On Robustness and Domain Adaptation using SVD for Word Sense Disambiguation

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 17–24 Manchester, August 2008 On Robustness and Domain Adaptation using SVD for Word Sense Disambiguation Eneko Agirre and Oier Lopez de Lacalle Informatika Fakultatea, University of the Basque Country 20018, Donostia, Basque Country {e.agirre,oier.lopezdelacalle}@ehu.es Abstract In this paper we explore robustness and domain adaptation issues for Word Sense Disambiguation (WSD) using Singular Value Decomposition (SVD) and unlabeled data. We focus on the semi-supervised do- main adaptation scenario, where we train on the source corpus and test on the tar- get corpus, and try to improve results us- ing unlabeled data. Our method yields up to 16.3% error reduction compared to state-of-the-art systems, being the first to report successful semi-supervised domain adaptation. Surprisingly the improvement comes from the use of unlabeled data from the source corpus, and not from the target corpora, meaning that we get robustness rather than domain adaptation. In addition, we study the behavior of our system on the target domain. 1 Introduction In many Natural Language Processing (NLP) tasks we find that a large collection of manually- annotated text is used to train and test supervised machine learning models. While these models have been shown to perform very well when tested on the text collection related to the training data (what we call the source domain), the performance drops considerably when testing on text from other domains (called target domains). In order to build models that perform well in new (target) domains we usually find two settings (Daum´e III, 2007): In the semi-supervised setting the goal is to improve the system trained on the source domain using unlabeled data from the tar- get domain, and the baseline is that of the system 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. trained on the source domain. In the supervised setting, training data from both source and tar- get domains are used, and the baseline is provided by the system trained on the target domain. The semi-supervised setting is the most attractive, as it would save developers the need to hand-annotate target corpora every time a new domain is to be processed. The main goal of this paper is to use unlabeled data in order to get better domain-adaptation re- sults for Word Sense Disambiguation (WSD) in the semi-supervised setting. Singular Value De- composition (SVD) has been shown to find corre- lations between terms which are helpful to over- come the scarcity of training data in WSD (Gliozzo et al., 2005). This paper explores how this ability of SVD can be applied to the domain-adaptation of WSD systems, and we show that SVD and unla- beled data improve the results of two state-of-the- art WSD systems (k-NN and SVM). For the sake of this paper we call this set of experiments the do- main adaptation scenario. In addition, we also perform some related exper- iments on just the target domain. We use unlabeled data in order to improve …

Content truncated to 500 words