Machine Learning Methods for Sentiment Analysis of Short Texts

In this paper, a study was conducted aimed at identifying toxic behavior on the Internet, expressed in writing comments in an angry and aggressive manner. With use of supervised machine learning methods, a multi-labeled classification of such comments by type of their toxicity (toxic, highly toxic, indecent, threat, racist) was solved on a real data. A theoretical overview of the approaches used in the practical part was given. In the practical part of the paper, it was shown that linear methods based on the “Bag of Words” notation, due to the small length of comments, cannot effectively solve the problem of classifying short texts. For the same reason, additional convolutional layers do not provide an increase in the quality of the popular basic models based on recurrent neural networks. Despite the high predictive accuracy of basic models based on the LSTM and GRU architectures, their metrics could be further improved by: including the “Attention” layer in the model, which highlights the most important words for evaluating an appropriate sentence class; use of GloVe model vectors with higher dimension.