Automatic Identification of Discourse Relations in Indian Languages

This paper describes the first effort on automatic identification of connectives and their arguments for three Indian languages Hindi, Malayalam and Tamil. We have adopted machine learning technique Conditional Random Fields (CRFs) for our work. We have used a corpus of 3000 sentences belonging to health domain. Domain independent features were extracted to improve the performance of the system. We mainly concentrated on the identification of explicit connectives and their arguments. Two sets of experiments were performed. First set of experiment was performed for the identification of connectives and next for the identification of argument boundaries. Using this approach we obtained encouraging results for all the three languages. Error analysis shows the presence of different structural patterns of discourse relations among three languages.