Structural Information Representation for Retrieval over Hierarchical Technical Documents: Insights from 3GPP Standards

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Retrieval-augmented generation provides an effective solution for domain-specific question answering by incorporating external knowledge. However, existing RAG systems typically represent retrieved knowledge as plain-text chunks, with limited attention to structural information in hierarchical technical documents. Different levels of structural information vary in semantic granularity, contextual scope, and their relationships with textual content, and adopting a unified representation strategy may limit retrieval effectiveness. This paper investigates structural information representation in hierarchical technical documents and its impact on end-to-end question answering performance, using 3GPP technical specifications as a representative scenario. First, a Structure-aware Knowledge Retrieval Framework is established to preserve document hierarchy and extract multi-level structural information. Based on this framework, three representation strategies, namely Independent Representation, Fusion Representation, and Hybrid Representation, are systematically compared through controlled experiments. The experimental results show that section-level structure exhibits strong semantic consistency with textual content and is therefore more suitable for Fusion Representation, whereas document-level structure is more suitable as an independent retrieval signal. Based on these findings, the Hybrid Representation strategy is adopted and combined with an answer fallback mechanism to construct the final framework. Evaluation on the 3GPP subset of TeleQnA, with Qwen3.7-Max as the generator, shows that the proposed framework achieves a question answering accuracy of 84.9%. These results indicate that differentiated representation according to the semantic level of structural information is beneficial for improving end-to-end question answering performance over hierarchical technical documents.