Semi-Supervised Learning for Word Sense Disambiguation: Quality vs. Quantity

International Conference RANLP 2009 - Borovets, Bulgaria, pages 197–202 Semi-Supervised Learning for Word Sense Disambiguation: Quality vs. Quantity Sandra K¨ubler Indiana University skuebler@indiana.edu Desislava Zhekova University of Bremen zhekova@uni-bremen.de Abstract In this paper, we discuss the importance of the quality against the quantity of automatically ex- tracted examples for word sense disambiguation (WSD). We first show that we can build a com- petitive WSD system with a memory-based clas- sifier and a feature set reduced to easily and efficiently computable features. We then show that adding automatically annotated examples improves the performance of this system when the examples are carefully selected based on their quality. Keywords word sense disambiguation, memory-based learning, semi-supervised learning 1 Introduction Word sense disambiguation (WSD) [7] is concerned with lexical ambiguity: It is the task of automatically assigning the correct meaning to occurrences of poly- semous words in their context. Thus, word sense dis- ambiguation is a necessary component for many appli- cations in Computational Linguistics, such as machine translation, information retrieval, information extrac- tion, or question answering. However, most of the WSD approaches reported so far rely on supervised learning techniques, relying on the data sets provided by the SensEval1 and SemEval2 competitions. This makes them susceptible for well known problems in supervised machine learning such as data sparseness or lacking domain independence. In the present paper, we will investigate a semi- supervised approach to WSD. Semi-supervised learn- ing starts with a supervised learner trained on avail- able data. In a second step, data are added from au- tomatically annotated sources. Semi-supervised ap- proaches, especially when they do not optimize for individual words, often result in no or minimal im- provements. Gonzalo and Verdejo [5] show that the good results for individual words cannot compare to supervised results. This either means that the qual- ity of the data is not good enough to be used for the given purpose, or that the approach in employing the data is not optimal. If the former is true, future ap- proaches to semi-supervised learning must concentrate 1 http://www.senseval.org 2 http://nlp.cs.swarthmore.edu/semeval/ on distinguishing reliably from unreliably annotated examples. Our hypothesis is that the automatically annotated examples that are added to the training material are not of high enough quality to improve the results. In order to investigate the problem, we start with the SensEval-3 English lexical sample task data set3 and then add examples extracted from a selection of lex- icons and corpora. The results show that only re- liably annotated examples should be considered and that they can be included in approaches different from the ones previously proposed (e.g. by Mihalcea [12]). We will show that a careful selection of automatically annotated examples gives a modest improvement over the supervised results, as compared to a significant drop in accuracy when all examples are added. 2 Related work There is a considerable amount of work published on word sense disambiguation for English. We will con- centrate here on some systems that took part in the SensEval-3 task and that …

Content truncated to 500 words