Abstractive Summarization

Automatic Meeting Summarization addresses the question if and how it is possible for a compute to analyze documents in order to generate a summary of their contents.

In contrast to extractive summarization, the abstractive approach models itself after the cognitive processes which occur in the human mind when people write a summary. First, an interpretation phase creates a symbolic representation of important contents of a given meeting. Then, this representation is condensed further by automatic inference processes, and lastly, a generation step produces the final summary text from the symbolic content representation.

Compared to the extractive approach, the quality of text output by an automatic abstractive summarizer resembles much closer that of manually written summaries, as a result of the explicit text generation phase at the end of the summarization pipeline. This is an especially important feature in the realm of multi-party interaction, where the spontaneous nature of dialog is typically quite different from written documents in coherence, conciseness, and structure.

However, the higher quality comes at a cost. Currently, the knowledge sources required for abstractive summarization need expensive modeling by an expert and are still limited to specific domains of discourse. Finding ways to relax these constraints is one of the most interesting and challenging questions of research in the field.

List of relevant Publications

Contact Person: