This thesis is concerned with automatic addressee identification in face-to-face meetings. The first part of the thesis is devoted to gaining theoretical insights into addressing based on the outcomes of the research in conversational and interaction analysis. The multimodal nature of addressing as well as its context sensitivity poses challenges not only for computational systems but also for humans in determining who is being addressed by the speaker. This thesis addresses both challenges. The second part of the thesis describes two meeting corpora employed in our study. The corpora were developed using annotation schemes designed in the interaction between data observed and theoretical insights into addressing obtained from the literature. To assess the credibility of the annotated data, the thesis provides an exhaustive reliability analysis of the annotation schemes. A detailed investigation of the problems human observers had in determining who is being addressed by the speaker shows that it is intrinsically difficult to distinguish between group and individual addressing. The third part of the thesis deals with the development of a computational model for automatic addressee dentification using Bayesian Networks. Features employed to model addressing are obtained from speech and gaze communication channels as well as from conversational and meeting contexts. Conversational context features are shown to be the most valuable, whereas utterance features are found to be the least reliable cues for addressee prediction. Evaluation of several types of Bayesian Network classifiers indicates that Bayesian Networks are effective computational models for the task of addressee prediction. The highest accuracies of addressee classifiers (about 80%) are achieved by combining features from all available resources. The features used to model addressing are in a sense ideal because they are obtained from human-annotated data. As the first step in the automation of the addressee detection process, the thesis proposes a Dynamic Bayesian Network model. The evaluation of performances of addressee classifiers relying on fully automatic features remains one of the tasks for future research.
|Award date||14 Mar 2007|
|Place of Publication||Enschede, The Netherlands|
|Publication status||Published - 14 Mar 2007|