Zero-Shot Cross-lingual Semantic Parsing

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4134 - 4153 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Zero-Shot Cross-lingual Semantic Parsing Tom Sherborne and Mirella Lapata Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB tom.sherborne@ed.ac.uk, mlap@inf.ed.ac.uk Abstract Recent work in cross-lingual semantic parsing has successfully applied machine translation to localize parsers to new languages. How- ever, these advances assume access to high- quality machine translation systems and word alignment tools. We remove these assump- tions and study cross-lingual semantic pars- ing as a zero-shot problem, without parallel data (i.e., utterance-logical form pairs) for new languages. We propose a multi-task encoder- decoder model to transfer parsing knowledge to additional languages using only English- logical form paired data and in-domain natu- ral language corpora in each new language. Our model encourages language-agnostic en- codings by jointly optimizing for logical-form generation with auxiliary objectives designed for cross-lingual latent representation align- ment. Our parser performs significantly above translation-based baselines and, in some cases, competes with the supervised upper-bound.1 1 Introduction Executable semantic parsing maps a natural lan- guage utterance to a logical form (LF) for execu- tion in some knowledge base to return a denotation. The parsing task renders an utterance as a semanti- cally identical, but machine-interpretable, expres- sion grounded in a denotation. The transduction between natural and formal languages has allowed semantic parsers to become critical infrastructure in building human-computer interfaces for question answering, (Berant et al., 2013; Liang, 2016; Kollar et al., 2018), dialog systems (Artzi and Zettlemoyer, 2011), and robotics (Dukes, 2014). Recent advances in semantic parsing have im- proved accuracy for neural parsers (Jia and Liang, 2016; Dong and Lapata, 2016; Wang et al., 2020a) and examined their generalization capabilities with new dataset challenges (Zhong et al., 2017; Yu 1Our code and data are available at github.com/tomsherborne/zx-parse. List flights from San Francisco to Pittsburgh? SELECT DI STI NCT f l i ght _1. f l i ght _i d FROM? encoder decoder ? ? ? ? ? ? ? ? ? ? ? ? ? ? SELECT SELECT SELECT WHERE * FROM SELECT; ; ; ; latent representation encoder decoder latent representation Figure 1: Accurate cross-lingual semantic parsing re- quires alignment of latent semantic representations across languages. The encoder generates a represen- tation of the English utterance (blue points) to condi- tion upon during decoding. Producing the same logical form from the equivalent Chinese utterance requires a similar encoding. However, without alignment, the rep- resentation may partially match (purple points) or not at all (red points), leading the decoder to generate an inaccurate, ill-formed query. et al., 2018), in addition to considering languages other than English (Duong et al., 2017; inter alia.). Prior work largely assumes that utterance-logical form training data is parallel in all languages (Jie and Lu, 2014), or must be created with human trans- lation (Susanto and Lu, 2017a). This entry barrier to localization for …

Content truncated to 500 words