It Takes Two to Tango: A Bilingual Unsupervised Approach for Estimating Sense Distributions using Expectation Maximization

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 695–704, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP It Takes Two to Tango: A Bilingual Unsupervised Approach for Estimating Sense Distributions Using Expectation Maximization Mitesh M. Khapra Salil Joshi Pushpak Bhattacharyya Department Of Computer Science and Engineering, IIT Bombay, Powai, Mumbai, 400076. {miteshk,salilj,pb}@cse.iitb.ac.in Abstract Several bilingual WSD algorithms which exploit translation correspondences be- tween parallel corpora have been pro- posed. However, the availability of such parallel corpora itself is a tall task for some of the resource constrained languages of the world. We propose an unsupervised bilingual EM based algorithm which relies on the counts of translations to estimate sense distributions. No parallel or sense annotated corpora are needed. The algo- rithm relies on a synset-aligned bilingual dictionary and in-domain corpora from the two languages. A symmetric general- ized Expectation Maximization formula- tion is used wherein the sense distributions of words in one language are estimated based on the raw counts of the words in the aligned synset in the target language. The overall performance of our algorithm when tested on 4 language-domain pairs is better than current state-of-the-art knowl- edge based and bilingual unsupervised ap- proaches. 1 Introduction Word Sense Disambiguation (WSD) is one of the central and most widely investigated problems in Natural Language Processing (NLP). A wide vari- ety of approaches ranging from supervised to un- supervised algorithms have been proposed. Of these, supervised approaches (Ng and Lee, 1996; Lee et al., 2004) which rely on sense annotated corpora have proven to be more successful, and they clearly outperform knowledge based and un- supervised approaches (Lesk, 1986; Walker and Amsler, 1986; Agirre and Rigau, 1996; Rada, 2005; Agirre and Soroa, 2009; McCarthy et al., 2004). However, creation of sense annotated cor- pora has always remained a costly proposition, es- pecially for some of the resource deprived lan- guages. In this context, “Disambiguation by Transla- tion” is a popular paradigm which tries to obvi- ate the need for sense annotated corpora without compromising on accuracy. Such algorithms rely on the frequently made observation that a word in a given source language tends to have differ- ent translations in a target language depending on its sense. Given a sentence-and-word-aligned par- allel corpus, these different translations in the tar- get language can serve as automatically acquired sense labels for the source word. Although these algorithms (e.g., (Diab and Resnik, 2002; Ng et al., 2003)) give high accuracies, the requirement of a significant amount of bilingual parallel cor- pora may be an unreasonable demand for many language pairs (perhaps more unreasonable than collecting sense annotated corpora itself). Recent work by Khapra et al. (2009) has shown that, within a domain, it is possible to leverage the annotation work done for WSD on one language (L2) for the purpose of another language (L1), by projecting parameters learned from wordnet and sense annotated corpus of L2 to L1. This method does not require a parallel corpus. However, it …

Content truncated to 500 words