Frustratingly Easy Cross-Lingual Transfer for Transition-Based Dependency Parsing

Proceedings of NAACL-HLT 2016, pages 1058–1063, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Frustratingly Easy Cross-Lingual Transfer for Transition-Based Dependency Parsing Oph´elie Lacroix1, Lauriane Aufrant1,2, Guillaume Wisniewski1 and Franc¸ois Yvon1 1LIMSI, CNRS, Univ. Paris-Sud, Universit´e Paris-Saclay, F-91405 Orsay 2DGA, 60 boulevard du G´en´eral Martial Valin, F-75509 Paris {ophelie.lacroix, lauriane.aufrant, guillaume.wisniewski, francois.yvon}@limsi.fr Abstract In this paper, we present a straightforward strategy for transferring dependency parsers across languages. The proposed method learns a parser from partially annotated data obtained through the projection of annotations across unambiguous word alignments. It does not rely on any modeling of the reliability of dependency and/or alignment links and is therefore easy to implement and parameter free. Experiments on six languages show that our method is at par with recent algorithmi- cally demanding methods, at a much cheaper computational cost. It can thus serve as a fair baseline for transferring dependencies across languages with the use of parallel corpora. 1 Introduction Cross-lingual learning techniques enable to transfer useful supervision information from well-resourced to under-resourced languages, helping the develop- ment of NLP tools for a large number of languages. In this work, we present a simple method for trans- ferring dependency parsers between languages. Two main strategies have been considered to transfer syntactic annotations: (a) direct model transfer and (b) annotation transfer. The first ap- proach assumes a common representation between the source and target languages (e.g. at the level of PoS tags), which enables to train a model on source data and to use it to parse target sentences. The per- formance of ‘pure’ delexicalized dependency trans- fer can be significantly improved using additional techniques such as self-training (Zeman and Resnik, 2008), smart data selection (Søgaard, 2011), relex- icalization and/or multi-source model transfer (Co- hen et al., 2011; Naseem et al., 2012; T¨ackstr¨om et al., 2013). The second approach (transfer of anno- tations) requires parallel sentences, in which word alignments are used to infer target syntactic struc- tures from source dependencies. The main difficulty here is to cope with cases of non-isomorphism be- tween the source and target structures as well as with the noise in source annotations and in alignments. Turning source trees into target trees indeed may require to filter poor alignments and to apply vari- ous heuristic transformation rules, such as the ones introduced in Hwa et al. (2005), later improved in Tiedemann (2014). In this study, we consider a simple, yet effective approach to transfer annotations, which entirely dispenses from the transfer rules of Hwa et al. (2005), the sharp filtering of partially annotated trees (Tiedemann, 2014), the inclusion of fake root dependencies for unattached words (Spreyer and Kuhn, 2009), or the multi-step process of Rasooli and Collins (2015). Our proposal is, in fact, quite as straightforward (apart from the use of parallel texts) as the delexicalized transfer method of McDonald et al. (2013) while achieving performances that surpass this state-of-the-art method by a wide margin, and competing with recent algorithmically costly methods: it globally outperforms the scores of …

Content truncated to 500 words