RC-RAG: Paraphrase-Guided RAG for Relation Completion

Paper context

Paper title: Bridging the Long-Tail Gap: Robust Retrieval-Augmented Relation Completion via Multi-Stage Paraphrase Infusion Abstract: Large language models (LLMs) struggle with relation completion (RC), both with and without retrieval-augmented generation (RAG), particularly when the required information is rare or sparsely represented. To address this, we propose a novel multi-stage paraphrase-guided relation-completion framework, RC-RAG, that systematically incorporates relation paraphrases across multiple stages. In particular, RC-RAG: (a) integrates paraphrases into retrieval to expand lexical coverage of the relation, (b) uses paraphrases to generate relation-aware summaries, and (c) leverages paraphrases during generation to guide reasoning for relation completion. Importantly, our method does not require any model fine-tuning. Experiments with five LLMs on two benchmark datasets show that RC-RAG consistently outperforms several RAG baselines. In long-tail settings, the best-performing LLM augmented with RC-RAG improves by 40.6 Exact Match (EM) points over its standalone performance and surpasses two strong RAG baselines by 16.0 and 13.8 EM points, respectively, while maintaining low computational overhead. Passages referencing this figure: ramework remains robust, providing an initial step toward addressing long-tail challenges in relation completion. (4) We introduce a reproducible evaluation framework in which RAG baselines and RC-RAG operate under identical settings, using the same retrieval corpus, benchmark datasets, and evaluation metrics to ensure fair comparison 2 2 2 We will release all software and data upon acceptance. . Figure 1: Overview of our multi-stage paraphrase-guided RAG framework, RC-RAG for relation completion. Given an entity–relation query, we generate relation paraphrases of the target relation and inject them across three stages: (1) paraphrase-infused hybrid retrieval to expand semantic coverage, (2) paraphrase-guided evidence aggregation to extract relation-relevant information, and (3) paraphrase ency conditions. In this work, we take a first step toward addressing these challenges by studying retrieval performance under varying data frequency conditions. 3 Proposed Method We propose RC-RAG , a novel RAG framework for relation completion that consists of three stages: (1) paraphrase-infused hybrid retrieval, (2) paraphrase-guided evidence aggregation , and (3) paraphrase-guided reasoning. Figure 1 illustrates the framework design in detail. Relation Paraphrases Generation: To capture lexical variations of relations, we construct a relation paraphrase set for each target relation. For a relation r r , we define a paraphrase set P r = { p 1 , p 2 , … , p k } P_{r}=\{p_{1},p_{2},\dots,p_{k}\} , where each p i p_{i} represents a natural language expression conveying the same relational se set P r = { p 1 , p 2 , … , p k } P_{r}=\{p_{1},p_{2},\dots,p_{k}\} , where each p i p_{i} represents a natural language expression conveying the same relational meaning as r r . For example, for