Cross-Lingual Dependency Parsing Using Code-Mixed TreeBank

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 997–1006, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 997 Cross-Lingual Dependency Parsing Using Code-Mixed TreeBank Meishan Zhang1 and Yue Zhang2∗ and Guohong Fu3 1. School of New Media and Communication, Tianjin University, China 2. Institute of Advanced Technology, Westlake Institute for Advanced Study 3. Institute of Artificial Intelligence, Soochow University, China mason.zms@gmail.com, zhangyue@westlake.edu.cn, ghfu@hotmail.com Abstract Treebank translation is a promising method for cross-lingual transfer of syntactic dependency knowledge. The basic idea is to map depen- dency arcs from a source treebank to its tar- get translation according to word alignments. This method, however, can suffer from im- perfect alignment between source and target words. To address this problem, we investigate syntactic transfer by code mixing, translating only confident words in a source treebank. Cross-lingual word embeddings are leveraged for transferring syntactic knowledge to the tar- get from the resulting code-mixed treebank. Experiments on University Dependency Tree- banks show that code-mixed treebanks are more effective than translated treebanks, giv- ing highly competitive performances among cross-lingual parsing methods. 1 Introduction Treebank translation (Tiedemann et al., 2014; Tiedemann, 2015; Tiedemann and Agi´c, 2016) has been considered as a method for cross-lingual syn- tactic transfer. Take dependency grammar for in- stance. Given a source treebank, machine transla- tion is used to find target translations of its sen- tences. Then word alignment is used to find mappings between source and target words, so that source syntactic dependencies can be pro- jected to the target translations. Following, a post- processing step is applied by removing unaligned target words, in order to ensure that the result- ing target syntax forms a valid dependency tree, The method has shown promising performance for unsupervised cross-lingual dependency pars- ing among transfer methods (McDonald et al., 2011; T¨ackstr¨om et al., 2012; Rasooli and Collins, 2015; Guo et al., 2016b). ∗Corresponding author. we are being accused unfairly here vi beskylls med or¨att d¨ar en sv root advmod advmod auxpass aux nsubjpass vi(we) beskylls(accused) or¨att(unfairly) d¨ar(there) root advmod advmod nsubjpass (a) full-scale translation. vi(we) being beskylls(accused) or¨att(unfairly) here sv en sv sv en root advmod advmod auxpass nsubjpass (b) this method, partial translation. Figure 1: An example to illustrate our method, where the source and target languages are English (en) and Swedish (sv), respectively. The treebank translation method, however, suf- fers from various sources of noise. For exam- ple, machine translation errors directly affect the resulting treebank, by introducing ungrammatical word sequences. In addition, the alignments be- tween source and target words may not be iso- morphic due to inherent differences between lan- guages or paraphrasing during translation. For ex- ample, in the case of Figure 1, the English words “are” and “being”, and the Swedish word “med”, do not have corresponding word-level translation. In addition, it can be perfect to express “as soon as they can” using “very quickly” in a translation, which looses word …

Content truncated to 500 words