S-LLM: Semi-Supervised Large Language Model for Chat Summarization

Abstract

As producing high-quality summaries of chat dialogues currently requires large labeled datasets, we propose a method to efficiently leverage unlabeled data. Using a pseudo-labeling approach and post-processing to improve the quality of the pseudo-summaries, we are able to improve the Rouge-2 score of DistilBART by more than 6 points when using only 1% of labeled data on the TWEETSUMM dataset.