Transparent Semantic Parsing with Universal Dependencies Using Graph Transformations

Proceedings of the 29th International Conference on Computational Linguistics, pages 4186–4192 October 12–17, 2022. 4186 Transparent Semantic Parsing with Universal Dependencies using Graph Transformations Wessel Poelman, Rik van Noord, Johan Bos Center for Language and Cognition University of Groningen The Netherlands contact@wesselpoelman.nl {r.i.k.van.noord,johan.bos}@rug.nl Abstract Even though many recent semantic parsers are based on deep learning methods, we should not forget that rule-based alternatives might offer advantages over neural approaches with respect to transparency, portability, and explainabil- ity. Taking advantage of existing off-the-shelf Universal Dependency parsers, we present a method that maps a syntactic dependency tree to a formal meaning representation based on Discourse Representation Theory. Rather than using lambda calculus to manage variable bind- ings, our approach is novel in that it consists of using a series of graph transformations. The re- sulting UD semantic parser shows good perfor- mance for English, German, Italian and Dutch, with F-scores over 75%, outperforming a neu- ral semantic parser for the lower-resourced lan- guages. Unlike neural semantic parsers, our UD semantic parser does not hallucinate out- put, is relatively easy to port to other languages, and is completely transparent. 1 Introduction Semantic parsing is the task of mapping natural language sentences to a formal meaning represen- tation such as Abstract Meaning Representations (Banarescu et al., 2013) or Discourse Representa- tion Structures (Bos et al., 2017). The current trend in this area is strongly geared towards using meth- ods based on deep learning. The best performing parsers use pre-trained language models (van No- ord et al., 2020; Zhou et al., 2021; Bevilacqua et al., 2021; Bai et al., 2022). But a good performance is perhaps not the only thing that matters. A draw- back of neural semantic parsers is that their output lacks explainability: why are the meaning represen- tations composed in the way they are? Moreover, they require vast amounts of training data, and are usually specific for a particular language. In addi- tion, their performance usually decreases for longer input sentences. In other words, it may look like we have made a lot of progress, but viewed from a different perspec- tive, we might actually have made a step back. This is especially so with regards to transparency and in- terpretability of semantic parsers. In this paper we describe a semantic parsing system for Discourse Representation Structures — the formal meaning representations proposed by Discourse Representa- tion Theory (Kamp and Reyle, 1993; Abzianidze et al., 2017) — that is based on Universal Depen- dencies (UD, de Marneffe et al., 2021). The first advantage of the UD framework is that it has been developed for numerous languages (using a cross- linguistically consistent annotation scheme) and that several high-performing parsers have been de- veloped for UD. This will make it easier, as we will show, to develop semantic parsers for languages other than English, in our case German, Italian and Dutch. The second advantage of using UD as input describing the syntactic structure of the sentence, is that it provides us with explainable support …

Content truncated to 500 words