Structural and philological features of text generative neural networks

Abstract

INTRODUCTION. Studying the features of text generative neural networks is an important step in the development of artificial intelligence. Despite the fact that the models have shown high efficiency in solving various problems in the field of journalism and media communications, they have a number of disadvantages. When working with neural networks, you can encounter both gross grammatical and semantic errors. To identify the leader in the most productive text generation, it is necessary to conduct a comparative analysis of the data produced by various services.MATERIALS AND METHODS. In the Russian segment, the most developed neural network services are GigaChat and YandexGPT. To conduct a comparative analysis, the most discussed and generally recognized service was selected – GhatGPT. The study was carried out over several months: September–December 2023. The methodology is based on philological analysis of generated texts and comparison of the accuracy of query output of selected models.RESULTS AND DISCUSSION. Philological and grammatical analysis of the three models allows us to determine the relevance of services for work in the field of journalism and media communications, as well as the software and technical limitations of neural networks. The analysis showed the presence of certain patterns in all neural network models. Generation is carried out according to a preprogrammed scenario. The result consists of a number of factors: the presence of names, abbreviations and wishes specified in the request. Only ChatGPT showed the absence of any censorship; other models refused to generate if the request contained words or names prohibited by the developer.CONCLUSION. The findings can be applied in practice in the media, blogging and media sphere. All three services have their positive and negative sides. According to the results of the study, ChatGPT is currently the leader in text generation and processing. Leadership of the service is ensured due to a wide range of capabilities and stability in issuing answers to requests. However, due to the availability of a large amount of information on the Internet necessary for rapid training of Russian networks, the situation may change in the near future.