DRLK: Dynamic Hierarchical Reasoning with Language Model and Knowledge Graph for Question Answering

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5123 - 5133 December 7-11, 2022 ©2022 Association for Computational Linguistics DRLK: Dynamic Hierarchical Reasoning with Language Model and Knowledge Graph for Question Answering Miao Zhang1,2,3, Rufeng Dai2,3,4, Ming Dong2,3,4∗, Tingting He2,3,4∗ 1National Engineering Research Center for E-Learning, 2Hubei Provincial Key Laboratory of Artificial Intelligence and Smart Learning, 3National Language Resources Monitoring and Research Center for Network Media, 4School of Computer, Central China Normal University, Wuhan, China {zmzhangmiao,dairufeng}@mails.ccnu.edu.cn {dongming,tthe}@ccnu.edu.cn Abstract In recent years, Graph Neural Network (GNN) approaches with enhanced knowledge graphs (KG) perform well in question answering (QA) tasks. One critical challenge is how to ef- fectively utilize interactions between the QA context and KG. However, existing work only adopts the identical QA context representa- tion to interact with multiple layers of KG, which results in a restricted interaction. In this paper, we propose DRLK (Dynamic Hi- erarchical Reasoning with Language Model and Knowledge Graphs), a novel model that utilizes dynamic hierarchical interactions be- tween the QA context and KG for reasoning. DRLK extracts dynamic hierarchical features in the QA context, and performs inter-layer and intra-layer interactions on each iteration, allowing the KG representation to be grounded with the hierarchical features of the QA con- text. We conduct extensive experiments on four benchmark datasets in medical QA and com- monsense reasoning. The experimental results demonstrate that DRLK achieves state-of-the- art performances on two benchmark datasets and performs competitively on the others1. 1 Introduction Question answering (QA) system is a hot research area in natural language processing, requiring the robot to clearly understand the scenario described in the question and then reason with relevant do- main knowledge (Jin et al., 2022). Recently, large- scale pre-trained language models (LMs) (Gu et al., 2022; Liu et al., 2021) have become a popular solution in several QA datasets (Mutabazi et al., 2021), achieving excellent performance. By train- ing on an ultra large-scale corpus, LMs learn the la- tent domain knowledge and perform well in down- stream tasks through fine-tuning (Lewis et al., 2020; ∗Corresponding author. 1Our code is available at https://github.com/MZ- MiaoZhang/DRLK KG Q A man presents it n rashes on face and also complains of decreased mental function. He is also having few macular lesions on his skin. On CT scan, intracranial calcification was seen. His 6-year old son is also having similar skin lesions. What would be the most likely diagnosis? A a) Neurofibrornatosis-1 b) Neurofibromatosis-2 c) Xeroderma pigmentosum d) Autosomal dominant inheritance Macular lesions Cerebral calcification Autosomal dominant inheritance Familial ganglia calcification Autosomal recessive Intracranial calcification Xeroderma pigmentosum Haw river syndrome Figure 1: An example from the MedMCQA dataset with KG, where the correct answer is indicated in bold. Chakraborty et al., 2020). However, fine-tuned LMs perform poorly when downstream tasks in- volve complex reasoning or require explicit knowl- edge. The fine-tuned approach relies on similar task patterns and sample forms, while black-box models result in uninterpretable behavior (McCoy et al., 2019). One research topic is to …

Content truncated to 500 words