CLASSIFICATION OF RUSSIAN-LANGUAGE SHORT TEXTS: COMPARISON OF TF-IDF+ML AND SIMPLE NEURAL NETWORK MODELS BASED ON UNIFIED ASSESSMENT PROTOCOLS
TL;DR
The results show that with a limited amount of training data, TF-IDF-based models provide quality comparable to simple neural network architectures, while neural network approaches demonstrate an advantage when increasing the size of the corpus.
Abstract
The paper considers the problem of automatic classification of Russian-language short texts using traditional statistical and neural network approaches. The aim of the study is to compare the effectiveness of TF-IDF models in combination with classical machine learning algorithms (SVM, logistic regression) and a simple neural network architecture when solving the same classification problem using unified assessment protocols. During the experiments, the influence of various methods of text preprocessing (cleaning, stemming, lemmatization) on the classification quality and the stability of models to noisy data is analyzed. The novelty of the work lies in a replicated comparison of traditional and neural network methods on Russian-language corpora, as well as in assessing the impact of morphological normalization on quality indicators. The results show that with a limited amount of training data, TF-IDF-based models provide quality comparable to simple neural network architectures, while neural network approaches demonstrate an advantage when increasing the size of the corpus. The results of the research can be used in the development of text analysis systems, content filtering and intelligent dialog interfaces for the Russian language.