Clustering algorithm used for automatic summarization model #summarization

The EDS task involves assigning a label to each sentence in a document to determine if it should be included in the summary. A cluster algorithm is introduced to reduce redundancy in the summary by using sentence embeddings. The Score-BERT model, based on BERT, scores sentences using an encoder and sigmoid layer. The Cluster-EDS model builds on Score-BERT by selecting sentences based on semantic similarity clusters. The K-means module maps sentence embeddings to a high-dimensional semantic space and selects sentences with the highest scores from different clusters for the summary. Text matching, introduced by Zhong et al., involves using Siamese-BERT architecture to match reference and candidate summaries by comparing their embeddings. The similarity score is calculated using cosine similarity. This method aims to improve the extractive summarization task by selecting sentences from different clusters to reduce redundancy and improve the overall quality of the summary.

Source link

Source link: https://www.nature.com/articles/s41598-024-66306-4