Identifying Emotion Labels from Psychiatric Social Texts Using a Bi-Directional LSTM-CNN Model

Abstract

Discussion features in online communities can be effectively used to diagnose depression and allow other users or experts to provide self-help resources to those in need. Automatic emotion identification models can quickly and effectively highlight indicators of emotional stress in the text of such discussions. Such communities also provide patients with important knowledge to help better understand their condition. This study proposes a deep learning framework combining word embeddings, bi-directional long short-term memory (Bi-LSTM), and convolutional neural networks (CNN) to identify emotion labels from psychiatric social texts. The Bi-LSTM is a powerful mechanism for extracting features from sequential data in which a sentence consists of multiple words in a particular sequence. CNN is another powerful feature extractor which can convolute many blocks to capture important features. Our proposed deep learning framework also applies word representation techniques to represent semantic relationships between words. The paper thus combines two powerful feature extraction methods with word embedding to automatically identify indicators of emotional stress. Experimental results show that our proposed framework outperformed other models using traditional feature extraction such as bag-of-words (BOW), latent semantic analysis (LSA), independent component analysis (ICA), and LSA+ICA.