METHOD OF USING A NEURAL NETWORK WITH A HYBRID ARCHITECTURE TO DETERMINE THE EMOTIONAL TONE OF TEXT MESSAGES

Abstract

The article reviews the current state of the scientific direction of determining emotional tone and presents a method for using a hybrid architecture neural network to determine the emotional tone of text messages. The method of using a hybrid architecture neural network to determine the emotional tone of text messages is intended for automated conversion of input data in the form of a trained hybrid architecture neural network model with a tokenizer and a text message for analysis into output data in the form of a membership class by emotional tone and its numerical evaluation. Method is based on the use of a hybrid neural network architecture that combines CNN and BiLSTM. The proposed combination contributes to the effective selection of local patterns, due to the properties of the CNN layer, and also allows to take into account long-term dependencies in the text, due to the properties of BiLSTM. The neural network model starts with an Embedding layer, which transforms text data into fixed-length numeric vectors. Next comes a layer that randomly “turns off” 20% of the neurons to reduce the risk of overfitting. Then comes a layer that uses convolutions to detect local patterns in the input data. Next comes a bidirectional LSTM layer, capable of taking into account context from both ends of the sequence, with mechanisms for randomly turning off neurons to improve generalization. This is followed by a layer that selects the maximum values from all features to reduce dimensionality. The final stage is a dense layer with a single neuron and sigmoid activation, which gives the probability that the text belongs to a class with positive tone. Experimental study of the effectiveness of method of using a hybrid architecture neural network to determine the emotional tone of text messages using the created software is presented. It was found that the use of the specified hybrid architecture allows you to achieve an accuracy of 0.974, which is higher than currently known analogues by more than 0.07 for the Accuracy metric.