我們提出一種方法，用於針對給定的原始與修正句對產生修正解釋。本研究所採用的方法結合句對與輔助的語言知識，旨在強化大型語言模型（LLMs）產出具教育意義的解釋之能力。該方法包含三個核心步驟：辨識句對中的詞彙層次修正位置、檢索與這些修正相關的語言知識，以及利用結合多來源檢索增強生成（retrieval-augmented generation, RAG）機制的LLM 生成針對性的解釋。我們開發了一個原型解釋生成系統WhyFix，並將該方法應用於LLM 上。在一項由具備自然語言處理專業知識的研究生所執行的評估中，系統針對代表性錯誤產生的解釋被判定為明顯優於兩類基準：一是既有錯誤參考資料中所提供的解釋（如Longman Dictionary of Common Errors），二是未進行檢索增強之GPT-4.1-nano 模型所產出的解釋。實驗結果顯示，結合語言知識可顯著提升解釋品質，使系統能提供具備清晰邏輯與語言根據的修正理由，進而將單純的錯誤修正轉化為具啟發性且具行動導向的學習契機。

We introduce a method for generating explanations for corrections within a given pair of corrected and original sentences. In our approach, we integrate the sentence pair with supplementary linguistic knowledge, aiming to maximizing ability of large language models (LLMs) to produce pedagogically valuable explanations. The method involves identifying lexical correction positions within the provided sentence pair, retrieving relevant linguistic knowledge concerning these corrections, and generating targeted explanations using an LLM enhanced with a composite, multi-source retrieval-augmented generation (RAG). We present a prototype explanation generator, WhyFix, that applies this method to LLMs. In an evaluation conducted by domain-aware graduate students in NLP, explanations generated by our system for a representative set of corrections were judged to be significantly superior to both the explanations found in established error reference books, such as the Longman Dictionary of Common Errors, and those generated by the baseline GPT-4.1-nano model without retrieval augmentation. Our methodology demonstrates that incorporating linguistic knowledge yields significant enhancements, enabling the system to deliver clear and well-grounded rationales for corrections, thereby transforming a simple correction into an insightful, actionable learning opportunity.