A More Cohesive Summarizer

We have developed a cohesive extraction based single document summarizer (COHSUM) based on coreference links in a document. The sentences providing the most references to other sentences and that other sentences are referring to, are considered the most important and are therefore extracted. Additionally, before evaluations of summary quality, a corpus analysis was performed on the original documents in the dataset in order to investigate the distribution of coreferences. The quality of the summaries is evaluated in terms of content coverage and cohesion. Content coverage is measured by comparing the summaries to manually created gold standards and cohesion is measured by calculating the amount of broken and intact coreferences in the summary compared to the original texts. The summarizer is compared to the summarizers from DUC 2002 and a baseline consisting of the first 100 words. The results show that COHSUM, aimed only at maintaining a cohesive text, performed better regarding text cohesion compared to the other summarizers and on par with the other summarizers and the baseline regarding content coverage.