Statistical pair pruning towards target class in learning-based anaphora resolution for Tamil

Anaphora resolution is an important task to be achieved in many natural language understanding (NLU) applications including machine translation. This paper proposes learning-based system to resolve pronouns in Tamil text built around various classification algorithms. To improve learning accuracy, the system is built in two folds. First is feature vector production where mentions are identified, characterised then a feature vectors of lexical, syntactic and semantic features are produced. Next is the pair pruning module where, number of non-target class pairs is reduced by deep statistical analysis of feature vector. Incorporating deeper pair pruning module dramatically increases the f-measure score when compared to training the same models but without the pruning module. On the tourism dataset of TDIL we trained the system with various classification algorithms and obtained encouraging results for a challenging language, Tamil. We discuss how varying the ratio of f-measure, precision and recall is between with and without the pruning module in comparative model.