A multi task learning framework using DeBERTa and BWO optimization for enhancing long term english vocabulary memory

English word memory is affected by semantic interference and context changes, which leads to rapid decline of long-term memory. This paper uses DeBERTa (Decoding-enhanced BERT with disentangled attention) to perform word part-of-speech classification and meaning matching, and uses BWO (Beluga Whale Optimization) to optimize hyperparameters and loss weights to alleviate the problem of memory decay and improve the depth and complexity of English word memory. A multi-task learning model based on DeBERTa is constructed to perform part-of-speech classification and meaning matching tasks on the target words respectively, and to introduce synonyms, antonyms and words with similar spellings for interference testing during the training process to simulate real memory scenarios. BWO is used to dynamically adjust the key hyperparameters of the DeBERTa model, such as the learning rate, batch size, and loss weight, to keep it in the optimal state at different learning stages. The precision of part-of-speech classification of the BWO optimized model reached 0.983, and the Top-5 accuracy of word meaning matching in the test set reached 0.990. In the interference test of synonyms, antonyms and words with similar spellings, the correct recognition rate of the BWO-optimized model for the target words was 92.5%, 90.8% and 88.3% respectively. Compared with PSO (Particle Swarm Optimization), GA (Genetic Algorithm) and GWO (Grey Wolf Optimizer), BWO showed better English word memory rate in cross-domain tests and under cognitive interference. This study shows that the BWO-optimized DeBERTa multi-task learning model can effectively improve the long-term memory retention of English words and has stronger anti-interference ability in complex memory environments.