Multilingual Spectral Clustering Using Document Similarity Propagation

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 871–879, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Multilingual Spectral Clustering Using Document Similarity Propagation Dani Yogatama and Kumiko Tanaka-Ishii Graduate School of Information Science and Technology, University of Tokyo 13F Akihabara Daibiru, 1-18-13 Kanda Chiyoda-ku, Tokyo, Japan yogatama@cl.ci.i.u-tokyo.ac.jp kumiko@i.u-tokyo.ac.jp Abstract We present a novel approach for multilin- gual document clustering using only com- parable corpora to achieve cross-lingual semantic interoperability. The method models document collections as weighted graph, and supervisory information is given as sets of must-linked constraints for documents in different languages. Recur- sive k-nearest neighbor similarity propa- gation is used to exploit the prior knowl- edge and merge two language spaces. Spectral method is applied to find the best cuts of the graph. Experimental results show that using limited supervisory in- formation, our method achieves promis- ing clustering results. Furthermore, since the method does not need any language dependent information in the process, our algorithm can be applied to languages in various alphabetical systems. 1 Introduction Document clustering is unsupervised classifica- tion of text collections into distinct groups of sim- ilar documents. It has been used in many in- formation retrieval tasks, including data organiza- tion (Siersdorfer and Sizov, 2004), language mod- eling (Liu and Croft, 2004), and improving per- formances of text categorization system (Aggar- wal et al., 1999). Advance in internet technology has made the task of managing multilingual docu- ments an intriguing research area. The growth of internet leads to the necessity of organizing docu- ments in various languages. There exist thousands of languages, not to mention countless minor ones. Creating document clustering model for each lan- guage is simply unfeasible. We need methods to deal with text collections in diverse languages si- multaneously. Multilingual document clustering (MLDC) in- volves partitioning documents, written in more than one languages, into sets of clusters. Simi- lar documents, even if they are written in differ- ent languages, should be grouped together into one cluster. The major challenge of MLDC is achieving cross-lingual semantic interoperability. Most monolingual techniques will not work since documents in different languages are mapped into different spaces. Spectral method such as Latent Semantic Analysis has been commonly applied for MLDC task. However, current techniques strongly rely on the presence of common words between different languages. This method would only work if the languages are highly related, i.e., languages that share the same root. Therefore, we need another method to improve the robustness of MLDC model. In this paper, we focus on the problem of bridg- ing multilingual space for document clustering. We are given text documents in different lan- guages and asked to group them into clusters such that documents that belong to the same topic are grouped together. Traditional monolingual ap- proach is impracticable since it is unable to pre- dict how similar two multilingual documents are. They have two different spaces which make con- ventional cosine similarity irrelevant. We try to solve this problem utilizing prior knowledge in …

Content truncated to 500 words