Universal Decompositional Semantic Parsing

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8427–8439 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 8427 Universal Decompositional Semantic Parsing Elias Stengel-Eskin Johns Hopkins University Aaron Steven White University of Rochester Sheng Zhang Johns Hopkins University Benjamin Van Durme Johns Hopkins University Abstract We introduce a transductive model for pars- ing into Universal Decompositional Semantics (UDS) representations, which jointly learns to map natural language utterances into UDS graph structures and annotate the graph with decompositional semantic attribute scores. We also introduce a strong pipeline model for pars- ing into the UDS graph structure, and show that our transductive parser performs compa- rably while additionally performing attribute prediction. By analyzing the attribute predic- tion errors, we find the model captures natural relationships between attribute groups. 1 Introduction A structured account of compositional meaning has been longstanding goal for both natural language understanding and computational semantics. To this end, a number of efforts have focused on en- coding semantic relationships and attributes in a semantic graph—e.g. Abstract Meaning Represen- tation (AMR; Banarescu et al., 2013), Universal Conceptual Cognitive Annotation (UCCA; Abend and Rappoport, 2013), and Semantic Dependency Parsing (SDP; Oepen et al., 2014, 2015, 2016). In these formalisms, semantic information is typ- ically encoded discretely, using nominal category labels for nodes and edges. This categorical en- coding can make such formalisms brittle when pre- sented with non-prototypical instances, and leads to challenges in coping with changing label ontolo- gies and new datasets (White et al., 2019). Further- more, they are difficult to annotate, often requiring trained linguists and large annotation manuals. The Decompositional Semantics framework presents an alternative to categorical formalisms that encodes semantic information in a feature- based scheme—using continuous scales rather than categorical labels. Starting with a feature-based se- mantic role representation rooted in Dowty 1991’s (1991) proto-role theory (Reisinger et al., 2015; White et al., 2016), this framework has expanded to cover a wide variety of phenomena: event factuality (Rudinger et al., 2018b), genericity (Govindarajan et al., 2019), entity types (White et al., 2016), and temporal relations (Vashishtha et al., 2019). While this rich array of annotation types has been separately modeled, no system yet exists for its joint prediction, which has only recently been made feasible by the introduction of Universal Decompositional Semantics v1.0 (UDS1.0). Pre- sented by White et al. (2019), UDS1.0 normalizes all of these annotations, and incorporates them as node- and edge-level attributes in a single semantic graph whose structure is deterministically extracted from Universal Dependencies (UD; Nivre et al., 2015) syntactic parses via the PredPatt tool (White et al., 2016; Zhang et al., 2017).1 An example graph can be seen in Fig. 1. We present the first joint UDS parser, which learns to extract both graph structures and attributes from natural language input. This parser is a sequence-to-graph transductive model which takes as input a sentence and outputs a UDS graph com- plete with node- and edge-level annotations. In contrast to the traditional semantic parsing …

Content truncated to 500 words