Generating new songs from lyrics: Artistic production using deep learning methods

Writing song lyrics is a creative process. Traditionally, this process relies on inspiration, intuition, and individual creativity. However, with the advancement of artificial intelligence technologies, this process has entered a new phase. In recent years, significant progress has been made in the field of Natural Language Processing (NLP), particularly in creative text generation. GAN architectures, in particular, have attracted attention for generating content that requires originality and randomness. Nonetheless, applying GANs to text generation remains more complex than their application in image generation. In agglutinative languages such as Turkish, generating text that is both meaningful and structurally coherent is even more challenging. Architectures such as TextGAN, SeqGAN, LeakGAN, RankGAN, MaliGAN, and RelGAN have not been systematically evaluated in the context of Turkish song lyric generation. This constitutes the fundamental problem of the study. The aim of this thesis is to identify the most suitable artificial intelligence model capable of generating original and artistically valuable lyrics by analyzing existing song lyrics. In this scope, emotion analysis, rhythmic structure analysis, and creative text generation techniques were employed alongside deep learning models. According to the results, the low BLEU and Jaccard scores indicate that the models struggled to establish surface-level similarity with reference texts. In contrast, the RelGAN model achieved the highest score in Cosine Similarity (0.0693), producing the most semantically coherent outputs. The MaliGAN model stood out in terms of language modeling with a low perplexity value (9588.23). However, the fact that diversity metrics (TTR and Unique Word Ratio) reached 1.0 suggests a lack of meaningful repetition and a weak natural flow. Based on literary analysis, only the RelGAN model was able to produce song lyrics that were satisfactory in terms of thematic consistency and emotional depth. In conclusion, considering both numerical performance and literary evaluation, RelGAN emerged as the most balanced and successful model. GAN architectures remain limited in generating poetry in agglutinative languages like Turkish, and future studies are recommended to incorporate Transformer-based models for improved outcomes.