Abstract: The authors present a novel weakly supervised framework to extract subevent relations (e.g., "voting" as a part of an "election") from text, resulting in the first large-scale subevent knowledge base with 239,000 tuples. This acquired knowledge significantly improves performance on downstream NLP tasks, including discourse parsing and identifying temporal or causal relations between events.
Introduction
Understanding that one event is a component of another—a subevent—is critical for discourse analysis and event-centric applications like summarization or tracking. For example, knowing that "nominating candidates" and "debating" are subevents of an "election" helps a system understand the internal structure and properties of that election.
However, subevent knowledge is notoriously scarce. Existing datasets are small, and subevent relations are rarely stated explicitly in text (people rarely write "X is a subevent of Y"). Furthermore, simple temporal overlap is a noisy signal; a "bombing" might happen during a "festival," but it is not a part of the festival. This research aims to bridge this gap by automatically building a massive, high-quality subevent knowledge base from unlabeled text.
Methodology
The authors propose a two-step "indirect" approach to generate weak supervision, followed by a contextual classifier to expand the knowledge base.
1. Seed Identification & Semantic Pruning
The process begins by extracting candidate pairs using prepositional patterns that indicate temporal containment (e.g., "$e_1$ during $e_2$"). To filter out "spurious" overlaps (like the bombing/festival example), the researchers implement a Definition-Guided Semantic Check. They build a graph where nodes are events and edges represent similarity, incorporating WordNet definitions of parent events as nodes to ensure subevents are semantically compatible with the parent's definition. A graph clustering algorithm (SLPA) then identifies valid clusters.
2. Contextual Classifier (BERT)
The validated seeds are used to find occurrences in the English Gigaword corpus. A BERT-based classifier is trained to recognize the "context" of a subevent relation. Crucially, the model replaces event trigger words with a [MASK] token during training, forcing it to learn the linguistic patterns of the surrounding context rather than just memorizing specific word pairs.
3. Knowledge Base Expansion
The trained classifier is applied to a massive corpus to identify new subevent pairs that did not appear in the initial seed patterns, allowing for significant scale-up.
Results and Knowledge Base Characteristics
The resulting knowledge base consists of 239,000 subevent tuples covering 10,318 unique parent events. Human evaluation of a random sample showed an overall accuracy of 90.1%. The KB demonstrates high coverage, including 73.8% of event types in the ACE dataset and 66.9% in KBP.
Downstream Impact
The acquired knowledge was tested across several downstream tasks with significant success:
- Implicit Discourse Relation Classification: Improved performance on PDTB 2.0 by 2.0 F1 points.
- Temporal and Causal Relations: Significantly boosted performance in identifying cross-sentence relations in datasets like RED, HiEve, and TimeBank. This is particularly notable because local sentence structure often fails to provide sufficient clues for cross-sentence relations, making external "common sense" knowledge essential.
Key Insights
- Temporal Containment $\neq$ Subevent: While subevents are temporally contained by parent events, the inverse is not true. Using dictionary definitions to prune coincidental overlaps is a highly effective way to clean up weak supervision.
- The "Language" of Subevents: By masking event triggers, the authors demonstrate that there is a specific way people describe subevent relationships in context, which can be generalized across different event types.
- Cross-Sentence Utility: Subevent knowledge is most powerful for relations spanning multiple sentences, where contextual clues are usually sparse.
Conclusion
This work provides a scalable method for acquiring subevent knowledge with high precision. By moving beyond simple temporal overlap and leveraging semantic definitions, the framework creates a rich resource that enhances a variety of event-centric NLP tasks. Future work may involve exploring more diverse text sources beyond news corpora to further increase the diversity of the acquired knowledge.
Summary updated