Synchronizing Graph and Embedding Representations of Lyrical Data with Graph Neural Networks and Developing a Classification-to-Generation Workflow For Genre-Specific Songs

Advancements in large language models and their ability to quickly generate prompt-based content has impacted the classification and generation of creative textual data, such as song lyrics. These advancements have generated attention within the machine learning research community and public domains. The unique structure of song lyrics, which follow spe- cific forms and exhibit a natural flow not evident in prose, poses a challenge to generating new songs. This project explores the integration of graph neural networks (GNNs) with transformer-based language models to improve genre-specific song lyric generation and pro- vides an extensive computational background for each technique utilized. Songs are repre- sented through a heterogeneous graph where song documents and lyric words are nodes con- nected by occurrence-based edges, and additionally as fine-tuned Bidirectional Encoder Rep- resentations from Transformers (BERT) embeddings. These feature representations are both used to train the graph neural network, which differs from traditional embeddings-only song classification techniques. The GNN is trained to classify songs by genre, capturing structural and semantic relationships between lyrics and musical genre, utilizing the dual representation of song lyric features. The classification features are used during prompted song generation to guide GPT-2, a transformer, in real time by modifying token selection based on genre- specific lyric relationships and overall genre proximity within the graph of the generated song. Thus, generation incorporates context-aware constraints to enhance genre-specificity. The classification-to-generation approach is evaluated using BLEU scores, cosine similarity, and perplexity to analyze genre consistency and creativity between original and generated lyrics. Results suggest that real-time graph-guided generation produces more coherent and genre-aligned lyrics, demonstrating a promising hybrid model for structured creative text gen- eration and proving the robustness of dual-representative features for song lyric classification.