Personalized language learning in digital education faces challenges in balancing personalization, data privacy, and limited user interaction data. To mitigate these issues, this paper proposes PERFED, a Privacy-Preserving Personalized Federated Learning paradigm with Reinforcement Learning and Meta-Learning for adaptive digital language education. PERFED introduces three key innovations: (i) a privacy-preserving federated optimization mechanism with adaptive differential privacy noise that adjusts based on client data sensitivity to ensure strong privacy protection while minimizing utility loss; (ii) a reinforcement learning-based adaptive policy that dynamically selects personalized learning strategies to improve learner engagement and accelerate convergence; and (iii) a meta-learning-driven module designed for quick adaptation to new users in low-data scenarios, improving cold-start performance in personalized language learning scenarios. Experiments on the archived multilingual Duolingo Second Language Acquisition Modeling corpus show that PERFED achieves $83.3\%\pm 0.47$ accuracy over five runs, outperforming FedAvg (78.5%), FedProx (80.2%), and FedPer (81.9%) under non-IID federated settings. PERFED also reduces convergence rounds while improving training stability under heterogeneous client distributions, all while maintaining a maximum cumulative privacy loss of $(ϵ=4.84, \delta ={10}^{-5})$ under a Rényi differential privacy accountant. These results demonstrate the effectiveness and robustness of PERFED for privacy-preserving intelligent language education. The evaluation is an offline synchronous simulation; the behavioral engagement signal is a log-derived proxy rather than a human-subject measure, and mobile deployment performance is not claimed.