Adaptive Assessment Of English Vocabulary Ability Based On Corpus Semantic Feature Representation And Deep Pattern Recognition

To address the coarse-grained semantic representation of items, static ability estimation, and singleobjective item selection in conventional English vocabulary tests, this study proposes the CSF-DPRCAT adaptive testing model. The model integrates corpus statistical features, contextual semantic representations, psychometric parameters, and a lexical relation graph, and combines learners' response outcomes, response times, prior practice, and forgetting information to estimate six dimensions of vocabulary ability: word-form recognition, basic meaning comprehension, contextual sense discrimination, collocation recognition, word-family formation, and semantic-relation transfer. A DQN-based policy is further employed to select items by jointly considering information gain, attribute coverage, semantic diversity, testing cost, item exposure, and difficulty smoothness. The experiments include response prediction and ability diagnosis on EVAT and two real external data subsets, together with offline CAT simulations conducted in a learner-response environment trained on real interaction data; the adaptive item-selection results therefore represent simulation-based administration rather than live online testing. The results show that CSF-DPR achieves AUC, ACC, and average attribute accuracy values of 0.879, 0.794, and 0.846, respectively, on EVAT, with an ability-estimation RMSE of 0.068; its AUC values on EdNet-Vocab and SLAM-English reach 0.886 and 0.869, respectively. The complete model obtains an AEMSE of 0.028 and an average test length of 14.2 items. Compared with GMOCAT, it reduces AEMSE by 17.6% and test length by 10.1%, demonstrating that the proposed method can obtain accurate, balanced, and linguistically interpretable estimates of vocabulary ability with fewer items.