基於評論面向之評論助益性預測

Abstract

在網路發達的時代，大部分的人在購物前都會上網尋求其他人的使用經驗，然而，隨著社群平台與電子商務網站發展熱絡，衍生出了資訊超載的問題，也有許多的電子商務網站或是評論網站開始採取機制來篩選出有幫助的評論，但若只是單純的做助益性票數的排名，並沒有辦法滿足使用者的資訊需求，因此本研究藉由探討評論之內容來建立合理的評論助益性評分系統，並將此系統應用於三種實務上的問題：偵測潛藏主題與情緒、預測各主題下最有幫助性的評論及預測整體評論之助益性分數，且與其他主題模型做應用於中英文四種不同來源資料集的比較。 本研究透過實作監督式主題情緒模型SJASM (Supervised Joint Aspect Sentiment Model)來作為評論助益性預測系統，其能夠探討評論中所討論的主題與使用者情緒，並用以預測評論助益性分數。接著，我們將SJASM模型和其他模型sLDA (Supervised Latent Dirichlet Allocation)、JST(Joint Sentiment/Topic Model)與ASUM(Aspect and Sentiment Unification Model) 做預測效果的比較。 研究結果中我們發現主題模型sLDA在預測整體的評論助益性分數上有不錯的表現，而在預測各主題下最有幫助性的評論時較適用於商品同質性較高的資料。主題情緒模型則適合使用於預測各主題下最有幫助性的評論，其中SJASM模型能夠在有多種助益性標籤的資料集中有較好的預測效果，因其能夠同時以目標變數來學習評論中的主題與情緒。

In the era of highly developed Internet, online reviews have become one of important resource for users to make purchase decision. However, with the development of social platforms and e-commerce websites, the problem of information overload has arisen, and there are many e-commerce websites or review websites have begun to adopt some mechanisms to filter helpful customer reviews. Nevertheless, if we simply rank the number of helpful votes, it is still hard to satisfy user needs. Therefore, the purpose of this research is to establish a reasonable review helpfulness prediction system by exploring the content of customer reviews. Also, we apply the system and other topic models to three practical issues, i.e. latent topic and sentiment detection, topic-based review helpfulness prediction and overall helpfulness score prediction, and we compare the result of these models for four different source datasets in Chinese and English. This research implement Supervised Joint Aspect Sentiment Model(SJASM) as review helpfulness prediction system, which can explore the topic and sentiment in a review and predict the helpfulness score. Next, we compare the SJASM model with other models: sLDA (Supervised Latent Dirichlet Allocation), JST (Joint Sentiment/Topic Model) and ASUM (Aspect and Sentiment Unification Model). In the result of this study, we found that the topic model sLDA has a good performance in overall helpfulness score prediction, and it is more suitable for the data with higher homogeneity of the product when ranking topic-based review helpfulness scores. We also found that topic sentiment models are suitable for the task of ranking topic-based review helpfulness scores. Among these models, the SJASM model can have a good predictive effect in a dataset with multiple helpfulness tags, because it can simultaneously learn the topics and sentiments of reviews with the target variables (helpfulness score).