H-RAG at SemEval-2026 Task 8: Hierarchical Parent-Child Retrieval for Multi-Turn RAG Conversations
| Authors | Passant Elchafei et al. |
| Year | 2026 |
| Field | NLP |
| arXiv | 2605.00631 |
| Download | |
| Categories | cs.CL, cs.IR |
Abstract
We present H-RAG, our submission to SemEval-2026 Task 8 (MTRAGEval), addressing both Task A (Retrieval) and Task C (Generation with Retrieved Passages). Task A evaluates standalone retrieval quality, while Task C assesses end-to-end retrieval-augmented generation (RAG) in multi-turn conversational settings, requiring both accurate answer generation and faithful grounding in retrieved evidence. Our approach implements a hierarchical parent-child RAG pipeline that separates fine-grained child-level retrieval from parent-level context reconstruction during generation. Documents are segmented into overlapping sentence-based child chunks, while full documents are preserved as parent units to provide coherent context. Retrieval combines hybrid dense-sparse search, tunable weighting, and embedding-based similarity rescoring over child chunks. Retrieved evidence is aggregated at the parent level and supplied to an instruction-tuned language model for response generation. H-RAG achieves an nDCG@5 score of 0.4271 on Task A and a harmonic mean score of 0.3241 on Task C (RB_agg: 0.2488, RL_F: 0.2703, RB_llm: 0.6508), underscoring the importance of retrieval configuration and parent-level aggregation in multi-turn RAG performance.
Engineering Breakdown
Plain English
This paper presents H-RAG, a hierarchical retrieval-augmented generation system submitted to SemEval-2026 Task 8, which tackles both standalone document retrieval and end-to-end RAG in multi-turn conversations. The key innovation is separating retrieval into fine-grained sentence-level chunks (children) while preserving full documents (parents) for coherent context during generation, using hybrid dense-sparse search with tunable weighting. The approach addresses a real production pain point: retrieval systems often return fragmented passages that lack sufficient context for accurate answer generation, while full-document retrieval sacrifices precision. By combining both levels hierarchically, H-RAG aims to improve both retrieval accuracy and generation faithfulness—two metrics that typically trade off against each other in RAG pipelines.
Core Technical Contribution
The core technical novelty is the hierarchical parent-child retrieval architecture that decouples retrieval granularity from generation context requirements. Unlike traditional RAG systems that retrieve uniform-sized passages or rely solely on dense or sparse retrieval, H-RAG explicitly models two levels: sentence-level child chunks for precise matching during retrieval, and full document parents for maintaining coherent narrative context during generation. This separation allows the system to achieve fine-grained retrieval precision without sacrificing the contextual coherence needed for faithful answer generation. The hybrid dense-sparse search with tunable weighting is a secondary contribution that lets practitioners balance lexical and semantic matching for their specific domain.
How It Works
The pipeline operates in three stages: (1) Document preprocessing segments source documents into overlapping sentence-based child chunks while maintaining pointers to their parent full documents; (2) Retrieval uses hybrid dense-sparse search where dense embeddings capture semantic similarity and sparse lexical matching captures keyword overlap, with a tunable weight parameter controlling their contribution; (3) Generation takes the retrieved child chunks, maps them back to their parents to reconstruct document-level context, and then feeds both the precise child passages and fuller parent context to a language model for answer generation. For multi-turn conversations, the system maintains conversational history and grounds each turn's answer in retrieved evidence, evaluating both answer correctness and whether the answer is actually supported by the retrieved passages. The overlapping sentence chunks ensure that important context spans are not lost at chunk boundaries, a common failure mode in chunked retrieval systems.
Production Impact
This approach directly addresses a major pain point in production RAG systems: the chunk-size dilemma where small chunks harm context quality and large chunks reduce retrieval precision. By separating retrieval targets (sentence chunks) from generation context (full documents), you get the best of both worlds without redesigning your entire pipeline. In practice, this means your retrieval component can remain accurate and fast—querying against sentence-level embeddings—while your generation component receives richer, more coherent context, improving answer quality and reducing hallucinations. The trade-off is moderate: you need to maintain both chunk-level and document-level indices (roughly 1.5-2x the storage), and the parent-mapping step adds minimal latency (microseconds per retrieved item). For conversational systems, the explicit grounding mechanism helps with explainability and trust, letting you show users exactly which source document passages support each answer.
Limitations and When Not to Use This
The paper's scope is limited to SemEval-2026 Task 8 benchmarks, which may not represent production query distributions—conversational evaluation datasets often skew toward factoid questions where retrieval-based grounding works well, but struggle with reasoning-heavy or multi-hop queries requiring synthesis across passages. The hierarchical approach assumes documents have meaningful sentence-level structure; it will degrade on poorly formatted documents, code snippets, or domains with non-standard formatting. The tunable weighting between dense and sparse search requires careful hyperparameter tuning per domain, and the paper likely doesn't fully explore sensitivity to these choices or provide guidance on setting them in new domains. Finally, the computational cost of maintaining dual indices and performing parent-child mapping at scale isn't discussed—this could become a bottleneck for systems with millions of documents.
Research Context
H-RAG builds on established RAG foundations (Karpukhin et al., 2020; Lewis et al., 2020) and extends them with hierarchical structure inspired by document-level retrieval work. The paper responds directly to observed limitations in multi-turn conversational RAG where retrieved snippets lack sufficient context for faithful generation—a problem highlighted in recent MTRAGEval benchmarks and conversational QA research. The hybrid dense-sparse retrieval approach draws from ColBERT and BM25 fusion techniques, applying them specifically to the hierarchical chunk structure. By focusing on both retrieval quality (Task A) and generation faithfulness (Task C), the paper addresses the increasingly important problem of evaluating whether RAG outputs are actually grounded in retrieved evidence, rather than just whether they're correct—a distinction becoming critical for enterprise and scientific applications.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::