TL;DR
This paper introduces MM-EMOG, a novel graph-based representation that captures multi-label emotions within social media posts to improve mental health classification. By leveraging emotion lexicons and graph convolutional networks, MM-EMOG effectively models the complex emotional landscape of user-generated content, leading to significant performance gains in identifying mental health conditions.
What Problem Does This Paper Address?
Social media has become a crucial, albeit sensitive, source of information for understanding mental health. A significant portion of individuals contemplating suicide disclose their intentions online, presenting an opportunity for early intervention. However, extracting meaningful insights from these platforms is challenging due to privacy concerns limiting access to personal data and the inherent complexity of human emotions. Unlike simpler sentiment analysis, a single social media post can convey multiple, nuanced emotions, making it difficult to accurately assess a user's mental state. Existing methods often focus on single emotions or rely on external contextual information, which is becoming increasingly restricted. This paper aims to address this by developing a method that can effectively capture and leverage multi-label emotions from text alone for mental health classification.
Methods
The core of the MM-EMOG approach lies in its novel representation of text that incorporates multi-label emotions. This is achieved in two main steps:
- Multi-label Emotion Embeddings:
- Graph Construction: A graph is built where nodes represent words (tokens) and documents (posts). Edges connect words to words, words to documents, and documents to documents, capturing co-occurrence relationships. These relationships are weighted using Pointwise Mutual Information (PMI), Term Frequency-Inverse Document Frequency (TF-IDF), and Jaccard similarity.
- Emotion Lexicons: Widely used emotion lexicons (EmoLex, TEC, SenticNet) are employed to associate words with one or more emotions. This allows for the generation of document-level, multi-label emotion classes.
- Graph Convolutional Network (GCN): A two-layer GCN is used to propagate information across the graph, learning contextual representations that harmonize heterogeneous emotional information. This process is trained using a multi-label emotion classification task.
- BERT Fine-tuning: The learned token representations from the GCN are then used to fine-tune a pre-trained BERT model for the same multi-label emotion classification task. This results in the final MM-EMOG embeddings.
- Mental Health Post Classification:
- Concatenated Embeddings: For the final mental health classification, the MM-EMOG embeddings are concatenated with standard BERT embeddings for token nodes.
- GCN for Classification: Another two-layer GCN is applied to the graph, this time for the mental health classification task. The output of this GCN is used to predict the mental health class of the post.
The key novelty lies in the integration of multi-label emotions within a graph-based framework, moving beyond single-emotion analysis and leveraging the rich contextual information embedded in word-document relationships.
Datasets
The proposed MM-EMOG model was evaluated on three publicly available social media datasets:
- TwitSuicide: A dataset of Twitter posts related to suicide.
- Reddit C-SSRS Dataset (CSSRS): A dataset from mental-health-related subreddits annotated using the Columbia-Suicide Severity Rating Scale.
- Twitter Depression Dataset (Depression): A dataset of tweets annotated for depression.
These datasets represent different platforms and annotation schemes, allowing for a comprehensive evaluation of the model's generalizability.
Key Insights and Ideas
- The Power of Multi-Label Emotions: The study strongly suggests that a single social media post can express multiple emotions simultaneously, and capturing this multi-label nature is crucial for accurate mental health classification. Ignoring this complexity leads to missed nuances.
- Emotion Lexicons Matter: The choice of emotion lexicon significantly impacts performance, with Twitter-specific lexicons (TEC, SenticNet) showing better results on Twitter-based datasets, highlighting the importance of domain-specific resources.
- De-identification and Emoticon Preservation Enhance Performance: Preprocessing steps like de-identifying user information and explicitly preserving emoticons as tokens improve model performance, especially on platforms where these elements are prevalent. This indicates that richer contextual signals, even seemingly minor ones, can be highly informative.
- Graph-based Learning for Context: The graph convolutional network approach effectively leverages the relationships between words and documents, allowing the model to learn a more holistic understanding of the text's emotional and semantic content.
Key Results
MM-EMOG consistently outperformed baseline models (BERT, RoBERTa, MentalBERT) across all three datasets.
- Overall Performance: The model achieved improvements of 8%, 21%, and 14% in weighted F1-score for TwitSuicide, CSSRS, and Depression, respectively.
- Concerning Classes: Notably, MM-EMOG showed significant improvements in identifying the most concerning classes (e.g., Strongly Concerning for TwitSuicide, Attempt/Behaviour/Ideation for CSSRS), demonstrating its ability to capture heightened negative emotions.
- Qualitative Analysis: Case studies revealed that MM-EMOG could better distinguish between expressions of empathy directed towards general negative situations versus those related to self-harm or hopelessness, a crucial distinction for mental health assessment.
Limitations and Future Work
The authors acknowledge several limitations:
- Language Dependency: The study primarily used English-based datasets and lexicons, limiting its applicability to other languages.
- Computational Resources: The graph network construction and training are computationally intensive, and the study was limited to a maximum of 256 words per post.
- Limited Public Benchmarks: There is a scarcity of publicly available state-of-the-art models specifically for single-post, text-based mental health classification.
Future work includes releasing a pre-trained MM-EMOG model with generalized emotion representations for downstream tasks and exploring low-resource languages. The authors also suggest investigating the use of more sophisticated graph structures and potentially incorporating other contextual information where privacy permits.
Summary updated