Labeling Unlabeled Data using Cross-Language Guided Clustering

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 383–391, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Labeling Unlabeled Data using Cross-Language Guided Clustering Sachindra Joshi IBM Research New Delhi, India jsachind@in.ibm.com Danish Contractor IBM Research New Delhi, India dcontrac@in.ibm.com Sumit Negi IBM Research New Delhi, India sumitneg@in.ibm.com Abstract The effort required to build a classifier for a task in a target language can be signifi- cantly reduced by utilizing the knowledge gained during an earlier effort of model building in a source language for a sim- ilar task. In this paper, we investigate whether unlabeled data in the target lan- guage can be labeled given the availabil- ity of labeled data for a similar domain in the source language. We view the problem of labeling unlabeled documents in the tar- get language as that of clustering them such that the resulting partitioning has the best alignment with the classes provided in the source language. We develop a cross language guided clustering (CLGC) method to achieve this. We also pro- pose a method to discover concept map- ping between languages which is utilized by CLGC to transfer supervision across languages. Our experimental results show significant gains in the accuracy of label- ing documents over the baseline methods. 1 Introduction The last few years have seen a rapid growth in the development of machine learning applications for non-English languages. This growth can be at- tributed to several factors such as increased Inter- net penetration (especially in non-English speak- ing countries) and wide adoption of Unicode stan- dards that allow people to generate content in their own language. A key guiding principal in the development of such applications for a new language (referred to as the target or resource-poor language) has been to leverage the existing models and linguistic re- sources available for a popular language such as English (also called source or resource-rich lan- guage). Existing literature examines two ways of utilizing this knowledge. The first way is to adapt an existing statistical model for a new tar- get language. Examples of this is the problem of cross-lingual sentiment classification (Xiaojun Wan 2009), or in a more general setting for cross language domain adaptation for classification (Pe- ter Prettenhofer and Benno Stein 2010). The sec- ond way is to develop linguistic resources for a target or resource-poor language by leveraging the resources available in a source or resource-rich language. An example of this is the work done for automatically transferring syntactic relations (in WordNet) from a source language (English) into a target language (Romanian) (Verginica Barbu Mi- titelu and Radu Ion 2005). In this paper, we investigate another way of uti- lizing the knowledge gained in one language for building machine learning applications in an an- other language. Our work focuses on generating training data (in contrast to adapting models and language resources) in the target language, given in-domain training data for the source language. The labeled data in the source language could be used …

Content truncated to 500 words