Reproduce “Advancing Mental Disorder Detection: A Comparative Evaluation of Transformer and LSTM Architectures on Social Media” on OpenResearch

About this paper

Khalid Hasan, Jamil Saquer, Mukulika Ghosh

Researchers at Missouri State University conducted a comparative evaluation of deep learning architectures for detecting mental health disorders from Reddit posts, finding that transformer models like RoBERTa achieve the highest F1 scores, reaching 99.54% on a hold-out dataset. The study also demonstrated that LSTM models augmented with contextualized BERT embeddings offer a competitive performance-to-computational-cost trade-off, outperforming LSTMs with static embeddings.

Key insights

• Transformer models consistently outperform LSTM-based models in detecting mental disorder posts, achieving higher accuracy and F1 scores.
• Contextualized word embeddings (BERT) are crucial for effective mental health discourse analysis, dramatically improving the performance of LSTM models to nearly match standalone transformers.
• While transformers offer top performance, LSTM models augmented with BERT embeddings provide a compelling balance between high performance and reduced computational overhead, suitable for resource-constrained deployments.

Summary by alphaXiv.