Customs risk rule generation is crucial for enhancing regulatory effectiveness. Existing methods over-rely on manual experience, making them difficult to cope with dynamic risks, and generally suffer from insufficient interpretability and lagging updates. To address these issues, this paper innovatively proposes a Semantic-Enhanced and Knowledge-Driven Rule Representation Learning framework (SEKD-RRL) for intelligent generation of customs risk rules. The core contributions of SEKD-RRL are: First, it designs a semantic enhancement module that integrates a semantic encoder into rule representation learning, effectively improving the model’s semantic understanding of unstructured customs texts and overcoming the limitations of traditional methods in text processing. Second, it constructs a knowledge-driven logical layer, fusing semantic similarity knowledge with a customs domain knowledge graph, realizing the transition from hard logical rules to soft logical rules, and significantly enhancing the domain knowledge relevance and interpretability of the generated rules. The SEKD-RRL framework mainly consists of a semantic enhancement module, a knowledge-driven logical layer, and a rule representation learning module. Experiments on a simulated customs dataset demonstrate that SEKD-RRL achieves a dual improvement in classification performance and interpretability in risk rule generation, and exhibits advantages in rule accuracy, simplicity, and expert evaluation. The research validates the effectiveness of the SEKD-RRL framework in the customs risk rule generation task, provides a technical path for building a new generation of intelligent customs risk regulation systems, and offers a reference for research in areas such as semantic-enhanced rule learning and knowledge-driven intelligent decision-making.