Vocabulary Generation and Automatic Evaluation in Chinese Reading Comprehension Using Deep Learning

Background: Chinese internet social neologisms change rapidly and create new difficulties for reading comprehension systems because accurate understanding requires both lexical diversity and cognitive semantic consistency. This study develops a Transformer-based Contrastive Generation and Evaluation (TCGE) framework combining deep learning and generative evaluation for vocabulary generation and automatic assessment of Chinese reading comprehension oriented to internet social neologisms, and constructs cognitive processing modeling aligned with human reading logic. Methods: TCGE combines a dynamic beam-width decoding strategy with BERTTransformer joint training, proposes the cognitive semantic consistency indicator, and adopts attention-enhanced generative evaluation; experiments are conducted on CMRC2018, DRCD, ZhihuQA and a self-built internet social-neologism test set. Results: With 182M parameters, TCGE achieves BLEU-4 of 0.43, ROUGE-L of 0.52 and cognitive semantic consistency of 0.78, improving over BART-large by 12.8%, 10.6% and 8.3%, while reducing inference time by 38.7%; the attention-enhanced evaluation model reaches 87% accuracy, 85% F1-score and 0.92 AUC. Conclusion: The proposed Transformer contrastive framework improves the balance between semantic accuracy, lexical diversity and deployable efficiency, and provides a reproducible basis for personalized Chinese reading evaluation in internet education scenarios based on deep learning and cognitive processing modeling of online neologisms.