IMPROVE Visiolinguistic Performance with Re-Query

We humans regularly ask for clarification if we are confused when discussing the visual world, yet the commonplace requirement in visiolinguistic problems like Visual Dialog, VQA, and Referring Expression Comprehension is to force a decision based on a single, static language input. Since this assumption does not match human practice, we relax it and allow our model to request new language inputs to refine the prediction for a task. Through the exemplar task of referring expression comprehension, we formalize and motivate the problem, introduce an evaluation method, and propose \textit{Iterative Multiplication of Probabilities for Re-query Of Verbal Expressions} (IMPROVE) -- a re-query method that updates the model's prediction across multiple queries. We demonstrate IMPROVE on two different referring expression comprehension models and show it can improve accuracy by up to 6.23% without additional training or modification to the model's architecture.