Fine-Tuning Large Language Models for Automatic Detection of Sexually Explicit Content in Spanish-Language Song Lyrics

The proliferation of sexually explicit content in popular music genres such as reggaeton and trap, consumed predominantly by young audiences, has raised significant societal concern regarding the exposure of minors to potentially harmful lyrical material. This paper presents an approach to the automatic detection of sexually explicit content in Spanish-language song lyrics by fine-tuning a Generative Pre-trained Transformer (GPT) model on a curated corpus of 100 songs, evenly divided between expert-labeled explicit and non-explicit categories. The proposed methodology leverages transfer learning to adapt the pre-trained model to the idiosyncratic linguistic features of urban Latin music, including slang, metaphors, and culturally specific double entendres that evade conventional dictionary-based filtering systems. Experimental evaluation on held-out test sets demonstrates that the fine-tuned model achieves 87% accuracy, 100% precision, and 100% specificity after a feedback-driven refinement loop, outperforming both its pre-feedback configuration and a non-customized baseline ChatGPT model. A comparative analysis reveals that the fine-tuned model agrees with expert human classification in 59.2% of cases versus 55.1% for the standard model, confirming that domain-specific adaptation enhances sensitivity to implicit and culturally embedded sexual references. These findings support the viability of deploying fine-tuned large language models as automated content moderation tools on music streaming platforms. Building on these technical results, the paper develops a public policy proposal for a multi-tier age-based content rating system for music analogous to the PEGI system for video games analyzed through the PESTEL framework and Kingdon’s Multiple Streams Framework, establishing both the technological feasibility and the policy pathway for systematic music content regulation.

💡 Research Summary

The paper addresses the growing societal concern that sexually explicit lyrics in popular urban Latin music—particularly reggaeton and trap—expose minors to potentially harmful content. Traditional keyword‑based filters fail because these genres rely heavily on slang, metaphors, and culturally specific double entendres that are absent from standard dictionaries. To overcome this limitation, the authors fine‑tune a Generative Pre‑trained Transformer (GPT) model on a curated corpus of 100 Spanish‑language songs, evenly split between expert‑labeled explicit and non‑explicit tracks.

The dataset was assembled by domain experts who not only assigned binary labels but also compiled a reference table of explicit phrases, euphemisms, and idiomatic expressions (e.g., “bellaquear,” “perreo,” “mamacita”). The fine‑tuning process employed supervised learning with binary cross‑entropy loss and the AdamW optimizer. An initial evaluation yielded modest performance (≈78 % accuracy), prompting the introduction of a feedback‑driven refinement loop: misclassifications identified by experts were fed back into the training set, and the model was re‑trained. After this iterative process the model achieved 87 % overall accuracy, with both precision and specificity reaching 100 %.

Beyond standard metrics, the authors measured agreement with human experts. The fine‑tuned model matched expert judgments in 59.2 % of cases, outperforming a non‑customized baseline ChatGPT (55.1 %). Notably, false negatives were dramatically reduced, a critical factor for protecting younger audiences.

The technical contribution is complemented by a policy component. Using a PESTEL analysis (Political, Economic, Social, Technological, Environmental, Legal) and Kingdon’s Multiple Streams Framework, the paper proposes a three‑tier age‑based music rating system analogous to the PEGI system for video games. The system would automatically assign “All‑Ages,” “Teen,” or “Adult” labels to tracks, enabling streaming platforms to enforce parental controls, curate age‑appropriate playlists, and provide transparent content warnings. This approach addresses the current regulatory gap in the music industry, where voluntary “Parental Advisory” labels are inconsistently applied.

In summary, the study demonstrates that (1) a large language model can be effectively adapted to a highly nuanced, low‑resource domain through modest expert‑annotated data, (2) a feedback loop substantially improves both sensitivity and trustworthiness of the classifier, and (3) coupling technical results with a structured policy framework yields a viable roadmap for deploying automated content moderation tools on music streaming services, thereby enhancing protection for vulnerable listeners while preserving artistic expression.