HNC: Leveraging Hard Negative Captions towards Models with Fine-Grained Visual-Linguistic Comprehension Capabilities

2026-05-06

AI Summary

This work addresses the limitations of existing image–text matching methods, which struggle to achieve fine-grained cross-modal semantic understanding due to the weak alignment inherent in web-collected data. To overcome this, the authors propose a hard negative caption generation mechanism that synthesizes a Hard Negative Caption (HNC) dataset for training, complemented by a human-designed, multi-level fine-grained mismatch evaluation benchmark. Integrating an image–text matching framework with zero-shot transfer and robust training strategies, the approach significantly enhances model performance in diagnostic mismatch detection and under noisy input conditions. Furthermore, it provides superior or competitive initialization for fine-tuning on downstream tasks, thereby advancing vision–language models toward more precise and nuanced semantic comprehension.

Abstract

Image-Text-Matching (ITM) is one of the defacto methods of learning generalized representations from a large corpus in Vision and Language (VL). However, due to the weak association between the web-collected image-text pairs, models fail to show a fine-grained understanding of the combined semantics of these modalities. To address this issue we propose Hard Negative Captions (HNC): an automatically created dataset containing foiled hard negative captions for ITM training towards achieving fine-grained cross-modal comprehension in VL. Additionally, we provide a challenging manually-created test set for benchmarking models on a fine-grained cross-modal mismatch task with varying levels of compositional complexity. Our results show the effectiveness of training on HNC by improving the models' zero-shot capabilities in detecting mismatches on diagnostic tasks and performing robustly under noisy visual input scenarios. Also, we demonstrate that HNC models yield a comparable or better initialization for fine-tuning

Problem

• Image-Text Matching
• Fine-Grained Comprehension
• Vision and Language
• Hard Negative Captions
• Cross-Modal Understanding

Innovation

• Hard Negative Captions
• Image-Text Matching
• Fine-Grained Comprehension
• Vision-Language Models
• Cross-Modal Mismatch