Semantic Composition with PSHRG for Derivation Tree Reconstruction from Graph-Based Meaning Representations

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5425 - 5439 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Semantic Composition with PSHRG for Derivation Tree Reconstruction from Graph-Based Meaning Representations Chun Hei Lo Wai Lam Hong Cheng Department of Systems Engineering and Engineering Management The Chinese University of Hong Kong {chlo, wlam, hcheng}@se.cuhk.edu.hk Abstract We introduce a data-driven approach to gener- ating derivation trees from meaning represen- tation graphs with probabilistic synchronous hyperedge replacement grammar (PSHRG). SHRG has been used to produce meaning rep- resentation graphs from texts and syntax trees, but little is known about its viability on the reverse. In particular, we experiment on Depen- dency Minimal Recursion Semantics (DMRS) and adapt PSHRG as a formalism that approx- imates the semantic composition of DMRS graphs and simultaneously recovers the deriva- tions that license the DMRS graphs. Consistent results are obtained as evaluated on a collec- tion of annotated corpora. This work reveals the ability of PSHRG in formalizing a syntax– semantics interface, modelling compositional graph-to-tree translations, and channelling ex- plainability to surface realization. 1 Introduction General graph-based meaning representations (MRs) that model sentence-level semantics aim to provide interpretable intermediate representa- tions that are application- and domain-independent (Koller et al., 2019). Recently, graph gram- mars and algebras that formalize semantic con- structions were introduced to MR processing (for example: (Koller, 2015; Drewes and Jonsson, 2017; Groschwitz et al., 2017; Chen et al., 2018; Groschwitz et al., 2018; Lindemann et al., 2019; Donatelli et al., 2019; Chen and Sun, 2020)). These formal grammars bridge between linguistic assump- tions and data-driven parsing, and offer the bene- fit of cross-framework adaptability. For instance, the Apply–Modify algebra was adopted in pars- ing across 5 MR frameworks (Oepen et al., 2019) (Lindemann et al., 2019; Donatelli et al., 2019). Another formalism that was adopted in generating semantic graphs from syntax trees is synchronous hyperedge replacement grammar (SHRG) (Peng et al., 2015; Chen et al., 2018; Chen and Sun, 2020). The use of SHRG in recovering syntax trees from MRs has however received scant research coverage. Empirical results of PSHRG’s application are lim- ited to Jones et al. (2012)’s work in semantic-based machine translation. An immediate application of MR-to-tree pars- ing is surface realization. Previous data-driven ap- proaches to it include rule-based (Flanigan et al., 2016; Song et al., 2017; Horvat, 2017; Ye et al., 2018) and neural methods (Song et al., 2018; Da- monte and Cohen, 2019; Hajdik et al., 2019). All these methods do not generate syntactic analyses. In contrast, the Answer Constraint Engine (ACE; Carroll et al., 1999; Carroll and Oepen, 2005; Vell- dal and Oepen, 2006), an HPSG grammar-based parser, generates both derivations and sentences from Minimal Recursion Semantics (MRS; Copes- take et al., 2005). If we can induce an SHRG from data, MRs can be translated into derivation trees without relying on a hand-engineered grammar, and natural language texts can be obtained by realizing the terminals. Combining the strengths of …

Content truncated to 500 words