Current language competence assessment relies on inefficient, subjective manual scoring and single tests, lacking accuracy in real-time prediction.This paper proposes a quantitative modelling framework integrating text, speech, and interactive data, transforming language features into computable variables via random forest ensemble learning (optimising feature selection, parameters, and tree structure).To tackle high-dimensional noise, a screening strategy enhances feature importance evaluation and parameter optimisation, boosting prediction robustness.In the sample, the average proficiency score was 57.3: 100% passed reading comprehension, 14.6% struggled with oral fluency, and 59% had tree structure flaws (yet 96.8% achieved high listening accuracy).With an average interaction frequency of 21.4, 79% of low-participation students showed a positive correlation between vocabulary reserve and comprehension ability.