近年來,以面向類別為基礎的評價分析逐漸被重視,在這個任務下,需要判斷評論句於某個預先定義的面向類別{食物, 價錢, 服務, 氣氛, 其他}所表達出的評價。在本論文中,我們致力於 SemEval-2014 任務 4 中的兩個子任務:面向類別偵測(Aspect Category Detection,ACD)及面向類別情緒偵測(Aspect Category Polarity Detection,ACP)。在 SemEval-2014 中,多數的隊伍使用 n-grams 與情緒字典作為機器學習主要特徵。然而,對於一些各類別通用的意見詞(例如,「讚」可以用來描述每一個面向類別),這些特徵難以此區分出這些意見所描述的類別。相對地,我們首先抓取意見詞所描述的面向詞,並使用(意見詞,面向詞)配對作為特徵來解決這個困難。做法上,首先利用意見詞辨識系統找出評論句中的意見詞,再用依存規則(dependency rule)來判斷出對應的面向詞。本系統目前已完成於餐廳評論領域。實驗顯示,使用 Word2Vec 作為特徵可以達到 87.5% 的正確率,加上(意見詞,面向詞)配對特徵可以達到 88.3% 正確率。所有的特徵一起使用可以從 84.4% 提升到 89.0%。實驗結果顯示出該配對於面向類別情緒偵測下是有效的。
In recent years, researches of aspect-category-based sentiment analysis have been approached in terms of predefined categories. In this paper, we target two sub-tasks of SemEval-2014 Task 4 dedicated to aspect-based sentiment analysis: detecting aspect category and aspect category polarity. Also, a pre-identified set of aspect categories {food, price, service, ambience, miscellaneous} defined by SemEval-2014 have been used in this paper. The majority of the submissions worked on these two sub-tasks with machine learning mainly with n-grams and sentiment lexicon features. The difficulty for these submissions is that some opinion word (e.g., "good") is general and cannot be referred to any particular category. By contrast, we use aspect-opinion pairs as one of the features in this paper to overcome this difficulty. To detect these pairs, we identify the opinion words in customer reviews, and then detect their related aspect terms by dependency rule. This system has been done on restaurant domain applying to Chinese customer reviews. Our experiment achieved 87.5% of accuracy by using Word2Vec to detect aspect category polarity. Aspect-opinion pair features employed in this system contribute to 88.3% of accuracy. When all features are employed, the accuracy is improved from 84.4% to 89.0%. Experimental results demonstrate the effectiveness of aspect-opinion pair features applied to the aspect-category-based sentiment classification system.