PaperlayerView exact revision on arXiv

Source-linked AI summary

Probing Cultural Signals in Large Language Models through Author Profiling

Valentin Lafargue, Ariel Guerra-Adames, Emmanuelle Claeys, Elouan Vuichard, Jean-Michel Loubes

arXiv:2603.16749v2cs.CLcs.LG

TL;DR

LLMs are increasingly used in socially consequential settings, motivating closer study of how they represent cultural identity. This paper evaluates zero-shot gender and ethnicity profiling from song lyrics across instruction-tuned models, finding above-chance performance alongside systematic, culturally uneven behavior. It introduces fairness metrics and uses model rationales to characterize these disparities and support interpretability.

  • Problem

    The paper addresses limited evidence about how LLMs infer and represent cultural identity, especially in demographic profiling from culturally rich song lyrics.

  • Method

    The study evaluates multiple open-source instruction-tuned LLMs on zero-shot singer gender and ethnicity profiling from lyrics, combining prompt variants, fairness metrics, and rationale analysis.

  • Results

    Models achieve above-chance profiling performance but show sensitivity to prompt design, genre, lyric length, and systematic disparities across demographic modalities.

  • Takeaways & Limitations

    LLMs encode rich stylistic knowledge while behaving non-uniformly across cultural contexts, so demographic inference tasks require careful evaluation of cultural representation.

  • Takeaways & Limitations

    Opaque training-data provenance limits causal interpretation of cross-model cultural-alignment comparisons.

Abstract

from arXiv · show

Large language models (LLMs) are increasingly deployed in applications with societal impact, raising concerns about the cultural biases they encode. We probe these representations by evaluating whether LLMs can perform author profiling from song lyrics in a zero-shot setting, inferring singers' gender and ethnicity without task-specific fine-tuning. Across several open-source models evaluated on more than 10,000 lyrics, we find that LLMs achieve non-trivial profiling performance but demonstrate systematic cultural alignment: most models default toward North American ethnicity, while DeepSeek-1.5B aligns more strongly with Asian ethnicity. This finding emerges from both the models' prediction distributions and an analysis of their generated rationales. To quantify these disparities, we introduce two fairness metrics, Modality Accuracy Divergence (MAD) and Recall Divergence (RD), and show that Ministral-8B displays the strongest ethnicity bias among the evaluated models, whereas Gemma-12B shows the most balanced behavior. Our code is available on [GitHub](https://github.com/ValentinLafargue/CulturalProbingLLM) and results on [HuggingFace](https://huggingface.co/datasets/ValentinLAFARGUE/AuthorProfilingResults).

1 Introduction

The paper probes whether LLMs can infer singers’ socially constructed gender and perceived ethno-cultural identity from lyrics, while testing systematic bias in those predictions. It evaluates zero-shot profiling and examines the factors, fairness disparities, and rationales associated with model decisions.

  • Motivation: LLMs are increasingly used in socially consequential domains where failures to represent cultural identity can reinforce inequity or silence minority voices.The paper distinguishes these representational harms from failures of factual knowledge.
  • Research framing: Author profiling infers sociodemographic and psychological attributes from an author’s creative output using cultural references and linguistic patterns.The paper adopts ethnicity as perceived ethno-cultural identity and gender as socially constructed gender.
  • Research framing: Song lyrics provide a rich, understudied profiling domain combining personal expression, genre conventions, and culturally embedded linguistic markers.Compared with shorter curated texts, lyrics are described as less censored, stylistically diverse, and deeply rooted in cultural contexts.
  • Research questions: The study asks whether LLMs can profile authors zero-shot, which factors influence decisions and performance, and whether predictions show gender or ethnic bias.These questions organize the evaluation of capability, decision factors, and systematic disparities.
  • Contributions: Multiple LLMs are evaluated without fine-tuning, using sociolinguistically informed prompts, statistical bias analysis with MAD and RD, and generated-rationale auditing.The rationale analysis supports interpretable explanations for prompt design and examination of cultural representations.

2 Related Work

Prior profiling research largely uses supervised classification or studies authorship, leaving limited evidence about zero-shot demographic inference and cultural bias in LLMs. This paper addresses that gap with song lyrics and analysis of predictions, biases, prompts, and explanations.

  • Existing profiling research: Earlier author-profiling work framed the task as text classification and established benchmarks primarily with supervised classifiers and explicit feature engineering.The cited tradition includes attributes such as gender and age.
  • Sociolinguistic foundations: Sociolinguistic studies document gender-linked differences in politeness, emotional expression, modality, and lexical choice, motivating linguistic cues as identity signals.These findings provide empirical grounding for gender inference from text.
  • Sociolinguistic foundations: Research links language patterns with census-based racial and ethnic categories, treating language as an expression of socially constructed ethnic identity.The paper notes that some specific associations have been contested.
  • LLM-based research: LLMs have been used for fine-tuned demographic prediction and zero-shot authorship attribution, but prior studies give limited attention to sociodemographic inference and cultural bias.The contrast concerns both task focus and analysis of bias.
  • This paper: This study examines zero-shot profiling from long-form, culturally rich song lyrics and analyzes prediction bias alongside prompt design, explanations, and sociolinguistic cues.Linguistic features serve as an analytical lens rather than an explicitly modeled feature set.

3 Data acquisition

The corpus merges lyric and metadata sources, harmonizes artist attributes, filters the analytical sample, and translates non-English lyrics into English before profiling. Dataset construction relies on solo-artist assumptions and has acknowledged representation limits.

  • Sources: Lyrics were obtained from Deezer and Spotify-related datasets, with metadata collected through Wasabi, Genius, and MusicBrainz resources.The sources were combined across the two acquisition pipelines.
  • Assumptions: The study excludes bands and assumes solo artists select songs aligned with their public artistic identity, even when they are not sole lyricists.Ghostwriters and multiple writers remain potential challenges for interpreting perceived lyricist ethnicity.
  • Filtering and harmonization: The corpus merges two lyric datasets and harmonizes metadata, retaining solo artists labeled man or woman while filtering groups, non-binary, and ambiguous cases.The unified schema initially includes multiple gender labels before analytical filtering.
  • Filtering and harmonization: The dataset excludes non-binary singers because five artists had fewer than ten songs each, leaving transgender representation underrepresented.The authors identify this as a limitation of available metadata and a direction for future work.
  • Translation: Non-English lyrics were translated into English before profiling to control overt grammatical gender cues and emphasize stylistic and semantic signals.Translations used Mistral Small 3.2 in a zero-shot configuration.

4 Author profiling methodology

The methodology evaluates small-to-medium open-source instruction-tuned LLMs on zero-shot author profiling from lyrics through multiple prompt variants. It combines model selection, prompt design, and bias/performance analysis with self-explanation auditing.

  • 4.1 Model Selection: The study selects open-source models ranging from 7B to 24B parameters to reduce memorization risk while retaining evaluation capability.The authors argue that smaller training samples make exposure to non-translated songs less likely than with much larger models.
  • Workflow: The workflow covers lyric collection, metadata harmonization, gender and macro-region balancing, translation, zero-shot profiling, prompt variants, and bias/performance analysis.Self-explanation auditing complements the main workflow.
  • 4.1 Model Selection: The model set spans Chinese, American, and European developers to examine behavior across distinct geographic regions.The listed examples include Qwen, DeepSeek, Llama, Gemma, Ministral, and Mistral models.
  • Prompt design: Five prompts form an incremental sequence from direct inference to informed cues, expressive rationales, sociolinguistic attributes, and corrected ethnicity instructions.The variants progressively add constraints, explanations, attribute evaluation, or guidance to avoid specific ethnicity errors.
  • Prompt design: The corrected informed prompt instructs models not to use theme or emotions when predicting ethnicity, based on rationale-derived errors.This variant adds the instruction to the informed prompt.

5 Bias evaluation methodology

The methodology evaluates whether balanced datasets reveal distributional and per-modality disparities in zero-shot LLM author profiling, using statistical tests and complementary fairness metrics.

  • Distributional bias: Balanced datasets enable comparison between inferred and ground-truth sociodemographic modality distributions without majority-class collapse.The analysis compares prediction distributions with balanced ground truth and uses one-vs-rest accuracy for modalities.
  • Distributional bias: Three complementary tests—chi-squared, Central Limit Theorem–based, and Wasserstein distance–based—assess whether predicted modality distributions match balanced ground truth.Stratified bootstrap resampling with 1,000 iterations accounts for sampling variability.
  • Modality Accuracy Divergence: Modality Accuracy Divergence (MAD) measures each modality’s relative accuracy deviation from the macro-average.MAD is scale-free and supports comparisons across models and experimental conditions, but cannot detect symmetric binary-classification bias.
  • Modality Accuracy Divergence: MAD uses one-vs-rest membership accuracy, capturing both false negatives and false positives for each modality.The metric summarizes uneven accuracy across categories of a sensitive attribute.
  • Recall Divergence: Recall Divergence (RD) quantifies disparities in per-modality recall and identifies groups that are consistently harder to recognize.Unlike overall accuracy, RD focuses on correct identification when the true modality is given and isolates group-wise under-recognition.
  • Recall Divergence: ROC curve points complement MAD and RD by comparing per-modality prediction behavior across gender and ethnicity for each LLM.MAD additionally reflects false-positive tendencies, whereas RD measures disparities in true-positive behavior.

6 Result

Instruction-tuned LLMs achieve above-chance zero-shot author profiling from song lyrics, but their predictions show systematic modality disparities shaped by model, prompt, and lyric characteristics.

  • Zero-shot profiling: 76%±2% gender accuracy and 44%±2% ethnicity accuracy were achieved by Mistral-24B without task-specific fine-tuning.The intervals are 95% confidence intervals obtained using stratified bootstrapping on songs.
  • Prediction distributions: DeepSeek-1.5B, Ministral-8B, and Mistral-24B significantly over-predict the woman modality.
  • Prediction distributions: North American ethnicity has high TPR and FPR for Llama-8B, Qwen-7B, Ministral-8B, and Mistral-24B, functioning as a default category.This aligns with the distributional shift observed in the ethnicity predictions.
  • Factors influencing performance: Larger models outperform smaller ones, lyric length is positively associated with accuracy, rap outperforms electronic dance music, and translation effects are broadly robust.The Well-informed prompt performs statistically worse because of its complexity, while the Corrected informed prompt slightly improves ethnicity inference.
  • Self-explanations: False ethnicity rationales disproportionately use emotional and thematic arguments, motivating a Corrected informed prompt that produces less biased and more accurate predictions.The corrected prompt tells models not to use theme or emotions for ethnicity prediction.
  • Self-explanations: −0.31 ± 0.02 was the overall correlation between North American ethnicity predictions and cultural references, compared with −0.43 ± 0.04 for Gemma-12B and −0.27±0.04 for Mistral-24B.
  • Bias evaluation: Every model had at least one over- or under-represented ethnicity modality, while Ministral-8B, Mistral-24B, and Qwen-7B rejected uniform gender predictions.
  • Bias evaluation: Ministral-8B is the most ethnicity-biased model, whereas Gemma-12B and Llama-8B obtain good ethnicity fairness results; DeepSeek-1.5B shows a bias-performance trade-off.For gender, DeepSeek-1.5B is the most biased, Gemma-12B has low bias, and Ministral-8B has high bias.

7 Conclusion

The paper evaluates zero-shot author profiling from song lyrics and introduces fairness-oriented analysis of demographic predictions. Models perform above chance but remain sensitive to prompts, genres, lyric length, and cultural context, with systematic disparities across modalities.

  • Instruction-tuned LLMs perform zero-shot author profiling from song lyrics without task-specific fine-tuning, achieving above-chance performance.
  • Models are sensitive to prompt design, musical genre, and lyric length, while translation effects are broadly robust.
  • Fairness-oriented metrics reveal systematic modality disparities, uneven cultural representation, and potential blind spots in demographic inference.
  • Rationale analysis suggests explanation prompting can improve interpretability, although it involves a small performance trade-off.
  • The findings support careful evaluation of demographic inference tasks and motivate future work on multilingual profiling and mitigation strategies.

8 Limitations

The study’s findings are constrained by the stylized lyric domain, simplified regional ethnicity categories, dataset-selection issues, translation artifacts, and opaque model training provenance.

  • Song lyrics are stylized and genre-dependent, so the findings may not generalize to other textual domains or communicative settings.
  • Regional operationalization approximates ethnicity and simplifies complex, fluid, and culturally varied social identities.
  • Some songs contain uninformative self-name mentions or localization references, and the dataset assumes singers wrote lyrics despite ghostwriters and co-writers.
  • Translated lyrics may distort stylistic, lexical, and discourse cues despite robustness analyses.
  • Opaque training-data origins and curation processes complicate interpretation of observed cultural alignment.

9 Ethical considerations

The paper treats ethnicity as a sociocultural construct and highlights ethical risks in inferring sensitive identities from lyrics, while calling for culturally aware evaluation and transparent reporting.

  • Ethnicity is treated as a sociocultural and analytical construct rather than a biological category.
  • Macro-regional categories simplify cultural variation and should not be understood as exhaustive or definitive.
  • Inferring gender and ethnicity from writing can reinforce stereotypes, essentialize cultural expression, and encourage reductive interpretations of identity.
  • The paper calls for culturally aware evaluation, clearer reporting of model training data, and interdisciplinary collaboration.
  • Featured artists introduced noise: performance dropped on Eminem’s collaboration with Sia, whose verses interfered with profiling.
  • Models sometimes rely on shallow lexical or thematic cues, producing culturally misleading rationales despite correct or incorrect predictions.
  • Miriam Makeba’s song was frequently misclassified as male, illustrating risks when gendered interpretations are inferred from lyrical narratives.

B Linguistic attribute evaluation with the Well-informed prompt as hypothesis foundation

The Well-informed prompt evaluates sociolinguistic attributes before or alongside profiling, revealing prompt-order effects, model-specific correlations, and descriptive links between language attributes and predictions.

  • B.1 Attribute evaluation: The Well-informed prompt has Attribute-first and Reasoning-first variants that change whether attribute evaluation precedes author profiling.
  • B.1 Attribute evaluation: Complex Well-informed instructions frequently yielded invalid predictions for DeepSeek models, while Ministral-8B failed on approximately 5% of balanced-dataset lyrics.
  • B.1 Attribute evaluation: Attributes are scored from 1 to 10; lyrics show higher Emotions and lower Politeness and Formal language than general written text.
  • B.2 Variant difference analysis: Gemma-12B is least affected by prompt ordering, whereas Llama-8B and Mistral-24B show the largest deviations, though overall variation remains modest.
  • B.3 Correlation study: Emotions, Romance, and Uncertainty correlate positively with feminine predictions, while Success, Slang usage, and Profanity correlate more strongly with masculine predictions.
  • B.3 Correlation study: Cultural references correlate negatively with North American predictions and positively with African, Asian, and South American predictions.
  • B.3 Correlation study: The analysis characterizes model behavior descriptively through assimilation, opposition, and blindness patterns without inferring causal mechanisms.
  • B.3 Correlation study: Mistral-24B has the most distinct correlation structure, with a more balanced tendency between North American and European predictions.

D Factors Influencing LLM Performance in Author Profiling

Profiling performance varies with prompt design, model scale, lyric length, emotion, translation, genre, and statistical distributional behavior, with several factors exposing uneven cultural sensitivity.

  • The analysis examines prompt design, model scale, lyric length, genre, translation, and emotions as factors influencing profiling performance.
  • Prompt sensitivity: Complex Well-informed prompts lower accuracy, while Corrected informed prompts modestly improve ethnicity prediction; explicit rationales slightly reduce performance.
  • Model scale: Larger models consistently outperform smaller ones, including DeepSeek-7B over DeepSeek-1.5B and Mistral-24B over Ministral-8B.
  • Lyric length: Profiling accuracy increases monotonically with lyric length, although length correlates with gender and ethnicity and may confound apparent gains.
  • Lyric emotions: Emotion-based classifiers perform significantly above chance on balanced datasets, but emotion vectors do not explain LLM-predicted labels.
  • Translation effects: Translation has limited gender impact, while ethnicity effects vary by model and prompt; overall shifts remain roughly 3% for ethnicity and 2% for gender.
  • Musical genre: Rap outperforms EDM, coinciding with substantially longer lyrics averaging 623 words versus 310 words.
  • Statistical tests: No model has a uniform ethnicity prediction distribution overall, with DeepSeek-1.5B closest to uniformity and Gemma-12B next closest.

F Non-Fairness Metric Evaluation

Standard fairness metrics are not directly applicable when the predicted label is itself the sensitive attribute, motivating distinct evaluation criteria. The comparison also separates predictive performance from bias, showing that low accuracy can coexist with low measured bias.

  • Metric applicability: Traditional fairness definitions cannot be meaningfully applied because the target label is the sensitive attribute itself.Their usual independence assumption between predictions and sensitive variables does not hold in this author-profiling setting.
  • Metric interpretation: Macro recall and macro F1-score measure predictive quality by averaging class-wise scores, rather than fairness.Higher values indicate better performance, unlike the proposed fairness metrics, where values closer to 0 indicate better fairness.
  • Results: The DeepSeek-1.5B model has the lowest macro F1-score and macro recall but is identified as the least biased model under the proposed fairness criterion.It was also the only model to consistently predict the Oceania modality.

G Comparison with other fairness metric on a non-author profiling task.

The paper contrasts Recall Divergence with established fairness metrics and tests it in standard binary-classification benchmarks. These experiments show that RD measures class-performance divergence outside author profiling, supporting its primary use when the target itself is a sensitive attribute.

  • Motivation: Statistical parity is incompatible with predictive performance when the sensitive variable is also the prediction target.This setting motivates fairness metrics designed specifically for author profiling.
  • Metric comparison: Disparate Impact and Equality of Odds evaluate group-related outcome or error disparities, whereas Recall Divergence captures a different form of divergence.DI and EoO use sensitive-attribute conditioning, while RD is evaluated here through class-conditioned predictive performance.
  • Metric interpretation: DI and EoO ratio values closer to 1 indicate greater fairness, while additive fairness scores closer to 0 indicate better parity.The paper emphasizes that metric values cannot be interpreted identically across these formulations.
  • Experimental setting: The comparison evaluates RD with a Random Forest on eight fairness datasets spanning population characteristics, credit approval, recidivism, and law-school admissions.The datasets include INC, MOB, EMP, TRA, PUC, CREDIT, COMPAS, and BAR, with varying sensitive attributes.
  • Results: In standard binary classification, RD measures divergence between positive- and negative-class predictive performance rather than fairness.Its value does not depend on the sensitive attribute, and it can become +∞ when the positive-class prediction expectation is zero.
  • Conclusion: The paper recommends RD primarily for author-profiling tasks, where the target variable itself corresponds to a sensitive attribute.The recommendation follows from the metric's differing interpretation in conventional classification benchmarks.
  • Dataset design: The study uses two balanced datasets because ethnicity and gender analyses require different balancing choices and produce different prompt-performance conclusions.Creating the gender-balanced dataset reduced the singer sample from about 3,000 to 2,000 while retaining a sample considered statistically relevant.
  • Preprocessing: Non-English lyrics were translated into English because grammatical gender markers can provide direct cues to author profiling.The pipeline detects languages at lyric and fragment levels, with a second pass intended to catch mixed-language or misclassified lyrics.
Loading 2603.16749v2…