Enhancing Mental Health Detection in Malay Social Media: A Comparative Study of RNN Architectures and Attention Mechanisms

Mental health detection using Natural Language Processing (NLP) is increasingly crucial with the vast data accessible from social media platforms. Despite the abundance of research on high-resource languages, there remains a sub-stantial gap for underrepresented languages such as Malay. This paper aims to bridge this gap by deploying a combination of Recurrent Neural Networks (RNNs), specifically Bidirectional Long Short-Term Memory (BiLSTM) and Bidirectional Gated Recurrent Unit (BiGRU), integrated with attention mechanisms, to detect mental health symptoms in the Malay language. The methodology utilises the MyDAS Corpus, a specialised dataset derived from Facebook, which includes various Malay-language interactions related to mental health. The models were enhanced with Word2Vec embeddings, meticulously fine-tuned for Malay, leading to significant performance improvements. Systematic experimentation yielded F1 scores of 0.83 and 0.82 with BiL-STM paired with Self and Multi-Head attention mechanisms, respectively, markedly surpassing the baseline F1 score of 0.78 via TF-IDF and Decision Tree classifier. The results highlight the efficacy of leveraging pre-trained embeddings with RNN-attention configurations to enhance the discrimination between psychological classes, particularly notable in ‘Depression,’ ‘Anxiety,’ and ‘No DAS Signal’ classes. The findings underscore the potential of these advanced NLP techniques for real-world mental health applications in languages with limited digital resources.