Abstract

Easy access of the internet generating tremendous amount of data every day. The data is generated at high volume with high velocity and to know the essence of data automatic text summarization is needed. The purpose of research is to design Abstractive Text Summarizer using deep learning techniques. In this study, we model the summarization task as a Sequenceto- Sequence (Seq2Seq) problem. The paper experiments with encoder decoder based different variants of RNN such as vanilla RNN, Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU). The problem with existing approach is that the context vector contains information from the last recent part of the sentence and forgets the relevance of the initial time step input. To solve this problem, the content-based attention mechanism is applied which uses cosine similarity to generate context vector at each time step. The CNN/Daily Mail dataset is used for the study. The performance of models is measured using the ROUGE score to select the ideal model for summarization. The experimental results show bidirectional LSTM combined with a content-based attention mechanism enhances summarization performance. The experimental data demonstrates that content-based attention outperforms over dot product based attention.