Generate Triggers in Neural Relation Extraction

Liu Yujiang

Keywords: relation triggers，evolutive mask， pointer network

Abstract: In the relation extraction task, the relationship between two entities is determined by some specific words in their source text. These words are called relation triggers, which are the evidence to explain the relationship; other words are called ir-relevant words. The current relationship extraction neural network model aims at identifying the relation type between two entities mentioned in source text by encoding the text and entities. However, these models cannot output the relation triggers, but only gives the result of relation classification. Although models can generate weights for every single word through the improvement of attention mechanism, the weights will be affected by irrelevant words essentially, which are not required by the relation extraction task. In order to output re-lation triggers accurately, we propose a novel training frame-work for Relation Extraction (RE) that reduces the negative effect of irrelevant words on them in the encoding stage. In specific, we leverage Evolutive Mask based Point Network (EMPN) as a decoder to generate relation triggers and encode these words again. For an ordered output in relation triggers, we utilize order loss to constrain the output order in them. Ex-tensive experiment results demonstrate that the effectiveness of our proposed model achieves state-of-the-art performance on three RE benchmark datasets.

One-sentence Summary: Predict relation type and generate trigger words to make the results reasonable with Evolutive Mask based Point Network.