Enhancements to Graph Based Methods for Single Document Summarization

Abstract

This paper focuses its attention on extractive summarization using popular graph based approaches. Graph based methods can be broadly classified into two categories: non-PageRank type and PageRank type methods. Of the methods already proposed-the Centrality Degree method belongs to the former category while LexRank and Continuous LexRank methods belong to later category. The paper goes on to suggest two enhancements to both PageRank type and non-PageRank type methods. The first modification is that of recursively discounting the selected sentences, i.e. if a sentence is selected it is removed from further consideration and the next sentence is selected based upon the contributions of the remaining sentences only. Next the paper suggests a method of incorporating position weight to these schemes. Thus we have experimented with 12 methods-six of non-PageRank type and six of PageRank type. To clearly distinguish between various schemes, we call the methods of incorporating discounting and position weight enhancements over Lexical Rank schemes as Sentence Rank (SR) methods. Intrinsic evaluation of all the 12 graph based methods were done using conventional Precision metric and metrics earlier proposed by us-Effectiveness1 (E1) and Effectiveness2 (E2). Experimental study brings out that the proposed SR methods are superior to all the other methods.

An Innovative Approach of Bangla Text Summarization by Introducing Pronoun Replacement and Improved Sentence Ranking

“…The Precision, Recall, and F-measure are brought into play here as these have long used as important evaluation metrics in information retrieval field [51]. If 'A' indicates the number of sentences retrieved by summarizer and 'B' shows the number of sentences that are relevant as compared to target set, Precision, Recall and F-measure are computed as:…”

A graph-based approach towards automatic text summarization.

“…These summaries are more likely to provide information about the article and they are longer in length as compared to the indicative summaries. Generic summaries [10] [8] are intended to convey the author's perspective, which also represents the theme of the document. These summaries inform the broad community of readers and include a general idea of the original document.…”

“…There are numerous other methods to calculate similarity between the sentences like discounting, cumulative sum method [13] and position weight [8]. In a Graph-based approach, sub-graphs are generated which may or may not be connected to each other.…”

“…This problem is more intense in multi-document summarization but can be avoided by replacing the pronouns with their antecedents [7]. On the other hand, abstractive summaries have a representation problem because they involve reformulation of contents, a process which makes it domain dependent and requires human supervision [8]. The biggest challenge for abstractive summarization is that the capabilities of current computing devices are limited because they cannot generate rich representations of a language.…”

“…Thus, abstractive summaries do not have high quality in terms of evaluation, and they can deviate from the topic of the document very easily. Simplicity, speed, domain independence, and non-requirement of background knowledge [8] are the reasons that we opted to focus on extractive text summarization.…”