Cross-Lingual Morphological Tagging for Low-Resource Languages

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1954–1964, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Cross-Lingual Morphological Tagging for Low-Resource Languages Jan Buys Department of Computer Science University of Oxford jan.buys@cs.ox.ac.uk Jan A. Botha Google Inc. London jabot@google.com Abstract Morphologically rich languages often lack the annotated linguistic resources required to develop accurate natural language pro- cessing tools. We propose models suitable for training morphological taggers with rich tagsets for low-resource languages without using direct supervision. Our approach extends existing approaches of projecting part-of-speech tags across lan- guages, using bitext to infer constraints on the possible tags for a given word type or token. We propose a tagging model us- ing Wsabie, a discriminative embedding- based model with rank-based learning. In our evaluation on 11 languages, on av- erage this model performs on par with a baseline weakly-supervised HMM, while being more scalable. Multilingual experi- ments show that the method performs best when projecting between related language pairs. Despite the inherently lossy pro- jection, we show that the morphological tags predicted by our models improve the downstream performance of a parser by +0.6 LAS on average. 1 Introduction Morphologically rich languages pose signifi- cant challenges for Natural Language Processing (NLP) due to data-sparseness caused by large vo- cabularies. Intermediate processing is often re- quired to address the limitations of only using sur- face forms, especially for small datasets. Common morphological processing tasks include segmenta- tion (Creutz and Lagus, 2007; Snyder and Barzi- lay, 2008), paradigm learning (Durrett and DeN- ero, 2013; Ahlberg et al., 2015) and morphologi- cal tagging (M¨uller and Schuetze, 2015). In this paper we focus on the latter. Parts-of-speech (POS) tagging is the most com- mon form of syntactic annotation. However, the granularity of POS varies across languages and annotation-schemas, and tagsets have often been extended to include tags for morphologically- marked properties such as number, case or de- gree. To enable cross-lingual learning, a small set of universal (coarse-grained) POS tags have been proposed (Petrov et al., 2012). For morphological processing this can be complemented with a set of attribute-feature values that makes the annotation more fine-grained (Zeman, 2008; Sylak-Glassman et al., 2015b). Tagging text with morphologically-enriched la- bels has been shown to benefit downstream tasks such as parsing (Tsarfaty et al., 2010) and seman- tic role labelling (Hajiˇc et al., 2009). In genera- tion tasks such as machine translation these tags can help to generate the right form of a word and to model agreement (Toutanova et al., 2008). Morphological information can also benefit au- tomatic speech recognition for low-resource lan- guages (Besacier et al., 2014). However, annotating sufficient data to learn ac- curate morphological taggers is expensive and re- lies on linguistic expertise, and is therefore cur- rently only feasible for the world’s most widely- used languages. In this paper we are interested in learning morphological taggers without the avail- ability of supervised data. A successful paradigm for learning without direct supervision is to make use …

Content truncated to 500 words