A Joint Model of Entity Recognition and Predicate Mapping for Chinese Knowledge Base Question Answering

Abstract

Knowledge base question answering(KBQA) is the key technology of natural language processing. How to understand the semantic information of the natural language problem and capture the semantic relationship between the problem and the structured triples are the problems that KBQA needs to solve. The boundary of subject entities in Chinese questions is not as clear as English, which increases the difficulty of entity recognition. Besides, the variable Chinese grammar makes predicate mapping more difficult for semantic analysis. Existing KBQA is usually implemented using a pipeline model, which has two disadvantages: (1) Errors caused by entity recognition will be propagated to predicate mapping. (2) Neither entity recognition nor predicate mapping can benefit from the information available to each other. So we propose a BERT-based KBQA to joint entity recognition and predicate mapping tasks that use their dependencies to improve model performance. BERT can solve the semantic ambiguity of the Chinese Q&A databases and improve the accuracy of Chinese Knowledge Base Question Answering(CKBQA). The model achieved an F1 score of 92.04% on the NLPCC 2016 KBQA dataset.