Implementation of a Kernel-Based Chinese Relation Extraction System

Entity relation extraction (RE) is an important task in information extraction. In this paper, a novel kernel-based Chinese entity relation extraction system is presented, which appies the improved sequence kernel function with KNN learning algorithm to fulfill the RE task. Experiments are carried out on 3 kinds of relation types and their 6 subtypes defined in the ACE guidelines. Results show that the new approach achieves an average precision up to 88%, significantly higher than feature-based approaches and traditional kernel methods. The new approach has a better generalization capability especially on small training sets. The system consists of 8 independent modules including named entity detection, candidate generation, etc. for easy maintenance and update. The system is implemented either as a Java application or plug-ins on gate platform. It extracts not only the binary relation, but also their description such as job in employment relation.