Data-Balanced Curriculum Learning for Audio Question Answering

2025-07-09

AI Summary

To address data imbalance and training instability in Audio Question Answering (AQA), this paper proposes a synergistic optimization framework integrating curriculum learning with statistical balancing. Methodologically, it (1) leverages large language models to automatically assess question difficulty and construct a progressive curriculum sequence; (2) applies class-level statistical filtering to mitigate long-tail bias inherent in audio modalities; and (3) incorporates format-guided decoding to constrain multiple-choice output generation, thereby enhancing reasoning consistency. Experiments on DCASE 2025 and five public benchmarks demonstrate substantial improvements in generalization: the proposed method achieves an average accuracy of 64.2%, outperforming strong baselines by +11.7 percentage points. These results establish a scalable, robust training paradigm for low-resource and non-stationary AQA tasks.

Technology Category

• Natural Language Processing: Question Answering
• Machine Learning: Calibration & Uncertainty Quantification
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Audio question answering (AQA) requires models to understand acoustic content and perform complex reasoning. Current models struggle with dataset imbalances and unstable training dynamics. This work combines curriculum learning with statistical data balancing to address these challenges. The method labels question difficulty using language models, then trains progressively from easy to hard examples. Statistical filtering removes overrepresented audio categories, and guided decoding constrains outputs to valid multiple-choice formats. Experiments on the DCASE 2025 training set and five additional public datasets show that data curation improves accuracy by 11.7% over baseline models, achieving 64.2% on the DCASE 2025 benchmark.

Problem

• Address dataset imbalances in audio question answering
• Stabilize training dynamics using curriculum learning
• Improve model accuracy through data balancing

Innovation

• Combines curriculum learning with data balancing
• Labels question difficulty using language models
• Uses statistical filtering for audio categories