Evaluating different similarity measures for automatic biomedical text summarization

Automatic biomedical text summarization is maturing and can provide a solution for biomedical researchers to access the information they need efficiently. Biomedical summarization approaches often rely on the similarity measure to model the source document, mainly when they employ redundancy removal or graph structures. In this paper, we examine the impact of the similarity measure on the performance of the summarization methods. We model the document as a weighted graph. Various similarity measures are used to build different graphs based on biomedical concepts, semantic types and a combination of them. We next use the graphs to generate and evaluate the automatic summaries. The results suggest that the selection of the similarity measure has a substantial effect on the quality of the summaries (≈37% improvement in ROUGE-2 metric, and ≈29% in ROUGE-SU4). The results also demonstrate that exploiting both biomedical concepts and semantic types yields slightly better performance.