Preliminary investigation of personalized next-sentence prediction for EFL writing with authentic context and personalized writing portfolio

Next-Sentence Prediction (NSP) systems based on LLMs frequently generate generic responses and factual inaccuracies. To address these issues, this preliminary investigation proposes a personalized NSP system for EFL writing assistance that integrates three modules. The Contextual Module processes user images to establish authentic context. The Information Extraction Module uses GraphRAG to retrieve factual knowledge, while the Writing Style Module analyzes historical writing data to match individual user patterns. These inputs are integrated into a Personalized and Contextual Prompt (PCP) designed using the Chain-of-Thought method. Automatic evaluation was conducted using data from 180 users, followed by two formal experiments involving high school (n=7) and undergraduate students (n=30). Results show that personalized NSP-PCP significantly improved semantic similarity and contextual correctness compared with the Baseline Prompt, demonstrating benefits for EFL writing. Ablation tests confirmed that all modules contributed to performance, with the Writing Style Module having the largest impact. The experiments also provided preliminary evidence that students actively used and evaluated the predicted sentences. Overall, integrating authentic context and personalized writing portfolio, including knowledge graphs constructed from writing histories and writing-style analysis, can improve NSP performance and personalization for EFL writing. However, hallucination was not directly evaluated, so its reduction remains unverified.