運用類神經網路方法分析基於面向的情感極性分類

Abstract

隨著時代以及科技技術的成長，人們不像過去一樣，需要查看報紙、購買雜誌、詢問左右鄰居的情報才能知道自己想要得知的資訊。在科技技術的成長下，不管是餐廳的評價、筆記型電腦的實用程度，大部分的人們都可以使用網際網路來查看是否有所想要的資訊。本論文使用的資料集由SemEval-2014 Task 4官方所提供，並且含有四項子任務：(一) Aspect term extraction、(二) Aspect term polarity、(三) Aspect category detection、(四) Aspect category polarity，本論文進行第二項子任務研究，判斷出句子中的面向詞是正面、負面或中立，評估方式採用Accuracy，並且與當年競賽結果相比較。本論文實驗方法將資料先進行前處理並且轉成詞向量作為輸入的來源，以及將極性做情感標籤，並且使用Bi-LSTM (Bi-directional Long Short-Term Memory)、Self-attention(自注意力機制)及使用Two-level encoding對資料進行訓練。最後去比對每種不同模型的準確率，結果顯示Two-level encoding預測準確率餐廳達82%，筆記型電腦則達78%。With the development of the times and technology, people are not the same as in the past, that they need to check newspapers, buy magazines, and ask their neighbors for information to know what they want to know. With the development of technology, most people can use the Internet to check whether they have the information they want, no matter it is restaurant reviews or the practicality of laptops.The dataset used in this study is officially provided by SemEval-2014 Task 4, which contains four subtasks: (1) Aspect term extraction, (2) Aspect term polarity, (3) Aspect category detection, (4) Aspect category polarity, This paper conducts the second sub-task study to judge whether the term words in the sentence are positive, negative or neutral.

The experimental method of this study pre-processes the data and converts it into the word vector as the source of input, and uses polarity asemotional label. Then Bi-LSTM (Bi-directional Long Short-Term Memory), Self-attention and Two-level encoding models are used to train the data.Finally, to compare the accuracy of each different model, the results show that using the two-level encoding method, the prediction accuracy rate is 82% for restaurants, and 78% for laptops.