PRISM: Fine-Grained Paper-to-Paper Retrieval with Multi-Aspect-Aware Query Optimization

2025-07-14

AI Summary

Current scientific paper retrieval predominantly relies on abstract-based modeling, which fails to capture multi-dimensional semantics across full texts, resulting in insufficient fine-grained matching capability. To address this, we propose a full-text segment-aware, multi-perspective document retrieval framework: it decomposes query papers into semantic views (e.g., method, experiment, conclusion) and performs aspect-aligned, fine-grained matching against corresponding sections of candidate papers; further, it integrates dense retrieval with paragraph-level semantic modeling via a multi-perspective query optimization mechanism and aspect-specific embedding strategy. We introduce SciFullBench—the first benchmark supporting full-text segment-level contrastive evaluation. On SciFullBench, our approach achieves an average 4.3% improvement in NDCG@10 over state-of-the-art baselines, significantly enhancing the discovery of deep cross-paper semantic associations.

Technology Category

• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Computer Vision: Image and Video Retrieval
• Data Mining & Knowledge Management: Intelligent Query Processing

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Graph Algorithms and Modeling for the Web: Querying, indexing, and retrieval in Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Scientific paper retrieval, particularly framed as document-to-document retrieval, aims to identify relevant papers in response to a long-form query paper, rather than a short query string. Previous approaches to this task have focused on abstracts, embedding them into dense vectors as surrogates for full documents and calculating similarity across them, although abstracts provide only sparse and high-level summaries. To address this, we propose PRISM, a novel document-to-document retrieval method that introduces multiple, fine-grained representations for both the query and candidate papers. In particular, each query paper is decomposed into multiple aspect-specific views and individually embedded, which are then matched against candidate papers similarity segmented to consider their multifaceted dimensions. Moreover, we present SciFullBench, a novel benchmark in which the complete and segmented context of full papers for both queries and candidates is available. Then, experimental results show that PRISM improves performance by an average of 4.3% over existing retrieval baselines.

Problem

• Enhances paper retrieval using multi-aspect fine-grained representations
• Addresses limitations of abstract-only dense vector embeddings
• Introduces segmented matching for multifaceted document dimensions

Innovation

• Fine-grained multi-aspect paper representations
• Aspect-specific embedding and matching
• Novel benchmark with full paper context