Can Visual Dialogue Models Do Scorekeeping? Exploring How Dialogue Representations Incrementally Encode Shared Knowledge
This work addresses the need for cognitively plausible dialogue models, though it is incremental as it focuses on evaluation rather than proposing a new method.
The paper tackled the problem of evaluating whether visual dialogue models can incrementally track shared knowledge, finding that models pretrained on VisDial show moderate but inconsistent scorekeeping ability, partially due to limited grounding interactions in the original task.
Cognitively plausible visual dialogue models should keep a mental scoreboard of shared established facts in the dialogue context. We propose a theory-based evaluation method for investigating to what degree models pretrained on the VisDial dataset incrementally build representations that appropriately do scorekeeping. Our conclusion is that the ability to make the distinction between shared and privately known statements along the dialogue is moderately present in the analysed models, but not always incrementally consistent, which may partially be due to the limited need for grounding interactions in the original task.