Dynamic Relevance Graph Network for Knowledge-Aware Question Answering

Proceedings of the 29th International Conference on Computational Linguistics, pages 1357–1366 October 12–17, 2022. 1357 Dynamic Relevance Graph Network for Knowledge-Aware Question Answering Chen Zheng Michigan State University zhengc12@msu.edu Parisa Kordjamshidi Michigan State University kordjams@msu.edu Abstract This work investigates the challenge of learn- ing and reasoning for Commonsense Question Answering given an external source of knowl- edge in the form of a knowledge graph (KG). We propose a novel graph neural network ar- chitecture, called Dynamic Relevance Graph Network (DRGN). DRGN operates on a given KG subgraph based on the question and an- swers entities and uses the relevance scores between the nodes to establish new edges dy- namically for learning node representations in the graph network. This explicit usage of rel- evance as graph edges has the following ad- vantages, a) the model can exploit the exist- ing relationships, re-scale the node weights, and influence the way the neighborhood nodes’ representations are aggregated in the KG sub- graph, b) It potentially recovers the missing edges in KG that are needed for reasoning. Moreover, as a byproduct, our model improves handling the negative questions due to con- sidering the relevance between the question node and the graph entities. Our proposed ap- proach shows competitive performance on two QA benchmarks, CommonsenseQA and Open- bookQA, compared to the state-of-the-art pub- lished results. 1 Introduction Solving Question Answering (QA) problems usu- ally requires both language understanding and hu- man commonsense knowledge. Large-scale pre- trained language models (LMs) have achieved suc- cess in many QA benchmarks (Rajpurkar et al., 2016, 2018; Min et al., 2019; Yang et al., 2018). However, LMs have difficulties in predicting the answer when reasoning over external knowledge is required (Yasunaga et al., 2021; Feng et al., 2020). Therefore, using the external sources of knowl- edge explicitly in the form of knowledge graphs (KGs) is a recent trend in question answering mod- els (Lin et al., 2019; Feng et al., 2020). Fig- ure 1, taken from the CommonsenseQA bench- The student practiced his guitar often, where is he always spent his free period? A. music room B. rock band C. toy store D. stage E. concert Music room guitar Playing instrument band concert instrument rock band Q free period UsedFor IsA AtLocation IsA UsedFor AtLocation UsedFor Figure 1: An example of the CommonsenseQA bench- mark. Given the question node Q, question entity nodes (blue boxes), correct answer entity node (red box), and wrong answer entity nodes (orange boxes), we predict the answer by reasoning over the question and the ex- tracted KG subgraph. mark, shows an example for which answering the question requires commonsense reasoning. In this example, the external KG provides the required background information to obtain the reasoning chain from question to answer. We highlight two issues in the previous approaches taken to solve this QA problem: a) the extracted KG subgraph some- times misses some edges between entities, which breaks the chain of reasoning that the current mod- els can not exploit the connections, b) the semantic …

Content truncated to 500 words