This thesis presents an intelligent tutoring system that enables students of Chinese to acquire active knowledge of words and grammatical constructions. The system relies on a Bayesian, linguistically motivated cognitive model that represents the estimated knowledge of the learner. This model is dynamically updated given observations about learner's behaviour and proficiency in the exercises. The model is then employed at run-time to select the exercises that are expected to maximise the learning outcome. The system was implemented together with a set of 100 English-to-Chinese translation exercises. Each exercise is associated with a set of solutions. If the student's answer is not correct, the system finds a solution with the shortest distance to the input and gives interactive feedback using a combination of error-specific and generic rules that provide relevant cues towards the closest correct translation. The system is integrated with a bilingual English-Chinese dictionary, and the student may look up unknown words at any stage. Both dictionary look-ups and learner's proficiency in the exercises serve as evidence that enables the system to infer probabilistic information about the learner's actual vocabulary knowledge. Compared with a baseline that randomly chooses exercises at the user's declared level, the proposed approach has shown a positive, statistically significant effect on the users' assessment of how much they have learnt. The results suggest that the cognitive system leads to improved learning outcomes. Experiments with larger groups of participants are required to detect potential differences in other effects.