類神經多任務學習結合情感詞典做屬性情感分析

Abstract

基於屬性的情感分析（ABSA）是一種較精細的情感分析任務，此任務會根據給定的文本去分析評論者對目標實體的不同屬性所抱持的情感態度。隨著網路和社群媒體的快速發展，人們已習慣於在網路上表達自己對於一些事物的想法，從而累積了大量的使用者原創內容（UGC），例如電影或飯店評論。對於不管是個人還是公司來說，能夠得知使用者對特定主題的想法是非常有價值的，因為這些資訊可以幫助他們做出更好的購買決策(對消費者而言)、或行銷與產品改善決策(對產品或服務的提供公司而言)。為了從大量的文本中自動提取出使用者觀點，ABSA被廣泛應用。其中，用來表達情感的情感詞資訊應有助於辨識在文本中使用者評論的屬性與情感傾向。然而，我們觀察到比起其他情感分析相關的任務，透過將情感詞典作為外部資源來將意見資訊納入模型中的ABSA研究較少。在這項研究中，我們提出了一個ABSA的整合模型，也就是將其兩個子任務整合進一個模型中訓練，並通過添加輔助任務的形式將情感詞辨識納入我們的多任務學習框架來提高其性能。為了提供較精準的情感詞，我們將建構特定領域的情感詞典，並且在訓練時嘗試採用不同的建構模型，評估它們對ABSA性能的影響。最後，我們在兩個SemEval ABSA挑戰的基準資料集合上進行實驗，實驗結果顯示我們提出的方法可以勝過現今最先進的方法。

Aspect-based sentiment analysis (ABSA) is a task of identifying fine-grained opinion polarity towards a specific aspect associated with an entity. With the rapid development of the Internet and social media, people have grown accustomed to expressing their opinions online, causing a massive accumulation of user-generated content (UGC), such as reviews on movies or hotels. Opinions toward specific topic from users are valuable to both individuals and corporations, since the information can help them make better purchase decisions (for individuals) or marketing and product improvement decisions (for corporations that provide services and products). To automatically extract opinions discussed in the huge amount of UGC, ABSA is widely applied. However, relatively few studies on ABSA have taken sentiment lexicons as an external resource and incorporated such information into the model, which should be helpful to support identifying the aspects and the sentiment attitudes expressed in the text. Therefore, we propose an integrated ABSA model that combines its two subtasks (i.e., aspect term extraction and aspect sentiment classification) together in a single model and attempt to improve its effectiveness by incorporating the sentiment term identification as an auxiliary task to the multi-task learning framework. Specifically, we will construct domain-specific sentiment lexicons by ourselves. Besides, we will adopt different models to construct sentiment lexicons and evaluate their effects on the effectiveness of ABSA. The experimental results over the two benchmark datasets of the SemEval ABSA challenge indicate that our proposed method can outperform the state-of-the-art performance benchmark.