Enhancing Meeting Summarization Models via Large Language Model Data Generation and Weighted Sampling

Meeting summaries improve team communication and decision-making efficiency by extracting key information and decisions. However, due to the complexity of organizing meeting records and the need to protect the privacy of participants, available datasets are often very scarce, which restricts the development and application of meeting summary technology. Existing data augmentation techniques usually expand the data set by transforming or reconstructing the existing data samples to improve the generalization ability of the model. This leads to a model that does not generalize well when faced with unseen data. Therefore, this paper proposes a data augmentation method for meeting summaries based on large language models, which is implemented by sequential expansion method and weight sampling strategy. First, considering that the length of text data generated by large language models cannot meet the requirements of meeting datasets, we propose a novel sequential expansion method to increase the length of generated meeting data and ensure the consistency of its context. Secondly, due to the imbalanced influence of the original dataset and the generated dataset during model training, we use a weighted sampling strategy to balance them and ensure that both can achieve stable learning. Experiments on the AMI dataset show that the performance of meeting summarization is significantly improved by applying this method to the current best pre-trained meeting summary model DIALOGLED-large.