Structuring Semantic Embeddings for Principle Evaluation: A Prototype-Guided Contrastive Learning Approach

2026-08-15

AI Summary

This study addresses the insufficient discriminability of general-purpose text embeddings in principled evaluation caused by semantic overlap. To mitigate this, we propose Prototype-Guided Contrastive Learning (PGCL), which integrates prototype-anchored attention with shifted margin regularization to generate compact, task-adaptive representations while keeping the encoder frozen. This approach effectively resolves confusion between semantically similar samples with distinct task labels. Experimental results demonstrate that PGCL outperforms original frozen embeddings across three benchmark datasets, achieving particularly significant improvements on Amazon Reviews and matching strong baselines on other tasks. These findings validate the effectiveness of PGCL for parameter-efficient adaptation, offering a robust solution for enhancing embedding distinctiveness without extensive model retraining.

Abstract

Reliable post-hoc evaluation asks whether already generated text satisfies a target criterion after generation. In this paper we study a focused frozen-embedding setting using principle-evaluation proxy tasks: toxicity detection, fine-grained emotion categorization, and ordinal review rating. General-purpose text embeddings are widely deployed for such tasks, but broad semantic similarity can place semantically similar yet task-distinct examples in overlapping regions of the representation space. We introduce Prototype-Guided Contrastive Learning (PGCL), a prototype-guided geometric regularization module built on top of frozen text embeddings. The module combines a semantic stream, a prototype-anchor attention stream, supervised contrastive learning, offset-based prototype-margin regularization, and stream regularization to produce a compact task-adapted representation without updating the base encoder. Controlled experiments show that PGCL improves over raw frozen embeddings on all three datasets and gives the clearest direct-baseline margin on AmazonReviews, while remaining competitive with strong direct frozen metric-learning baselines on GoEmotions and ToxicComment. We also add supervised residual-adapter, encoder-LoRA, full fine-tuning, objective ablation, sensitivity, and fully logged few-shot LLM protocol diagnostics to define the boundary of the claim. The theoretical analysis is revised as a sufficient-condition account for prototype-margin behavior under explicit assumptions in the prototype-mapping space, rather than as an unconditional training or final-embedding separation guarantee.

Problem

• Principle Evaluation
• Frozen Embeddings
• Semantic Overlap
• Representation Space

Innovation

• Prototype-Guided Contrastive Learning
• Frozen Embeddings
• Geometric Regularization
• Principle Evaluation
• Prototype-Margin