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

TL;DR

The paper tackles automatic detection of sexually explicit content in Spanish-language reggaeton and trap lyrics by fine-tuning a GPT model on a carefully annotated corpus of 100 songs. It demonstrates that domain-specific transfer learning yields high precision and strong overall accuracy, especially after a feedback-driven refinement loop, and outperforms a generic ChatGPT baseline in aligning with expert judgments. Beyond the technical results, the authors outline a public policy framework (MCAR) modeled after PEGI, supported by PESTEL and Kingdon’s MSF analyses, to enable a multi-tier age-based rating system for music content. The work highlights practical implications for parental controls and content-aware recommendations on streaming platforms while acknowledging limitations (small dataset, single annotator) and proposing future work on dataset expansion, multi-annotator labeling, and deployment considerations. Overall, it provides a viable technical path and policy blueprint for automated, culturally aware moderation of explicit music lyrics.

Abstract

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.

Figures (7)

• Figure 1: Pathways through which sexually explicit content in music can influence adolescents and children. Repeated exposure through streaming platforms leads to normalization of explicit content during critical periods of identity formation, potentially resulting in behavioral changes, attitude shifts, and value distortion. Research has documented associations between exposure to degrading sexual content in music and shifts in adolescent sexual behavior Martino2006.
• Figure 2: Overview of the proposed three-phase methodology. Phase 1 involves the construction and expert annotation of a balanced corpus of Spanish-language reggaeton and trap lyrics. Phase 2 applies transfer learning to fine-tune a pre-trained GPT model for binary classification of explicit content. Phase 3 evaluates the model through standard classification metrics and applies a feedback-driven refinement loop (dashed arrow) to improve performance iteratively.
• Figure 3: Overview of the fine-tuning process for explicit content detection. A pre-trained GPT model with general linguistic knowledge is adapted to the specific task of classifying reggaeton and trap lyrics using a labeled corpus of 100 songs. The fine-tuning process employs the AdamW optimizer with cross-entropy loss and a low learning rate to preserve pre-trained knowledge while learning domain-specific patterns.
• Figure 4: Hypothesis testing framework for evaluating the fine-tuned transformer model against traditional classification methods. The experimental comparison on 50 songs shows that the fine-tuned GPT model (59.2% agreement with expert) outperforms the standard ChatGPT baseline (55.1%), leading to rejection of the null hypothesis $H_0$.
• Figure 5: Proposed Music Content Age Rating (MCAR) framework with five age-based tiers and associated content descriptors. The system is modeled on PEGI's proven multi-tier structure, adapted for the specific content dimensions relevant to music lyrics.