To provide a more meaningful feedback, the system needs to know not only the topic of the conversation, but also have an idea of the emotional content of the conversation and the conversational style of the interaction. Negative and positive attitudes can be detected based on the word choice and would require a two-class lexical modeling, similarly to the method of topic classification.
In addition, an interaction can take a formal (a presentation with a single principal speaker, a form-based dialog) or an informal character (group discussion). Assessment of the degree of formality can aid the system in learning the appropriate strategy of the interrupt control (see section 5.3).
These cues can help the system learn when it is appropriate to intervene in the interaction and more accurately rate its feedback to the users. We are currently exploring these aspects of the conversation modeling for use in our system.