Exploring Graph-Algebraic CCG Combinators for Syntactic-Semantic AMR Parsing

Proceedings of Recent Advances in Natural Language Processing, pages 112–121, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_014 112 Exploring Graph-Algebraic CCG Combinators for Syntactic-Semantic AMR Parsing Sebastian Beschke University of Hamburg, Germany beschke@informatik.uni-hamburg.de Abstract We describe a new approach to semantic parsing based on Combinatory Categorial Grammar (CCG). The grammar’s se- mantic construction operators are defined in terms of a graph algebra, which allows our system to induce a compact CCG lex- icon. We introduce an expectation max- imisation algorithm which we use to fil- ter our lexicon down to 2500 lexical tem- plates. Our system achieves a semantic triple (Smatch) precision that is competit- ive with other CCG-based AMR parsing approaches. 1 Introduction Parsing sentences to formal meaning representa- tions, known as Semantic Parsing, is a task at the frontier of Natural Language Understanding. Ab- stract Meaning Representation (AMR) is a mean- ing representation language that represents sen- tence semantics in the form of graphs. Research on AMR parsing systems has been very productive in recent years with many competing approaches. Current AMR parsers vary regarding the extent to which they rely on a formal grammar. Some of the most successful systems generate AMRs through an end-to-end neural architecture, with no explicit symbolic derivations (Zhang et al., 2019). Other parsers employ transition systems with lim- ited explanatory power (Peng et al., 2018). Con- structing grammar-based semantic analyses that can be understood in terms of linguistic theory is a more difficult task than end-to-end AMR parsing because of the additional structural requirements on the output and the algorithmic constraints im- posed thereby. In this paper, we explore how semantic parsers can be built to be interpretable and transparent. Interpretability means that our system exposes rich symbolic information in the form of CCG de- rivations. Transparency means that it works with a compact and intuitively plausible lexicon. The lexicon is itself an artifact that can be inspected. We achieve these goals by equipping CCG with graph-based semantics. Meaning reprentations are constructed through the operations of a simple graph algebra, which effectively constrains the search space for parsing and lexicon induction and makes the available operations and resulting lex- ical items easy to understand. Technical contributions of this paper include a modified expectation-maximisation (EM) al- gorithm to induce compact delexicalised CCG lex- ica, a technique for training a syntactic-semantic supertagger with incomplete labels, and a hybrid update mechanism for training the linear parsing model. 1.1 Related Work This work builds upon the concept of graph- algebraic CCG, which has so far been tested only in the context of lexicon induction (Beschke and Menzel, 2018). We extend the lexicon induction process by delexicalisation and EM filtering and demonstrate the first end-to-end parsing system based on graph-algebraic CCG. The idea of ap- plying graph algebras to AMR parsing has also been applied in the context of Interpreted Regular Tree Grammar (Groschwitz et al., 2018). Further- more, improved definitions of graph-composing CCG combinators have been proposed (Blodgett and Schneider, 2019) to cover a wider range of …

Content truncated to 500 words