An Analysis of Neural Network Predictions from the Linguistic Standpoint

This paper provides the evaluation of the most widely used methods of sentiment analysis on the dataset of Russian social networks texts ‘Rusentiment’. In this research Logistic Regression, Multinomial Naïve Bayes, Support Vector Machine, Random Forest and LSTM models are used for the purpose of sentiment classification. The main goal of the research is to extract and compare the most valuable features of each model and to find out what algorithms and methods of text representation are the most appropriate for this task. This research makes a comparison of models by comparing metrics values, predictions and the most significant features of them. The feature weights and predictions of the models are visualized with ELI5 Python library toolkits [Korobov M., Lopuhin K., 2019]. The linguistic interpretation of predictions is based on the distributive and statistical analysis of the most valuable features and extraction of the most frequent structural schemes of both classes.