Social Signal Processing and Computational Modelling for Small Groups

点击翻译按钮获取中文摘要

The goal of this project is to develop intelligent systems and computational models for the analysis and support of small group interaction and team collaboration in multiple modalities******Social Signal Recognition Models******The task in this part of the project is to take utterances or sentences from a group discussion, extract verbal and nonverbal features from them, and automatically recognize social signals using machine learning models. These social signals are observable behaviours that relate to group decision-making, such as expression of positive and negative sentiment, proposals, agreement, decisions, questions, and floor-holding. The features we utilize for machine learning will primarily relate to vocal behaviour, including the linguistic content of utterances, and nonverbal features such as prosody, laughter, and pauses. We will investigate the usefulness of these language features in conjunction with multi-layer neural networks. Since such deep learning models typically require a large amount of training data to be most effective, and the amount of available group interaction data is limited, we will look at domain adaptation from related domains such as email, social media, and instant messaging. We will develop information visualization techniques for explaining deep learning predictions.******Group Prediction Models / Complex Adaptive Systems******This project component involves making temporal predictions about group phenomena, such as overall group performance, group productivity, member satisfaction, group cohesion, emergent leadership, and overall dominance levels. We will use social network analysis (SNA) algorithms for modelling how social structure evolves during the decision-making task. We will use agent-based modelling (ABM) to simulate how individuals with particular conversational strategies can lead to complex group behaviour. Both SNA and ABM are particularly useful computational tools for treating small groups as complex adaptive systems. In such systems, individual agents adapt to each other and influence each other, leading to complex group behaviour that can include self-organization, emergent leadership, abrupt phase transitions, and sensitivity to initial conditions. ******Retrodiction Models / Hidden Interactions******In this project component, we will investigate methods for analyzing group interaction and dynamics in situations where direct observation of the group is limited. In those cases, we may want to use the limited data to predict what must have occurred in the past in order to bring the group to that state, requiring retrodictive models. For example, we may have artifacts from a meeting, such as participant notes, emails, and presentation slides, and want to predict how active different members had been in the meeting, or what their expressed sentiment was. ******The project will also include collection and annotation of group interaction data.**