A Contrastive Supervised Learning Framework for Reliable Reviewer-Manuscript Matching

Automated reviewer-manuscript matching is a core element of contemporary peer-review systems, often posed as a similarity-based retrieval task on pretrained text embeddings. Although the current methods perform well on ranking measures like Mean Reciprocal Rank (MRR) and NDCG, current methods implicitly posit that the closer the similarity, the greater the assignment confidence. In this work, we demonstrate that this assumption is actually flawed because of the role of embedding geometry that is often ignored. With a thorough analysis of various transformer-based encoders and a variety of scientific fields, we show that pretrained embeddings possess strong inter-intra domain overlap, anisotropy, and weak separability, which results in ambiguous similarity distribution and unreliable assignment. We also demonstrate that traditional ranking measures do not reflect confidence in margins and absolute separability, and tend to hide geometric flaws. In order to overcome these constraints, we introduce a Contrastive Geometry Refinement Framework, which explicitly re-structures the embedding space on the basis of supervised contrastive learning. The proposed solution creates intra-domain compactness and inter-domain separation, which leads to better centroid separation, larger similarity margins and higher discriminability. Theoretical studies prove that contrastive optimization enhances inter-domain separation and the intra-domain variance is decreased. Empirical findings show that there are significant improvements in many metrics, such as almost 2 times improvement in MRR and NDCG, amplification in margins, and thereby geometric disentanglements in domains. We find that incorporating geometry and not similarity magnitude is the key predictor of trustworthy reviewer assignment. This paper offers a principled basis to enhance performance and trust in scholarly recommendation systems by redefining reviewer recommendation as a geometry-guided contrastive learning problem.