Parsing English into Abstract Meaning Representation Using Syntax-Based Machine Translation

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1143–1154, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Parsing English into Abstract Meaning Representation Using Syntax-Based Machine Translation Michael Pust, Ulf Hermjakob, Kevin Knight, Daniel Marcu, Jonathan May Information Sciences Institute Computer Science Department University of Southern California {pust, ulf, knight, marcu, jonmay}@isi.edu Abstract We present a parser for Abstract Meaning Representation (AMR). We treat English- to-AMR conversion within the framework of string-to-tree, syntax-based machine translation (SBMT). To make this work, we transform the AMR structure into a form suitable for the mechanics of SBMT and useful for modeling. We introduce an AMR-specific language model and add data and features drawn from semantic re- sources. Our resulting AMR parser signif- icantly improves upon state-of-the-art re- sults. 1 Introduction Abstract Meaning Representation (AMR) is a compact, readable, whole-sentence semantic an- notation (Banarescu et al., 2013). It includes en- tity identification and typing, PropBank semantic roles (Kingsbury and Palmer, 2002), individual en- tities playing multiple roles, as well as treatments of modality, negation, etc. AMR abstracts in nu- merous ways, e.g., by assigning the same concep- tual structure to fear (v), fear (n), and afraid (adj). Figure 1 gives an example. AMR parsing is a new research problem, with only a few papers published to date (Flani- gan et al., 2014; Wang et al., 2015) and a publicly available corpus of more than 10,000 English/AMR pairs.1 New research problems can be tackled either by developing new algo- rithms/techniques (Flanigan et al., 2014; Wang et al., 2015) or by adapting existing algo- rithms/techniques to the problem at hand. In this paper, we investigate the second approach. The AMR parsing problem bears a strong for- mal resemblance to syntax-based machine transla- tion (SBMT) of the string-to-tree variety, in that 1LDC Catalog number 2014T12 inst fear-01 die-01 ARG0 ARG1 polarity - inst soldier inst ARG1 The soldier was not afraid of dying. The soldier was not afraid to die. The soldier did not fear death. Figure 1: An Abstract Meaning Representation (AMR) with several English renderings. a string is transformed into a nested structure in both cases. Because of this, it is appealing to ap- ply the substantial body of techniques already in- vented for SBMT2 to AMR parsing. By re-using an SBMT inference engine instead of creating cus- tom inference procedures, we lose the ability to embed some task-specific decisions into a custom transformation process, as is done by Flanigan et al. (2014) and Wang et al. (2015). However, we reap the efficiency gains that come from work- ing within a tested, established framework. Fur- thermore, since production-level SBMT systems are widely available, anyone wishing to generate AMR from text need only follow our recipe and retrain an existing framework with relevant data to quickly obtain state-of-the-art results. Since SBMT and AMR parsing are, in fact, distinct tasks, as outlined in Figure 2, to adapt the SBMT parsing framework to AMR parsing, we develop novel representations and techniques. …

Content truncated to 500 words