Normalizing Compositional Structures Across Graphbanks

Proceedings of the 28th International Conference on Computational Linguistics, pages 2991–3006 Barcelona, Spain (Online), December 8-13, 2020 2991 Normalizing Compositional Structures Across Graphbanks Lucia Donatelli, Jonas Groschwitz, Alexander Koller, Matthias Lindemann, Pia Weißenhorn Department of Language Science and Technology Saarland Informatics Campus Saarland University, Germany {donatelli,jonasg,koller,mlinde,piaw}@coli.uni-saarland.de Abstract The emergence of a variety of graph-based meaning representations (MRs) has sparked an impor- tant conversation about how to adequately represent semantic structure. MRs exhibit structural differences that reflect different theoretical and design considerations, presenting challenges to uniform linguistic analysis and cross-framework semantic parsing. Here, we ask the question of which design differences between MRs are meaningful and semantically-rooted, and which are superficial. We present a methodology for normalizing discrepancies between MRs at the compositional level (Lindemann et al., 2019), finding that we can normalize the majority of divergent phenomena using linguistically-grounded rules. Our work significantly increases the match in compositional structure between MRs and improves multi-task learning (MTL) in a low-resource setting, serving as a proof of concept for future broad-scale cross-MR normalization. 1 Introduction Graph-based representations of sentence meaning offer an expressive and flexible means of modeling natural language semantics. In recent years, a number of different graphbanks have annotated large corpora with graph-based semantic representations of various types (Oepen and Lønning, 2006; Ivanova et al., 2012; Banarescu et al., 2013; Abend and Rappoport, 2013). Because of differences in graphbank design principles, individual graphs can differ greatly and often in fundamental strategies (Oepen et al., 2019). Fig. 1 illustrates distinct graphs from the three graphbanks of the SemEval 2015 Shared Task on Semantic Dependency Parsing, which we take as our focus in this paper. The graphs visibly differ with respect to edge labels, edge directions, the treatment of a periphrastic verb construction, and coordination. In this paper, we develop a methodology to (i) understand the nature of the differences across graphbanks and (ii) normalize annotations from different graphbanks at the level of their compositional structures. This approach allows us to identify which design differences between graphbanks are linguistically meaningful and important for the design of future corpora; as well as to identify more unified approaches to certain structures to potentially facilitate multi-task learning (MTL). In Section 2 we build upon Lindemann et al. (2019), who used AM dependency trees to represent the compositional structure of graph-based meaning representations (MRs) based on the AM algebra (Groschwitz et al., 2018). We detail a new methodology for identifying and quantifying mismatches between the compositional structures assigned to the same sentence by different graphbanks; we then present our extended AM+ algebra, with which we can systematically reshape these compositional structures to align them across graphbanks (Section 3). Finally, we provide key examples of how our methodology normalizes specific linguistic phenomena that pose challenges to MR design and differ in representation across graphbanks (Section 4). Using our methods, we increase the match between the AM dependency trees (compositional structure) of the different graphbanks to 76.3 (DM-PAS), 78.8 (DM-PSD), and 82.0 (PAS-PSD) directed unlabeled F-score, compared to …

Content truncated to 500 words