A Simple and Effective Model for Answering Multi-span Questions

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 3074–3080, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 3074 A Simple and Effective Model for Answering Multi-span Questions Elad Segal1, Avia Efrat1, Mor Shoham1, Amir Globerson1,3, Jonathan Berant1,2 1Tel Aviv University 2Allen Institute for AI 3Google Research elad.segal@gmail.com, {aviaefra,morshoham}@mail.tau.ac.il amir.globerson@gmail.com, joberant@cs.tau.ac.il Abstract Models for reading comprehension (RC) com- monly restrict their output space to the set of all single contiguous spans from the in- put, in order to alleviate the learning problem and avoid the need for a model that generates text explicitly. However, forcing an answer to be a single span can be restrictive, and some recent datasets also include multi-span ques- tions, i.e., questions whose answer is a set of non-contiguous spans in the text. Naturally, models that return single spans cannot answer these questions. In this work, we propose a simple architecture for answering multi-span questions by casting the task as a sequence tagging problem, namely, predicting for each input token whether it should be part of the output or not. Our model substantially im- proves performance on span extraction ques- tions from DROP and QUOREF by 9.9 and 5.5 EM points respectively. 1 Introduction The task of reading comprehension (RC), where given a question and context, one provides an an- swer, has gained immense attention recently. In most datasets and models (Rajpurkar et al., 2016; Trischler et al., 2016; Seo et al., 2017; Yu et al., 2018; Kwiatkowski et al., 2019), RC is set up as an extractive task, where the answer is constrained to be a single span from the input. This makes learning easier, since the model does not need to generate text abstractively, while still being expres- sive enough to capture a large set of questions. However, for some questions, while the answer is indeed extractive, i.e., contained in the input, it is not a single span. For example, in Figure 1 the answer includes two people who appear as non- contiguous spans in the context. Existing models (Seo et al., 2017; Dua et al., 2019) are by design unable to provide the correct answer to such multi- span questions. B I B I B I I Question: “Who was able to receive over 50% of the vote?” Answer: {“Barack Obama”, “George W. Bush”} Passage: “… In 2012, Barack Obama narrowly won with 48.4% … … In 2008, Barack Obama won the county with 50.5% … Republican George W. Bush carried Clallam twice, defeating John Kerry by 51.3% to 46.3% in …” Figure 1: A multi-span question from DROP, and a BIO tagging for it (O tags omitted). The first occur- rence of Barack Obama does not answer the question. While most work has largely ignored this issue, recent work has taken initial steps towards handling multi-span questions. Hu et al. (2019) proposed to predict the number of output spans for each ques- tion, and used a non-differentiable inference proce- dure to find them in the text, leading to …

Content truncated to 500 words