Question Answering by Reasoning Across Documents with Graph Convolutional Networks

Proceedings of NAACL-HLT 2019, pages 2306–2317 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2306 Question Answering by Reasoning Across Documents with Graph Convolutional Networks Nicola De Cao University of Edinburgh University of Amsterdam nicola.decao@uva.nl Wilker Aziz University of Amsterdam w.aziz@uva.nl Ivan Titov University of Edinburgh University of Amsterdam ititov@inf.ed.ac.uk Abstract Most research in reading comprehension has focused on answering questions based on in- dividual documents or even single paragraphs. We introduce a neural model which integrates and reasons relying on information spread within documents and across multiple docu- ments. We frame it as an inference problem on a graph. Mentions of entities are nodes of this graph while edges encode relations between different mentions (e.g., within- and cross- document coreference). Graph convolutional networks (GCNs) are applied to these graphs and trained to perform multi-step reasoning. Our Entity-GCN method is scalable and com- pact, and it achieves state-of-the-art results on a multi-document question answering dataset, WIKIHOP (Welbl et al., 2018). 1 Introduction The long-standing goal of natural language under- standing is the development of systems which can acquire knowledge from text collections. Fresh in- terest in reading comprehension tasks was sparked by the availability of large-scale datasets, such as SQuAD (Rajpurkar et al., 2016) and CNN/Daily Mail (Hermann et al., 2015), enabling end-to-end training of neural models (Seo et al., 2016; Xiong et al., 2016; Shen et al., 2017). These systems, given a text and a question, need to answer the query relying on the given document. Recently, it has been observed that most questions in these datasets do not require reasoning across the doc- ument, but they can be answered relying on in- formation contained in a single sentence (Weis- senborn et al., 2017). The last generation of large-scale reading comprehension datasets, such as a NarrativeQA (Kocisky et al., 2018), Trivi- aQA (Joshi et al., 2017), and RACE (Lai et al., 2017), have been created in such a way as to ad- dress this shortcoming and to ensure that systems query: country Thorildsplan candidates: {Denmark, Finland, Sweden, Italy, ...} answer: Sweden Thorildsplan is a small park in Kristineberg in Stockholm, named in 1925 after the writer [..] Stockholm is the capital of Sweden and the most populous city in [..] Figure 1: A sample from WIKIHOP where multi-step reasoning and information combination from different documents is necessary to infer the correct answer. relying only on local information cannot achieve competitive performance. Even though these new datasets are challeng- ing and require reasoning within documents, many question answering and search applications re- quire aggregation of information across multiple documents. The WIKIHOP dataset (Welbl et al., 2018) was explicitly created to facilitate the devel- opment of systems dealing with these scenarios. Each example in WIKIHOP consists of a collec- tion of documents, a query and a set of candidate answers (Figure 1). Though there is no guaran- tee that a question cannot be answered by relying just on a single sentence, the authors ensure that it …

Content truncated to 500 words