情感分析於電影推薦與評論展現系統之應用

隨著以文字資訊為主的社交平台興起，例如：微博、推特、部落格…等微型網誌，消費者對於購買商品或服務品質的評價可以在網路世界中迅速傳播，對於其他消費者的購買意願造成很大的影響，同時也加深大眾對於該產品的品牌形象。對於電影產業更是如此，消費者只能透過片商剪輯的預告片，觀賞部分電影片段，就必須決定是否要進電影院觀賞，事後也沒有退換貨的服務，因此民眾在購買電影票之前，會更加注重網路上對於該部電影的相關評論以及心得分享。有鑑於此，如何從巨量的網路資訊當中，正確且有效率地辨別顧客所表達的語意與情緒，成為近年來文字探勘學者致力於探討的議題。本論文實作出一個有效的電影評價系統，蒐集2019年Yahoo!奇摩電影網頁中網友滿意榜的短評資料，透過意見提取、屬性擷取、情緒分析、語意指向、特徵分群與機器學習分類法等技術，對評論按照其極性做分類，實驗結果顯示正確率為83.74%，F1-Measure也達84.29%，代表本研究在評論極性的判別上，確實有達到預期的效果。最終評論呈現的方式有兩個特點，首先，評論會依據其情緒強度由大至小排序，讓使用者優先瀏覽情緒與內容最豐富的評論；再者，藉由呈現意見詞與屬性詞搭配的結果，提供使用者搜尋電影多面向的情緒分析成果，了解該電影在各個屬性類別的各自評價，藉此推薦合適的電影給消費者觀賞。

關鍵字

文字探勘；情緒分析；特徵分群；語意指向；機器學習

並列摘要

Following the rise of social media platforms for text information, such as Weibo, Twitter and Blog. Consumers’ rating for purchasable commodity and service quality can be rapidly spread in social media. It causes significant effect to other consumers’ desire to purchase. It also impresses the public about the product’s brand imagine. Furthermore, in movie industry, consumers have to decide whether to go into theater only through watching the segments of movie trailer. They can’t get a refund when they feel regrettable. So consumers will pay more attention on related comments and knowledge-sharing. For this reason, how to identify consumer’s expression of mood and semantization correctly becomes the subject for dedicated scholars. This essay produces an efficient movie evaluation system. It collected netizen’s satisfactory list of comments from 2019 Yahoo movie web page. Through Feature Extraction, Attribute Capture, Sentiment Analysis, Semantic Orientation, Feature Clustering, Machine Learning Classification to classify comments in accord with polarity. This experiment proves that the accuracy reaching 83.74% and the F1-Measure reaching 84.29%. It means that this study has achieved its anticipative result in identifying the polarization of comments. There are two characters appearing in final comments. First, comments will be listed in sequence according to sentiment intensity that let users browse the most abundant ones at first place. Secondly, by matching opinion keywords and feature keywords to offer users the outcome of multi-faceted analysis which could let them know the evaluation of each film’s attribute. Through it to recommend the suitable movie to consumers.