Multilingual Semantic Parsing And Code-Switching

Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 379–389, Vancouver, Canada, August 3 - August 4, 2017. c⃝2017 Association for Computational Linguistics Multilingual Semantic Parsing and Code-switching Long Duong, Hadi Afshar, Dominique Estival, Glen Pink, Philip Cohen, Mark Johnson Voicebox Technologies {longd,hadia,dominiquee,glenp,philipc,markj}@voicebox.com Abstract Extending semantic parsing systems to new domains and languages is a highly ex- pensive, time-consuming process, so mak- ing effective use of existing resources is critical. In this paper, we describe a transfer learning method using crosslin- gual word embeddings in a sequence-to- sequence model. On the NLmaps corpus, our approach achieves state-of-the-art ac- curacy of 85.7% for English. Most im- portantly, we observed a consistent im- provement for German compared with several baseline domain adaptation tech- niques. As a by-product of this approach, our models that are trained on a com- bination of English and German utter- ances perform reasonably well on code- switching utterances which contain a mix- ture of English and German, even though the training data does not contain any code-switching. As far as we know, this is the first study of code-switching in se- mantic parsing. We manually constructed the set of code-switching test utterances for the NLmaps corpus and achieve 78.3% accuracy on this dataset. 1 Introduction Semantic parsing is the task of mapping a natu- ral language query to a logical form (LF) such as Prolog or lambda calculus, which can be ex- ecuted directly through database query (Zettle- moyer and Collins, 2005, 2007; Haas and Riezler, 2016; Kwiatkowksi et al., 2010). Semantic parsing needs application or domain specific training data, so the most straightforward approach is to manufacture training data for each combination of language and application domain. However, acquiring such data is an expensive, lengthy process. This paper investigates ways of leveraging ap- plication domain specific training data in one lan- guage to improve performance and reduce the training data needs for the same application do- main in another language. This is an increasingly common commercially important scenario, where a single application must be developed for multi- ple languages simultaneously. In this paper, we investigate the question of transferring a seman- tic parser from a source language (e.g. English) to a target language (e.g. German). In particular, we examine the situation where there is a large amount of training data for the source language but much less training data for the target language. It is important to note that, despite surface language differences, it has long been suggested that logi- cal forms are the same across languages (Fodor, 1975), motivating transfer learning for this paper. We conceptualize our work as a form of do- main adaptation, where we transfer knowledge about a specific application domain (e.g. naviga- tion queries) from one language to another. Much work has investigated feature-based domain adap- tation (Daume III, 2007; Ben-David et al., 2007). However, it is a non-trivial research question to apply these methods to deep learning. We experiment with several deep learning meth- ods for supervised …

Content truncated to 500 words