Context-Aware Text Re-Ranking through Unsupervised Manifold Learning

Abstract

Information Retrieval (IR) enables efficient knowledge discovery within large document collections. Despite the evolution from sparse lexical models such as BM25 to dense neural embeddings, retrieval ranking still relies predominantly on independent query–document scoring, and most re-ranking strategies require supervised relevance signals. In image retrieval, unsupervised methods that exploit manifold structure and contextual neighborhood relationships have proven highly effective, yet their transfer to text remains unexplored. This study investigates whether such rank-based contextual re-ranking techniques can enhance text retrieval without any modality-specific adaptation. We present a systematic pipeline integrating data collection, sparse and dense retrieval, contextual re-ranking, and effectiveness analysis, and conduct 972 experiments spanning three unsupervised methods across nine diverse public datasets and four retrieval models. Results measured by MAP, Precision, and Recall show consistent improvements over baseline retrievers, with relative Precision@20 gains reaching up to +19.8%. These findings support the hypothesis that unsupervised contextual re-ranking can significantly improve text retrieval effectiveness and suggest that the underlying techniques are modality-agnostic.