COMPARISON OF RNN AND LSTM NEURAL NETWORKS

Abstract

The research aims to identify the advantages and disadvantages of different approaches to sequential data processing, which is an important aspect in natural language processing tasks such as sentiment analysis, machine translation, and text generation. The purpose of the work. The purpose of the work is to investigate the effectiveness of different neural network architectures for the problem of sentiment classification, with an emphasis on comparing RNN and LSTM models. Methodology. The paper examines the theoretical aspects of the functioning of recurrent neural networks (RNN) and long-term short-term memory (LSTM) networks, which are specialized variants of RNN. An experimental comparison of four different neural network models, including simple recurrent networks (RNNs), LSTM networks, and convolutional neural networks (CNNs), applied to the sentiment classification task was conducted. For the experiment, the imdb_reviews dataset was chosen, which contains movie reviews intended for binary sentiment classification (positive or negative feedback). The implementation and training of the models was done using the TensorFlow and Keras libraries, which provide a toolkit for efficient machine learning. The process of training and testing the models took place using standard approaches to preprocessing textual data, such as tokenization and sequence preparation. Scientific novelty. It is shown that the main advantage of LSTM is the ability to solve the problem of long-term dependencies, which makes them more effective for tasks where it is important to take into account the context of long data sequences. It has been experimentally confirmed that the training time of recurrent neural networks is significantly longer compared to non-recurrent models, but they demonstrate slightly better accuracy. Conclusions. The results of the study indicate that the use of LSTM networks is a more effective approach for solving complex problems that require consideration of the context at the level of sequences exceeding in length typical fragments of text. LSTMs are superior to them due to the ability to preserve long-term dependencies, which is especially important in tasks where it is necessary to take into account the relationship between distant data elements.