The Meaning Factory at SemEval-2017 Task 9: Producing AMRs with Neural Semantic Parsing

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 929–933, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics The Meaning Factory at SemEval-2017 Task 9: Producing AMRs with Neural Semantic Parsing Rik van Noord CLCG University of Groningen r.i.k.van.noord@rug.nl Johan Bos CLCG University of Groningen johan.bos@rug.nl Abstract We evaluate a semantic parser based on a character-based sequence-to-sequence model in the context of the SemEval- 2017 shared task on semantic parsing for AMRs. With data augmentation, super characters, and POS-tagging we gain major improvements in performance compared to a baseline character-level model. Although we improve on previ- ous character-based neural semantic pars- ing models, the overall accuracy is still lower than a state-of-the-art AMR parser. An ensemble combining our neural se- mantic parser with an existing, traditional parser, yields a small gain in performance. 1 Introduction Traditional open-domain semantic parsers often use statistical syntactic parsers to derive syntac- tic structure on which to build a meaning repre- sentation. Recently there have been interesting attempts to view semantic parsing as a transla- tion task, mapping a source language (here: En- glish) to a target language (a logical form of some kind). Dong and Lapata (2016) used sequence-to- sequence and sequence-to-tree neural translation models to produce logical forms from sentences, while Barzdins and Gosko (2016) and Peng et al. (2017) used a similar method to produce AMRs. From a purely engineering point of view, these are interesting attempts as complex models of the se- mantic parsing process can be avoided. Yet little is known about the performance and fine-tuning of such parsers, and whether they can reach perfor- mance of traditional semantic parsers, or whether they could contribute to performance in an ensem- ble setting. In the context of SemEval-2017 Task 9 we aim to shed more light on these questions. In particular we participated in Subtask 1, Parsing Biomedical Data, and work with parallel English-AMR train- ing data comprising extracts of scientific articles about cancer pathway discovery. More specifically, our objectives are (1) try to reproduce the results of Barzdins and Gosko (2016), who used character-level models for neu- ral semantic parsing; (2) improve on their re- sults by employing several novel techniques; and (3) combine the resulting neural semantic parser with an existing off-the-shelf AMR parser to reach state-of-the-art results. 2 Neural Semantic Parsing 2.1 Datasets Our training set consists of the second LDC AMR release (LDC2016E25) containing 39,620 AMRs, as well as the training set of the bio AMR corpus that contains 5,452 AMRs. As development and test set we use the designated development and test partition of the bio AMR corpus, both con- taining 500 AMRs. HTML-tags are removed from the sentences. 2.2 Basic Translation Model We use a seq2seq neural translation model to translate English sentences into AMRs. This is a bi-LSTM model with added attention mechanism, as described in Bahdanau et al. (2014). Similar to Barzdins and Gosko (2016), but contrasting with Peng et al. (2017), we train the model only …

Content truncated to 500 words