Loan Question Answering Platform Based on ERNIE and Knowledge Graph

Abstract

At present, the excessive amount of loan consultation has brought great pressure to manual customer service. However, the existing loan question answering (QA) platforms cannot solve this problem well because of their poor understanding ability. Therefore, the authors constructs a loan QA platform based on ERNIE and knowledge graph (KG). Firstly, they use semi-automatic methods to construct KG with data from a loan company. Secondly, they use token-level random mask strategy (TRM), word-level fixed mask strategy (WFM), and fine-tuning strategy integrating knowledge (IK) to train ERNIE. Finally, they construct a QA platform based on KG and trained ERNIE and experiment with proprietary datasets. The results show that ERNIE trained after three strategies achieve average improvements of 14.7% on judging intention similarity of sentence pairs and 14.28% on retrieving the most similar intention problem compared with the baseline. It also shows that their platform achieves an average improvement of 13% on question answering compared with the customer service app of the loan company.