Construction of a Retrieval-Enhanced and Dynamically Updated Collocation Knowledge Base System for Second Language Learning

To address the limitations of static collocation knowledge bases in Intelligent Systems for second language learning, an L2 collocation domain Adaptation framework PL2CDAF is proposed. The framework introduces a hierarchical two-stage adaptation process, combining error-aware pretraining and reinforcement to transform a Transformer-based model into a dynamic collocation knowledge base. It integrates a multi-head neural memory network for storing and updating collocation patterns, enabling real-time error detection and contextually appropriate feedback generation. Experimental validation demonstrates its effectiveness in enhancing collocation awareness and corrective feedback, offering a scalable and integrable solution for personalized language learning.