Using Emoji as an Emotion Modality in Text-Based Depression Detection

Text-based depression detection has long been investigated by exploring useful handcrafted linguistic features and word embeddings. This paper focuses on utilizing emoji as an emotional modality to detect whether a subject is depressed or not based on text. In particular, we propose to extract sentence-level emotional information with model pretrained to predict emoji of text on social media and semantic information with widely used embedding model. The embeddings are then input to the classification model to predict one's mental state. Experiments are conducted on user-generated posts from three datasets and clinical conversational data from DAIC-WOZ. Results on social media data indicate emojis' superior performance in general, with further enhancement derived from modality fusion. Furthermore, emoji outperforms contextual text embeddings in sparse scenarios like clinical interview dialogues. We also provide a detailed analysis showing that the emojis extracted from healthy and depressed subjects are significantly different, suggesting that emoji can be a reliable emotion representation in such implicit yet complex sentiment analysis settings.