基於意見詞修飾關係之微網誌情感分析技術

Microblog Sentiment Analysis Based on Opinion Target Finding and Modifying Relation Identification

10.6841/NTUT.2013.00350

葉庭瑋

情感分析 ； 意見傾向 ； 微網誌 ； 意見詞 ； 修飾關係 ； opinion analysis ； opinion orientation ； microblog ； sentiment word ； modifying relation

臺北科技大學資訊工程系研究所學位論文

2013年

碩士

王正豪

繁體中文

如何有效分析文件的意見傾向，一直是熱門的研究議題之一。若能準確分類評論文章、網誌內容，將有助於產品或服務上的競爭分析或了解大眾在公共議題上的意見傾向。相關研究著重於如何建立情感詞彙資源並利用各種分類方法，對文件內容進行情感分析，然而微網誌有其挑戰:短句的字數有限且欠缺上下文的語義；另一方面，短句常直接帶有一些情緒性的字眼，可能有助於情感分析。本論文提出一個基於評論目標發掘及意見詞修飾關係之微網誌評論內容意見傾向計算方法。首先 ，從微網誌收集主題相關評論及句子簡化處理。接著根據評論主題以及意見詞的修飾關係，發掘出主題相關的評論目標以判斷其意見傾向。最後利用網友評論實作推薦系統，可以有效克服cold start的問題。 實驗針對50部電影在Twitter上的1000篇英文評論進行分析，結果顯示本論文方法平均準確率accuracy為84.44%，同時最高precision可達88.89%，優於SVM及Naive Bayes分類法。由此可驗證意見詞修飾關係的規則判斷能有效提高意見傾向分類的準確率。

Opinion analysis has grown to be one of the most active research areas in natural language processing. If we can classify reviews and messages of blogs correctly, it will help to analyze product and service competition and to realize the opinion orientations of the people on public issues. Existing research focuses on establishing lexical resources, and utilizing text classification for opinion analysis. Nevertheless, there are some challenges for microblogs. First, text messages are limited to 140 characters. Second, it’s difficult to know what users want to express without suitable contexts; on the other hand, short messages usually contain sentiment words, which could help opinion analysis. In this paper, we propose an opinion orientation estimation approach based on target finding and opinion modifying relations in microblog reviews. First, it collects reviews from microblog and preprocess the source data. Then, by extracting any entity or aspect of the entity about which an opinion has been expressed according to opinion modifying relations, we calculate the overall score of opinion orientation. Finally, by utilizing the opinion orientation in microblog reviews for recommendation, the cold start problem can be overcome. In our experiment on the 1000 movie reviews of 50 movies from Twitter, the average accuracy of the proposed method is 84.44%, and the highest precision is 88.89%, which is better than SVM and Naive Bayes. This validates the higher precision from modifying relation identification for opinion orientation classification.

1. 陳仕堯(2017)。結合意見探勘之電影推薦系統的研究。淡江大學資訊管理學系碩士班學位論文。2017。1-66。