Good Neighbors Make Good Senses: Exploiting Distributional Similarity for Unsupervised WSD

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 65–72 Manchester, August 2008 Good Neighbors Make Good Senses: Exploiting Distributional Similarity for Unsupervised WSD Samuel Brody School of Informatics University of Edinburgh s.brody@sms.ed.ac.uk Mirella Lapata School of Informatics University of Edinburgh mlap@inf.ed.ac.uk Abstract We present an automatic method for sense- labeling of text in an unsupervised manner. The method makes use of distributionally similar words to derive an automatically labeled training set, which is then used to train a standard supervised classifier for distinguishing word senses. Experimental results on the Senseval-2 and Senseval-3 datasets show that our approach yields sig- nificant improvements over state-of-the-art unsupervised methods, and is competitive with supervised ones, while eliminating the annotation cost. 1 Introduction Word sense disambiguation (WSD), the task of identifying the intended meaning (sense) of words in context, is a long-standing problem in Natural Language Processing. Sense disambiguation is of- ten characterized as an intermediate task, which is not an end in itself, but has the potential to improve many applications. Examples include summariza- tion (Barzilay and Elhadad, 1997), question an- swering (Ramakrishnan et al., 2003) and machine translation (Chan and Ng, 2007). WSD is commonly treated as a supervised clas- sification task. Assuming we have access to data that has been hand-labeled with correct word senses, we can train a classifier to assign senses to unseen words in context. While this approach often achieves high accuracy, adequately large sense labeled data sets are unfortunately difficult to obtain. For many words, domains, languages, and sense inventories they are unavailable, and c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. in most cases it is unreasonable to expect to ac- quire them. Ng (1997) estimates that a high accu- racy domain-independent system for WSD would probably need a corpus of about 3.2 million sense tagged words. At a throughput of one word per minute (Edmonds, 2000), this would require about 27 person-years of human annotation effort. SemCor (Fellbaum, 1998) is one of the few cor- pora that have been manually annotated for all words — it contains sense labels for 23,346 lem- mas. In spite of being widely used, SemCor con- tains too few tagged instances for the majority of polysemous words (typically fewer than 10 each). Supervised methods require much larger data sets than this to perform adequately. The problem of obtaining sufficient labeled data, often referred to as the data acquisition bot- tleneck, creates a significant barrier to the use of supervised WSD methods in real world applica- tions. In this work we wish to take advantage of the high accuracy and strong capabilities of supervised methods, while eliminating the need for human an- notation of training data. Our approach exploits a sense inventory such as WordNet (Fellbaum, 1998) and corpus data to automatically create a collec- tion of sense labeled instances which can subse- quently serve to train any supervised classifier. The key premise of our work is that a …

Content truncated to 500 words