A BERT-Bi-LSTM-Based Knowledge Graph Question Answering Method

Abstract

With the development of knowledge graph, the research of question answering methods based on knowledge graph has gradually become a hot spot. However, in the current mainstream question answering methods, there is insufficient mining of the semantic information of question sentences, resulting in poor entity recognition and relation recognition effects, and many question answering methods rely on predefined rules, which have low transferability and high labor costs. To solve this problem, this paper proposes a knowledge graph question answering method based on BERT word vector, which mainly includes entity recognition and relation recognition. In the entity recognition part, the BERT-Bi-LSTM-CRF model is built, which can fully mine the semantic information contained in the question and improve the accuracy. In the part of relation recognition, the traditional method of semantic relation between multiple entities in the question is transformed into a text classification problem, which can simplify the model complexity and improve the accuracy. Finally, experiments were performed on entity recognition and relation recognition using related data sets. The results show that compared with traditional question answering methods, this method can achieve higher accuracy in entity recognition and relation recognition.