Parsing Dependency Paths to Identify Event-Argument Relation

Abstract

Mentions of event-argument relations, in particular dependency paths between event-referring words and argument-referring words, can be decomposed into meaningful components arranged in a regular way, such as those indicating the type of relations and the others allowing relations with distant arguments (e.g., coordinate conjunction). We argue that the knowledge about arrangements of such components may provide an opportunity for making event extraction systems more robust to training sets, since unseen patterns would be derived by combining seen components. However, current state-of-the-art machine learning based approaches to event extraction tasks take the notion of components at a shallow level by using n-grams of paths. In this paper, we propose two methods called pseudo-count and Bayesian methods to semi-automatically learn PCFGs by analyzing paths into components from the BioNLP shared task training corpus. Each lexical item in the learned PCFGs appears in 2.6 distinct paths on average between event-referring words and argument-referring words, suggesting that they contain recurring components. We also propose a grounded way of encoding multiple parse trees for a single dependency path into feature vectors in linear classification models. We show that our approach can improve the performance of identifying event-argument relations in a statistically significant manner.