A Redundancy Elimination Approach towards Summary Refinement

Abstract

A summary generated by a machine, in contrast to human-generated summaries are produced in less time, unbiased, not time or mood dependent and reliable. However many commonly used approaches are feature based methods that look out for important sentences or phrases by observing features or cues. Such feature based methods may end up producing summaries that contain sentences which are similar in meaning, mostly which depend on sentence scoring and hence not a desirable factor. The proposed work takes a machine generated summary as rough summary and uses binomial distribution to identify importance of every sentence in the rough summary. The semantic similarity between sentences is identified and the sentences are removed thereby refining the summary. By eliminating similar sentences the summary is refined so that only informative sentences are left in the summary. The proposed redundancy elimination approach is applied on summaries obtained from an existing summarization system with the fuzzy based summarization model as a case study. Evaluation of the summary refinement approach is done on DUC2002 dataset and the results are promising.