With the development of deep learning, distributed word vector technology based on neural networks shows strong potential. However, the English corpus is diverse in type and large in volume, and traditional word vector computation is difficult to adapt to the semantic evolution and semantic association of words in long textual contexts in dynamic environments, and the effect varies significantly on different sizes of corpora. In this paper, a word vector computation method Transformer-DSSM based on the Transformer model for English corpus is proposed to solve the problem that traditional word vector computation cannot effectively deal with polysemous words and contextual relevance. The study is based on a deep semantic model, utilizing the Transformer coding layer for feature extraction, and calculating lexical semantic relations through the self-attention mechanism and cosine similarity. Experiments on the SICK and MSRP datasets show that the Transformer-DSSM model obtains excellent performance with Pearson’s coefficient of 0.887, Spearman’s coefficient of 0.845, and mean squared error of 0.2286 on the SICK test set, and reaches an accuracy of 77.8% and an F1 value of 83.8% on the MSRP test set. In addition, simulation experiments on the English news recommendation datasets MIND and Adressa verify the effectiveness and practicality of the model in word vector representation, providing a new solution for English corpus processing.
