QA-Align: Representing Cross-Text Content Overlap by Aligning Question-Answer Propositions

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9879–9894 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 9879 QA-Align: Representing Cross-Text Content Overlap by Aligning Question-Answer Propositions Daniela Brook Weiss 1 Paul Roit 1 Ayal Klein 1 Ori Ernst 1 Ido Dagan1 1Computer Science Department, Bar-Ilan University {dani.b.weiss,plroit,ayal.s.klein,oriern}@gmail.com dagan@cs.biu.ac.il Abstract Multi-text applications, such as multi- document summarization, are typically required to model redundancies across related texts. Current methods confronting consolida- tion struggle to fuse overlapping information. In order to explicitly represent content over- lap, we propose to align predicate-argument relations across texts, providing a potential scaffold for information consolidation. We go beyond clustering coreferring mentions, and instead model overlap with respect to re- dundancy at a propositional level, rather than merely detecting shared referents. Our setting exploits QA-SRL, utilizing question-answer pairs to capture predicate-argument relations, facilitating laymen annotation of cross-text alignments. We employ crowd-workers for constructing a dataset of QA-based align- ments, and present a baseline QA alignment model trained over our dataset. Analyses show that our new task is semantically challenging, capturing content overlap beyond lexical similarity and complements cross-document coreference with proposition-level links, offering potential use for downstream tasks. 1 Introduction End-to-end neural methods have become the de- facto standard for natural language understanding models. While these often work well for single document tasks, tasks that consider multiple tex- tual inputs remain more challenging. A key dif- ficulty concerns consolidating information from different, possibly redundant texts, which is crucial for applications such as multi-document summa- rization (MDS), sentence fusion (McKeown et al., 2010; Thadani and McKeown, 2013) or multi-hop question-answering (Welbl et al., 2018; Feldman and El-Yaniv, 2019). Previous works show that MDS methods for example, often just concatenate rather than merge inputs (Lebanoff et al., 2019a), or erroneously consolidate on non-coreferring ele- ments (Lebanoff et al., 2020b). Recognizing such Figure 1: An example of two alignments expressing the same propositions captured by QASRL question- answers. challenges, a few earlier MDS works (Liu et al., 2015; Liao et al., 2018; Shapira et al., 2017) at- tempted at leveraging semantic structures for con- solidation, such as Abstract Meaning Representa- tion (Banarescu et al., 2013) or Open Knowledge Graph (Wities et al., 2017), however, these initial means were found either too fine-grained (AMR) or immature (OKR) for efficient downstream con- solidation. In this paper, we propose that a useful step to- ward effective consolidation is to detect and align minimal propositions that refer to the same informa- tion. Therefore, we choose to capture propositions at the fine-grained level of predicate-argument rela- tions, each representing a single predication involv- ing two sentence elements, along with the seman- tic role that relates them. Identifying these types of alignments at the propositional level would fa- cilitate recognizing information redundancies and salience. Aligning propositions across documents may also prove useful to a variety of cross-text tasks, such as knowledge-base population, Multi- hop QA and cross document event extraction. Consider the example in Figure 1; our align- ments represent that …

Content truncated to 500 words