Improving Coherence by Reordering the Output of Extractive Summarization using Centering Theory through Genetic Algorithm

Extractive summarization is a widely studied and fairly easy to implement technique. It works by choosing the most important parts of a document(s) as a summary. However, this can lead to a lack of coherence in the summary itself. In this study, the principle of continuity in Centering Theory is used to maintain the entity coherence between subsequent sentences obtained from extractive news summarizer. Simultaneously, the relative order of sentences belonging to the same source document is maintained. These two considerations are implemented as fitness functions for a genetic algorithm that is used to obtain the optimal ordering of sentences in the summary. Based on the results of our study involving human judgment, a weighted fitness function combining 75% continuity and 25% relative order yields the most acceptable sentence ordering.