Is Multi-Hop Reasoning Really Explainable? Towards Benchmarking Reasoning Interpretability

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8899–8911 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 8899 Is Multi-Hop Reasoning Really Explainable? Towards Benchmarking Reasoning Interpretability Xin Lv1,2, Yixin Cao3, Lei Hou1,2∗, Juanzi Li1,2 Zhiyuan Liu1,2, Yichi Zhang4, Zelin Dai4 1Department of Computer Science and Technology, BNRist 2KIRC, Institute for Artificial Intelligence, Tsinghua University, Beijing 100084, China 3Nanyang Technological University, Singapore 4Alibaba Group, Hangzhou, China lv-x18@mails.tsinghua.edu.cn, yixin.cao@ntu.edu.sg {houlei,lijuanzi,liuzy}@tsinghua.edu.cn Abstract Multi-hop reasoning has been widely studied in recent years to obtain more interpretable link prediction. However, we find in exper- iments that many paths given by these mod- els are actually unreasonable, while little work has been done on interpretability evaluation for them. In this paper, we propose a unified framework to quantitatively evaluate the inter- pretability of multi-hop reasoning models so as to advance their development. In specific, we define three metrics, including path recall, local interpretability, and global interpretabil- ity for evaluation, and design an approximate strategy to calculate these metrics using the in- terpretability scores of rules. Furthermore, we manually annotate all possible rules and estab- lish a Benchmark to detect the Interpretability of Multi-hop Reasoning (BIMR). In exper- iments, we verify the effectiveness of our benchmark. Besides, we run nine representa- tive baselines on our benchmark, and the ex- perimental results show that the interpretabil- ity of current multi-hop reasoning models is less satisfactory and is 51.7% lower than the upper bound given by our benchmark. More- over, the rule-based models outperform the multi-hop reasoning models in terms of perfor- mance and interpretability, which points to a direction for future research, i.e., how to bet- ter incorporate rule information into the multi- hop reasoning model. Our codes and datasets can be obtained from https://github. com/THU-KEG/BIMR. 1 Introduction Multi-hop reasoning for knowledge graphs (KGs) has been extensively studied in recent years. It not only infers new knowledge but also provides reasoning paths that can explain the prediction re- sults and make the model trustable. For example, ∗ Corresponding Author Bob Seger Pianist Beethoven Piano occupation P106 occupation P106 instrument P1303 instrument P1303 Triple Query: (Bob Seger, instrument, ? ) Chuck Connors Los Angeles Jack Kemp Natural Causes place of death P20 place of birth P19 manner of death P1196 manner of death P1196 Triple Query: ( Chuck Connors, manner of death, ? ) Figure 1: Illustration of reasoning paths for link predic- tion. Although both paths succeed in finding the cor- rect tail entity, only the upper path is reasonable and the lower one is unreasonable. Figure 1 shows two inferred triples and their rea- soning paths. Conventional KG embedding mod- els (e.g., TransE (Bordes et al., 2013)) implicitly find the target entity piano given the query (Bob Seger, instrument, ?), while multi-hop reasoning models complete the triple and explicitly output the reasoning path (the purple solid arrows). Hence, multi-hop reasoning is expected to be more reliable in real systems, as we can safely add an inferred triple to KG by justifying …

Content truncated to 500 words