Briefing Assistant: Learning Human Summarization Behavior over Time

We describe a system intended to help report writers produce summaries of important activities based on weekly interviews with members of a project. A key element of this system is to learn different user and audience preferences in order to produce tailored summaries. The system learns desired qualities of summaries based on observation of user selection behavior, and builds a regression-based model using item features as parameters. The system’s assistance consists of presenting the writer with a successively better ordered list of items from which to choose. Our evaluation study indicates a significant improvement in average precision (and other metrics) by the end of the learning period as compared to baseline of no learning. We also describe our ongoing work on automatic feature extraction to make this approach domain independent.