DRS Parsing as Sequence Labeling

Proceedings of the 11th Joint Conference on Lexical and Computational Semantics, pages 213 - 225 July 14-15, 2022 ©2022 Association for Computational Linguistics DRS Parsing as Sequence Labeling Minxing Shen Heinrich Heine University Düsseldorf Universitätsstraße 1 40225 Düsseldorf, Germany minxing.shen@hhu.de Kilian Evang Heinrich Heine University Düsseldorf Universitätsstraße 1 40225 Düsseldorf, Germany kilian.evang@hhu.de Abstract We present the first fully trainable semantic parser for English, German, Italian, and Dutch discourse representation structures (DRSs) that is competitive in accuracy with recent sequence-to-sequence models and at the same time compositional in the sense that the out- put maps each token to one of a finite set of meaning fragments, and the meaning of the ut- terance is a function of the meanings of its parts. We argue that this property makes the system more transparent and more useful for human- in-the-loop annotation. We achieve this simply by casting DRS parsing as a sequence label- ing task, where tokens are labeled with both fragments (lists of abstracted clauses with rela- tive referent indices indicating unification) and symbols like word senses or names. We give a comprehensive error analysis that highlights areas for future work.1 1 Introduction Semantic parsing is the task of mapping natural- language sentences to symbolic representations of their meaning. Although most current natural language understanding (NLU) applications are handled by end-to-end systems that solve specific tasks (such as machine translation, conversation, or sentiment analysis) without intermediate symbolic meaning representations, semantic parsing contin- ues to attract research interest for good reasons: first, next-generation NLU systems may become more accurate and certainly more easily explain- able and debuggable by combining symbolic rep- resentations with end-to-end techniques. Second, symbolic meaning representations are amenable to symbolic reasoning, which may be instrumental in enabling, e.g., digital assistants to solve more complex tasks. Third, better and more transparent computational models of text-meaning mapping 1Our system is available at https://github.com/ ShenMinX/DRS-parser can be a useful tool for semantics, i.e., to under- stand how natural-language semantics works. In recent years, most work on annotating natural- language text with comprehensive, broad-coverage meaning representations has been performed in three frameworks: Abstract Meaning Represen- tations (Banarescu et al., 2013), Universal Cogni- tive Conceptual Annotation (Abend and Rappoport, 2013), and Discourse Representation Structures (Abzianidze et al., 2017). Accurate parsers exist for all three (e.g., Lindemann et al., 2020; Oepen et al., 2020; van Noord et al., 2020). Each for- malism has its specific strength: AMRs go very far in abstracting away from surface variation in how a certain meaning is expressed, UCCA has a clear mapping between form and meaning and a modular architecture, and DRSs ground natural language meaning in first-order logic, by explicitly representing the scopes of negation, quantification, disjunction, etc. In this paper, we focus on parsing to DRSs. State-of-the-art DRS parsers follow the encoder- decoder paradigm pioneered for machine transla- tion by Sutskever et al. (2014): the input sequence is encoded by a neural network into a vector, then another network predicts the output sequence (or in this case: output …

Content truncated to 500 words