A Lightweight Videogame Dialogue Manager

Key insights

4 slides

The problem space

The paper targets conversation between videogame non-player characters.

It presents a lightweight procedural dialogue manager for videogame NPCs. The system extends the Talk of the Town framework.

Key number

Generated conversations scored 4.05 on naturalness, versus 1.34 for random.

In a preliminary evaluation, procedurally generated conversations were rated substantially more natural than randomly assembled ones. The difference was statistically significant.

Scope of result. The evaluation measures conversational flow controlled by the dialogue manager rather than content quality or other system aspects.

1 / 4

Ask AI about this report

Summary

This paper presents a lightweight procedural dialogue manager designed for videogame non-player characters, extending the Talk of the Town framework so that knowledge exchanges between NPCs render as naturalistic conversations rather than purely symbolic transfers. Its central argument is that this approach avoids the authoring burden, limited player agency, and weak state-responsiveness associated with conventional branching dialogue, while producing socially believable exchanges. The architecture rests on extensible "dialogue moves" that resemble speech acts, preauthored conversational goal plans, and stackable conversational frames activated by world-state preconditions, with obligations handling local coherence and goals handling global coherence.

The system integrates four knowledge-propagation evidence types, namely statement, declaration, lie, and eavesdropping, and draws on per-line metadata covering preconditions, moves, obligations, lie conditions, and effects. Lines are selected through disjoint preconditions and an Expressionist-based natural language generation module, after which conversation state, world state, and character beliefs are updated upon deployment. The paper positions this work against related systems including Versu, Façade, Bot Colony, and Inform 7's Threaded Conversation.

The principal empirical result comes from a preliminary human evaluation with seven respondents, who rated procedurally generated conversations as substantially more natural (mean 4.05, σ = 0.97) than randomly assembled ones (mean 1.34, σ = 0.82), a statistically significant difference (t(418) = 30.96; p < 0.0001). The authors are careful to note that this measures conversational flow controlled by the dialogue manager rather than content quality or other system aspects. Architectural sections are supported by illustrative examples rather than broader data, and the closing references situate the work within interactive narrative, combinatorial dialogue authoring, emergent narrative, and information-state approaches to dialogue.