Machine learning based automatic text summarization system

Abstract

In this research work, two deep-learning-based text summarization techniques named WL-AttenSumm and AttSum-Hybrid have been proposed. WL-AttenSumm is the first deep learning-based summarization model proposed in this study for extractive summarization of single documents that learns the syntactic and semantic relationships from the text. It employs a Word-level Attention mechanism that focuses more on the important parts of the input sequence so relevant semantic features are captured at the word level. This model employs CNN and Bi-GRU. Experimentation has been done with three different pre-trained word embedding models: GloVe, word2vec, and fasttext. newlineAnother deep-learning-based text summarization technique proposed as a part of this study is a novel hybrid summarization system, AttSum-Hybrid that takes into consideration language context and relationship between the text as well as captures structural information of the sentences. In this hybrid summarization framework, a contextual model based on a deep learning approach is combined with the statistical feature-based model. BERT works as a newlinefeature extractor in the contextualized representation model such that it generates a vector representation for each word depending upon the context in which the word appears. The Convolutional Bi-LSTM network is used in this contextual model. On the other hand, a statistical feature representation framework has been developed as a part of this hybrid summarization system that incorporates a few better-performing sentence scoring features so that the structural aspects of the text can also be captured while creating an extractive summary of the document. Experimental results demonstrate that AttSum-Hybrid has performed better than WL-AttenSumm. newline newline