Bootstrapping a Crosslingual Semantic Parser

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 499–517 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 499 Bootstrapping a Crosslingual Semantic Parser Tom Sherborne, Yumo Xu and Mirella Lapata Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB {tom.sherborne,yumo.xu}@ed.ac.uk, mlap@inf.ed.ac.uk Abstract Recent progress in semantic parsing scarcely considers languages other than English but professional translation can be prohibitively expensive. We adapt a semantic parser trained on a single language, such as English, to new languages and multiple domains with minimal annotation. We query if machine translation is an adequate substitute for training data, and extend this to investigate bootstrapping using joint training with English, paraphrasing, and multilingual pre-trained models. We develop a Transformer-based parser combining para- phrases by ensembling attention over multiple encoders and present new versions of ATIS and Overnight in German and Chinese for eval- uation. Experimental results indicate that MT can approximate training data in a new lan- guage for accurate parsing when augmented with paraphrasing through multiple MT en- gines. Considering when MT is inadequate, we also find that using our approach achieves parsing accuracy within 2% of complete trans- lation using only 50% of training data.1 1 Introduction Semantic parsing is the task of mapping natural language utterances to machine-interpretable ex- pressions such as SQL or a logical meaning repre- sentation. This has emerged as a key technology for developing natural language interfaces, especially in the context of question answering (Kwiatkowski et al., 2013; Berant et al., 2013; Liang, 2016; Kollar et al., 2018), where a semantically complex ques- tion is translated to an executable query to retrieve an answer, or denotation, from a knowledge base. Sequence-to-sequence neural networks (Sutskever et al., 2014) are a popular approach to semantic parsing, framing the task as sequence transduction from natural to formal languages (Jia and Liang, 2016; Dong and Lapata, 2016). 1Our code and data can be found at github.com/ tomsherborne/bootstrap. Recent proposals include learning intermediate logic representations (Dong and Lapata, 2018; Guo et al., 2019), constrained decoding (Yin and Neubig, 2017; Krishnamurthy et al., 2017; Lin et al., 2019), and graph-based parsing (Bogin et al., 2019; Shaw et al., 2019). Given recent interest in semantic parsing and the data requirements of neural methods, it is unsurpris- ing that many challenging datasets have been re- leased in the past decade (Wang et al., 2015; Zhong et al., 2017; Iyer et al., 2017; Yu et al., 2018, 2019). However, these widely use English as synonymous for natural language. English is neither linguis- tically typical (Dryer and Haspelmath, 2013) nor the most widely spoken language worldwide (Eber- hard et al., 2019), but is presently the lingua franca of both utterances and knowledge bases in seman- tic parsing. Natural language interfaces intended for international deployment must be adaptable to multiple locales beyond prototypes for English. However, it is uneconomical to create brand new datasets for every new language and domain. In this regard, most previous work …

Content truncated to 500 words