利用向量組合方法改善網路服務匹配

Abstract

在text-based的網路服務比對方法中，由於將網路服務表示為純文字描述，並 使用文字來表示網路服務和進行比對，因此文字比對的準確性會對服務比對的表 現造成很大的影響。本研究中我們透過下面四個步驟提升服務比對的準確性: 1. 從網路服務描述文件中取得關鍵字，並將關鍵字透過預先訓練的詞向量模型轉換 為向量表示，2. 從參考資料中取得文字關係，3. 將文字關係用於向量結合，改 善預訓練詞向量中的不足，4. 計算關鍵字詞向量的cosine similarity來得到網路服 務相似度。在實驗中我們使用了網路服務比對的benchmark OWLS-TC V4來評估 提出方法的表現，並利用假設檢定將我們的方法與現有的服務比對方法iSeM做比 較，在比較結果中我們的方法 (MAP=0.9242) 表現優於iSeM (MAP=0.8529)。

In text-based service matchmaking approach, since the web service is treated as a plain text and use term tokens as the internal representation to match services, the accuracy of the text comparison will affect the performance of service matchmaking. In this research, we improve the performance of service matchmaking through the following four steps: 1. extract keywords from WSDL and convert them into vector representations through pre-trained word vector model, 2. extract word relations from reference data, 3. use word relations for vector combination to improve the quality of pre-trained word vectors, and 4. calculate the cosine similarity between keyword word vectors to get the similarity of two web services. An experiment is also conducted based on an OWLS-TC V4 service matchmaking benchmark with hypothesis testing to compare our proposed approach with the iSeM approach. The result of the experiment shows that our approach (MAP=0.9242) excels iSeM (MAP=0.8529) by.