Abstract
Grounded conversational agents are a fascinating research line on which important progress has been made lately thanks to the development of neural network models and to the release of visual dialogue datasets. The latter have been used to set visual dialogue games which are an interesting test bed to evaluate conversational agents. Researchers’ attention is on building models of increasing complexity, trained with computationally costly machine learning paradigms that lead to higher task success scores. In this paper, we take a step back: We use a rather simple neural network architecture and we scrutinize the GuessWhich task, the dataset, and the quality of the generated dialogues. We show that our simple Questioner agent reaches state-of-the art performance, that the evaluation metric commonly used is too coarse to compare different models, and that high task success does not correspond to high quality of the dialogues. Our work shows the importance of running detailed analyses of the results to spot possible models’ weaknesses rather than aiming to outperform state-of-the-art scores