Concise Answers to Complex Questions: Summarization of Long-form Answers

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 9709–9728 July 9-14, 2023 ©2023 Association for Computational Linguistics Concise Answers to Complex Questions: Summarization of Long-form Answers Abhilash Potluri∗ Fangyuan Xu∗ Eunsol Choi Department of Computer Science The University of Texas at Austin {acpotluri, fangyuan, eunsol}@utexas.edu Abstract Long-form question answering systems provide rich information by presenting paragraph-level answers, often containing optional background or auxiliary information. While such compre- hensive answers are helpful, not all information is required to answer the question (e.g. users with domain knowledge do not need an explana- tion of background). Can we provide a concise version of the answer by summarizing it, while still addressing the question? We conduct a user study on summarized answers generated from state-of-the-art models and our newly proposed extract-and-decontextualize approach. We find a large proportion of long-form answers (over 90%) in the ELI5 domain can be adequately summarized by at least one system, while com- plex and implicit answers are challenging to compress. We observe that decontextualization improves the quality of the extractive summary, exemplifying its potential in the summarization task. To promote future work, we provide an extractive summarization dataset covering 1K long-form answers and our user study annota- tions. Together, we present the first study on summarizing long-form answers, taking a step forward for QA agents that can provide answers at multiple granularities. 1 Introduction Long-form answers (Fan et al., 2019), as com- pared to span-based short answers (Rajpurkar et al., 2016), can provide comprehensive answers to a broader set of questions (Cao and Wang, 2021; Fan et al., 2019). While providing comprehensive infor- mation in multiple sentences is helpful, users often prefer short and concise answers to their questions when possible (Choi et al., 2021). Today’s search engines already present concise answers by high- lighting the most relevant parts from the passage excerpts. In this paper, we present the first study on summarizing long-form answers. ∗Equal contribution. Summarizing long-form answers introduces a new challenge in addition to the faithfulness and fluency challenges of generic summarization which mostly focus on news articles (Nallapati et al., 2016; Narayan et al., 2018): the summary out- put should still provide a reasonable answer to the original question. We take inspiration from a recent study (Xu et al., 2022) that reports that up to 40% of sentences in long-form answers con- tain non-essential information, such as providing background information or examples (Wang et al., 2022), which demonstrates the potential for com- pressing long-form answer. We first aim for an extractive summarization model and collect sentence-level annotations on long-form answers, where annotators identify sen- tences that address the question directly and can serve as the “summary”.1 We collect a dataset cov- ering 1,134 examples, each consisting of a question, a long-form answer, and a set of summary sen- tences. To improve the extractive summaries col- lected, we propose a simple and yet novel summa- rization approach, extract-and-decontextualize, which first extracts summary sentences and re- writes them …

Content truncated to 500 words