Methods of Literary Text Generation and Their Testing (Based on Material of Russian Short Story)

At the current stage of language technology development, it is possible in a short period of time to generate text that is similar to human-written. The aim of this study is to investigate how language models perform in the task of creating a literary text in a genre of short story. This research examines two approaches to short story generation. The first approach utilizes pre-trained GPT-architecture models, such as GPT-3.5 from the ChatGPT chatbot and ruGPT-3.5 from GigaChat. The second approach involves fine-tuning a ruGPT-3-Small model using a corpus of short stories written by authors and a corpus of Russian Short Stories. The results of the study are analyzed using a combination of quantitative and qualitative methods. Quantitative analysis considers metrics such as the average sentence length, while qualitative analysis evaluates the generated stories based on human feedback. Overall, the results of the study suggest that, at the moment, it is often possible to distinguish between human-written and machine-generated text.