CRII: RI: Towards Abstractive Summarization of Meetings

Northeastern University, Boston MA

Investigators

Abstract

Meeting is a common way to collaborate, share information and exchange opinions. Many available meeting transcripts, however, are lengthy, unstructured, and thus difficult to navigate. It would be time-consuming for users to access important meeting output by reading the full transcripts. Consequently, automatically generated meeting summaries is of great value to people and businesses alike by providing quick access to the essential content of past meetings. The core objective of this research project is to automatically generate abstract-style focused meeting summaries to help users digest the vast amount of meeting content in an easy manner. It helps the research community to better understand the characteristics of the meeting domain, define the summarization task in meetings in a more consistent way, improve speech summarization evaluation metrics, and allow the wide use of speech summarization techniques in many applications (such as generating meeting minutes or lecture outlines). The broader impacts of this project includes sharing insights on conversational text with social scientists, providing natural language processing research training to students, and contributing effective methods for meeting summarization to the general public. This research project aims at constructing abstractive summaries of meetings by developing computational models for important outcome identification and natural language summary generation, as well as designing objective summary evaluation methods. Existing meeting summarization systems remain largely extractive: Their summaries are comprised exclusively of patchworks of utterances selected directly from the meetings to be summarized. Although relatively easy to construct, extractive approaches simply present a set of utterances as the final summary, and fall short of producing concise and readable summaries, largely due to the spontaneous nature of spoken dialogue. This project formulates a new framework that accounts for the special aspects of meetings and use them to identify the utterances that contain important outcomes. A discriminative learning-based latent variable model trained with rich features is utilized to jointly capture topic shifting and extract utterances with important outputs. To perform sentence planning and surface realization in one single process, a neural network-based natural language generation model is developed. Objective evaluation methods are designed to measure various aspects for the quality of generated summaries.