應用興趣點辨識技術從Web中挖掘新商家資訊

Mining POIs from Web via POI recognition and Relation Verification

本論文提出一套系統能從網頁中自動化的挖掘新的店家資訊的方法。透過地址相關的特殊的關鍵字（如：台北市+新開幕）進行搜尋，找到可能包含地址及新開幕店家的網頁，再利用地址辨識模型先從結果中擷取地址，並從周圍透過興趣點辨識模型擷取商家名稱（Store Name Recognition），最終使用地址與興趣點關聯配對（POI Relation）模型推斷該商家名稱是否位於該地址。我們特別著重在商家名稱辨識以及POI Relation的模型建立。針對興趣點辨識模型的資料準備，我們將黃頁上的商家名稱透過實體篩選以及資料前處理，應用Distant Learning及序列標記，可以訓練出F1值0.816的興趣點辨識模型。其次關於POI Relation預測則是針對反例的準備進行研究，其中效能最好的模型有0.754的準確率。整體系統效能則使用兩個興趣點辨識模型搭配三種關聯分類模型，共進行六次實驗並分析，最好的組合平均每個IP每天能找到約49個新的興趣點。

This paper presents a system that could automatically extract new POIs from Web. First, we use special queries (e.g. Taipei+New Open) to find Web pages that might contain addresses for new stores. For web pages that contain addresses, we then apply store name recognition model to extract possible POIs. Finally, we train a model to find the most possible POI for the address found in the page. In this paper, we focus on POI name recognition and POI relation prediction. For POI recognition, we use store names from yellow pages as seed to prepare the training data via distant learning. Through entity selection and data processing, we obtain a model with 0.816 F1-measure as opposed to 0.432 F1-measure for a dictionary-based baseline. As for POI relation prediction, we compare three different strategies for negative example preparation. The best model could get 0.754 accuracy. We combine two POI recognition models with three classification models to test the overall performance. The best combination could extract 49 POIs every day with a single IP.