---
title: "基于规则提示的知识图谱通用推理预训练模型 — 崔员宁 2024"
description: "Knowledge graphs are usually incomplete. Most completion models are trained separately for each graph and learn embeddings tied to that graph's own…"
authors: ["崔员宁", "孙泽群", "胡伟"]
year: 2024
venue: "计算机研究与发展 (Journal of Computer Research and Development)"
source: "https://crad.ict.ac.cn/article/doi/10.7544/issn1000-1239.202440133"
doi: "10.7544/issn1000-1239.202440133"
licence: "Free access (crad.ict.ac.cn), no open licence stated — readonly; © 计算机研究与发展"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/papers/cui-2024-pre-trained-universal/"
---
[Elamur Research](https://md.research.elamur.ai/) · [this page in HTML](https://research.elamur.ai/free/papers/cui-2024-pre-trained-universal/)

[home](https://md.research.elamur.ai/) / [papers](https://md.research.elamur.ai/free/papers/) / [2024](https://md.research.elamur.ai/free/papers/2024/page/4/) / 基于规则提示的知识图谱通用推理预训练模型

# 基于规则提示的知识图谱通用推理预训练模型

**Authors:** 崔员宁, 孙泽群, 胡伟 · **Year:** 2024 · **Venue:** 计算机研究与发展 (Journal of Computer Research and Development) · **Licence:** Free access (crad.ict.ac.cn), no open licence stated — readonly; © 计算机研究与发展

**Links:** [Publisher's page](https://doi.org/10.7544/issn1000-1239.202440133) · [PDF](https://crad.ict.ac.cn/cn/article/pdf/preview/10.7544/issn1000-1239.202440133.pdf) · [DOI 10.7544/issn1000-1239.202440133](https://doi.org/10.7544/issn1000-1239.202440133)

## Summary
Knowledge graphs are usually incomplete. Most completion models are trained separately for each graph and learn embeddings tied to that graph's own entities and relations, so they cannot transfer to new graphs. Recent pre-trained "universal" reasoners such as ULTRA avoid the need for entity alignment by encoding relations over an adjacency-based relation graph. However, that graph cannot tell useful reasoning paths apart from noisy ones.

The authors propose RulePreM, which uses mined logical rules as prompts. Rules are extracted quickly with AnyBURL, and only relation-only rules above a confidence threshold are kept. Rules that share a head relation are merged into a relational IO graph, in which every relation has separate input and output nodes. This captures direction and connects each relation to its inverse. A GNN encoder, RuleGNN, uses confidence-aware attention to produce relation representations. These representations then condition entity encoding and candidate scoring.

Evaluated on 43 KGs across inductive, fully-inductive and transductive settings, RulePreM achieves higher average scores than both supervised and pre-trained baselines. The result matters because a single model can be reused across graphs without costly alignment.

## Key points
- RulePreM is a pre-train–prompt model for universal KG link prediction. It requires no pre-annotated entity alignment between graphs.
- Rules come from AnyBURL with a 100-second mining budget. Entity-containing rules are discarded, and a confidence threshold ε filters out weak relational rules.
- In the relational IO graph, each edge carries the highest confidence among the rules that contain it. This confidence feeds an attention mechanism in RuleGNN.
- Only the query relation gets a learnable initial vector, while all other relations and nodes start as zeros. This keeps the model free of graph-specific parameters.
- The evaluation covers 43 datasets, and the work was supported by NSFC grant 62272219 (Nanjing University).

*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{2024,
  title = {{基于规则提示的知识图谱通用推理预训练模型}},
  author = {崔员宁 and 孙泽群 and 胡伟},
  journal = {计算机研究与发展 (Journal of Computer Research and Development)},
  year = {2024},
  doi = {10.7544/issn1000-1239.202440133},
  url = {https://doi.org/10.7544/issn1000-1239.202440133}
}
```

- [← previous 基于短语及依存的标注规则和短语识别算法研究](https://md.research.elamur.ai/free/papers/liu-2024-annotation-rules-phrase/)
- [next → 大语言模型的幻觉问题研究综述](https://md.research.elamur.ai/free/papers/liu-2024-hallucinations-large-language/)
- [↑ back to papers of 2024](https://md.research.elamur.ai/free/papers/2024/page/4/)