Fully-Semantic Parsing and Generation: the BabelNet Meaning Representation

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1727 - 1741 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Fully-Semantic Parsing and Generation: the BabelNet Meaning Representation Abelardo Carlos Martínez Lorenzo1,2, Marco Maru1, and Roberto Navigli1 1Sapienza NLP Group Sapienza University of Rome 2Babelscape, Rome first.lastname@uniroma1.it, surname@babelscape.com Abstract A language-independent representation of meaning is one of the most coveted dreams in Natural Language Understanding. With this goal in mind, several formalisms have been proposed as frameworks for meaning represen- tation in Semantic Parsing. And yet, the de- pendencies these formalisms share with respect to language-specific repositories of knowledge make the objective of closing the gap between high- and low-resourced languages hard to ac- complish. In this paper, we present the Ba- belNet Meaning Representation (BMR), an in- terlingual formalism that abstracts away from language-specific constraints by taking advan- tage of the multilingual semantic resources of BabelNet and VerbAtlas. We describe the ra- tionale behind the creation of BMR and put forward BMR 1.0, a dataset labeled entirely according to the new formalism. Moreover, we show how BMR is able to outperform previous formalisms thanks to its fully-semantic framing, which enables top-notch multilingual parsing and generation. We release the code at https: //github.com/SapienzaNLP/bmr. 1 Introduction Natural Language Understanding (NLU) enables machines to understand human language. A key enabling task in NLU is that of Semantic Pars- ing, whose longed-for dream is that of developing a formalism that can be used as an interlingual representation of meaning, i.e., one that, indepen- dently of the language, can explicitly embed sen- tence meaning into a machine- and human-readable form (Navigli, 2018). To this end, different for- malisms such as Abstract Meaning Representation (Banarescu et al., 2013, AMR), Universal Concep- tual Cognitive Annotation (Abend and Rappoport, 2013, UCCA) and Universal Meaning Represen- tation (Van Gysel et al., 2021, UMR), have been proposed over the years. As of now though, AMR is the most popular formalism for Semantic Parsing, being widely ap- plied to a variety of areas of NLP, such as Machine Translation (Song et al., 2019), Question Answer- ing (Lim et al., 2020; Bonial et al., 2020b; Kapani- pathi et al., 2021), Human-Robot Interaction (Bo- nial et al., 2020a), Text Summarization (Hardy and Vlachos, 2018; Liao et al., 2018) and Information Extraction (Rao et al., 2017). The primary precept of AMR is that different sentences carrying the same meaning should have the same graph representation. Nonetheless, a few inherent properties of AMR make it inappro- priate for the purpose of providing a language- agnostic representation of meaning. In fact, nodes within AMR graphs are represented by means of either English lemmas or OntoNotes frames (Hovy et al., 2006) which, in turn, are based on PropBank (Kingsbury and Palmer, 2002). The issue with lem- mas is that they are merely surface forms devoid of semantics, whereas, with respect to frames, even though analogous repositories exist in other lan- guages such as AnCora for Spanish (Aparicio et …

Content truncated to 500 words