Skip to content

probing cultural signals in large language models through author profiling

content_hash: 052410d0b54af657 ---# Probing Cultural Signals in Large Language Models through Author Profiling

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/vlafargue/author_profiling_llm
Area: Social Computing
Keywords: Author Profiling, Cultural Bias, Large Language Models, Song Lyrics, Fairness Metrics

TL;DR

This paper evaluates the zero-shot author profiling capabilities of 7 open-source LLMs on over 10,000 English-translated song lyrics to infer artist gender and continental ethnicity, introduces Modality Accuracy Divergence (MAD) and Recall Divergence (RD) to quantify asymmetric subgroup bias when the target itself is sensitive, reveals a pervasive North American default alignment across most models (with DeepSeek-1.5B skewing Asian), and presents a rationale-guided debiasing prompt.

Background & Motivation

Large language models (LLMs) are increasingly deployed in high-stakes socio-technical systems such as automated educational grading and cross-border content moderation. In educational applications, LLM evaluators often mischaracterize diverse, dialectal, or culturally distinct writing styles as lower quality, directly penalizing minority students. Similarly, in automated content moderation, systems frequently over-flag vernacular and non-standard dialects as toxic or abusive, suppressing marginalized voices. Crucially, these harms do not stem from simple deficits in factual world knowledge, but rather from fundamental failures of LLMs to interpret and fairly represent diverse cultural identities—a cognitive failure termed cultural blindness, where models remain incapable of recognizing cultural perspectives distinct from their training defaults.

Author profiling—the task of inferring sociodemographic and psychological attributes of an author from their creative textual output—serves as a natural diagnostic probe for these latent cultural representations. Previous research has primarily approached author profiling via supervised classification with hand-engineered stylistic features on short, curated texts like social media posts, or framed LLM evaluations around binary authorship attribution. In contrast, creative song lyrics represent a dense, highly normative, and deeply culturally embedded linguistic medium that combines personal expression, vernacular idioms, and literary conventions while avoiding the heavy censorship and brevity of microblogs. However, studying cultural profiling on song lyrics encounters two major tensions: first, multilingual lyrics contain explicit grammatical gender markers (such as gendered adjective inflections in Romance languages) that offer superficial shortcuts rather than probing deep stylistic reasoning; second, standard group fairness criteria (e.g., Disparate Impact and Equality of Odds) assume that model predictions should be statistically independent of sensitive demographic attributes, making them mathematically inapplicable when the prediction target itself is the sensitive attribute.

To resolve these challenges, this paper curates a standardized benchmark of 10,808 songs with metadata mapped to macro-continental regions, translating all non-English lyrics into English to strip surface grammatical gender markers and isolate stylistic, thematic, and cultural signals. Core idea: probe LLM cultural representations through zero-shot lyric author profiling, introduce Modality Accuracy Divergence (MAD) and Recall Divergence (RD) to measure asymmetric subgroup parity under self-sensitive prediction, and leverage rationale error word-cloud contrasts to formulate debiased prompts.

Method

Overall Architecture

The evaluation pipeline comprises four core stages: data acquisition and grammatical debiasing, zero-shot multi-model author profiling, asymmetric fairness evaluation, and self-explanation error auditing. The system gathers 10,808 tracks from Spotify and Deezer, deduplicates them using TF-IDF title cosine similarity, and maps artist origins into six macro-regions (Africa, Asia, Europe, North America, Oceania, South America) alongside binary gender. Non-English lyrics are translated into English via Mistral-24B under zero-shot deterministic decoding to remove gendered inflection shortcuts. Seven open-source instruction-tuned LLMs originating from the US, Europe, and China (DeepSeek-1.5B/7B, Qwen-7B, Llama-8B, Gemma-12B, Ministral-8B, Mistral-24B) are evaluated across five incremental prompt variants to predict artist gender and macro-continental ethnicity. Model predictions are evaluated on ethnicity-balanced and gender-balanced subsets via stratified bootstrapping using MAD, RD, and statistical distribution tests, culminating in differential word-cloud auditing of generated rationales to formulate debiasing prompt corrections.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Lyric Data Sanitization & Grammatical Debiasing<br/>Spotify/Deezer crawl + TF-IDF deduplication + English translation"] --> B["Progressive Prompt Engineering & Sociolinguistic Probes<br/>Regular / Informed / Expressive / Well-informed / Corrected"]
    B --> C["Multi-Model Zero-Shot Author Profiling<br/>7 US/EU/CN open-source LLMs infer gender & continental ethnicity"]
    C --> D["Dual Asymmetric Fairness Evaluation<br/>MAD (Modality Accuracy Divergence) + RD (Recall Divergence)"]
    D --> E["Rationale Error Auditing & Reflective Debiasing<br/>Word-frequency subtraction isolates theme/emotion shortcuts"]

Key Designs

1. Grammatical Gender Stripping and Macro-Regional Metadata Harmonization

Directly inferring demographic attributes from multilingual song lyrics risks letting models exploit explicit morphological gender inflections (e.g., Spanish cansado vs. cansada) rather than reasoning over stylistic and cultural signals. To force models to rely strictly on semantic perspective, cultural references, and socio-affective tone, an automated language filtering and normalization pipeline is established: lyrics are segmented into sentence fragments and classified using langid and langdetect (lyrics with \(\ge 80\%\) English fragments and \(< 15\%\) out-of-vocabulary words against standard English lexicons are retained as English). All non-English sections are translated locally using Mistral-24B on dual A100 GPUs under greedy decoding (\(\text{temperature} = 0.0\)), strictly preserving stanza breaks, formatting, and non-lexical vocalizations (la la la). Metadata from Genius, Spotify, and MusicBrainz are standardized, mapping artist nationalities onto six macro-regions (Africa, Asia, Europe, North America, Oceania, South America) and constructing stratified balanced subsets to eliminate natural geographic class skews.

2. MAD and RD Fairness Metrics for Self-Sensitive Prediction

Conventional algorithmic fairness criteria like Disparate Impact quantify independence between a classification decision and an extraneous sensitive attribute \(S_i\). In author profiling, however, the target variable \(\hat{S}_i = f_i(X)\) is identical to the sensitive attribute itself, causing traditional statistical parity to conflict directly with model accuracy. To address this, the authors introduce two scale-free fairness diagnostics. First, Modality Accuracy Divergence (MAD) measures how unevenly One-vs-Rest accuracy is distributed across the \(K\) categories of sensitive attribute \(S_i\) relative to the macro-average accuracy \(Acc_{f,i} := \frac{1}{K}\sum_{k=1}^K Acc_{f,i,k}\):

\[MAD_{f,i} := \frac{1}{K} \sum_{k=1}^K \frac{|Acc_{f,i,k} - Acc_{f,i}|}{Acc_{f,i}}\]

where \(Acc_{f,i,k} := P(f_{i,k}^W(X) = W_k^i)\) captures both false-positive over-assignment and false-negative under-recognition. Second, Recall Divergence (RD) specifically isolates true-positive recognition parity across subgroups, quantifying the relative dispersion of per-modality recall from the macro-average recall \(Rec_{f,i} := \frac{1}{K}\sum_{k=1}^K Rec_{f,i,k}\):

\[RD_{f,i} := \frac{1}{K} \sum_{k=1}^K \frac{|Rec_{f,i,k} - Rec_{f,i}|}{Rec_{f,i}}\]

where \(Rec_{f,i,k} := P(\hat{S}_i = s_k^i \mid S_i = s_k^i)\). By combining MAD (which reflects both over-allocation and under-recognition) with RD (which focuses on recognition failure for true group members), researchers can diagnose whether an LLM exhibits group-wise blindness or systematic over-projection.

3. Progressive Prompt Exploration and Rationale-Guided Reflective Debiasing

To systematically examine how task specification and reasoning prompts modulate demographic inference, five progressive prompt variants were designed: Regular directly requests demographic labels; Informed instructs the model to utilize lyrical content, tone, perspective, cultural references, and language patterns; Informed & Expressive requires generating supporting keywords and natural-language explanations; Well-informed elicits quantitative 1–10 ratings across 20 sociolinguistic attributes (e.g., emotions, slang, politeness, pronouns) prior to or following classification. Crucially, observing that chain-of-thought rationales slightly degraded accuracy, the authors analyzed generated explanations during classification failures by computing differential word frequencies between incorrect and correct rationales:

\[\Delta \text{Freq}(w) = \text{Freq}_{\text{incorrect}}(w) - \text{Freq}_{\text{correct}}(w)\]

Word-cloud visualizations revealed that incorrect ethnicity inferences were heavily dominated by high-level abstractions like theme and emotional (e.g., misattributing universal themes of oppression or blues-like sorrow to North American Black culture). To eliminate this shortcut, the authors introduced the Corrected Informed prompt, which explicitly instructs the model: "Do NOT use the theme or emotion of the song to decide." This constraint successfully compelled models to ground classifications in verifiable geopolitical, linguistic, and cultural markers, substantially reducing demographic bias.

Key Experimental Results

Main Results

Evaluations were performed across the balanced evaluation subsets (drawing 300 songs per ethnicity modality and 500 per gender modality, evaluated with 1,000 stratified bootstrap iterations to construct 95% confidence intervals). The random guess baselines are 50% for binary gender and 16.7% for six-way continental ethnicity. All models achieve non-trivial profiling performance well above chance, with Mistral-24B and Gemma-12B attaining the strongest predictive capabilities.

Model Prompt Type Gender Acc (%) Ethnicity Acc (%) Ethnicity MAD (↓) Ethnicity RD (↓) Gender RD (↓)
DeepSeek-1.5B Informed & Expressive 55 ± 2 19 ± 2 0.06 ± 0.01 0.40 ± 0.09 0.55 ± 0.05
DeepSeek-1.5B Corrected 56 ± 2 21 ± 2 0.06 ± 0.01 0.24 ± 0.08 -
DeepSeek-7B Informed & Expressive 56 ± 3 27 ± 2 0.11 ± 0.01 0.60 ± 0.07 0.21 ± 0.05
DeepSeek-7B Corrected 59 ± 3 29 ± 2 0.11 ± 0.01 0.54 ± 0.06 -
Gemma-12B Informed & Expressive 73 ± 3 41 ± 2 0.11 ± 0.01 0.45 ± 0.05 0.08 ± 0.04
Gemma-12B Corrected 72 ± 3 41 ± 2 0.10 ± 0.01 0.40 ± 0.04 -
Llama-8B Informed & Expressive 61 ± 3 32 ± 2 0.14 ± 0.01 0.69 ± 0.05 0.20 ± 0.05
Llama-8B Corrected 67 ± 3 42 ± 2 0.10 ± 0.01 0.46 ± 0.05 -
Ministral-8B Informed & Expressive 65 ± 3 30 ± 2 0.16 ± 0.01 0.77 ± 0.06 0.41 ± 0.04
Ministral-8B Corrected 70 ± 3 32 ± 2 0.15 ± 0.01 0.69 ± 0.05 -
Mistral-24B Informed & Expressive 76 ± 2 41 ± 2 0.12 ± 0.01 0.49 ± 0.05 0.21 ± 0.04
Mistral-24B Corrected 76 ± 2 44 ± 2 0.11 ± 0.01 0.47 ± 0.05 -
Qwen-7B Informed & Expressive 68 ± 3 33 ± 2 0.14 ± 0.01 0.60 ± 0.05 0.25 ± 0.05
Qwen-7B Corrected 68 ± 2 33 ± 2 0.13 ± 0.01 0.61 ± 0.05 -

Ablation Study

Aggregated ablation and prompt complexity comparisons across models (macro-averaged accuracy and 95% confidence intervals from bootstrap resampling) illustrate the structural impact of instruction depth, cognitive overhead, and debiasing constraints.

Prompt Configuration Gender Acc (%) Ethnicity Acc (%) Empirical Observation & Mechanism
Regular (Baseline direct inference) 68.8 ± 1.6 33.7 ± 1.1 Simple input-output classification prompt with minimal framing
Informed (Sociolinguistic guidance) 68.5 ± 1.2 35.2 ± 0.8 Adds guidance on content, tone, perspective; modest gain in ethnicity
Informed & Expressive (Rationale elicitation) 64.5 ± 1.0 31.8 ± 0.7 Generating keywords and CoT introduces slight degradation across models
Well-informed (20-attribute rating first) 62.8 ± 1.1 27.2 ± 0.6 Multi-step rating instruction overloads smaller LLMs, causing severe drops
Corrected (Negative theme/emotion constraint) 68.5 ± 1.4 36.8 ± 0.9 Forbidding theme/emotion shortcuts achieves top accuracy and fairness parity

Key Findings

  • North American Default vs. Regional Skews: On the balanced ethnicity benchmark, Llama-8B, Ministral-8B, Mistral-24B, and Qwen-7B treat North America as the default category, exhibiting simultaneous high true-positive and high false-positive rates on ROC curves. In stark contrast, DeepSeek-1.5B defaults toward Asian ethnicity, leading to grotesque misattributions—such as interpreting South African anti-apartheid singer Miriam Makeba's song A Piece Of Ground (which describes gold mining and racial dispossession) as a "narrative from Asia."
  • Negative Correlation with "Cultural References": Sociolinguistic attribute ratings reveal that the presence of explicit cultural references is strongly negatively correlated with North American predictions (Pearson \(r = -0.31 \pm 0.02\), dropping to \(-0.43 \pm 0.04\) in Gemma-12B), but positively correlated with African (\(+0.29\)), Asian (\(+0.19\)), and South American (\(+0.14\)) predictions. This indicates that North America functions as an "unmarked baseline," with models requiring explicit, overt regional markers (e.g., the Waiapu River for New Zealand or the game of parqués for Colombia) to predict non-Western regions.
  • Fairness-Accuracy Trade-off: Ministral-8B exhibited the strongest ethnicity bias among all models (RD of \(0.77 \pm 0.06\)). Gemma-12B displayed the most balanced performance across modalities (gender RD of \(0.08 \pm 0.04\), ethnicity RD of \(0.45 \pm 0.05\)). While DeepSeek-1.5B achieved a low MAD (\(0.06\)), this is a byproduct of uniformly low accuracy (\(19\%\)), illustrating that low divergence without sufficient capacity reflects low utility rather than equitable representation.
  • Text Length and Genre Mediation: Profiling accuracy increases monotonically with lyric length up to the 95th percentile (964 words), as longer texts provide richer discourse markers. Consequently, word-dense genres like Rap (averaging 623 words) yield substantially higher accuracy than repetitive genres like Electronic Dance Music (EDM, averaging 310 words).

Highlights & Insights

  • Asymmetric Metric Framework for Profiling Tasks: The paper smartly identifies the mathematical limitation of traditional group fairness metrics (e.g., Disparate Impact) when the target prediction itself is sensitive, formalizing MAD and RD to measure false-positive over-assignment and true-positive under-recognition across demographic groups.
  • Probing Implicit Western Hegemony via Literary Stylometry: By moving beyond standard questionnaire benchmarks into creative, unstructured song lyrics, the study empirically demonstrates how open-source foundation models internalize North American culture as the default human condition.
  • Diagnostic Error Clouds Informing Prompt Architecture: Rather than guessing prompt modifications, the authors utilize vocabulary frequency subtraction (\(\Delta \text{Freq}\)) on erroneous vs. correct explanations, cleanly isolating spurious affective associations (theme, emotional) and converting that empirical finding into an effective negative constraint prompt.

Limitations & Future Work

  • Stylistic Domain Specificity and Author Attribution Noise: Experiments are conducted exclusively on commercial song lyrics, which feature heightened figurative language, rhyming constraints, and genre tropes that may not generalize to formal discourse; furthermore, commercial songwriting frequently involves uncredited ghostwriters and collaborative teams, meaning the singer's demographic metadata may not align with the actual lyricist.
  • Translation Normalization Artifacts: While zero-shot translation via Mistral-24B eliminates grammatical gender markers in Romance languages, machine translation inherently normalizes syntax and dampens regional colloquialisms, potentially stripping subtle non-Western cultural markers.
  • Broad Continental Aggregations: The operationalization of ethnicity into six macro-continental categories is a coarse sociological simplification that flattens immense intra-continental cultural, ethnic, and national diversity.
  • vs. PAN Shared Tasks and Classical Profiling (Argamon et al., 2009; HaCohen-Kerner et al., 2018): Classical author profiling relies on supervised classification with explicit stylistic lexicons and n-grams over short blogs; this work pioneer zero-shot profiling with general-purpose instruction-tuned LLMs over long-form, artistic lyrics.
  • vs. LLM Authorship Verification (Huang et al., 2024): While recent LLM authorship research focuses on binary attribution (verifying whether two texts share an author), this paper probes deep sociodemographic inference and the fairness disparities embedded within those predictions.
  • vs. Cultural Alignment Benchmarks (AlKhamissi et al., 2024; Tao et al., 2024): Existing cultural alignment studies evaluate explicit surveys or cultural trivia; this work uses author profiling as an implicit probe, demonstrating how default cultural assumptions bias model outputs even when no explicit cultural question is asked.

Rating

  • Novelty: ⭐⭐⭐⭐☆ Creative formulation of lyric author profiling as a cultural probe, supported by the dedicated MAD and RD metric definitions for self-sensitive prediction.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous evaluation across 7 models from three continents, 5 prompt variants, bootstrap confidence intervals, translation robustness checks, and differential explanation analyses.
  • Writing Quality: ⭐⭐⭐⭐⭐ Lucid sociological grounding, crisp mathematical formulations of fairness metrics, and compelling qualitative case studies (e.g., Miriam Makeba and Shakira).
  • Value: ⭐⭐⭐⭐☆ Highly impactful for understanding and mitigating demographic biases in NLP systems deployed for automated essay evaluation, content moderation, and sociolinguistic profiling.