Development of a Classification System for Texts by Scientific Specialties Using Machine Learning Methods

Abstract

In the process of preparing a dissertation, the researcher is faced with the problems of determining the research topic, as well as the problems of writing a text for a particular specialty. To increase the objectivity of the classification of scientific texts by specialties, a system has been developed using machine learning algorithms. In the process of building the classification system, the following tasks were solved: development tools were identified, the collection and processing of initial data was carried out, machine learning models were built, and a web application was developed. The initial dataset is a sample of texts for the group of Russian scientific specialties “Informatics and Computer Engineering”. For the objectivity of the study, the initial data was filtered - the least representative classes were removed. Conversion of texts for vectorization was also carried out. For the vectorization of the source texts, the TF-IDF model was used, which allowed loading the entire data volume with limited technical capabilities. Multiclass logistic regression is chosen as a machine learning model used to classify a scientific specialty. For training, the initial data divided into 2 parts - training and test in a ratio of 80 to 20. Accuracy is used as a metric of the quality of the machine learning model. The choice of the metric is due to the sufficient balance of the classes. The model trained on the training data made it possible to achieve the accuracy of determining the scientific specialty on test data equal to 0.87. To use a ready-made machine learning model for classification of a scientific specialty, a web application has been developed using Flask. The web app is currently located at http://predict-spec.herokuapp.com/ The most urgent tasks for finalizing the system at the moment are: transferring a web application to a more powerful server, finalizing machine learning models, displaying visual information on the analyzed work.