AdvSumm: Adversarial Training for Bias Mitigation in Text Summarization

2025-06-06

AI Summary

Large language models (LLMs) often inherit spurious correlations and framing biases from training data, leading to unfair text summarization outputs. To address this, we propose a domain-agnostic adversarial training framework that enables universal debiasing of abstractive summarization without requiring bias labels or domain-specific knowledge. Our method, built upon a Seq2Seq architecture, introduces a novel Embedding-layer Gradient-guided Perturber (EGP) module that injects targeted perturbations into token embeddings via gradient-based guidance; integrates an adversarial loss to suppress bias-confounded representations; and employs a multi-dimensional bias evaluation protocol for robust assessment. Experiments across multiple benchmarks demonstrate significant mitigation of name–nationality and political stance biases, while preserving ROUGE scores—outperforming standard Transformer baselines and back-translation–enhanced methods.

Technology Category

• Natural Language Processing: Summarization
• Machine Learning: Adversarial Learning & Robustness
• Multiagent Systems: Adversarial Agents

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Large Language Models (LLMs) have achieved impressive performance in text summarization and are increasingly deployed in real-world applications. However, these systems often inherit associative and framing biases from pre-training data, leading to inappropriate or unfair outputs in downstream tasks. In this work, we present AdvSumm (Adversarial Summarization), a domain-agnostic training framework designed to mitigate bias in text summarization through improved generalization. Inspired by adversarial robustness, AdvSumm introduces a novel Perturber component that applies gradient-guided perturbations at the embedding level of Sequence-to-Sequence models, enhancing the model's robustness to input variations. We empirically demonstrate that AdvSumm effectively reduces different types of bias in summarization-specifically, name-nationality bias and political framing bias-without compromising summarization quality. Compared to standard transformers and data augmentation techniques like back-translation, AdvSumm achieves stronger bias mitigation performance across benchmark datasets.

Problem

• Mitigates bias in text summarization models
• Reduces name-nationality and political framing biases
• Enhances robustness without compromising summarization quality

Innovation

• Adversarial training framework for bias mitigation
• Gradient-guided perturbations at embedding level
• Reduces bias without compromising summarization quality