結合卷積神經網路與遞迴神經網路於推文極性分類

Combining Convolutional Neural Network and Recurrent Neural Network for Tweet Polarity Classification

隨著網路日漸發達，不分男女老少均常常使用社群網站來分享生活瑣事或是評論時事，每天產生的訊息量十分可觀，若我們去分析這些資訊得到社會大眾對事件的好惡，便可以更容易地去做較佳的決策。本文以Twitter做為研究對象，針對英語推文進行情緒分析，我們使用Tweepy從Twitter上收集推文，並用於訓練詞向量（Word Vector），然後使用卷積神經網路（Convolutional Neural Network，CNN）對訓練好的詞向量進行微調，使其帶有情緒特徵，之後再用於遞迴神經網路（Recurrent Neural Network，RNN）進行訓練，得出最後極性分類的結果，我們的系統使用SemEval-2018 Task1: Affect in Tweets子任務V-oc的資料集進行訓練，且與競賽結果做比較，落在大約第五名的成績。

With the development of the Internet, male and female, old and young, often use social network to share the trivia of everyday things and comment on current affairs. The amount of information generated every day is very considerable. If we analyze those data to get the impressions from society, we can easier to make better decisions. This paper chooses Twitter as the research subject and conduct sentiment analysis on English tweets. We use Tweepy to collect tweets on Twitter and use them to train word vector. After that, the trained word vector is fine-tuned to have emotional features by Convolutional Neural Network (CNN) . Then, the fine-tuned vector is used for training in the Recurrent Neural Network (RNN) to get the final polarity classification results. Our system uses the dataset of the subtask V-oc of SemEval-2018 Task1: Affect in Tweets for training. Compared to the results of competition, we are in the fifth place.