DeepLegal-CN: Research and Application of a DeepSeek-Based Large Language Model for the Legal Domain

The legal domain is one of the most linguistically intensive professional fields, characterized by complex structures, specialized terminology, and intricate logical reasoning. Legal professionals must frequently handle diverse legal documents such as regulations, judgments, contracts, and legal opinions. This complexity poses significant challenges for natural language processing (NLP), especially in understanding and generating legal texts. Recent advancements in large language models (LLMs) have demonstrated remarkable generalization capabilities across various NLP tasks, sparking interest in their application to legal scenarios. However, most existing LLMs face limitations in legal applications, especially in Chinese contexts. Proprietary models entail high inference costs and raise privacy concerns, while open-source models often lack domain-specific training on legal corpora. Moreover, the majority are trained on English datasets, limiting their adaptability to Chinese legal language, which differs significantly in expression and logic. To address these challenges, this paper proposes a legal LLM framework based on DeepSeek, a powerful open-source Chinese LLM. The framework integrates feature engineering and task-driven prompt design to enhance the model’s capability in understanding and reasoning over Chinese legal texts. We evaluate the model across multiple tasks, including legal question answering and case matching, demonstrating its effectiveness and identifying existing limitations. This study provides a novel approach for legal AI development in the Chinese context and offers valuable insights into building adaptable, domain-aware LLMs for legal applications.