Cohesive Summary Extraction From Multi-Document Based On Artificial Neural Network

Abstract

Increasing growth in the volume of digital data from documents performed the difficulty of accessing important information. The solution is using automatic summarization systems that aim to extract important information in a short time. Usually, these systems work to extract a single summary from a single document or multi-documents. However, extracting a summary from a multi-document may encounter some obstacles. This work focuses on how to overcome these obstacles and extract an appropriate and cohesion summary by presenting and suggesting four important contributions with a qualitative leap in the process of extracting the important information that seeks to extract a candidate summary that matches the sentences of the golden summary. The first suggestion is a set of features to extract important sentences and easy to understand, the second, build a Backpropagation Multi-layer Perceptron Neural Network (BMPNN) based on these features input to extract the score for each sentence, the third, using the Random oversampling method and its effective role in rebalancing the data during the training process in BMPNN, and finally, solving the problem of reordering sentences in the candidate summary according to the importance of the sentence depending on one of the features. The results of the evaluation Rouge-1, Rouge-2, and Rouge-L measures showed that the candidate's summary is very close to the golden summary in terms of matching sentences, and it achieved very good results.