A Dynamic Answering Path Based Fusion Model for KGQA

The Knowledge Graph Question Answering (KGQA) task is useful for information retrieval systems, intelligent customer service systems, etc., which has attracted the attention of a large number of researchers. Although the performance of KGQA has been further improved by introducing the Deep Learning models, there are still some difficulties to be solved, such as the representation of questions and answers, the efficient construction way of candidate path set, etc. In this paper, we propose a complete approach for KGQA task. Firstly, we devise a novel candidate path generation process, which effectively improves computation performance by reducing the number of candidate paths corresponding to a question and at the same time guarantees the accuracy of results. Secondly, considering the textual expression diversity of questions and stochastic of candidate paths, we present four models to learn semantic features of Chinese sequence with different focuses. Finally, in order to combine the advantages of each presented model, we propose a dedicated fusion policy which can get the most suitable path from the path set predicted by our presented models. We conduct experiments on Chinese Knowledge Base Question Answering (CKBQA) dataset. Experiment results show that our approach achieves better performance than the best one published in CCKS2019 competition.