Towards a Unified Model for Generating Answers and Explanations in Visual Question Answering

TL;DR

This paper introduces UMAE, a unified multimodal model that jointly generates answers and explanations in VQA, improving answer accuracy and explanation quality through multitask learning and prompt-based fine-tuning.

Contribution

The paper presents UMAE, a novel multitask learning framework that unifies answer and explanation generation in VQA using prompt tokens and a multimodal encoder-decoder model.

Findings

Surpasses prior answer accuracy on A-OKVQA by 10-15%.

Achieves new state-of-the-art explanation scores on A-OKVQA and VCR.

Demonstrates promising out-of-domain performance on VQA-X.

Abstract

The field of visual question answering (VQA) has recently seen a surge in research focused on providing explanations for predicted answers. However, current systems mostly rely on separate models to predict answers and generate explanations, leading to less grounded and frequently inconsistent results. To address this, we propose a multitask learning approach towards a Unified Model for Answer and Explanation generation (UMAE). Our approach involves the addition of artificial prompt tokens to training data and fine-tuning a multimodal encoder-decoder model on a variety of VQA-related tasks. In our experiments, UMAE models surpass the prior state-of-the-art answer accuracy on A-OKVQA by 10~15%, show competitive results on OK-VQA, achieve new state-of-the-art explanation scores on A-OKVQA and VCR, and demonstrate promising out-of-domain performance on VQA-X.