GGatrieval: Fine-grained Grounded Alignment Retrieval for Verifiable Generation

ACL ARR 2025 May Submission2330 Authors

Abstract: Retrieval-Augmented Generation (RAG) systems enhance large language models (LLMs) by incorporating retrieval mechanisms to address limitations in knowledge coverage. However, conventional retrieval mechanisms primarily operate at the sentence level, leading to semantic incompleteness at finer grammatical constituent granularities and thereby diminishing generation quality. To address this limitation, we propose GGatrieval (Fine-grained Grounded Alignment Retrieval for Verifiable Generation), a novel framework that refines retrieval by focusing on interactions at the level of grammatical constituents. Drawing from human cognitive processes, GGatrieval introduces a document selection criterion and assigns categorical labels using a Fine-grained Grounded Alignment strategy. These labels enable document reranking and support a Semantic Compensation Query Augmentation strategy, producing enriched queries that retrieve documents closely aligned with the original query. Experimental results on the ALCE benchmark and the extended Natural Questions datasets demonstrate GGatrieval's superior performance relative to mainstream baselines, with ablation studies confirming the effectiveness of our selection criteria and classification methodology.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: Dialogue and Interactive Systems,Generation,Information Retrieval and Text Mining,Language Modeling,Question Answering

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 2330