Looking Beyond Sentence-Level Natural Language Inference for Question Answering and Text Summarization

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1322–1336 June 6–11, 2021. ©2021 Association for Computational Linguistics 1322 Looking Beyond Sentence-Level Natural Language Inference for Question Answering and Text Summarization Anshuman Mishra, Dhruvesh Patel, Aparna Vijayakumar, Xiang Lorraine Li College of Information and Computer Sciences, University of Massachusetts Amherst Pavan Kapanipathi, Kartik Talamadupula IBM Research Abstract Natural Language Inference (NLI) has gar- nered significant attention in recent years; however, the promise of applying NLI break- throughs to other downstream NLP tasks has remained unfulfilled. In this work, we use the multiple-choice reading comprehension (MCRC) and checking factual correctness of textual summarization (CFCS) tasks to inves- tigate potential reasons for this. Our find- ings show that: (1) the relatively shorter length of premises in traditional NLI datasets is the primary challenge prohibiting usage in downstream applications (which do bet- ter with longer contexts); (2) this challenge can be addressed by automatically converting resource-rich reading comprehension datasets into longer-premise NLI datasets; and (3) mod- els trained on the converted, longer-premise datasets outperform those trained using short- premise traditional NLI datasets on down- stream tasks primarily due to the difference in premise lengths. 1 Introduction Large-scale, open Natural Language Inference (NLI) datasets (Bowman et al., 2015; Williams et al., 2018) have catalyzed the recent development of NLI models that exhibit close to human-level per- formance. However, the use of these NLI models for other downstream Natural Language Process- ing (NLP) tasks has met with limited success. Two of the most popular downstream tasks where NLI models’ use has been explored are Multiple-choice Question Answering (MCRC) and Checking Fac- tual Correctness of Summaries (CFCS) (Trivedi et al., 2019; Falke et al., 2019; Clark et al., 2018) – both of which can easily be cast into the NLI form, as shown in Figure 1. Looking closely at the composition of these datasets, it is evident that there is a stark difference in the lengths of the con- texts/premises when compared to NLI datasets. As seen in Table 1, traditional NLI datasets have much Passage/Premise/Full�text The�first�time�my�father�and I�ever�went�fishing�became a�family�legend.�…�We�were hot,�sticky,�and�mad�that�the fish�refused�to�suck�up�our night�crawlers�.�…�While driving�out�we�saw�a�truck with�a�boat�trailer�and�boat that�was�stuck�in�the�mud. …�my�dad�helped�pull�the man�from�the�mud.�In�return, this�fellow�gave�dad�some fish�…�we�agreed�to�take�in the�fish�as�if�we�had caught�them.�…�As�we�got up�to�do�the�dishes,�mom cleared�her�throat.�"I�just have�one�question�of�you two�great�fishermen,�How was�it�again�that�you�two managed�to�not�only�clean your�fish,�but�also�freeze them�before�you�got home." Hypothesis�(NLI) The�fishing�became�a family�legend�because they�make�themselves�a fool�in�front�of�the�mother. Question�(QA) Why�did�fishing�become a�family�legend? They�make�themselves�a fool�in�front�of�the�mother� Answer Summary�(Summarization) The�father�and�son pretended�catching�a�fish which�was�given�to�them making�a�fool�in�front�of the�mother.�That�is�how the�fishing�trip�became�a family�legend. Figure 1: The tasks of Question Answering and Check- ing Factual Consistency of Text-Summaries can natu- rally be transformed into the Natural Language Infer- ence problem. shorter premises than the context texts from these downstream tasks. Prior research has shown that the capabilities required for handling local infer- ence are very different from those required to per- form inference over longer forms of text (Cooper et al., 1996; Lai et al., 2017a). In this work, we ex- plore this conflict as a major bottleneck in the utility of NLI models (trained on traditional NLI datasets) for downstream NLP tasks. We compare the us- age of long …

Content truncated to 500 words