Extractive Text Summarization Using K-Medoid Clustering on BERT

Extractive summarization involves choosing sentences directly from the source text document to produce a summary. This method is improved by employing clustering techniques and a scoring function algorithm, which facilitates the identification of pertinent information in the input documents. In this research, we propose a hybrid methodology for extractive summarization that incorporates Bidirectional Encoder Representations from Transformers (BERT) and the K-Medoid Clustering algorithm. The entire input document is initially encoded with the BERT encoder, which generates sentence embeddings for the text. Then, we apply the K-Medoid clustering algorithm to these sentence embeddings to generate clusters. A final consideration is given to sentences near the centre of each cluster, and these sentences are incorporated into the final summary. We evaluate the effectiveness of the proposed model for extractive summarization tasks using the CNN/DM news dataset. In continuation to new dataset we also evaluate our framework on medical scientific papers long document PubMed dataset. Our method outperforms both clustering and non-clustering models, as demonstrated by the experimental results.