Contrastive Learning for Large-scale Color-Name Dataset: Tackling Sparsity with Negative Sampling

Paper Title

Contrastive Learning for Large-scale Color-Name Dataset: Tackling Sparsity with Negative Sampling

Publication Info

• Topic area: Computational modeling of color-name relationships for recommendation and generation tasks.
• Keywords: Contrastive learning, color naming, RGB encoding, multi-task learning, negative sampling, semantic embeddings, human-computer interaction, perceptual modeling, recommendation systems, generative models.

Background and Problem

• Problem / challenge: Existing models for color naming struggle with data sparsity, imbalanced distributions, and the many-to-many mapping between colors and names. They often fail to generalize under these conditions and lack bidirectional capabilities (color-to-name and name-to-color).
• Significance: Accurate color-name modeling supports applications in visualization, design, accessibility, and human-computer interaction, enabling intuitive and semantically consistent communication between humans and systems.
• Motivation and related work: Prior work has explored probabilistic models, rule-based systems, and neural approaches like Text2Color, but these methods are limited by one-directional mappings, sparse datasets, and insufficient embedding alignment. This paper addresses these gaps with a unified contrastive learning framework.

Solution

• Proposed approach: A contrastive learning framework that unifies color-to-name recommendation and name-to-color generation tasks within a shared embedding space, leveraging negative sampling and multi-task optimization.
• Novelty:
  1. Introduction of a multi-task architecture combining a pre-trained Transformer-based name encoder, an RGB encoder, and an RGB generator.
  2. Enhanced RGB representation using multi-color space features (e.g., HSL, LAB, HCL, CMYK) for richer semantic understanding.
  3. Use of negative sampling and contrastive learning to address data sparsity and imbalanced distributions.
  4. Dual-task design enabling both high-accuracy recommendations and fast, lightweight generation.
• Procedure and key techniques:
  • Preprocessing of the XKCD dataset to clean and normalize over 3 million color-name pairs.
  • Training a shared embedding space using contrastive learning with binary cross-entropy loss (for collaborative filtering) and mean squared error loss (for RGB generation).
  • Incorporating multi-color space features into the RGB encoder and leveraging a pre-trained BERT model for semantic name embeddings.
  • Introducing a Neural Collaborative Filtering (NCF) module for similarity computation and an RGB generator for fast inference.

Results

• Concrete findings:
  • Achieved 71.26% Top-10 accuracy in color-to-name recommendation (vs. 53.34% for Heer & Stone).
  • Reduced CIELAB perceptual error to 25.34 for name-to-color recommendation and 26.61 for name-to-color generation (vs. 27.24 for Text2Color).
  • Inference times of 0.43 ms/query (color-to-name) and 0.8 ms/query (name-to-color generation).
• Advantage over baselines:
  • +33.6% relative improvement in Top-10 accuracy for color-to-name recommendation compared to Heer & Stone.
  • 2% reduction in CIELAB error for name-to-color generation compared to Text2Color.
  • Faster inference for name-to-color generation compared to retrieval-based methods.
• Experiments / evaluation:
  • Conducted on a cleaned XKCD dataset with 10,000 test samples.
  • Metrics: Top-K accuracy for color-to-name, CIELAB distance for name-to-color, and inference time.
  • Ablation studies demonstrated the importance of enhanced RGB representation, negative sampling, NCF, and MSE loss.
• Limitations and future work:
  • Current work focuses on English color names; future work could extend to multilingual and cross-cultural datasets.
  • Further user studies are needed to validate practical utility.
  • Potential improvements in handling compositional and contextual color descriptions.

Summary

This paper introduces a contrastive learning framework for color-name recommendation and generation, addressing challenges of data sparsity, imbalanced distributions, and many-to-many mappings. By integrating enhanced RGB representations, negative sampling, and a multi-task architecture, the model achieves state-of-the-art performance, with significant improvements in accuracy and perceptual quality over baselines. The framework supports both high-accuracy recommendations and fast generation, making it suitable for real-world applications in design, visualization, and human-computer interaction. Future work includes multilingual extensions, user studies, and adaptive learning.