An experimental study of game theory with various word embeddings for automatic extractive text summarization

Text Summarization is a process to abridge a long-size document into condensed form by comprising all the prime information and central theme. Researchers proposed numerous approaches for automatic text summarization, but still state of art methods are not able to capture semantic information of input text and preserve the text cohesiveness properly. To incorporate these features the proposed novel hybrid approach integrates evolutionary game theory with word embedding models. Here each sentence acts as a player and the meaning (senses) of words of sentences form strategy for the respective games. The payoff is determined using word embedding similarity measures between words and senses of sentences. The proposed hybrid method is capable of having the sentences (in summary) from the sentences (in the document) that share a strong relationship. This idea preserves textual cohesiveness effectively by modelling the interdependence between the sentences. The proposed work is evaluated and compared with existing state-of-the-art approaches on DUC02 (Task 1) datasets using the ROUGE-1, ROUGE-2, ROUGE-SU4 and cohesion score metric. The experimental results show the proposed method outperforms the state-of-art approaches.