Metric Learning for Synonym Acquisition

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 793–800 Manchester, August 2008 Metric Learning for Synonym Acquisition Nobuyuki Shimizu Information Technology Center University of Tokyo shimizu@r.dl.itc.u-tokyo.ac.jp Masato Hagiwara Graduate School of Information Science Nagoya University hagiwara@kl.i.is.nagoya-u.ac.jp Yasuhiro Ogawa and Katsuhiko Toyama Graduate School of Information Science Nagoya University {yasuhiro,toyama}@kl.i.is.nagoya-u.ac.jp Hiroshi Nakagawa Information Technology Center University of Tokyo n3@dl.itc.u-tokyo.ac.jp Abstract The distance or similarity metric plays an important role in many natural language processing (NLP) tasks. Previous stud- ies have demonstrated the effectiveness of a number of metrics such as the Jaccard coefficient, especially in synonym acqui- sition. While the existing metrics per- form quite well, to further improve perfor- mance, we propose the use of a supervised machine learning algorithm that fine-tunes them. Given the known instances of sim- ilar or dissimilar words, we estimated the parameters of the Mahalanobis distance. We compared a number of metrics in our experiments, and the results show that the proposed metric has a higher mean average precision than other metrics. 1 Introduction Accurately estimating the semantic distance be- tween words in context has applications for machine translation, information retrieval (IR), speech recognition, and text categorization (Bu- danitsky and Hirst, 2006), and it is becoming clear that a combination of corpus statistics can be used with a dictionary, thesaurus, or other knowl- edge source such as WordNet or Wikipedia, to in- crease the accuracy of semantic distance estima- tion (Mohammad and Hirst, 2006). Although com- piling such resources is labor intensive and achiev- ing wide coverage is difficult, these resources to some extent explicitly capture semantic structures 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. of concepts and words. In contrast, corpus statis- tics achieve wide coverage, but the semantic struc- ture of a concept is only implicitly represented in the context. Assuming that two words are semanti- cally closer if they occur in similar contexts, statis- tics on the contexts of words can be gathered and compared for similarity, by using a metric such as the Jaccard coefficient. Our proposal is to extend and fine-tune the latter approach with the training data obtained from the former. We apply metric learning to this task. Al- though still in their infancy, distance metric learn- ing methods have undergone rapid development in the field of machine learning. In a setting simi- lar to semi-supervised clustering, where known in- stances of similar or dissimilar objects are given, a metric such as the Mahalanobis distance can be learned from a few data points and tailored to fit a particular purpose. Although classification meth- ods such as logistic regression now play impor- tant roles in natural language processing, the use of metric learning has yet to be explored. Since popular current methods for synonym ac- quisition require no statistical learning, it seems that supervised machine learning should easily outperform them. Unfortunately, there are obsta- cles to overcome. Since metric learning algorithms usually learn the parameters of …

Content truncated to 500 words