下一句預測（NSP）已成為大型語言模型在支援英語為外語（EFL）寫作上的重要應用方向。然而，現有的 NSP 常缺乏個人化與事實可靠性。為了改善這些限制，本研究提出一套系統：結合個人化與情境化提示（PCP）的個人化 NSP 寫作助理。該系統包含三個模組：情境模組（CM），用以擷取如圖片輸入等真實情境；資訊擷取模組（IEM），利用 GraphRAG 檢索相關資訊；以及寫作風格模組（WSM），用以捕捉學生的寫作風格。這三個模組的輸出將整合成 PCP，用來引導預測句的生成。系統以 OpenAI o3 為主要模型。本研究之評估包含以相似度與正確性指標進行的自動化測試，以及針對七位印尼 EFL 高中生進行的正式實驗，採用前實驗研究設計，分析其預測功能的使用行為並評估預測句的表現。 結果顯示，本研究所提出的系統能有效提升預測句的表現。三個模組皆對系統性能有顯著貢獻，其中 WSM 對預測句整體表現的影響最大，證實符合學生寫作風格是實現個人化的關鍵。IEM 在學生寫作深度不足時，提供必要的內容細節，且縮小高低成就學生之間預測句性能差距的最重要因素。正式實驗中，G-Eval 分數亦證實系統生成的預測句表現優異，學生反映系統提供的預測句有助於構思並加快文章寫作速度。此外，越積極提供情境圖片的學生對預測句的接受度越高。 本研究證明，結合個人化、真實情境與 GraphRAG 及 OpenAI o3，為改進 NSP 寫作助理提供了一種可行的方法。研究結果為開發更具適應性與可持續性的 EFL 寫作助理，提供了實證依據與設計見解。;Next-sentence prediction (NSP) for EFL writing support often lacks personalization and factual reliability. To address these limitations, this study proposed a personalized NSP with Personalized and Contextual Prompt (NSP-PCP) for writing assistant, which proposed three modules such as the Contextual Module (CM), which extracts authentic context such as image inputs; the Information Extraction Module (IEM), which retrieve relevant information using GraphRAG; and the Writing Style Module (WSM), which captures individual writing styles. These modules are integrated into a PCP that guides the generation of predicted sentences. The system is powered by OpenAI o3 as the primary model. The evaluation comprised an automatic test employing similarity and correctness metrics, as well as a formal experiment conducted with seven Indonesian EFL high school students using a pre-experimental design, in which they engaged in writing tasks with our proposed system. The results indicate that the proposed system improved the performance of predicted sentences. All three modules made significant contributions. WSM had the largest impact on overall performance, confirming that matching a student′s writing style is a big portion to personalization. IEM provided essential content details when students′ writing lacked depth, which was the most influential factor in reducing the performance gap between higher and lower achieving students. In the formal experiment, G-Eval scores confirmed the higher performance of predicted sentences, and students reported that the predictions supported idea development and drafting efficiency. Moreover, students who more actively provided contextual images showed higher acceptance of the predicted sentences. Therefore, this study demonstrates that combining personalization, authentic context, and GraphRAG with OpenAI o3 model provides a practical approach to improving NSP mechanism for writing assistance. The findings offer empirical evidence and design insights for developing more adaptive and sustainable writing support systems of EFL writing in the future.
