Event Schema Miner with locally contrastive optmization

ACL ARR 2025 February Submission3855 Authors

Abstract:

Event Schema Induction is an important task
in natural language processing (NLP) that aims
to summarize event types and their associated
argument roles from a corpus. However, the
task remains challenging due to several issues:
limited coverage of event element extraction,
ambiguous semantics of event reprensentation,
and insufficient semantic distinctiveness in the
event embedding space. In this paper, we
propose Event Schema Miner (ESM), a novel
framework with locally contrastive optmization for mining event schemas. The framework effectively addresses these challenges
through three key components, each promoting the next: scenario-aware event extraction to
improve the coverage, instruction-driven event
respresentaion to resolve semantic ambiguity,
and target-centric model optimization to refine embedding space. Experimental results
show that ESM surpasses state-of-the-art methods on standard evalution metrics, excelling in
discovering high-quality, high-coverage event
schemas from rather complicated contexts with
severe semantic ambiguity.

Paper Type: Long

Research Area: Machine Learning for NLP

Research Area Keywords: knowledge-augmented methods, representation learning, optimization methods

Contribution Types: NLP engineering experiment

Languages Studied: Chinese, English

Submission Number: 3855