QASem Parsing: Text-to-text Modeling of QA-based Semantics

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7742 - 7756 December 7-11, 2022 ©2022 Association for Computational Linguistics QASem Parsing: Text-to-text Modeling of QA-based Semantics Ayal Klein Eran Hirsch Ron Eliav Valentina Pyatkin Avi Caciularu Ido Dagan Computer Science Department, Bar-Ilan University {ayal.s.klein,hirsch.eran,roneliav1,valpyatkin,avi.c33}@gmail.com dagan@cs.biu.ac.il Abstract Various works suggest the appeal of incorpo- rating explicit semantic representations when addressing challenging realistic NLP scenarios. Common approaches offer either comprehen- sive linguistically-based formalisms, like AMR, or alternatively Open-IE, which provides a shal- low and partial representation. More recently, an appealing trend introduces semi-structured natural-language structures as an intermediate meaning-capturing representation, often in the form of questions and answers. In this work, we further promote this line of research by considering three prior QA-based semantic representations. These cover verbal, nominalized and discourse-based predications, regarded here as jointly providing a compre- hensive representation of textual information — termed QASem. To facilitate this perspective, we investigate how to best utilize pre-trained sequence-to-sequence language models, which seem particularly promising for generating rep- resentations that consist of natural language expressions (questions and answers). In par- ticular, we examine and analyze input and out- put linearization strategies, as well as data aug- mentation and multitask learning for a scarce training data setup. Consequently, we release the first unified QASem parsing tool, easily applicable for downstream tasks that can bene- fit from an explicit semi-structured account of information units in text. 1 Introduction A traditional line of research in NLP has been de- voted to designing various kinds of semantic rep- resentations, that aim to explicate textual meaning with a formal, consistent annotation schema. Rep- resentations such as Semantic Role Labeling (SRL; e.g. Baker et al., 1998), Discourse Representation Theory (Kamp et al., 2011) and others (Copestake et al., 2005; Banarescu et al., 2013; Abend and Rappoport, 2013; Oepen et al., 2015; White et al., 2016; Bos et al., 2017) provide applications with an explicit account of semantic relations in a text. Numerous recent works illustrate how leveraging explicit representations facilitate downstream pro- cessing of challenging tasks (Lee and Goldwasser, 2019; Huang and Kurohashi, 2021; Mohamed and Oussalah, 2019; Zhu et al., 2021; Chen and Dur- rett, 2021; Fan et al., 2019). While traditional representations rely on pre-defined schemata or lexica of linguistic classes (e.g. semantic roles), the popular approach of Open Information Ex- traction (OpenIE; Etzioni et al., 2008) aims for more loosely-structured, easily attainable represen- tations, comprised of tuples of natural language fragments. These light-weight structures, however, come with a cost of lacking consistency and com- prehensive coverage, and do not capture deeper semantic information like semantic roles. In a recent trend, which can be seen as an emerging mid-point between full-fledged semantic formalisms and bare-bone textual fragments, re- searches leverage question-answer pairs (QAs) as a representation of textual information (Michael et al., 2018). For example, several works proposed using QAs as an intermediate structure for assess- ing information alignment between texts, e.g. for evaluating summarization quality (Eyal et al., 2019; Gavenavicius, …

Content truncated to 500 words