Text mining (NLP) Abstractive text summarization using deep sequence models (T-0399) (MFN 8758)

Abstract

The reading of the long text is time consuming and sometimes understanding the context becomes difficult. Summaries are important specifically when we need to save our time and to understand the actual context of a long text corpus. Summarization is a technique to create a concise and accurate summary of a large script or a set of articles. In recent years abstractive text summarization tasks are most challenging in natural language processing. The existing encoder-decoder approaches have a potential issue. For the longer sequence of reviews, they need to compress all the necessary information into a fixed-length vector. This thesis aims to solve a very inherent task in data mining that is review summarization. Summary of the reviews has challenges that are dealing with variable length reviews, freestyle writing, and unstructured behavior. Our aim to create a shorter version of the review in abstractive manners while preserving the sentiment and points. In the decision-making process, it helps online customers to judge the product or service. To generate an optimal summary we have used a BRNN with LSTM's in the encoding layer. In the decoding layer, the attention mechanism is applied to the decoding cell that is just a two-layer LSTM with dropout. We have used Concept Net Number-Batch 3.0 word embeddings and Amazon Food reviews dataset. To reduce training loss and compute the learning rate of each parameter, we have used Adam Optimizer to reduce the loss function and for faster converge. We have achieved R1 38.75, R2 16.5, RL 36.25, and reduced the training loss with a new value of 0.031 for the whole dataset after removing the duplications.