資訊管理學報
楊亨利;林青峰;
頁: 335-361
日期: 2018/07
摘要: 隨著電子商務的發展，現在一個虛擬商場能夠查詢到的商品資訊常常超過人們所能負擔；如何從玲瓏滿目的商品中找到最符合需要，對使用者而言是非常重要的。傳統上的推薦系統是由個人資料、社會關係、購買或點閱記錄等資料來進行相似性的比對，進而完成推薦。本研究則採用網路上關於商品使用者評價文的意見傾向資料來進行推薦。因應消費者可能對商品功能屬性有不同的權重偏好，我們的系統考量到各個屬性構面的網路評價，並具有擴充彈性，會以代理人系統去網路上自動讀入可能的新商品與新屬性資料。在訓練好意見分類器之後，我們建立了一個推薦雛型系統，最後還進行了可用性及易用性的實驗。我們採用SVM分類器，在測試階段中，在巨觀平均上，我們的「非常負面、負面、中立、正面、非常正面」五類正面分類器的平均精準率、召回率、F1值分別為80.85%、79.81%、80.32%；五類負面分類器的平均精準率、召回率、F1值分別為81.66%、79.17%、80.39%。若如傳統一樣只分為「正面、中立、負面」三類的程度的話，其分類各項績效均為80%以上、或甚至超過90%。而在可用性、易用性實驗亦顯示這個系統在有用性、易用性與總評的各項目上，均可令使用者滿意。
關鍵字: 意見挖掘;情感分析;網路評論;推薦系統;
Abstract: Purpose - This study aims to propose a prototype system based on reviews opinion mining for functional product recommendation. Design/methodology/approach - A multi-level SVM classifier was proposed and a prototype was also built. Finally, we designed an experiment for evaluating usability and ease of use. Findings - In classifying five categories, i.e., {very negative, negative, neutral, positive, very positive}, our positive classifier had precision 80.85%, recall 79.81% and F1-score 80.32%; and negative classifier had precision 81.66%, recall 79.17% and F1- score 80.39%. If using traditional three categories, all the above performance were above 80% or even above 90%. The experiment also indicated that our system is easy to use and more accurate than searching information manually. Research limitations/implications - This study ignored the posted date of product reviews. Future research may further explore the possibly opinions in the different stages of the life cycle of a product. Practical implications - It is very important for people to find the products meeting their needs. Traditional recommendation system might use personal data, social relationship, purchase data or click-through records to calculate users' similarity for recommendation. This study tries to apply sentimental orientation of product reviews on each feature crawled from internet for recommendation. The proposed framework is also extensible since an automated agent would automatically process new products or new features. Originality/value - We proposed a new hybrid-based recommendation system by applying web opinion mining to combine content-based and collaborative filtering.
Keywords: opinion mining;sentiment analysis;internet review;recommendation system;
瀏覽次數: 14433 下載次數: 1334
引用 導入Endnote
頁: 335-361
日期: 2018/07
摘要: 隨著電子商務的發展，現在一個虛擬商場能夠查詢到的商品資訊常常超過人們所能負擔；如何從玲瓏滿目的商品中找到最符合需要，對使用者而言是非常重要的。傳統上的推薦系統是由個人資料、社會關係、購買或點閱記錄等資料來進行相似性的比對，進而完成推薦。本研究則採用網路上關於商品使用者評價文的意見傾向資料來進行推薦。因應消費者可能對商品功能屬性有不同的權重偏好，我們的系統考量到各個屬性構面的網路評價，並具有擴充彈性，會以代理人系統去網路上自動讀入可能的新商品與新屬性資料。在訓練好意見分類器之後，我們建立了一個推薦雛型系統，最後還進行了可用性及易用性的實驗。我們採用SVM分類器，在測試階段中，在巨觀平均上，我們的「非常負面、負面、中立、正面、非常正面」五類正面分類器的平均精準率、召回率、F1值分別為80.85%、79.81%、80.32%；五類負面分類器的平均精準率、召回率、F1值分別為81.66%、79.17%、80.39%。若如傳統一樣只分為「正面、中立、負面」三類的程度的話，其分類各項績效均為80%以上、或甚至超過90%。而在可用性、易用性實驗亦顯示這個系統在有用性、易用性與總評的各項目上，均可令使用者滿意。
關鍵字: 意見挖掘;情感分析;網路評論;推薦系統;
Functional Product Recommendation System Based on Sentiment Mining from Web Reviews
Abstract: Purpose - This study aims to propose a prototype system based on reviews opinion mining for functional product recommendation. Design/methodology/approach - A multi-level SVM classifier was proposed and a prototype was also built. Finally, we designed an experiment for evaluating usability and ease of use. Findings - In classifying five categories, i.e., {very negative, negative, neutral, positive, very positive}, our positive classifier had precision 80.85%, recall 79.81% and F1-score 80.32%; and negative classifier had precision 81.66%, recall 79.17% and F1- score 80.39%. If using traditional three categories, all the above performance were above 80% or even above 90%. The experiment also indicated that our system is easy to use and more accurate than searching information manually. Research limitations/implications - This study ignored the posted date of product reviews. Future research may further explore the possibly opinions in the different stages of the life cycle of a product. Practical implications - It is very important for people to find the products meeting their needs. Traditional recommendation system might use personal data, social relationship, purchase data or click-through records to calculate users' similarity for recommendation. This study tries to apply sentimental orientation of product reviews on each feature crawled from internet for recommendation. The proposed framework is also extensible since an automated agent would automatically process new products or new features. Originality/value - We proposed a new hybrid-based recommendation system by applying web opinion mining to combine content-based and collaborative filtering.
Keywords: opinion mining;sentiment analysis;internet review;recommendation system;
瀏覽次數: 14433 下載次數: 1334
引用 導入Endnote