Various research questions can be addressed within this general topic.
1. Searching for Patterns in Turn Sequences in Meeting Data
In the European AMI project four participant face to face meetings have been recorded (audio and video), transcribed, and annotated with dialogue acts, relations between dialogue acts (for instance question answer relations) and addressee labels (group, individual,undecided).
A simple sequential model of a meeting is a list of actions where an action is a triple
with S speaker, A addressee and D dialogue act (type), and T a temporal duration class (for example: short, middle, long). For each of the annotated meetings these sequences can be computed using available software packages.
In meetings several types of activities take place: sometimes there are group discussions about some particular topic, sometimes a single person has the floor giving a presentation, now and then interrupted by some question for elaboration. Some meetings or parts of meetings are well organized and one leading person plays a central role, some meetings are less well organized. These types of activities are expected to be related to particular patterns in action sequences. The question is whether we can identify clusters of action sequences that are related to different types of meeting activities.
What has to be done:
* select a number of features for modelling action sequences.
* evaluate the result of automatic clustering: do the clusters correspond to meaningful labels for various types of meeting activities?
Use matlab, weka or other available software for clustering and machine learning.
2. How Influential Are You?
In meetings some people are more dominant than others. From a positive point of view the dominant person may be seen as very knowledgeable and the other attendants may learn a lot from what he/she says. On the other hand, it could also be the case that the dominant behaviour has a negative effect on the meeting process by preventing others from voicing their opinions.
At the Human Media Interaction group, we have developed a simple model that is able to predict with a reasonable accuracy for each participant in a meeting whether this person is relatively 'high', 'moderate' or 'low' dominant. This model is based on three main features: the frequency of interruptions, the number of floor grabs and the total number of turns a participant took.
The research study aims at the creation of a system that implements the model described above for real-time application. The features mentioned above are to be obtained from the voice signals obtained from real meetings. The system can be developed using pre-recorded meetings.
Time permitting, an experiment can be carried out to in which the technology is applied to real-time meetings in which the influence and dominance levels that are calculated by the system are fed back into the meeting. The purpose of this experiment is to find out if and how the behaviour of the participants is affected by this kind of information.
For more information about the devised model, see:
For more information about a similar experiment, see: http://web.media.mit.edu/~joanie/second-messenger/chi04-doccon-dimicco.pdf
More information about this assignment? Contact:
University of Twente (HMI) is part of the location Twente