Semantic graph parsing with recurrent neural network DAG grammars

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2769–2778, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 2769 Semantic Graph Parsing with Recurrent Neural Network DAG Grammars Federico Fancellus and Sorcha Gilroyd and Adam LopezL and Mirella LapataL sSamsung AI Centre (SAIC), Toronto, Canada dPeak AI, Manchester, United Kingdom LUniversity of Edinburgh, Edinburgh, United Kingdom Abstract Semantic parses are directed acyclic graphs (DAGs), so semantic parsing should be mod- eled as graph prediction. But predicting graphs presents difficult technical challenges, so it is simpler and more common to pre- dict the linearized graphs found in seman- tic parsing datasets using well-understood se- quence models. The cost of this simplicity is that the predicted strings may not be well- formed graphs. We present recurrent neural network DAG grammars, a graph-aware se- quence model that ensures only well-formed graphs while sidestepping many difficulties in graph prediction. We test our model on the Parallel Meaning Bank—a multilingual se- mantic graphbank. Our approach yields com- petitive results in English and establishes the first results for German, Italian and Dutch. 1 Introduction Semantic parsing is the task of mapping natural language to machine interpretable meaning rep- resentations, which in turn can be expressed in many different formalisms, including lambda cal- culus (Montague, 1973), dependency-based com- positional semantics (Liang et al., 2011), frame se- mantics (Baker et al., 1998), abstract meaning rep- resentations (AMR; Banarescu et al. 2013), mini- mal recursion semantics (MRS; Copestake et al. 2005), and discourse representation theory (DRT; Kamp 1981). Explicitly or implicitly, a representation in any of these formalisms can be expressed as a directed acyclic graph (DAG). Consider the sentence “Ev- ery ship in the dock needs a big anchor”. Its meaning representation, expressed as a Discourse Representation Structure (DRS, Kamp 1981), is shown in Figure 1.1 A DRS is drawn as a box with 1For simplicity, our examples do not show time represen- tations, though these are consistently present in our data. x2 b4 dock(x2) b1 x1 b2 ship(x1) PARTOF(x1, x2) ⇒ e1, s1, x3 b3 need(e1) PIVOT(e1, x1) THEME(e1,x3) anchor(x3) big(s1) TOPIC(s1, x3) Figure 1: The discourse representation structure for “Every ship in the dock needs a big anchor”. For ease of reference in later figures, each box includes a vari- able corresponding to the box itself, at top right in gray. two parts: the top part lists variables for discourse referents (e.g. x1, e1) and the bottom part can con- tain unary predicates expressing the type of a vari- able (e.g. ship, need), binary predicates specify- ing relationships between variables (e.g. PARTOF, TOPIC), logical operators expressing relationships between nested boxes (e.g. ⇒, ¬), or binary dis- course relations (e.g., RESULT, CONTRAST). To express a DRS as a graph, we represent each box as a node labeled □; each variable as a node la- beled by its associated unary predicate; and each binary predicate, logical operator, or discourse re- lation …

Content truncated to 500 words