Reinforcement Learning Models for Abstractive Text Summarization
Sergiu Buciumas · 2019
Abstractive text summarization is an active research topic in Natural Language Understanding. We live in a digital world where the information for every topic in Internet is increasing considerable, and users would benefit by generating summaries. Summaries are categorized in the following: summary generated through extractive methods and summary generated through abstractive methods. Abstractive methods are highly complex, due the fact that generates an abstract summary consisting of ideas or concepts listed in the documents but are repeated or interpreted with different words or phrases (models that are not restricted to selecting and rearranging phrases from the original documents). Most of today's tasks in Natural Language Processing (NLP) are performed with semi-supervised or supervised learning models. In this research we are looking to perform the text summarization using a new architecture that take advantage of Pre-trained models and Reinforcement Learning (RL). The goal of the research is to extend the summarization to multiple datasets at the time of training to generate summaries. The evaluation is performed on the CNN/Daily Mail, New York Times and Wikihow datasets.