CENTRANK: A GRAPH CENTROID BASED RANKING FOR EXTRACTIVE TEXT SUMMARIZATION

With the increase in text data generated online, an automatic summarizer has become essential. The challenges in creating summaries automatically are: semantic relatedness to the topic, coverage, redundancy reduction, relevance, coherence, relatedness to the central concept of the document. In order to tackle these challenges, we propose CENTRANK, a graph centroid based ranking for extractive, single-document, generic summarization. Here, the document is modelled as a text graph. The words represent the nodes of the graph and the relation between the words represents the edges. The centroid of the graph represents the important words, in turn, important sentences. The word2vec model has been used to identify the semantic relatedness of the words. A feature matrix is constructed by considering the semantic features of the sentences like similarity of the sentence to the centroid. Further the coverage, redundancy reduction, relevance and coherence scores are used to enhance the feature matrix. The sentences with the highest scores are selected. A parametrization has been performed to record the performance of the proposed system, namely, CENTRANK, against various summary sizes. Also, a parametrization on the values of the objective function for various summary sizes is also recorded. It is inferred that a size of half the document size yields the maximum ROUGE score.