Implementation of Priority Algorithm-Based KoBART News Summary System for Improving Performance of Abstract Summary

This study implemented a Korean news summary system based on the KoBART model using a priority algorithm. First, the KoBART model was fine-tuned using the artificial intelligence (AI) hub’s dataset to make it suitable for Korean news summaries. The priority algorithm was then applied to the original news articles to determine the similarity between all sentences. Finally, sentences with a low similarity to the original news were excluded, and the remaining sentences with a relatively high similarity were input into the fine-tuned KoBART model, which generated a news summary. The proposed model improved the ROUGE-L score by an average of 9.7% compared with the fine-tuned KoBART model. In addition, even when the token length was increased for long newsitems, the proposed model showed relatively low performance degradation. Consequently, we confirmed that using a priority algorithm to exclude sentences of low importance improved the performance of the abstractive summary.