Faithful Knowledge Graph Explanations in Commonsense Question Answering

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10811 - 10817 December 7-11, 2022 ©2022 Association for Computational Linguistics Faithful Knowledge Graph Explanations in Commonsense Question Answering Guy Aglionby and Simone Teufel Department of Computer Science and Technology University of Cambridge United Kingdom {guy.aglionby,sht25}@cl.cam.ac.uk Abstract Knowledge graphs are commonly used as sources of information in commonsense ques- tion answering, and can also be used to express explanations for the model’s answer choice. A common way of incorporating facts from the graph is to encode them separately from the question, and then combine the two represen- tations to select an answer. In this paper, we argue that highly faithful graph-based explana- tions cannot be extracted from existing models of this type. Such explanations will not include reasoning done by the transformer encoding the question, so will be incomplete. We confirm this theory with a novel proxy measure for faith- fulness and propose two architecture changes to address the problem. Our findings suggest a path forward for developing architectures for faithful graph-based explanations. 1 Introduction In commonsense question answering, many ap- proaches incorporate knowledge from external re- sources in addition to using large pre-trained lan- guage models. Most often this is done to improve performance, however explanations can also be ex- pressed by highlighting a subset of this information. Making the model output the facts used to answer a particular question can increase trustworthiness and help with debugging. For this kind of explanation to be helpful, it must faithfully represent the model’s reasoning; that is, the explanation must accurately reflect the facts used to answer the question. Faithfulness is in- dependent of whether or not the explanation is a reasonable justification of why an answer was cho- sen (Jacovi and Goldberg, 2020). We refer to how convincing an explanation is for an answer as plau- sibility.1 It is useful to separate the two concepts as a faithful representation of the model’s reasoning 1Terminology for these concepts varies, e.g. faithfulness is also known as trustworthiness. See Jacovi and Goldberg (2020) for discussion. Text Encoder Graph Encoder M L P Figure 1: The architecture of one class of question an- swering models that use knowledge graphs. The orange dashed line is removed in our first ablation. process will be the most useful to a developer who is debugging a model, regardless of whether it is plausible or not. We argue that explanations from a broad class of models are of limited faithfulness (K M et al., 2018; Mihaylov and Frank, 2018; Lin et al., 2019; Feng et al., 2020; Yasunaga et al., 2021; Madaan et al., 2021, illustrated in figure 1). Explanations extracted from the graph encoder (figure 2) are unlikely to reflect the full set of facts used because the text encoder also independently reasons about the question and contributes to answer selection. In addition to our theoretical arguments, we pro- pose a proxy method for empirically measuring faithfulness. Our results confirm that, because per- formance does not …

Content truncated to 500 words