This study develops a machine learning approach to score business English writing and generate feedback that is tied to the actual text. Existing automated scoring systems are usually reliable for surface features such as grammar, vocabulary, and sentence form, but they are less consistent when the task involves tone, coherence, or how well a response fits a business context. In many cases, the feedback they produce is also too broad to help with revision. To address this problem, we organize the model into three parts. The first part learns a representation of writing quality so that responses with similar wording but different communicative effectiveness can be separated more clearly. The second part segments the text and examines local structure, including sentence-level variation and the progression of ideas. The third part ranks candidate feedback and removes comments that are repetitive or weakly connected to the relevant passage. The three parts are used jointly during scoring and feedback generation, rather than as separate stages with little interaction. This reduces overlap between modules and helps keep the model stable across different writing samples. In the experiments, the model performed better than the baselines on both scoring and feedback. Correlation with human ratings increased from 82% to 89%, and feedback precision rose from 71% to 86%. User satisfaction also improved from 3.6 to 4.4 out of 5, suggesting that the revised feedback was more specific and easier to use in practice.
