| 摘要: | 精神疾病是源自於大腦生病導致的情緒、知覺、思考、認知及行為上的異常，本論 文的研究主題是精神疾病文本的多標籤分類，專注於開發深度學習模型，用於自動理解 文本語意內涵，預測一個至多個事先定義好的標籤。我們提出標籤強化超圖注意力網路(Label-Enhanced Hypergraph Attention Networks, LE-HyperGAT) 模型，藉由超圖構對 文本進行建模抽取特徵，並使用超圖注意力網路來捕捉文本之間的語意關係，最後利用 標籤的嵌入向量，強化標籤與文本間的關係，達到更好的文本標籤類別預測成效。實驗 資料來自心靈園地網站 (http://www.psychpark.org)，民眾可以在這個平台上提出各式各 樣與心理情緒、精神疾病相關的問題，專業的精神科醫師會根據問題提供回覆，以及針 對問題意涵的種類給予適當的多個標籤。我們蒐集的資料集(簡稱 PsychPark) 包括 2,752 篇民眾的留言，每篇留言的平均字數為 264.96 字以及平均有 1.58 個標籤。實驗結果顯 示我們提出的 LE-HyperGAT 模型有最高的 Macro-averaging F1 分數 0.3713，比相關研究模型(包含 CNN, BiLSTM, LSAN, BERT, GraphCNN, TextGCN 以及HyperGAT)表現 更好。之外，錯誤分析實驗可以進一步發現 LE-HyperGAT 可以解決低頻標籤的問題，有效提升多標籤文本分類的效能。;Mental illness stems from brain maladies leading to abnormalities in emotions, perceptions, thoughts, cognition, and behavior. This paper focuses on the multi-label classification of psychiatric texts, emphasizing the development of deep learning models for automatically understanding textual semantics and predicting one or multiple predefined labels. We propose the Label-Enhanced Hypergraph Attention Networks (LE-HyperGAT) method, which models texts by extracting features using hypergraph structures and captures semantic features using hypergraph attention networks. Finally, we reinforce the relationship between labels and text by utilizing label embedding vectors to achieve improved category predictions. Experimental data is sourced from the PsychPark website (http://www.psychpark.org), where individuals can pose various questions related to mental and emotional health issues. Professional psychiatrists respond and annotate appropriate multiple labels based on the nature of the questions. Our dataset (PsychPark) comprises 2,752 posts, with an average of 264.96 words and 1.58 labels. Experimental results demonstrate that our proposed LE-HyperGAT model achieves the highest Macro-averaging F1 score of 0.3713, outperforming related research models (including CNN, BiLSTM, LSAN, BERT, GraphCNN, TextGCN, and HyperGAT). Additionally, error analysis further reveals that LE-HyperGAT addresses low-frequency label issues, effectively enhancing the performance of multi-label text classification. |
