Rewriting this content in English:
—
**Title: Enhancing Large Language Models with Knowledge Graphs for Specific Domains**
### Abstract
A new study from the University of Science and Technology of China introduces a novel framework called KG-SFT, which significantly improves the performance of large language models (LLMs) in specific domains by utilizing knowledge graphs. The research demonstrates that this approach can achieve remarkable results even with minimal training data.
### Key Contributions
1. **Framework Overview**: The KG-SFT method integrates knowledge graphs into the fine-tuning process of LLMs, enhancing their ability to understand and manipulate domain-specific knowledge.
2. **Data Efficiency**: In low-data scenarios, such as medical question answering in English, KG-SFT achieves a 14% increase in accuracy using only 5% of the training data compared to traditional methods.
### Methodology
– **Knowledge Graph Integration**: The framework leverages structured knowledge from graphs to guide the learning process of LLMs, enabling them to better capture domain-specific patterns.
– **Explainability Focus**: KG-SFT emphasizes generating high-quality explanations, which not only improves model performance but also provides interpretable insights into decision-making processes.
### Experimental Results
1. **Performance Across Domains**: The framework shows strong performance across multiple domains, including medicine and accounting, with particularly impressive results in low-data settings.
2. **Comparison with Baselines**: KG-SFT consistently outperforms existing methods, especially in tasks requiring complex reasoning.
### Innovation Highlights
– **Quality Over Quantity**: Unlike traditional approaches that focus on data quantity, KG-SFT prioritizes the quality of explanations to drive improved model performance.
– **Modular Design**: The framework can be easily integrated as a plugin with other data augmentation techniques, offering flexibility for various applications.
### Conclusion
KG-SFT represents a significant advancement in leveraging external knowledge to enhance LLM capabilities. By focusing on high-quality data and domain-specific insights, this approach opens new possibilities for improving model performance in specialized fields.
—
**Author Information**
– **Lead Author**: Han Zhu Chen, a Ph.D. student at the University of Science and Technology of China since 2021, under the supervision of Professor Wang Jie.
– **Research Interests**: Large language models, knowledge graphs, and data synthesis for reasoning tasks.
**Paper Link**: [https://openreview.net/pdf?id=oMFOKjwaRS](https://openreview.net/pdf?id=oMFOKjwaRS)
—
This English version maintains the technical details and key insights from the original Chinese article while ensuring clarity and readability for an international audience.
相关文章
