isolation. However, if a system recognizes that it refers to a previously mentioned "merger," that coreference link provides strong evidence that "combination" is indeed an event. Pipelined models also suffer from "error propagation": if the first stage misses a trigger, the second stage cannot link it.
Methodology: Joint Structured Learning
The researchers propose a joint structured learning model based on the Structured Perceptron. This model captures the natural interactions between trigger identification and coreference resolution by treating the entire document as a single structure.
Key Innovations
- Event Graph Representation: The model represents a document as a graph where nodes are potential event triggers and edges represent coreference links.
- Joint Decoding Algorithm: To navigate the massive search space of possible graphs, the authors implemented a custom incremental decoding algorithm. This combines segment-based decoding (to handle multi-token triggers) with a best-first clustering approach to identify the most likely previous mention for a new trigger.
- Global Decision Making: Unlike a pipeline that makes local "yes/no" decisions for triggers first, this model evaluates coreference features while deciding if a word is a trigger. This allows discourse context to inform semantic identification.
Experimental Setup
The study utilizes the ProcessBank corpus, which consists of 200 paragraphs from the textbook Biology by Campbell and Reece. This domain is particularly challenging due to complex, technical event descriptions. Unlike other common datasets, ProcessBank provides explicit coreference links between individual mentions, which is essential for the "best-first" clustering approach used in this research.
Results and Discussion
The joint model was compared against a strong pipelined baseline that utilized the same features but processed the tasks sequentially.
Coreference Resolution Performance
The joint model demonstrated significant improvements in coreference resolution. As shown in the results table, the model achieved a gain of 6.9 BLANC F1 points and 1.8 CoNLL F1 points over the baseline.
Trigger Identification and Interaction
While the joint model did not drastically improve overall trigger detection (60.21 F1 vs 60.68 F1 for the baseline), it significantly improved the precision of the coreference links it created. The most significant finding is that coreference information helps "rescue" difficult triggers. If a pronoun (like "it") or an abstract noun (like "process") refers to a clear event (like "phosphorylation"), the joint model can correctly identify the vague term as an event.
Furthermore, the joint model effectively reduced false positives. By training on the joint task, the model explores a larger space of potential links, learning to penalize spurious coreference connections more strictly than a pipelined model.
Conclusion
The research highlights the power of interaction between semantic and discourse tasks. Future work aims to incorporate more event components into the joint model, such as event arguments and sub-events, to create a more holistic understanding of document-level discourse. Additionally, a larger corpus would likely allow the joint model to show even greater advantages in trigger identification, as the current ProcessBank dataset is relatively small.
Summary updated