Investigate alternative tree encoding mechanisms for dialogue generation

Investigate encoding mechanisms for representing dialogue subtrees in knowledge-constrained NPC dialogue generation beyond linearization, including graph encoder approaches, to better capture branching and cyclic dialogue structures.

Background

The authors currently linearize dialogue subtrees into maximal coverage paths to accommodate sequence-based LLMs, acknowledging that this simplification may not fully capture the complex branching and cycles in RPG dialogue graphs.

They explicitly state that exploring alternative encoding strategies, such as graph encoders, is deferred to future work, indicating an unresolved methodological direction that could improve modeling fidelity for tree-structured dialogue generation.