Automatic Persian Multi-Text Summarization Techniques based on Meta-Heuristic Algorithms

The main objective of this study is to present a pattern for standard summarization of Persian texts with the approach of converting the problem to optimization problem by compatible meta-heuristic algorithms. In this research, standard multi-text "Pasokh" collection, which contains 50 different types of news from the most popular news agencies in Iran, each containing 20 documents, as well as 5 summaries of abstractive and 5 extractive, used for evaluation. First, the preprocessing performed on the input texts and the initial summary generated with TF-ISF benchmark, readability and consistency criteria of the sentences, similarity to the title, position of the sentence in the text, and the length of the sentence. With respect to each of these criteria, weighting function assigned to extracted sentences and a similarity matrix created. Then, output of the extraction system processed by Genetic algorithm and Cuckoo search algorithm for the final summary. Eventually, the output obtained from the previous step analyzed using the Rouge evaluation tools and the comparison with the human abstracts. The average of all values obtained in Rouge evaluation tools for calculation the overlapping of common samples of human summaries and machine summaries by Cuckoo search algorithm were higher than the values obtained by Genetic algorithm as well as Ijaz online summarizer system. Meanwhile, among the eight criteria, the longest common sub-sentence with a value of 0.33 and the number of common words in the text with 0.40 were better than the rest. The results of the comparison of two algorithms indicate that the Cuckoo search algorithm is better in the entire criteria. On the other hand, comparing the results suggests that the average time calculated for summarizing by the proposed system is also less.