Dependency graphs as a generic interface between parsers and relation extraction rule learning

Abstract

In this paper, we propose to use dependency graphs rather than trees as the interface between a parser and the rule acquisition module of a relation extraction (RE) system. Dependency graphs are much more expressive than trees and can easily be adapted to the output representations of various parsers, in particular those with richer semantics. Our approach is built on top of an existing minimally supervised machine learning system for relation extraction. We extend its original tree-based interface to a graph-based representation. In our experiments, we make use of two different dependency parsers and a deep HPSG parser. As expected, switching to a graph representation for the parsers outputting dependency trees does not have any impact on the RE results. But using the graph-based representation for the extraction with deep HPSG analyses improves both recall and f-score of the RE and enables the system to extract more relation instances of higher arity. Furthermore, we also compare the performance among these parsers with respect to their contribution to the RE task. In general, the robust dependency parsers are good in recall. However, the fine-grained deep syntactic parsing wins when it comes to precision. © 2011 Springer-Verlag.