Abstract
As a paratactic language, sentence-level argument extraction in Chinese suffers much from the frequent occurrence of ellipsis with regard to inter-sentence arguments. To resolve such problem, this paper proposes a novel global argument inference model to explore specific relationships, such as Coreference, Sequence and Parallel, among relevant event mentions to recover those intersentence arguments in the sentence, discourse and document layers which represent the cohesion of an event or a topic. Evaluation on the ACE 2005 Chinese corpus justifies the effectiveness of our global argument inference model over a state-of-the-art baseline.
The Challenge of Ellipsis in Chinese Event Extraction
In Information Extraction (IE), "Event Extraction" involves identifying a trigger and its associated arguments (e.g., participants, location, time). While sentence-local systems perform well for English, Chinese presents a unique challenge due to its paratactic nature. As a "pro-drop" language, Chinese frequently omits arguments mentioned in previous sentences to avoid redundancy.
Analysis of the ACE 2005 Chinese corpus reveals that a significant 20.8% of arguments are inter-sentence, meaning they reside outside the sentence containing the trigger. In contrast, this figure is less than 1% for English. Standard models restricted to single sentences are thus inherently limited in recall.
Global Argument Inference Model
To address this, the authors propose a Global Argument Inference Model that operates across sentence, discourse (three-sentence window), and document layers. The model identifies three primary relationships between event mentions to "bridge" missing arguments:
- Coreference: Mentions referring to the same event share participants.
- Parallel: Morphological structures (e.g., "刺死" - stab to death) where two triggers share subjects/objects.
- Sequence: Logical progressions, such as causal or temporal links between events.
The system utilizes Integer Linear Programming (ILP) to optimize a global objective function, ensuring consistency across these relationships while incorporating Argument Semantics to weight the likelihood of specific entity types filling certain roles.
Experimental Evaluation
The model was evaluated using the ACE 2005 Chinese corpus, consisting of 567 training, 33 development, and 66 testing documents. To isolate the performance of argument extraction, the authors utilized ground-truth entities and values.
Key Results
The global inference model (BIM+RE+AS) demonstrated substantial improvements over state-of-the-art baselines, particularly in recall.
- Argument Identification: F1-score increased from 58.1% to 64.4%, with recall jumping from 57.2% to 66.1%.
- Role Determination: F1-score improved from 53.9% to 58.7%.
- Relational Impact: The Sequence relation provided the most significant individual boost (+1.2% F1), highlighting the power of logical event flow in identifying missing participants.
Limitations and Future Directions
Despite its success, the model faces challenges with pseudo-triggers; when triggers are identified automatically, false positives (25.6%) degrade argument extraction performance. Additionally, the current three-sentence discourse window may miss very long-distance dependencies. Future work will explore cross-document information and deeper semantic modeling to refine precision and expand the scope of inference.
Summary updated