基於本體論與規則的食譜推薦專家系統

Abstract

近年來食譜網站與社群媒體提供了大量的料理資訊，但使用者在實際決策時往往需要同時考量料理風格偏好、過敏或禁忌限制，以及家中現有食材是否足以完成料理等諸多條件，使得「選到喜歡的食譜」未必等同於「能立刻做出來」。為回應此一現實與理想的落差，本研究設計一套以知識為核心的食譜推薦專家系統，強調推薦結果的可製作性與可解釋性。 本研究利用 Protégé 來進行系統本體論（ontology）的建立與管理，以OWL/RDF（Web Ontology Language/ Resource Description Framework）建立食譜領域本體論，定義食譜（Recipe）、食材（Ingredient）、料理風格（Cuisine）與使用者（User）等概念，並以 requiresIngredient、isEquivalentTo 與 isAllergicTo 等關係描述食譜內容、食材可替代性與使用者條件。推理部分採用 SWI-Prolog 程式語言，並以 semweb 套件作為本體論知識的存取介面；規則庫在推理過程中透過 rdf/3 直接查詢 RDF store，將其作為唯一的資料來源，使系統能在本體論內容更新後自動反映於推理結果，同時維持資料與規則的分離，降低重複建置與同步維護的成本。 在結果驗證方面，本研究以使用者情境設計展示系統的推理與輸出，並依序執行：偏好匹配、安全性過濾與可製作性評估。實驗呈現顯示，系統能將推薦分為「可立即製作」與「採買導向」兩種輸出：前者提供具有完整食材的食譜，後者則列出缺少的食材清單以支援採買規劃；此外，當本體論中存在食材替代關係且使用者持有可替代食材時，系統可調整可製作性的判定並使推薦類型產生翻轉，提升推薦在真實料理情境下的應用。最後，本研究討論與生鮮超市會員資料與商品資料庫整合的應用構想，說明如何將缺少食材的結果進一步連結到商品資訊，使推薦不僅停留在推薦食譜清單上，而能延伸為可執行的採買與烹飪決策。

In recent years, recipe websites and social media have provided an abundance of cooking information. However, real-world recipe selection must simultaneously consider cuisine preferences, allergy or dietary restrictions, and whether the ingredients available at home are sufficient, meaning that “finding a liked recipe” does not necessarily imply “being able to cook it right away.” To bridge this gap, this study develops a knowledge-centric recipe recommendation expert system that emphasizes executability and explainability. An OWL/RDF ontology is constructed to model the recipe domain, defining concepts such as Recipe, Ingredient, Cuisine, and User, and relations including requiresIngredient, isEquivalentTo, and isAllergicTo to represent recipe requirements, ingredient substitutability, and user constraints. SWI-Prolog is adopted as the reasoning engine, with the semweb package serving as the ontology access layer. During inference, the rule base queries the RDF store via rdf/3 as the single source of truth, allowing recommendations to automatically reflect ontology updates while keeping data (knowledge) separated from logic (rules), thereby reducing duplication and synchronization costs. Scenario-based evaluations demonstrate the reasoning process and outputs through preference matching, safety filtering, and cookability assessment. The system produces two actionable outputs: “ready-to-cook” recipes with complete ingredients and “shopping-oriented” recipes with missing-ingredient lists for purchase planning. Moreover, when substitution relations exist and substitutes are available, the system revises cookability decisions and can flip recommendation types, improving feasibility. Finally, an integration scenario with grocery membership and product databases is discussed to link missing ingredients to purchasable items and support end-to-end shopping and cooking decisions.