Graph based extractive text summarization based on triangle counting approach

Abstract

Currently, with the exponential rising quantity of automated textual data available on the Web, end users require the ability to get information in summary form, while keeping the most vital information in the document. As a result of this, the necessity for the creation of Summarization systems became vital. Summarization systems, collect and focus on the most important ideas of the papers and help the users to find and understand the main ideas of the text faster and in a simpler way from the dispensation of information. Compelling set of such systems are those that create summaries of extracts. This type of summary, which is called Extractive Summarization, extracts the most applicable sentences from the main document. The used methods, usually assign a score for every sentence in the text, based on specific features. Then choose the most important sentences, according to the degree of score for each sentence. These features include but not limited to, the sentence length, its similarity with the title, the position of the sentence in the main document, and the frequency of the words in the sentence. Nevertheless, not have been achieved quality summaries corresponds with the ones made by humans, and therefore proposing the techniques continue to be raised, for the aims of improving the outcomes. One methodology for creating extractive summary is using the graph theory. One field in graph theory called graph pruning / reduction, which aims to find the greatest illustration of the original graph with less number of nodes and edges. This paper proposes a method of extractive summarization based on a graph reduction technique called the triangle counting approach. This method has three main phases. The first phase is graph representation, where nodes are the sentences and edges are the similarity between the sentences. The second phase is triangles construction, and the third phase is bit vector representation for the triangles nodes and finally create the summary based on the values of bit vector. The proposed method was evaluated, using ROUGE measures on the dataset DUC2002. The results showed that by using triangle counting as a reduction technique, it performs better than the state of the art methods.