ERIC: an embodied real-time intelligent commentary agent

Future work includes reflecting the character's affect in the prosody of his speech, generating gestures based on the semantic content of his speech, and generalising the agent's production of coherent discourse.

The prosody module will take the character's current affective state, and produce prosodic specifications to depict the affective state. The prosody will be generated according to established relationships between affective state and speech prosody in humans.

The gesture module currently generates some gestures in response to events it receives from the knowledge reasoning module. It would be desirable to replace this link with a link to semantic information coming from the NLG module, thus making the gesture generation module domain-independent.

Discourse coherence is currently maintained by the fusion module, integrated with priority-based and random selection of a single utterance from all the candidates. We wish to have the discourse coherence selection separate from the final selection, such that an intermediate module between the NLG and fusion modules forms a shortlist of templates according to the discourse theory rules, and the fusion module only chooses from this shortlist. Then the role of maintaining discourse coherence could be filled by several interchangeable modules modelling different theories of discourse coherence.