Abstract:Event Detection (ED) is a crucial information extraction task that aims to identify the event triggers and classify them into predefined event types. However, most existing methods did not perform well when processing events with implicit triggers. And most methods considered ED as a sentence-level task, lacking effective context for event semantics. Moreover, how to maintain good performance under low resource conditions still needs further study. To address these problems, we propose a novel end-to-end ED model called DE3TC, which D etects E vents with E ffective E vent T ype Information and C ontext. We construct an event type-specific Clue to capture the interaction between event type name and trigger words, providing event type information for implicit triggers. For accessing the effective context of event semantics for sentence-level ED, we consider the correlations between types and select similar types' descriptions as context. With contextualized representation from a contextual encoder, DE3TC learns the event type information for all events including implicit ones. And it performs sentence-level ED efficiently with effective contexts. The empirical results on ACE 2005 and MAVEN datasets show that: (i) DE3TC obtains state-of-the-art performance compared with previous methods. (ii) DE3TC is also excelled under low-resource conditions.
What problem does this paper attempt to address?
This paper attempts to solve the following three main problems:
1. **How to efficiently detect events with implicit trigger words?**
- In real - life situations, events in texts are often implicit, that is, trigger words do not clearly convey the semantics of the events. For example, in the sentence "Stewart's marriage to Alana Hamilton in 1979 lasted for 5 years and they had 2 children", "marriage" is obviously the trigger word for the "Life.Marry" event. However, in another example, "deployed" is the trigger word for the "Movement.Transport" event, which is not easy to identify. Therefore, event - type information is required to identify these implicit trigger words. Most existing methods perform poorly in dealing with this situation, while the method proposed in this paper effectively identifies implicit trigger words by capturing the interaction between event - type names and trigger words and providing event - type information.
2. **How to obtain effective context for sentence - level event detection?**
- Context is crucial for semantic understanding, but sentence - level event - detection tasks often lack effective context. It is difficult to identify trigger words solely based on the semantics of a single sentence. For example, in Example 2 in Figure 1, it is difficult to identify "deployed" as the trigger word for the "Movement.Transport" event based on the sentence itself. Moreover, most existing methods ignore the associations between similar event types, and these associations provide valuable context information. The method proposed in this paper solves this problem by selecting descriptions of similar event types as context.
3. **How to achieve good performance with limited resources?**
- Most existing event - detection methods rely on a large amount of high - quality annotated data. When the training data is insufficient, the performance of these methods will decline significantly. In practical applications, obtaining high - quality annotated data is very expensive. Therefore, how to design an event - detection model that can still perform well with a small amount of annotated data has become a key challenge. The DE3TC model proposed in this paper can achieve good performance with limited resources by using event - type information and effective context.
To solve these problems, the author proposes a new end - to - end event - detection model named DE3TC, which is implemented in the following ways:
- **Event Type Information Constructor**: Build clues for each event type, capture the interaction between event types and trigger words, and provide event - type information.
- **Context Selector**: Select descriptions of similar event types as context and provide effective event - semantic context.
- **Contextual Encoder**: Use a pre - trained BERT encoder to encode the event - semantic modeling sequence and generate context representations.
Experimental results show that DE3TC achieves state - of - the - art performance on the ACE 2005 and CASIE datasets, especially performing excellently with limited resources.