An Interpretable Reasoning Network for Multi-Relation Question Answering

Proceedings of the 27th International Conference on Computational Linguistics, pages 2010–2022 Santa Fe, New Mexico, USA, August 20-26, 2018. 2010 An Interpretable Reasoning Network for Multi-Relation Question Answering Mantong Zhou Minlie Huang∗ Xiaoyan Zhu State Key Lab. of Intelligent Technology and Systems, National Lab. for Information Science and Technology, Dept. of Computer Science and Technology, Tsinghua University, Beijing, PR China zmt.keke@gmail.com, aihuang@tsinghua.edu.cn, zxy-dcs@tsinghua.edu.cn Abstract Multi-relation Question Answering is a challenging task, due to the requirement of elaborated analysis on questions and reasoning over multiple fact triples in knowledge base. In this paper, we present a novel model called Interpretable Reasoning Network that employs an interpretable, hop-by-hop reasoning process for question answering. The model dynamically decides which part of an input question should be analyzed at each hop; predicts a relation that corresponds to the current parsed results; utilizes the predicted relation to update the question representation and the state of the reasoning process; and then drives the next-hop reasoning. Experiments show that our model yields state-of-the-art results on two datasets. More interestingly, the model can offer traceable and observable intermediate predictions for reasoning analysis and failure diagnosis, thereby allowing manual manipulation in predicting the final answer. 1 Introduction Open-domain Question Answering (QA) has always been a hot topic in AI and this task has recently been facilitated by large-scale Knowledge Bases (KBs) such as Freebase (Bollacker et al., 2008). However, due to the variety and complexity of language and knowledge, open-domain question answering over knowledge bases (KBQA) is still a challenging task. Question answering over knowledge bases falls into two types, namely single-relation QA and multi- relation QA, as argued by Yin et al. (2016). Single-relation questions, such as “How old is Obama?”, can be answered by finding one fact triple in KB, and this task has been widely studied (Bordes et al., 2015; Xu et al., 2016; Savenkov and Agichtein, 2017). In comparison, reasoning over multiple fact triples is required to answer multi-relation questions such as “Name a soccer player who plays at forward position at the club Borussia Dortmund.” where more than one entity and relation are mentioned. Compared to single-relation QA, multi-relation QA is yet to be addressed. Previous studies on QA over knowledge bases can be roughly categorized into two lines: semantic parsing and embedding-based models. Semantic parsing models (Yih et al., 2014; Yih et al., 2016) obtain competitive performance at the cost of hand-crafted features and manual annotations, but lack the ability to generalize to other domains. In contrast, embedding-based models (Bordes et al., 2014b; Hao et al., 2017; Yavuz et al., 2017) can be trained end-to-end with weak supervision, but existing methods are not adequate to handle multi-relation QA due to the lack of reasoning ability. Recent reasoning models (Miller et al., 2016; Wang et al., 2017) mainly concentrate on Reading Comprehension (RC) which requires to answer questions according to a given document. However, transferring existing RC methods to KBQA is not trivial. For one reason, the focus of reasoning in RC is usually …

Content truncated to 500 words