In NLP, the quality of training depends on the " Standardization " of input data. Informal language (slang, typos) and mixed languages create high " noise, " making it difficult for the model to learn semantic meaning.
Using dictionaries to normalize text and translation services to create a coherent dataset is the most foundational step. While tokenization (Option C) and entity recognition (Option B) are standard parts of the NLP pipeline, they cannot function effectively if the underlying text is inconsistent or fragmented. Removing non-English data (Option D) would result in a non-representative model that fails to meet the needs of a diverse customer base.