---
title: "基于解耦与组合的大模型知识问答方法 — 施圣泽 2026"
description: "Classic knowledge base question answering (KBQA) systems need large annotated datasets and heavy training, and they adapt poorly to unseen knowledge bases."
authors: ["施圣泽", "秦利", "胡军", "鲁晶晶", "王丹丹"]
year: 2026
venue: "中文信息学报 (Journal of Chinese Information Processing)"
source: "http://jcip.cipsc.org.cn/article/doi/10.3969/j.issn.1003-0077.2026.08.008"
doi: "10.3969/j.issn.1003-0077.2026.08.008"
licence: "Free access (jcip.cipsc.org.cn), no open licence stated — readonly; © 中文信息学报"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/papers/shi-2026-leveraging-decoupled-composition/"
---
[Elamur Research](https://md.research.elamur.ai/) · [this page in HTML](https://research.elamur.ai/free/papers/shi-2026-leveraging-decoupled-composition/)

[home](https://md.research.elamur.ai/) / [papers](https://md.research.elamur.ai/free/papers/) / [2026](https://md.research.elamur.ai/free/papers/2026/page/5/) / 基于解耦与组合的大模型知识问答方法

# 基于解耦与组合的大模型知识问答方法

**Authors:** 施圣泽, 秦利, 胡军, 鲁晶晶, 王丹丹 · **Year:** 2026 · **Venue:** 中文信息学报 (Journal of Chinese Information Processing) · **Licence:** Free access (jcip.cipsc.org.cn), no open licence stated — readonly; © 中文信息学报

**Links:** [Publisher's page](https://doi.org/10.3969/j.issn.1003-0077.2026.08.008) · [PDF](http://jcip.cipsc.org.cn/cn/article/pdf/preview/10.3969/j.issn.1003-0077.2026.08.008.pdf) · [DOI 10.3969/j.issn.1003-0077.2026.08.008](https://doi.org/10.3969/j.issn.1003-0077.2026.08.008)

## Summary
Classic knowledge base question answering (KBQA) systems need large annotated datasets and heavy training, and they adapt poorly to unseen knowledge bases. LLMs can generate code, but they struggle to write complete KB queries in one step. Their drafts often fail on syntax, logic type and hop count, and they are easily misled by too many or irrelevant facts. The authors propose DC-KBQA, a training-free framework that splits query generation into two parts. The first is a reasoning skeleton, a logical form whose entities, relations, classes and literals are replaced by placeholders. The second is the set of semantic components that fill it. An LLM-guided retriever pulls a few candidate skeletons from a code base of example queries and a few components from the knowledge base. Ontology- and instance-level schema checks prune invalid components. An LLM-guided composer then fills the skeletons and picks the best query by logical soundness and semantic consistency. On GrailQA, WebQSP and GraphQA, DC-KBQA beats prompting-only LLMs and the frameworks KB-Binder and KB-Coder. The largest gains come in zero-shot generalization, showing that retrieval plus composition makes queries more executable than free generation.

## Key points
- **Data:** evaluated on three Freebase-based benchmarks, GrailQA (64,331 questions), WebQSP (4,737) and GraphQA (5,166), using S-expression logical forms. Metrics are EM and F1 on GrailQA and F1 on the other two.
- **Setup:**
  - BM25 retrieves 50 similar example questions, and the 5 most frequent skeletons are kept.
  - The LLM labels each question's logic type as plain, count, superlative or comparison, and skeletons are filtered to match that type.
  - 100 candidate relations are retrieved, schema-checked and narrowed by the LLM to the 10 most relevant.
- **GrailQA results (overall):** EM/F1 of 48.0/50.4, versus 40.9/46.3 for KB-Coder and 30.1/31.7 for GPT-4. On the zero-shot split, F1 is 50.8 against KB-Coder's 43.2. On the compositional split, F1 (49.1) is slightly below KB-Coder (49.6).
- **WebQSP and GraphQA:** F1 is 61.3 on WebQSP (KB-Coder 60.5, KB-Binder 56.6) and 37.5 on GraphQA (KB-Coder 35.8).
- **Ablations and parameters:** performance drops when any of these is removed: retrieved skeletons (LK-KBQA), logic-type filtering (LT-KBQA), schema checks (SC-KBQA) or LLM-based composition (CB-KBQA). On GrailQA, 100–200 semantic components works best. Too many skeletons or components adds noise.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/shelves/linguistics/) · [Natural language processing](https://md.research.elamur.ai/free/shelves/nlp/)

## Cite

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@article{2026,
  title = {{基于解耦与组合的大模型知识问答方法}},
  author = {施圣泽 and 秦利 and 胡军 and 鲁晶晶 and 王丹丹},
  journal = {中文信息学报 (Journal of Chinese Information Processing)},
  year = {2026},
  doi = {10.3969/j.issn.1003-0077.2026.08.008},
  url = {https://doi.org/10.3969/j.issn.1003-0077.2026.08.008}
}
```

- [← previous 基于检索增强思维提示的汉语框架语义解析方法](https://md.research.elamur.ai/free/papers/li-2026-retrieval-augmented-chain/)
- [next → 基于解题思维机制的大语言模型多项选择阅读理解增强框架](https://md.research.elamur.ai/free/papers/su-2026-facilitating-large-language/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/papers/2026/page/5/)