QA Reasoning Enhancement Model Based on the Fusion of Dictionary and Hierarchical Directed Graph

In the study of question and answer system, using pre-training and pre-training language model and knowledge graph for joint reasoning still faces two challenges: one is how to effectively solve the problem of knowledge lack in the pre-training stage; the other is how to capture more local evidence after external knowledge enhancement. To address these two pain points, This paper presents a new quiz inference enhancement model——Inference Enhancement Model (InferEM), From the perspective of knowledge enhancement and the interpretability of the reasoning process, First, the dictionary information is integrated into the pre-training through synonym replacement, translation enhancement, Improve the model’s ability to predict low-frequency words; Then, we propose the hierarchical digraph method, Using the hierarchical directed graph (HD-GNN) extracted from the knowledge graph, Query related neighbors’ attention selection strongly related edges capture local information, Enhance the reliability of the evidence chain. In this paper, we evaluate the newly proposed model InferEM model on two dataset benchmarks in the field of common sense quiz, which outperforms the existing single quiz inference model and the existing pre-trained language and knowledge graph joint model.