Automatic Text Summarizing System Using Reinforcement Learning Technique

Abstract

Automatic document summarization, the process of retrieving key information from a source document is a relevant as well as challenging art. Summarization can be done based on different aspects, on which different works have been done. Extractive summarization and abstractive summarization are such two major types of summarization problems. Extractive summarization is a task in which both machine learning and deep learning techniques have been tested aiming at better results. One of the less used type of machine learning technique called reinforcement learning (RL) was previously used mostly for the task of developing video games such as Atari games and this method is later being used for different NLP tasks too. Au- tomatic summarization using reinforcement learning techniques has become one of the major tasks in the modern research area. Beginning with the traditional reinforcement learning techniques, the research has reached to a state where combination of both deep learning as well as reinforcement learning methods got used together. In this work, we propose a novel abstractive text summarizer which is built using both deep neural network and reinforcement learning technique. The generated summaries are evaluated manually and using average semantic similarity score as well. The system is tested on Amazon fine food review dataset and the results show that the generated summaries are meaningful and context dependent in most of the cases.