Research on the Application of Large Language Model in Data Integration

A Symbolic Regression Paradigm for Interpretable Data Governance

Zhanfang Chen; Yuan Ren; Xiaoming Jiang; Rui Qi· EAI Endorsed Transactions on Scalable Information Systems· 2026· DOI 10.4108/eetsis.10245

The core problem

Large Language Models (LLMs) have emerged as a major breakthrough in artificial intelligence, with widespread applications across diverse domains. Their powerful capabilities in language comprehension and generation enable effective handling of natural language processing tasks such as text generation, question answering, machine translation, and information retrieval. This paper investigates the application of LLM technology in data integration, a core aspect of data governance. In contrast to end-to-end black-box approaches, the authors reframe data integration as a problem of discovering interpretable mapping rules through symbolic regression. The fundamental problem of data integration is first defined, and then a general-purpose large model framework for data governance is proposed, built on a deep symbolic regression foundation. The framework comprises a symbolic expression generator and a metadata-enhanced executor, aiming to achieve both high accuracy and interpretability.

The proposed model is evaluated on two representative data integration tasks. Experimental results demonstrate its effectiveness, validating practical quality. The results highlight the advantage of the symbolic regression paradigm in enhancing interpretability. While specific quantitative metrics are not detailed in the abstract, the authors report that the framework achieves both high accuracy and interpretability, as evidenced by the successful discovery of interpretable mapping rules. The metadata fine-tuning approach, where generated symbolic expressions serve as key metadata, contributes to improved performance. The evaluation confirms that the model can handle diverse data integration scenarios, providing a foundation for further research in data governance.

Large Language Models (LLMs) have emerged as a major breakthrough in artificial intelligence, with widespread applications across diverse domains. Their powerful capabilities in language comprehension and generation enable effective handling of natural language processing tasks such as text generation, question answering, machine translation, and information retrieval. This paper investigates the application of LLM technology in data integration, a core aspect of data governance. In contrast to end-to-end black-box approaches, the authors reframe data integration as a problem of discovering interpretable mapping rules through symbolic regression. The fundamental problem of data integration is first defined, and then a general-purpose large model framework for data governance is proposed, built on a deep symbolic regression foundation. The framework comprises a symbolic expression generator and a metadata-enhanced executor, aiming to achieve both high accuracy and interpretability.

The proposed model is trained using a combination of recurrent neural networks and reinforcement learning techniques for expression generation. The execution of the discovered rules is structured based on a Transformer encoder architecture enhanced with a dedicated metadata embedding layer. To enhance performance, the authors incorporate metadata fine-tuning, where the generated symbolic expressions serve as key metadata to guide the integration process. The overall architecture can be represented as follows:

The paper's key contribution is the reframing of data integration as a symbolic regression problem, which contrasts with end-to-end black-box approaches. By discovering interpretable mapping rules, the framework addresses a critical need in data governance for transparency and explainability. The combination of recurrent neural networks and reinforcement learning for expression generation, along with a metadata-enhanced Transformer executor, represents a novel architecture. The use of metadata fine-tuning further enhances the integration process. The results validate the practical quality of the approach and underscore the potential of symbolic regression in improving interpretability. Future work may explore scaling the framework to larger and more complex data integration tasks, as well as integrating additional metadata sources. The taxonomy candidates—Architecture, Cybersecurity, Network, Cryptography—suggest potential application domains where interpretable data integration is crucial.

CS practitioners and researchers

Opening member content…