A Framework for Abstractive Summarization of Conversational Meetings

Unlike extractive summarization, abstractive summarization creates summaries by synthesizing new words and sentences that maintain the original meaning of the source. This presents new challenges that researchers and developers face when developing language processing models for text generation. Utilizing advanced models in automatic speech recognition and natural language processing such as OpenAI’s Whisper and Meta’s BART, we simplify the process of speech recognition and abstractive summarization of long meetings. We propose a system framework and conduct analysis of speech to text specifically in conversations with more than two speakers in a meeting environment. Through both quantitative and qualitative analysis we evaluate the proposed model performance compared to the BART base model, and show that with summarizing long meeting dialogues our model improved summarization by 139.6% over the base model in the ROUGE-LSUM metric. Our proposed framework for abstractive summarization is a practical and accurate solution providing accessibility accommodations to hard-of-hearing people and accurate and insightful analysis to industry and academia.