In natural language processing, Event Detection (ED) is the task of identifying "triggers"—the specific words or phrases that anchor an event—and classifying them into types. However, language is rarely straightforward. A single word like "fire" can describe a weapon discharge, a professional termination, or a literal blaze.
This phenomenon, known as trigger word polysemy, represents a fundamental hurdle for information extraction. As Heng Ji et al. (2008) note, analyzing sentences in isolation often leads to local errors because the "big picture" is missing. To solve this, researchers have moved away from simple keyword matching toward models that treat context as the primary signal for disambiguation.
The Limits of the Trigger Word
Traditional systems often suffer from what Jiawei Chen et al. (2021) call the "trigger curse." Models tend to over-rely on specific keywords, memorizing that "attack" usually means a Conflict event. This leads to failure when the word is used metaphorically or in a different domain. For example, in cybersecurity, common terms often fail to represent actionable events, leading to high false-positive rates (Trong et al. 2020).
Jian Liu et al. (2022) distinguish between two types of events:
- Trigger-dependent events: Anchored by concrete verbs with high semantic specificity.
- Context-dependent events: Anchored by "light verbs" (like take or go) or abstract terms that require external evidence to be understood.
When a trigger is semantically vague, the model must look elsewhere to resolve the ambiguity.
Disambiguation through Arguments
The most immediate source of clarity is the event's arguments—the entities (who, what, where) involved in the action. Shulin Liu et al. (2017) demonstrated that by using supervised attention to force a model to "pay attention" to participants, the system can distinguish between a professional termination and physical violence.
The identity of the entity itself provides a statistical prior for the event type. Yu Hong et al. (2011) show that entity subtypes are powerful discriminators; for instance, the verb "leave" signifies a different event when the subject is a "presidential candidate" versus a "software company." By conditioning the trigger's meaning on its participants, the authors effectively resolve lexical ambiguity that local sentence features alone cannot handle.
Leveraging Global and Cross-Document Context
Sometimes, even the immediate sentence is not enough. A word's sense might only become clear when looking at the entire document or even a cluster of related articles.
- Document-Level Reasoning: Yue Zhao et al. (2018) developed DEEB-RNN, which uses a hierarchical structure to create a document-level summary. This "global" view helps the model understand the overarching theme (e.g., a financial report vs. a crime story), which in turn clarifies ambiguous triggers within individual sentences.
- Cross-Sentence Dependencies: Events rarely happen in isolation. Dongfang Lou et al. (2021) argue that word sense ambiguity is best resolved by looking at neighboring sentences, as the necessary context often resides outside the immediate clause.
- Cross-Document Consistency: Heng Ji et al. (2008) leverage the fact that related news articles tend to describe the same events consistently. If one article clearly identifies an "attack," that information can be used to fix an ambiguous label in a related, more vague article.
Structural and Causal Solutions
Recent work has moved toward more radical reformulations of the problem to bypass trigger fragility entirely. Jianye Xie et al. (2021) treat event detection as a graph parsing problem rather than a classification task. By generating event graphs, the model captures correlations between multiple events, reducing the reliance on any single, potentially ambiguous anchor word.
To combat the "reasoning shortcuts" models take, the authors of a 2020 EMNLP study proposed Context-Selective Mask Generalization. By masking the trigger word during training, the model is forced to learn the event type solely from the surrounding context. This ensures that even if a trigger is polysemous, the model has developed the "muscles" to look at the environment to find the truth.
Ultimately, these diverse approaches converge on a single intuition: in natural language, the meaning of an action is not contained within the word itself, but in the web of relationships—entities, themes, and sequences—that surround it.