Poetry production from Karacaoğlan poems with artificial intelligence methods

This thesis aims to develop a machine learning model that can generate new poems in the style of one of the important figures of Turkish folk poetry, Karacaoğlan. The study analyzes the linguistic structure, themes, imagery, emotional depth, and technical features of Karacaoğlan's poetry to be utilized during the model training process through machine learning algorithms. A dataset was created using 665 poems by Karacaoğlan obtained from the website "antoloji.com," and preprocessing steps such as cleaning and normalization were applied. Long Short-Term Memory (LSTM) networks and encoder-decoder methods were employed in the study. LSTM networks were preferred for effectively capturing the time-series and linguistic features of the poems. Encoder-Decoder Models were used to learn the structural elements and meaning of the poems to generate new lines. Starting from an initial sentence, coherent poems were produced around a specific theme. The performance of the model was evaluated by adjusting the number of verses in the generated poems (2, 4, 6, 8, 10, 16 verses). The evaluation of the results utilized metrics of thematic analysis, emotional analysis, stylistic analysis, and historical context analysis, yielding formal results. However, the metric analyses of the generated poems did not reach the desired level, primarily due to the insufficient size of the generated poetry dataset. When the produced poems were reviewed by expert literary figures, it was noted that they somewhat reflected Karacaoğlan's style and emotions. However, it was also emphasized that they did not create a complete emotional impact and exhibited a mechanical expression. This situation demonstrates that artificial intelligence cannot fully express human emotions and modes of expression. The study opens new horizons for the use of machine learning in the field of literature. Future research is recommended to carry out similar studies across different languages and cultures and to examine content produced by both humans and artificial intelligence.