Wiki-ly Supervised Part-of-Speech Tagging

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1389–1398, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Wiki-ly Supervised Part-of-Speech Tagging Shen Li Computer & Information Science University of Pennsylvania shenli@seas.upenn.edu Jo˜ao V. Grac¸a L2F INESC-ID Lisboa, Portugal javg@l2f.inesc-id.pt Ben Taskar Computer & Information Science University of Pennsylvania taskar@cis.upenn.edu Abstract Despite significant recent work, purely unsu- pervised techniques for part-of-speech (POS) tagging have not achieved useful accuracies required by many language processing tasks. Use of parallel text between resource-rich and resource-poor languages is one source of weak supervision that significantly improves accu- racy. However, parallel text is not always available and techniques for using it require multiple complex algorithmic steps. In this paper we show that we can build POS-taggers exceeding state-of-the-art bilingual methods by using simple hidden Markov models and a freely available and naturally growing re- source, the Wiktionary. Across eight lan- guages for which we have labeled data to eval- uate results, we achieve accuracy that signifi- cantly exceeds best unsupervised and parallel text methods. We achieve highest accuracy re- ported for several languages and show that our approach yields better out-of-domain taggers than those trained using fully supervised Penn Treebank. 1 Introduction Part-of-speech categories are elementary building blocks that play an important role in many natu- ral language processing tasks, from machine trans- lation to information extraction. Supervised learn- ing of taggers from POS-annotated training text is a well-studied task, with several methods achieving near-human tagging accuracy (Ratnaparkhi, 1996; Toutanova et al., 2003; Shen et al., 2007). How- ever, while English and a handful of other languages are fortunate enough to have comprehensive POS- annotated corpora such as the Penn Treebank (Mar- cus et al., 1993), most of the world’s languages have no labeled corpora. The annotated corpora that do exist were costly to build (Abeill´e, 2003), and are often not freely available or restricted to research- only use. Furthermore, much of the annotated text is of limited genre, normally focusing on newswire or literary text. Performance of treebank-trained sys- tems degrades significantly when applied to new do- mains (Blitzer et al., 2006). Unsupervised induction of POS taggers offers the possibility of avoiding costly annotation, but de- spite recent progress, the accuracy of unsupervised POS taggers still falls far behind supervised sys- tems, and is not suitable for most applications (Berg- Kirkpatrick et al., 2010; Grac¸a et al., 2011; Lee et al., 2010). Using additional information, in the form of tag dictionaries or parallel text, seems unavoid- able at present. Early work on using tag dictionaries used a labeled corpus to extract all allowed word-tag pairs (Merialdo, 1994), which is quite an unrealis- tic scenario. More recent work has used a subset of the observed word-tag pairs and focused on gener- alizing dictionary entries (Smith and Eisner, 2005; Haghighi and Klein, 2006; Toutanova and Johnson, 2007; Goldwater and Griffiths, 2007). Using corpus- based dictionaries greatly biases the test results, and gives little …

Content truncated to 500 words