Reinforcement Learning for Sinhala Named Entity Recognition

Abstract

Named Entity Recognition is a subtask of Natural Language Processing. In the literature various machine learning methods have been proposed to tackle this task such as supervised learning, unsupervised learning and semi supervised learning. It is established that, for supervised learning, a large number of annotated data is needed for a system to have a good generalization capability. For Indic languages such as Sinhala it is challenging to build a Named Entity Recognition system due to its inherent features such as lack of capitalization. Even though there have been several researches conducted for Sinhala, most of them have used supervised learning methods with small quantities of training data. The resurgence of Reinforcement Learning demonstrates its effectiveness in problems that humans solve effectively in an incremental manner over time. Even though it is hard to find research that has applied Reinforcement Learning to Natural Language Processing tasks, there have been attempts to map tasks such as Part of Speech tagging to the Reinforcement Learning paradigm. This dissertation first casts the Sinhala Named Entity Recognition task into the Reinforcement Learning paradigm and then proposes a language specific Named Entity Recognition system that can be trained to generalize better using the current annotated data available.