食品汙染之網路媒體即時分類系統-以三聚氰胺為例
A Web-based Online Document Classification System for Food contamination News – Take Melamine as the Example
中文摘要
食品安全對於人民健康的影響甚大，因此保障人民食品安全提升其品質，是重要 的一項議題。隨著國際間食品及食品原料進出口貿易的往來日漸頻繁，世界各地 食品及原料的流動變得無遠弗屆，所以只要一有食品及原料遭受汙然，隨時都有 可能流通到世界各地，導致社會大眾的不安與恐慌（如 2008 年爆發的三聚氰胺 食品污染事件）。
因此，本研究建構即時的食品汙染新聞監測及自動分類系統，此系統能定時且自 動化搜尋 Google News 有關台灣食品污染的新聞報導，監測到在台灣各地發生的 食品汙染相關新聞事件(本研究以三聚氰胺相關新聞為例)，能在第一時間獲得最 完整的資訊，並同時將新聞儲存在系統資料庫中。
本研究以程式語言 C/C++撰寫而成的潛藏狄利克雷分配(Latent Dirichlet
Allocation, LDA)方法的 Gibbs LDA++為發展工具，將訓練文件做多次的非監督 式學習（Unsupervised Learning）與監督式（Supervised Learning）的訓練，訓練 所得之參數結果就是系統分類器的分類依據。本論文將含有三聚氰胺關鍵字之新 聞分為「汙染事件、檢驗、醫療健康、法律政策、其他」五大類別，以 521 篇訓 練文件和 793 篇測試文件做相關之分類訓練。而分類器的成效評估方式是以精確 率(precision)和召回率(recall)及經過精確率和召回率換算後的 F-measure 來進行 效能評估，最後經過系統分類器所得分類結果之平均精確率為 69.66%，平均召 回率為 64.52%，所得之 F-measure 為 0.68。
依據成效結果探討其原因及未來研究改進之方向，期望本系統能節省監測人員閱 讀大量繁雜新聞的時間，提早進行相關準備，預防遭受汙染之食品在市場上流動 擴散之目標。
英文摘要
Food safety is essential to human health, and also to guard the food safety as well as the quality of food is an important issue. As long as the international food and food materials trade being more frequently, the flow of them are far-reaching, and will spread all over the world once they get contaminated. Turns out, it will bring unrest and panic to the public (such as the 2008 outbreak of melamine food contamination).
This study is to build an online food contamination monitoring and automatic classification system to regularly search Google news about food contamination (we took melamine-related news as an example) in Taiwan and obtain the most
complete information and stored them in the database. The system will classify news into correct categories and can help users find relevant information.
In this study, we used Gibbs LDA++, which is a C/C++ implementation of Latent Dirichlet Allocation (LDA) to train news documents by unsupervised learning and supervised learning. The classifier was built by the parameter estimations and inferences from LDA training results and then adjusted manually by human expert.
We defined the melamine news as five categories including "contamination",
"analysis", "medical and health", "law and policy" and “others”. 521 news documents were used as training data to train the classifier and 793 documents were used to test and evaluate the classifier. The assessment of the effectiveness for the classifier is based on precision, recall and F-measure. According to the
evaluation for the classifier, the macro-precision is 69.66%, the macro-recall is 64.52% and the F-measure is 0.68.
According to the evaluation results, we estimate the performance of classification system and will improve the system in the future research. We expect the system could save time for reading complexity news, and help people get prepared for food contamination.
