Extractive based email summarization: An unsupervised hybrid approach using graph based sentence ranking and K-means clustering algorithm

Abstract

Over the years, Automatic Text Summarization is widely studied by many researchers. Here, an attempt is made to generate an automatic summary of a given text document based on an unsupervised hybrid model. The model comprises of an extractive method: a Graph-based text ranking and K-means: a clustering algorithm. Ranked sentences are obtained using the graph-theoretic ranking model here word frequency, word position, and string pattern based ranking are calculated. The K-Means algorithm generates the coherent topic clusters. Using the output of Graph-based method and K-means clusters, Sentence Importance Score(SIS) is calculated for each sentence, where top 70 percent ranked sentences and centralised topics of each cluster (excluding those topics which fall in the outlier zone) are used. The unsupervised hybrid approach is an attempt to inherit one of the human practice of reading and then summarizing the text in short while keeping the original insight of that text by the virtue of important sentences and keywords. The system is tested on dataset for Summarization and Keyword Extraction from Emails which on evaluation gives an average of 0.57 score on ROUGE 2.0 tool.