An Extended Sequence Labeling Approach For Relation Extraction

Abstract

A relation extraction approach based on sequence labeling has been proposed to extract entities and relation triples jointly. That approach does not take triple overlapping into consideration. In this paper, the approach is improved to become more friendly to overlapped triples. First, the sequence labeling model is extended to make it possible to predict more than one tags for a token. And all gold tags of a token are used for supervision. Then a more effective algorithm is designed to construct triples. Experiments on CoNLL04 dataset show that our approach achieves a much better overall performance than our baselines.