Cross-lingual CCG Induction

Proceedings of NAACL-HLT 2019, pages 1577–1587 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1577 Cross-lingual CCG Induction Kilian Evang University of Düsseldorf Germany evang@hhu.de Abstract Combinatory categorial grammars are linguis- tically motivated and useful for semantic pars- ing, but costly to acquire in a supervised way and difficult to acquire in an unsupervised way. We propose an alternative making use of cross-lingual learning: an existing source- language parser is used together with a par- allel corpus to induce a grammar and parsing model for a target language. On the PASCAL benchmark, cross-lingual CCG induction out- performs CCG induction from gold-standard POS tags on 3 out of 8 languages, and un- supervised CCG induction on 6 out of 8 lan- guages. We also show that cross-lingually in- duced CCGs reflect known syntactic proper- ties of the target languages. 1 Introduction Combinatory Categorial Grammar (CCG) (Steed- man, 2001) is a grammar formalism known for its linguistic elegance and computational effi- ciency. It has been successfully used for statis- tical syntactic parsing (Clark and Curran, 2004; Lewis et al., 2016) and has emerged as a lead- ing grammar formalism in semantic parsing (Cur- ran et al., 2007; Zettlemoyer and Collins, 2007; Kwiatkowski et al., 2011, 2013; Reddy et al., 2014; Artzi et al., 2015; Beschke and Men- zel, 2018). Semantic parsing is important be- cause it translates natural language utterances to something that a computer can understand, e.g., database queries, computer commands, or logi- cal formulas, enabling next-generation informa- tion systems and knowledge extraction from text, among other applications. CCGs used in most work to date are ei- ther hand-crafted (Zettlemoyer and Collins, 2007; Kwiatkowski et al., 2013; Artzi et al., 2015) or extracted from large syntactically annotated cor- pora (Curran et al., 2007; Reddy et al., 2014). In He NP1 had (S[dcl]2\ NP1)/ NP3 three N4 / N5 sons N5 N4 >0 NP3 ∗ S[dcl]2\ NP1 >0 S[dcl]2 <0 Aveva S[dcl]2/ NP3 tre N4 / N5 figli N5 N4 >0 NP3 ∗ S[dcl]2 >0 Figure 1: Projection of an English CCG derivation to an Italian translation. The indices distinguish different instances of categories. either case language-specific human effort is re- quired. Acquiring CCGs in an unsupervised way is difficult and does not reach the performance of supervised methods (Bisk and Hockenmaier, 2013). As a result, most research focuses on En- glish and other languages are neglected, meaning that speakers of other languages have delayed or no access to CCG-based semantic parsing techno- logy. We propose to overcome this bottleneck by in- ducing CCGs cross-lingually, i.e., transferring an existing grammar from English to other languages via unannotated parallel data. The process is il- lustrated for one English-Italian sentence pair in Figure 1: the English sentence is parsed by an ex- isting CCG parser and word-aligned to the Italian sentence. Italian words receive categories equiva- lent to those of the aligned English words, and a semantically equivalent derivation is built for the Italian sentence. With enough derivations pro- …

Content truncated to 500 words