Improving LLM-based Unified Event Relation Extraction via Multiple Answer Questions
Abstract: Extracting event relations that deviate from known schemas has proven challenging for previous methods based on multi-class classification, MASK prediction, or prototype matching. While the LLM-based method can devise diverse instructions to alleviate these issues, it is also accompanied by certain limitations: the need to create a large number of training and inference samples, heightened sensitivity to the sequence of event relation generation, and difficulties in extracting scattered event relations. To tackle these challenges, we present an improved unified event relation extraction framework based on LLM named MAQERE. Firstly, we transform the pair-based extraction issue in LLM-based methods into a multiple answer question problem, which reduces the number of samples required for training and inference. Additionally, by incorporating a bipartite matching loss, we have reduced the dependency of the LLM-based method on the generation sequence. Then, we employ Parse-CoT to extract structured information for enhancing the connections between event mentions. Our experimental results demonstrate that MAQERE can significantly improve the performance of the LLM-based method in the task of event relation extraction.
Paper Type: Long
Research Area: Information Extraction
Research Area Keywords: Information Extraction; event relation extraction
Contribution Types: NLP engineering experiment
Languages Studied: English
Section 2 Permission To Publish Peer Reviewers Content Agreement: Authors grant permission for ACL to publish peer reviewers' content
Submission Number: 74
Loading