Cross-lingual Learning of an Open-domain Semantic Parser

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 579–588, Osaka, Japan, December 11-17 2016. Cross-lingual Learning of an Open-domain Semantic Parser Kilian Evang University of Groningen The Netherlands k.evang@rug.nl Johan Bos University of Groningen The Netherlands johan.bos@rug.nl Abstract We propose a method for learning semantic CCG parsers by projecting annotations via a parallel corpus. The method opens an avenue towards cheaply creating multilingual semantic parsers mapping open-domain text to formal meaning representations. A first cross-lingually learned Dutch (from English) semantic parser obtains f-scores ranging from 42.99% to 69.22% depend- ing on the level of label informativity taken into account, compared to 58.40% to 78.88% for the underlying source-language system. These are promising numbers compared to state-of-the-art semantic parsing in open domains. 1 Introduction Scarceness of manually annotated corpora for training dependency parsers has led researchers to explore more indirect forms of supervision, such as cross-lingual learning, where annotations in one language are used in training a system for another language. Semantic parsers, which map sentences directly to logically interpretable meaning representations, equally suffer from a lack of annotated training corpora, despite recent efforts like the Groningen Meaning Bank (Basile et al., 2012) or AMR Bank (Banarescu et al., 2013). The lack is especially pronounced for languages other than English. This paper aims to show that cross-lingual learning can help create semantic parsers for new languages with little knowledge about those languages and minimal human intervention. We present a method that takes an existing (source-language) semantic parser and parallel data and learns a semantic parser for the target language. Our method is in principle applicable to all parsers producing interpreted derivations (i.e., parse trees) of Combinatory Categorial Grammar (CCG; Steedman 2001) . It is independent of the concrete meaning representation formalism used, as long as meaning representations are assembled in the standard CCG way using the lambda calculus. We evaluate our method by applying it to English as source language, Dutch as target language and Discourse Representation Theory (DRT; Kamp and Reyle 1993) as meaning representation formalism, and measuring the performance of the obtained Dutch semantic parser. 2 Related Work Semantic parsing has been tackled from a wide variety of angles. Systems that add a semantic interpre- tation component to an existing supervised syntactic parser (Curran et al., 2007; Le and Zuidema, 2012; Lewis and Steedman, 2013) have wide coverage but require much syntactically annotated training data. Other approaches are restricted to relatively narrow linguistic domains but manage to do without strong syntactic supervision. Forms of supervision used include sentence/meaning representation pairs (Wong and Mooney, 2007; Zettlemoyer and Collins, 2007) and even weaker forms of supervision (Clarke et al., 2010; Liang et al., 2011; Kwiatkowski et al., 2013; Goldwasser and Roth, 2011; Chen and Mooney, 2011; Krishnamurthy and Mitchell, 2012; Reddy et al., 2014; Artzi and Zettlemoyer, 2011; Poon, 2013). Only recently have approaches not relying on explicit syntactic supervision successfully been applied to This work is licenced under a Creative Commons Attribution 4.0 International …

Content truncated to 500 words