Assessing Modality Bias in Video Question Answering Benchmarks with Multimodal Large Language Models

TL;DR

This work defines the modality importance score (MIS) to quantify how each modality contributes to answering VidQA questions, using both formal definitions and MLLM-derived estimates. Applying MIS to TVQA, LifeQA, and AVQA reveals substantial unimodal bias and a paucity of truly multimodal questions that require cross-modal integration. Human validation shows MIS aligns with perceived modality relevance, suggesting MIS can guide the curation of modality-balanced datasets and inform model design for better cross-modal reasoning. The study highlights practical implications for constructing robust VidQA benchmarks and advancing multimodal learning in MLLMs by prioritizing complementary questions and dynamic modality fusion strategies.

Abstract

Multimodal large language models (MLLMs) can simultaneously process visual, textual, and auditory data, capturing insights that complement human analysis. However, existing video question-answering (VidQA) benchmarks and datasets often exhibit a bias toward a single modality, despite the goal of requiring advanced reasoning skills that integrate diverse modalities to answer the queries. In this work, we introduce the modality importance score (MIS) to identify such bias. It is designed to assess which modality embeds the necessary information to answer the question. Additionally, we propose an innovative method using state-of-the-art MLLMs to estimate the modality importance, which can serve as a proxy for human judgments of modality perception. With this MIS, we demonstrate the presence of unimodal bias and the scarcity of genuinely multimodal questions in existing datasets. We further validate the modality importance score with multiple ablation studies to evaluate the performance of MLLMs on permuted feature sets. Our results indicate that current models do not effectively integrate information due to modality imbalance in existing datasets. Our proposed MLLM-derived MIS can guide the curation of modality-balanced datasets that advance multimodal learning and enhance MLLMs' capabilities to understand and utilize synergistic relations across modalities.

Figures (10)

• Figure 1: Example of a video clip with multimodal questions demonstrating different modality importance. $Q_1$ is answerable using either subtitle or video information, while $Q_2$ requires integrating information from both modalities. (Sec. \ref{['sec:category']})
• Figure 2: Question categorization based on human study vs MLLM-derived MIS
• Figure 3: Proportion of MIS based Question types per Annotated Answer Type
• Figure 4: Example from AVQA where annotated answer type is different from our categorization. For this video, the subtitle is "civil defense siren".
• Figure 5: Human modality perception study