ConvAttRecurNet: An Attention-based Hybrid Model for Suicidal Thoughts Detection

Suicide constitutes a serious societal concern, but it rarely represents the initial phase of a person's battle. There are frequently apparent signs of suffering prior to a suicide attempt, which, if identified, can offer a chance for prompt assistance. Changes in food or sleeping habits, increased risk-taking, or estrangement from loved ones are all indicators of a potential mental health crisis. Individuals may display emotions such as pessimism, worthlessness, or feeling like a burden to others, which are frequently accompanied by severe mood swings. It is feasible to offer the requisite assistance and potentially save lives by promptly recognizing these indicators. Thus, this paper proposes a novel model, ConvAttRecurNet, which is an attention-based hybrid model that integrates convolutional neural networks (CNN) and bidirectional long-term memory (BiLSTM). The publicly available datasets from Kaggle are used to train the proposed model. Before using the data for training, the data is pre-processed to clean the text data, convert the text into tokens, and create word embeddings using the word2vec method. The word embeddings are provided to the ConvAttRecurNet for training with the customized 10 layers. The proposed model outperforms benchmarked models like BiLSTM and BiLSTM+Attention, achieving the highest accuracy of 94.6 on test data. However, the BiLSTM and BiLSTM+Attention models are overfitting by achieving the highest accuracy in training while performing poorly on the validation data. Hence, using the proposed ConvAttRecurNet in real life can benefit society in detecting suicidal thoughts at an early stage from social media texts.