Enhancing Extractive Summarization in Student Assignments Using BERT and K-Means Clustering

Evaluation through learning assessments is a fundamental factor in determining student’s success in achieving specified competencies. During the essay evaluation, the lecturer needs to check each assignment individually. However, when dealing with long essay answers, extra attention is necessary to extract the key points effectively. As a result, this process takes a lot of time and may potentially lead to carelessness or boredom during the assignment-checking process. The way to overcome this problem is by using Extractive summarization. The Extractive Summarization method summarizes by extracting key points from the student assignments and creating a summary without making any changes to the original text. Currently, Extractive Summarization widely uses deep learning technology, such as Bidirectional Encoder Representations from Transformers (BERT). BERT can effectively recognize contextual information within sentences. Output from BERT is sentence embeddings that can serve as valuable features for clustering. The summary result generated by applying k-means clustering to group sentences that have similarities and then selecting one primary sentence from each cluster to represent the cluster. This research proposed an approach for selecting sentence candidates for each cluster using TF-IDF weighting. The proposed method achieved the best ROUGE score on ROUGE-1 recall with 0.73003. We compare our results with the previous study’s BERT k-means approach, which selects sentence candidates from the closest sentences to the centroid for summary selection. The experimental results show that the proposed method achieves slightly better ROUGE scores than the previous study. Furthermore, in terms of execution time comparison, the proposed method has a shorter execution time.