The Exploration of the Reasoning Capability of BERT in Relation Extraction

EasyChair Preprint 3394

Abstract

Relation classification task is to predict relation between the entity pair in a given sentence. Most of these sentences have certain words or schema that can help to do relation classification. However, we also found there are some sentences do not have such structure, they require model to have certain reasoning capability to predict relation correctly, we call them "reasoning instances". In this paper, we mainly aim to explore the reasoning capability of BERT in these instances. We first propose a BERT-based relation classification model based on MG Lattice model, then we test whether BERT could infer the relation between entities in reasoning instances correctly. Then we explore what kind of information can help BERT to predict relation in these instances. Through various comparison experiment, we conclude that BERT can not infer the relation between entities by the meaning of sentence, it mainly uses the concept information of the entity itself and the information learned on previous instances to help the model to do relation classification. The conclusion inspires us that we can use BERT to predict the relation between entities which defined by multiple sentences in the future.

Keyphrases: BERT, BLSTM, Relation Classaction, Relation Reasoning

@booklet{EasyChair:3394,
  author    = {Lili Li and Xin Xin and Ping Guo},
  title     = {The Exploration of the Reasoning Capability of BERT in Relation Extraction},
  howpublished = {EasyChair Preprint 3394},
  year      = {EasyChair, 2020}}