We investigated how the natural language processing (NLP) model enhances coherence and cohesion in college-level writing through cross-lingual computational analysis. Conventional automated writing evaluation systems emphasize surface-level grammar while neglecting structural discourse patterns. Therefore, we developed a unified transformer-based pipeline employing paraphrase-multilingual-MiniLM-L12-v2, which generates dense vector embeddings to compute semantic similarity matrices and trace discourse trajectories across English and Chinese essays. Quantitative profiles were validated through structured student and instructor interviews, forming a mixed-methods framework that integrates algorithmic metrics. The results showed substantial improvements, a 34.10% increase in local coherence, an 18.37% rise in global thematic alignment, and a 56.45% reduction in redundant sentence components, indicating enhanced semantic density and structural consistency. Improvements in embedding distributions corresponded with cognitive interventions observed during interviews, confirming the pedagogical relevance of NLP-driven feedback. In further studies, keystroke and behavioral metrics must be integrated to enrich multimodal evaluation, and open-domain text styles and diverse populations need to be included to validate and refine discourse architecture through advanced NLP modeling.