Meeting decision detection: multimodal information fusion for multi-party dialogue understanding

Abstract

Modern advances in multimedia and storage technologies have led to huge archives of human conversations in widely ranging areas. These archives offer a wealth of information in the organization contexts. However, retrieving and managing information in these archives is a time-consuming and labor-intensive task. Previous research applied keyword and computer vision-based methods to do this. However, spontaneous conversations, complex in the use of multimodal cues and intricate in the interactions between multiple speakers, have posed new challenges to these methods. We need new techniques that can leverage the information hidden in multiple communication modalities – including not just “what” the speakers say but also “how” they express themselves and interact with others. In responding to this need, the thesis inquires into the multimodal nature of meeting dialogues and computational means to retrieve and manage the recorded meeting information. In particular, this thesis develops the Meeting Decision Detector (MDD) to detect and track decisions, one of the most important outcomes of the meetings. The MDD involves not only the generation of extractive summaries pertaining to the decisions (“decision detection”), but also the organization of a continuous stream of meeting speech into locally coherent segments (“discourse segmentation”). This inquiry starts with a corpus analysis which constitutes a comprehensive empirical study of the decision-indicative and segment-signalling cues in the meeting corpora. These cues are uncovered from a variety of communication modalities, including the words spoken, gesture and head movements, pitch and energy level, rate of speech, pauses, and use of subjective terms. While some of the cues match the previous findings of speech segmentation, some others have not been studied before. The analysis also provides empirical grounding for computing features and integrating them into a computational model. To handle the high-dimensional multimodal feature space in the meeting domain, this thesis compares empirically feature discriminability and feature pattern finding criteria. As the different knowledge sources are expected to capture different types of features, the thesis also experiments with methods that can harness synergy between the multiple knowledge sources. The problem formalization and the modeling algorithm so far correspond to an optimal setting: an off-line, post-meeting analysis scenario. However, ultimately the MDD is expected to be operated online – right after a meeting, or when a meeting is still in progress. Thus this thesis also explores techniques that help relax the optimal setting, especially those using only features that can be generated with a higher degree of automation. Empirically motivated experiments are designed to handle the corresponding performance degradation. Finally, with the users in mind, this thesis evaluates the use of query-focused summaries in a decision debriefing task, which is common in the organization context. The decision-focused extracts (which represent compressions of 1%) is compared against the general-purpose extractive summaries (which represent compressions of 10-40%). To examine the effect of model automation on the debriefing task, this evaluation experiments with three versions of decision-focused extracts, each relaxing one manual annotation constraint. Task performance is measured in actual task effectiveness, usergenerated report quality, and user-perceived success. The users’ clicking behaviors are also recorded and analyzed to understand how the users leverage the different versions of extractive summaries to produce abstractive summaries. The analysis framework and computational means developed in this work is expected to be useful for the creation of other dialogue understanding applications, especially those that require to uncover the implicit semantics of meeting dialogues.