English

A First Look: Towards Explainable TextVQA Models via Visual and Textual Explanations

Computer Vision and Pattern Recognition 2021-05-07 v1 Artificial Intelligence Computation and Language Machine Learning

Abstract

Explainable deep learning models are advantageous in many situations. Prior work mostly provide unimodal explanations through post-hoc approaches not part of the original system design. Explanation mechanisms also ignore useful textual information present in images. In this paper, we propose MTXNet, an end-to-end trainable multimodal architecture to generate multimodal explanations, which focuses on the text in the image. We curate a novel dataset TextVQA-X, containing ground truth visual and multi-reference textual explanations that can be leveraged during both training and evaluation. We then quantitatively show that training with multimodal explanations complements model performance and surpasses unimodal baselines by up to 7% in CIDEr scores and 2% in IoU. More importantly, we demonstrate that the multimodal explanations are consistent with human interpretations, help justify the models' decision, and provide useful insights to help diagnose an incorrect prediction. Finally, we describe a real-world e-commerce application for using the generated multimodal explanations.

Keywords

Cite

@article{arxiv.2105.02626,
  title  = {A First Look: Towards Explainable TextVQA Models via Visual and Textual Explanations},
  author = {Varun Nagaraj Rao and Xingjian Zhen and Karen Hovsepian and Mingwei Shen},
  journal= {arXiv preprint arXiv:2105.02626},
  year   = {2021}
}

Comments

This paper is done when Xingjian was an intern in Amazon PARS group, summer 2020. This paper is accepted by NAACL-MAI-Workshop, 2021

R2 v1 2026-06-24T01:50:14.686Z