An Intelligent System for Assessing the Emotional Connotation of Textual Statements
E. S. Yusifov, Синева Ирина Сергеевна · 2022 Wave Electronics and its Application in Information and Telecommunication Systems (WECONF) · 2022
The objective was to use classical machine learning algorithms (without neural networks) to train a model capable of recognizing emotions in a text with an accuracy as close as possible to transformers. The following methods were considered: support vector method, linear models with stochastic gradient descent, K-nearest neighbor method, Random Forest with boosting (CatBoost). To reduce outliers, the GoEmotions dataset was pre-processed by removing annotations with consistency less than two, selecting the most popular annotations, cleaning the text from punctuation and other outliers, and stemming. The paper specified suboptimal parameters for four methods of model building based on machine learning. The study concluded that GoEmotions texts in the feature space have a complex connected structure within classes and therefore are difficult to partition using classical methods.