Relational Facts Extraction with Splitting Mechanism

Abstract

Relational fact extraction is aimed to extract triples from sentences. Recent years, Sequence-to-sequence learning has been utilized for this task because of its advantage of modeling three different entity overlapping types. In their model, they utilized the same RNN cell to decode entities and relation in a triplet. Actually, the information required to predict entities and relation are different. So we shouldn’t mix the process of extracting entities from the original sentence and predicting the relation between entities. Based on the above observation, we propose a novel mechanism to split the process of decoding entities and relation. We conduct extensive experiments on NYT and WebNLG datasets. The experimental results show that our Splitting-Mechainsm (SM) helps to promote performance.