CNN based binary classification to recognize suicide related tweets

As one deep learning based algorithm commonly used in various computer vision tasks [25], CNN also demonstrated excellent performance in many NLP tasks, including various text classification tasks [26–29]. We leveraged a classic CNN model for short text classification proposed by Kim et al. [26] to build the tweets binary classifier. We cleaned the tweets using the script from Stanford [30]. Then, we converted the tokens in each tweet to one-hot vectors and mapped the one-hot vectors to pre-trained GloVe Twitter embedding. The mapped embedding was used as the initial input feature to the CNN model. For the CNN model training, various filters were applied to generate the convolutional layers. We applied the max pooling strategy on the feature maps generated by different filters. We added dropout on the pooling layer to avoid overfitting. The pooling layer was connected to a fully connected layer with softmax output. The architecture of the CNN framework is shown in Fig. 2.

The architecture of CNN based binary classifier for suicide related labels prediction