Improving Automatic Meeting- Understanding by Leveraging Meeting Participant Behavior

Most office workers participate in multiple meetings on a daily basis. Although surveys show that large parts of these meetings are often not useful to all the participants, it has been shown (Banerjee, Rose, & Rudnicky, 2005) that participants do sometimes need to retrieve information discussed at previous meetings, and that this is usually a difficult task. The human-impact goal of this thesis is to help humans retrieve the information they need from past meetings. Several approaches have been explored in the past to help humans with this retrieval task. These approaches include meeting recording and browsing systems (Cutler, et al., 2002), and systems that automatically detect and extract pieces of useful information from the speech such as action items (Purver, Ehlen, & Niekrasz, 2006). These approaches are often examples of either classic supervised learning (with offline data collection, annotation and model training) or unsupervised learning with some adaptation to the meeting participants. While these approaches make use of the expertise of offline human annotators, we believe that little effort has been made to effectively harness the knowledge that the meeting participants have. Specifically, meeting participants will be the best judges of what information is important in a meeting. This judgment, if properly leveraged, can provide high quality information with which to improve automatic meeting-understanding systems. The challenge of leveraging meeting participant knowledge, however, is that they may have little motivation to provide labeled data to the system without some perceptible and immediate benefit. Moreover providing such information may be distracting and thus undesirable. Our hypothesis is that despite this challenge, it is possible to motivate the human users of an interactive system to provide supervision. We propose to extract this supervision by designing services that provide the user with immediate benefit, but that are designed in such a way that as the user interacts with the system, his actions can be interpreted as labeled data. Given this labeled data, the system can improve its performance over time. We propose two mechanisms: passive and active supervision extraction. In the passive approach, the system cannot select data points to query labels for, and data acquisition from user actions occurs entirely due to the design of the interface. In the active approach, the system selects data points and queries the user for their labels. Although this is similar to active learning, we are interested in motivating ordinary users to provide data (as opposed to giving the data to labelers). We create this motivation by embedding the queries in an interactive service that gives the user immediate benefit every time a query is made. The user’s responses are then interpreted as labels. We apply these