以情緒導引強化注意力機制於情緒分析之研究

Abstract

情緒分析 (sentiment analysis) 是一個找出文字中的情緒與情感的重要任務，常用於分析句子中的情緒與感情。此問題常被視為一種分類的問題，利用深度神經網路模型可以達到很好的成果，注意力機制也被證實有很好的效果。再者，先前研究也指出情緒辭典對情緒分析問題上有很好的成效。然而，情緒辭典並沒有適當地被應用在先前的研究中。 本篇論文探索了情緒引導之注意力機制，以完整利用情緒辭典並將情緒辭典結合在注意力機制中，藉此幫助分類。我們提出兩種結合方法，第一，為了有效利用情緒辭典，我們轉換情緒詞典中的情緒值，使其成為一組增強注意力權重係數，以最小化原本模型內的注意力權重係數之錯誤。第二，我們提出了情緒多頭注意力機制，我們使用從情緒值轉換而來的注意力權重係數，做為第二組注意力頭，以協助模型關注更多資訊。我們實驗在六組情緒分析資料集上，結果顯示此方法準確度皆超越先前最佳的模型，相較於先前的分數提升0.12%到8.12%。

Sentiment analysis is an important task, which extracts sentiment, emotion or affect in text. The problem is often treated as a classification problem for which deep neural methods have been well explored and attention mechanisms have generated promising performance. Studies have shown that lexicon is highly effective for sentiment analysis. However, lexicon has not been fully utilized by the previous methods. No existing method integrates lexicon into the attention mechanism effectively to solve the problem. This thesis explores the sentiment-guided attention mechanism, which integrates lexicon into attention mechanism and proposes two approaches. First, to utilize sentiment lexicons, we transform lexicon values into guiding weights to minimize the error of attention weights. Second, we propose sentiment multi-head attention to help the model jointly attend to sentiment information provided by the transformed lexicon values. Experiments show that our models outperform state-of-the-art models on six sentiment analysis benchmarks with improved accuracy of 0.12% to 8.12%.