Learning when to trust distant supervision: An application to low-resource POS tagging using cross-lingual projection

Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL), pages 178–186, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Learning when to trust distant supervision: An application to low-resource POS tagging using cross-lingual projection Meng Fang and Trevor Cohn Department of Computing and Information Systems, The University of Melbourne meng.fang@unimelb.edu.au, t.cohn@unimelb.edu.au Abstract Cross lingual projection of linguistic an- notation suffers from many sources of bias and noise, leading to unreliable annota- tions that cannot be used directly. In this paper, we introduce a novel approach to sequence tagging that learns to correct the errors from cross-lingual projection using an explicit debiasing layer. This is framed as joint learning over two corpora, one tagged with gold standard and the other with projected tags. We evaluated with only 1,000 tokens tagged with gold stan- dard tags, along with more plentiful par- allel data. Our system equals or exceeds the state-of-the-art on eight simulated low- resource settings, as well as two real low- resource languages, Malagasy and Kin- yarwanda. 1 Introduction Part-of-speech (POS) tagging is a critical task for natural language processing (NLP) applica- tions, providing lexical syntactic information. Au- tomatic POS tagging has been extremely suc- cessful for many rich resource languages through the use of supervised learning over large train- ing corpora (McCallum et al., 2000; Lafferty et al., 2001; Ammar et al., 2016). However, learn- ing POS taggers for low-resource languages from small amounts of annotated data is very challeng- ing (Garrette and Baldridge, 2013; Duong et al., 2014). For such problems, distant supervision via heuristic methods can provide cheap but inaccu- rately labelled data (Mintz et al., 2009; Takamatsu et al., 2012; Ritter et al., 2013; Plank et al., 2014). A compromise, considered here, is to use a mix- ture of both resources: a small collection of clean annotated data and noisy “distant” data. A popular method for distant supervision is to use parallel data between a low-resource language and a rich-resource language. Although annotated data in low-resource languages is difficult to ob- tain, bilingual resources are more plentiful. For example parallel translations into English are of- ten available, in the form of news reports, nov- els or the Bible. Parallel data allows annotation from a high-resource language to be projected across alignments to the low-resource language, which has been shown to be effective for sev- eral language processing tasks including POS tag- ging (Yarowsky and Ngai, 2001; Das and Petrov, 2011), named entity recognition (Wang and Man- ning, 2014) and dependency parsing (McDonald et al., 2013). Although cross-lingual POS projection is popu- lar it has several problems, including errors from poor word alignments and cross-lingual syntac- tic divergence (T¨ackstr¨om et al., 2013; Das and Petrov, 2011). Previous work has proposed heuris- tics or constraints to clean the projected tag before or during learning. In contrast, we consider com- pensating for these problems explicitly, by learn- ing a bias transformation to encode the mapping between ‘clean’ tags and the kinds of tags pro- duced …

Content truncated to 500 words