AI RESEARCH
Beyond Cosine Similarity: Zero-Initialized Residual Complex Projection for Aspect-Based Sentiment Analysis
arXiv CS.CL
•
ArXi:2603.28205v1 Announce Type: new Aspect-Based Sentiment Analysis (ABSA) is fundamentally challenged by representation entanglement, where aspect semantics and sentiment polarities are often conflated in real-valued embedding spaces. Furthermore, standard contrastive learning suffers from false-negative collisions, severely degrading performance on high-frequency aspects. In this paper, we propose a novel framework featuring a Zero-Initialized Residual Complex Projection (ZRCP) and an Anti-collision Masked Angle Loss,inspired by quantum projection and entanglement ideas.