| 摘要: | 維度型情感分析任務的目標是從輸入文本中，分析出作者的情緒??負面程度 (Valence) 以及激動程度 (Arousal)，可以應用於眾多情境中，例如:在心理諮詢和臨床 心理學領域，幫助心理師和臨床醫生更好地了解患者的情感狀態和心理需求，進而提供 更加準確且有效的心理輔導和治療。本研究旨在探討中文心理諮詢領域的維度型情感分 析。我們提出了一個情感強化門控圖神經網路模型 (Sentiment-enhanced Gated Graph Neural Networks, SentiGGNN)，用於分析中文心理諮詢文本的情感強度與激動程度。首 先，我們為每一篇輸入文本建構依存句法分析圖以及情感關聯圖。然後，經由門控圖神 經網路來學習圖的節點表示。接著，經由雙向長短期記憶-卷積神經網路學習序列的表示， 再透過注意力機制將節點表示與序列表示融合得到一個新的文本表示向量。最後，經過 多層感知器得到文本的維度情感 (Valence-Arousal)預測值。
我們蒐集線上心理諮詢的民眾留言共 4,163 筆，然後人工標記維度情感取平均值， 最終建置了第一個中文心理諮詢領域的維度型情感分析資料集 (Psycho-VASentiment)。 藉由實驗與效能評估分析得知，我們提出的SentiGGNN模型優於其他相關研究模型 (包 含 RNN, CNN, LSTM, Attention LSTM, Regional CNN-LSTM, Word-level BERT, HyperGAT, TextGCN, ADGCN, UGformer 以及 TextING)。??外，我們將維度型情感分析技術應用在 社群媒體輿情分析，藉由案例分析從大數據下找到有用的觀點，藉以驗證維度情感分析 技術的實務價值。
;Dimensional sentiment analysis focuses on predicting sentiment intensity in the valence arousal domains, which can be applied to help psychologists and clinicians understand the emotional state and psychological needs of patients, thereby providing more accurate and effective psychological counseling and therapy. This study aims to explore dimensional sentiment analysis in Chinese psychological counseling texts. We propose a Sentiment- enhanced Gated Graph Neural Networks (SentiGGNN) model to analyze sentiment intensities in the valence and arousal domains. Firstly, we construct sentiment and dependency graphs for each input text. Then, we learn node representations through GGNN architecture and sequence representations using BiLSTM-CNN networks. Subsequently, we fuse node representations with sequence representations based on the attention mechanism. Finally, we obtain the valence and arousal prediction values through a multi-layer perceptron.
We collected 4,163 online psychological counseling texts and manually annotated them to obtain average valence-arousal values, resultig in the first Chinese dimensional sentient analysis dataset in the psychological counseling domain, Psycho-VASentiment. Experimental results and performance evaluations revealed that our proposed SentiGGNN model performed other related methods, including RNN, CNN, LSTM, Attention LSTM, Regional CNN-LSTM, Word-level BERT, HyperGAT, TextGCN, ADGCN, UGformer, and TextING. In addition, we apply our dimensional sentiment analysis techniques to implement a social media analysis platform, providing valuable insights into the collected big data and confirming the effectiveness of our proposed model. |
