In the field of Natural Language Processing, Word Sense Disambiguation (WSD) is a critical task aimed at determining the correct meaning of a polysemous word within a specific context. Traditional WSD methods, which rely on a static set of word senses, struggle with new or polysemous words and fail to adequately address the needs of English learners. To overcome these limitations, we adopt a Definition Modeling approach that utilizes the fine-tuning of the Collins Dictionary for the Flan-T5 model, guiding the model to focus on the target word through prompt templates. This method allows for the dynamic generation of the target word’s meaning based on the context, thereby capturing the diversity and dynamism of word senses. Additionally, to address the limitations of current mainstream evaluation methods, we propose a new evaluation metric called the Semantic Perplexity-BERT Composite Score (SPBS), which is designed to comprehensively assess the performance of word sense generation models. We validated the effectiveness of the proposed method across multiple datasets, and experimental results indicate that our method achieves significant improvements in generating more contextually appropriate word senses. Furthermore, the SPBS metric demonstrates superior differentiation ability and accuracy in evaluating the quality of word sense representations.