Reading Comprehension with Graph-based Temporal-Casual Reasoning

Proceedings of the 27th International Conference on Computational Linguistics, pages 806–817 Santa Fe, New Mexico, USA, August 20-26, 2018. 806 Reading Comprehension with Graph-based Temporal-Casual Reasoning Yawei Sun, Gong Cheng, Yuzhong Qu National Key Laboratory for Novel Software Technology, Nanjing University, China ywsun@smail.nju.edu.cn, gcheng@nju.edu.cn, yzqu@nju.edu.cn Abstract Complex questions in reading comprehension tasks require integrating information from multiple sentences. In this work, to answer such questions involving temporal and causal relations, we generate event graphs from text based on dependencies, and rank answers by aligning event graphs. In particular, the alignments are constrained by graph-based reasoning to ensure temporal and causal agreement. Our focused approach self-adaptively complements existing solutions; it is automatically triggered only when applicable. Experiments on RACE and MCTest show that state-of-the-art methods are notably improved by using our approach as an add-on. 1 Introduction The task of reading comprehension has received wide attention from the research community. A ma- chine’s ability to understand natural language is tested by answering multiple-choice questions based on a given passage (Hirschman et al., 1999), as illustrated in Fig. 1. State-of-the-art methods have achieved encouraging results on simple factoid questions, but still, it is a challenge to answer a complex ques- tion that requires extracting information from more than one sentence in the passage. Typical examples are those involving temporal or causal relations spanning multiple sentences (Sugawara et al., 2017; Trischler et al., 2016), like Questions 1 and 2 in Fig. 1. They constitute a considerable proportion of the questions in popular datasets like MCTest (Richardson et al., 2013) and RACE (Lai et al., 2017). To integrate multiple sentences for answering a complex question, various techniques have emerged, but many of them ignore the semantic relations between sentences (Wang et al., 2016) or superficially use a sliding window scanning over the words of the passage without sentence breaks (Smith et al., 2015; Trischler et al., 2016). To meet the challenge, we explicitly model the event structure of the passage by a graph representing temporal and causal relations between events extracted from the passage. Rep- resenting a candidate answer in a similar way, we align the two graphs subject to temporal and causal agreement which is verified by graph-based reasoning, and rank candidate answers by their degrees of alignment. Our approach, called Graph-based Temporal-Casual Reasoning (GTCR) which focuses on temporal and casual questions, can be used as an effective add-on to other methods. Our research contribution in this paper is threefold. • To answer complex questions in reading comprehension tasks, we represent the passage and each candidate answer by an event graph to explicitly model temporal and causal relations, which are crucial to the understanding and integration of multiple sentences for answering questions. We generate event graphs from text based on dependencies produced by an off-the-shelf parser, thereby benefiting from methodological progress and tools on dependency parsing. • To score a candidate answer, we align it with the passage using their event graphs. In an alignment, event matching is constrained by temporal and causal …

Content truncated to 500 words