Skip to content

Shaped by Values: Language Through the Lens of Culture and Morality

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/neelefalk/shaped-by-values
Area: LLM Reasoning
Keywords: Moral Foundations Theory, Cultural Dimensions, Perspectivism, Egocentric Bias, Annotation Misalignment

TL;DR

This paper introduces MCWR, a two-stage multilingual dataset spanning US English, Mexican Spanish, Chinese, and Arabic that couples generated texts with self-reported value profiles for both writers and readers, uncovering that reader value inferences are heavily driven by an egocentric projection bias and that reader-writer value distance is the primary source of perceptual misalignment.

Background & Motivation

Morals and cultural values serve as fundamental cognitive blueprints that govern social impression formation, decision-making, and everyday communication across societies. Nevertheless, human beings rarely articulate their core normative tenets explicitly in everyday conversations. This pervasive scarcity gives rise to an acute value inference gap: even when textual signals lack explicit ethical statements, human recipients spontaneously construct elaborate mental representations regarding the writer's underlying character, moral standing, and social orientation. Crucially, such inferences are inherently subjective and intuitive, heavily anchored in the perceiver's own internal moral schema rather than in the linguistic tokens alone.

Within the landscape of NLP and value alignment, existing paradigms for analyzing moral and cultural discourse suffer from a critical conceptual blind spot. Prevailing annotation frameworks treat value attributes as objective, intrinsic properties of the text itself, frequently collapsing subjective disagreements into majority-vote aggregates while discarding the sociocognitive context of both the speaker and the receiver. As a consequence, existing resources capture nothing more than an uncontextualized reader's impression of an unknown writer's prose. Furthermore, prior benchmarks remain overwhelmingly constrained to Western, Educated, Industrialized, Rich, and Democratic (WEIRD) demographics, predominantly centered around US English. NLP systems and foundation models trained on such writer- and reader-agnostic corpora risk encoding systematically skewed interpretations while mistaking audience-specific projections for universal human values.

Overcoming these structural impediments requires moving beyond conventional single-sided labeling by establishing a grounded communicative loop that jointly captures the writer's authentic values, the produced text, and the reader's contextualized interpretation. This work bridges cultural psychology and NLP by operationalizing validated psychometric instruments within a rigorous cross-cultural, multi-agent annotation framework. Core idea: construct MCWR, the first two-stage multilingual dataset encompassing ground-truth self-reported values for both writers and readers across four typologically diverse languages, and employ hierarchical mixed-effects modeling to prove that subjective value inferences are predominantly shaped by egocentric projection, with value distance between interlocutors serving as the primary driver of communicative misalignment.

Method

Overall Architecture

The methodological pipeline of MCWR consists of a two-stage crowdsourced collection followed by multilevel psychometric decomposition and regression modeling. In Stage I (Writers), 200 participants across four cultural cohorts (US, Mexico, China, and Arabic-speaking populations) write open-ended responses to 10 value-salient prompts (arguments, interviews, personal experiences, and dialogues) and complete a 50-item psychometric inventory establishing their ground-truth moral and cultural dimensions. In Stage II (Readers), nearly 500 blind annotators evaluate 20 texts each from various writers, inferring the author's values on a refined 25-item Likert scale while providing their own self-reported profiles. Finally, 44 linear mixed-effects models and mixed-effects logistic regressions decompose annotator-versus-text variance and quantify the causal impact of value discrepancy on perceptual misalignment.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Value-Eliciting Prompts<br/>5 types: arguments/interviews/experiences/dialogues"] --> B["Two-Stage Multilingual Coordinated Sampling<br/>US English / Mexican Spanish / Chinese / Arabic"]
    B --> C["Dual-Perspective Psychometric Mapping<br/>MFQ2 Moral Foundations + CVSCALE Dimensions"]
    C --> D["Linear Mixed-Effects Variance Decomposition<br/>ICC analysis decoupling writer vs. reader variance"]
    D --> E["Bidirectional Value Distance & Misalignment Modeling<br/>Quantifying egocentric projection & cognitive mismatch"]

Key Designs

1. Two-Stage Multilingual Coordinated Sampling: Escaping the WEIRD Monolingual Silo To transcend the pervasive English-centric and Western-biased footprint of existing NLP datasets, MCWR implements a rigorous sampling design across four distinct sociolinguistic regions: US English (stratified across political ideology), Mexican Spanish, Chinese, and diverse Arabic varieties. On the writer side, 200 participants authored 1,981 texts across five value-triggering prompt genres: controversial social arguments, structured personal interviews, episodic autobiographical narratives, and conversational replies sourced from ValueNet and Reddit AITA. On the reader side, each generated text received approximately 5 independent blind judgments from native annotators with zero biographical knowledge of the writer, culminating in 8,023 dense, high-validity annotations that comprehensively connect linguistic expressions with sender and receiver identities.

2. Dual-Perspective Psychometric Mapping: Unifying Ground-Truth and Inferred Value Spaces To quantitatively bridge communicative intentions with reader perceptions, the framework integrates Moral Foundations Theory (MFT 2.0: Care, Equality, Proportionality, Authority, Loyalty, Purity) with Hofstede's Cultural Dimensions (CVSCALE: Collectivism, Power Distance, Uncertainty Avoidance, Long-Term Orientation, Gender Egalitarianism). Both writers and readers self-report their full baseline profiles via 50 standardized items at the end of the survey. For the reading phase, the inventory is condensed into a 25-item Likert instrument with built-in attention and consistency checks (e.g., dual collectivism items worded in opposite directions) to minimize cognitive fatigue without degrading reliability. This places the writer's authentic values \(\mathbf{v}_w\), the reader's self-reported values \(\mathbf{v}_r^{\text{self}}\), and the inferred values \(\mathbf{v}_r^{\text{inferred}}\) into a shared continuous metric space.

3. Linear Mixed-Effects Variance Decomposition: Decoupling Subjective Projection from Textual Signal To establish the extent to which annotations reflect textual reality versus annotator disposition, the authors specify 44 linear mixed-effects models across each value and language cohort. In each model, the reader's inferred score serves as the continuous dependent variable, while the reader's eleven self-reported values and demographic attributes enter as fixed effects, conditioned on crossed random intercepts for Annotator ID and Text ID: $\(y_{ijk} = \beta_0 + \sum_{m=1}^{11} \beta_m v_{r,i}^{(m)} + \boldsymbol{\gamma}^\top \mathbf{x}_{r,i} + u_i^{\text{reader}} + w_j^{\text{text}} + \epsilon_{ijk}\)$ By inspecting Intraclass Correlation Coefficients (ICCs), the framework separates text-driven variance \(\text{ICC}_{\text{text}}\) from annotator-driven variance \(\text{ICC}_{\text{annotator}}\). Furthermore, semi-partial \(R^2\) (\(sr^2\)) isolates the unique marginal contribution of the reader's own matching value, statistically demonstrating that subjective projection dominates value attribution when textual evidence is ambiguous.

4. Bidirectional Value Distance & Misalignment Modeling: Uncovering the Roots of Evaluative Misinterpretation Addressing when and why readers misjudge writers, the methodology employs mixed-effects logistic regression to model binary stance misalignment events \(Y_{ij}\) (instances where the inferred value category contradicts the writer's ground truth). Fixed effects incorporate the absolute distance along each value dimension \(d_{ij}^{(m)} = |v_{w,j}^{(m)} - v_{r,i}^{(m)}|\) alongside demographic alignment indicators. The empirical findings demonstrate that value distance along the matching diagonal axis is the strongest predictor of misreading: readers systematically project their own ethical inclinations onto authors, inevitably generating cognitive misalignment when their internal moral anchors diverge from the author's true perspective.

Benchmark & Evaluation Protocol

Analyses were executed across the unified MCWR dataset comprising 8,023 annotations, 200 writers, 498 readers, and 1,981 texts. All continuous features were z-score standardized within language. Significance testing strictly controlled for multiple hypotheses: the 11 confirmatory diagonal projection tests were adjusted via Holm correction, while the 110 exploratory off-diagonal and demographic terms were corrected using Benjamini–Hochberg False Discovery Rate (FDR) adjustments. Absolute human performance against writer ground truth was benchmarked against random guessing using global Mean Squared Error (MSE), binned 3-class F1 metrics, and Krippendorff's \(\alpha\).

Key Experimental Results

Main Results

The table below details model fits and variance decompositions from Analysis 1 (linear mixed-effects models predicting perceived writer values from reader profiles, from Table 15 in the paper). Values report marginal \(R^2\) (\(R^2_m\)), conditional \(R^2\) (\(R^2_c\)), text intraclass correlation (\(\text{ICC}_t\)), and annotator intraclass correlation (\(\text{ICC}_a\)), in percentages (%).

Value Dimension US \(R^2_m\) / \(R^2_c\) US \(\text{ICC}_t\) / \(\text{ICC}_a\) MEX \(R^2_m\) / \(R^2_c\) MEX \(\text{ICC}_t\) / \(\text{ICC}_a\) ZH \(R^2_m\) / \(R^2_c\) ZH \(\text{ICC}_t\) / \(\text{ICC}_a\) AR \(R^2_m\) / \(R^2_c\) AR \(\text{ICC}_t\) / \(\text{ICC}_a\)
Care 5.4 / 45.1 23.1 / 19.9 10.0 / 52.1 24.4 / 25.4 3.0 / 49.7 31.1 / 16.5 12.6 / 45.7 13.4 / 29.4
Equality 8.3 / 49.9 12.5 / 34.1 17.5 / 49.6 12.0 / 35.4 7.3 / 41.1 16.2 / 22.5 12.2 / 46.4 5.5 / 37.2
Proportionality 12.1 / 57.7 9.6 / 44.8 8.1 / 52.5 8.9 / 39.3 8.7 / 51.9 13.4 / 35.4 15.0 / 48.9 3.6 / 41.0
Authority 17.3 / 59.5 9.6 / 47.2 11.0 / 50.8 11.9 / 35.4 7.4 / 42.7 14.0 / 25.7 22.6 / 58.2 7.9 / 47.8
Loyalty 14.4 / 55.1 6.0 / 45.8 9.3 / 47.6 15.8 / 28.9 7.7 / 46.8 13.5 / 29.7 18.0 / 53.5 9.9 / 40.5
Purity 21.5 / 54.5 7.1 / 45.8 11.0 / 48.2 6.6 / 37.8 6.2 / 50.1 6.5 / 39.4 15.4 / 58.0 3.5 / 50.0
Collectivism 7.0 / 35.4 13.2 / 20.2 9.0 / 43.3 19.4 / 21.6 4.5 / 31.6 18.9 / 10.8 14.3 / 39.0 11.2 / 25.2
Uncertainty Avoidance 13.3 / 57.7 6.2 / 48.1 14.3 / 56.8 5.9 / 47.2 3.9 / 42.2 13.0 / 25.3 21.9 / 54.8 4.5 / 47.5
Long-Term Orientation 13.4 / 60.1 5.0 / 51.6 13.9 / 55.6 5.8 / 46.2 7.2 / 52.1 8.6 / 39.8 19.9 / 54.2 5.6 / 45.5
Gender Egalitarianism 10.7 / 44.7 3.4 / 38.1 10.3 / 43.1 4.7 / 34.6 4.8 / 43.5 3.4 / 35.6 7.5 / 42.4 – / 37.9
Power Distance 18.5 / 54.0 8.0 / 44.3 5.3 / 43.6 7.0 / 32.5 7.6 / 35.2 9.9 / 22.4 15.3 / 45.1 3.2 / 38.7
Mean 12.9 / 52.2 9.4 / 40.0 10.9 / 49.4 11.1 / 34.9 6.2 / 44.3 13.5 / 27.6 15.9 / 49.7 6.8 / 40.1

Ablation Study

The table below summarizes findings from Analysis 2 (logistic regression predicting reader-writer conflict), showcasing base misinterpretation rates, marginal explanatory power (\(R^2_m\)), and the unique semi-partial contribution (\(sr^2\), in %) of the matching value distance for the US and Mexico cohorts (from Tables 22, 23, 24, and 25).

Value Dimension US Misalign Rate (%) US \(R^2_m\) (%) US Matching Dist \(sr^2\) (%) MEX Misalign Rate (%) MEX \(R^2_m\) (%) MEX Matching Dist \(sr^2\) (%)
Care 9.3 4.0 0.7 16.0 2.7 0.4
Equality 19.6 1.7 0.9 22.0 4.3 2.7
Proportionality 12.6 2.4 0.2 7.9 5.9 1.6
Authority 13.9 8.7 1.1 17.2 6.2 3.3
Loyalty 24.4 3.7 2.0 28.3 4.4 1.7
Purity 20.7 8.7 5.4 21.5 2.8 1.2
Collectivism 17.9 2.0 0.5 17.5 2.8 1.6
Uncertainty Avoidance 6.7 3.6 0.6 6.9 3.7 0.2
Long-Term Orientation 6.2 6.3 0.1 5.6 7.3 2.7
Gender Egalitarianism 19.9 4.4 1.1 26.4 2.5 0.5
Power Distance 19.0 6.5 4.7 19.2 2.7 0.6

Key Findings

  • Annotators Outweigh Texts in Determining Labels: Across all linguistic groups, random annotator variance (\(\text{ICC}_a\) mean 27.6%–40.1%) consistently dwarfs text-specific variance (\(\text{ICC}_t\) mean 6.8%–13.5%). In subjective moral annotation, "who reads" is far more consequential than "what was written."
  • Egocentric Projection Fills Information Vacuums: When texts lack definitive markers (such as for Purity, Uncertainty Avoidance, and Power Distance, where \(\text{ICC}_t\) plunges below 8%), annotators systematically fall back on their own internal values to infer the author's stance (unique \(sr^2\) reaching 5.4%–7.2%). Conversely, transparent dimensions like Care exhibit high inter-annotator consensus (\(\text{ICC}_t\) up to 31.1%) and negligible projection.
  • Value Distance Causes Evaluative Mismatch: In Analysis 2, absolute distance between author and reader on the same dimension accounts for more than half of the total explained variance in multiple settings (e.g., \(sr^2 = 5.4\%\) of \(8.7\%\) for Purity in the US), demonstrating that interpersonal ideological divergence directly generates cognitive misattribution.

Highlights & Insights

  • First Dual-Sided Human Ground Truth: Resolves a long-standing impasse in data perspectivism by establishing a closed-loop dataset tracking both the communicator's authentic identity and the interpreter's cognitive disposition.
  • Sociodemographics Fall Short of Value Profiles: Empirical likelihood ratio tests show that standard demographic features (age, gender, education) explain negligible incremental variance once personal moral values are accounted for, indicating that NLP user modeling should prioritize psychological profiles over demographic categories.
  • Critical Implications for LLM Value Alignment: RLHF and DPO pipelines relying on crowdsourced preference rankings may inadvertently embed annotator-specific egocentric projections into reward models, cautioning against treating human feedback as an objective ground truth without reader profiling.

Limitations & Future Work

  • Sampling Constraints in Non-Western Cohorts: Due to regional Prolific availability, Chinese and Arabic participants included expatriates and international students, introducing potential acculturation confounders despite localized dialect screening.
  • Psychometric Scale Pruning Under Crowdsourcing Constraints: Condensing questionnaires to 25 items degraded internal consistency for certain dimensions in Arabic (e.g., Cronbach's \(\alpha \approx 0.56\)–\(0.58\) for Purity and Proportionality).
  • Future Directions: Evaluating multi-text aggregations to see if expanded author context reduces reader projection bias, and applying the dual-profiling framework to probe ideological sycophancy in LLM-as-a-Judge architectures.
  • vs. Conventional Moral Corpora (MFTC, ValueNet, Social-Chem-101): Existing resources only collect third-party inferences on ungrounded texts; MCWR uniquely anchors annotations against the writer's authentic self-reported baseline, revealing substantial perceptual divergence.
  • vs. Data Perspectivism (e.g., Davani et al., 2022; Gordon et al., 2022): Prior perspectivist literature predominantly attributes annotator disagreement to demographic variables; this work demonstrates that deep-seated moral foundations and cultural values exert far greater predictive power over interpretive variation.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Pioneering closed-loop dual-profile dataset design and cross-cultural psychometric integration in computational linguistics.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Exceptional statistical rigor with 44 linear mixed-effects models, variance decomposition via ICCs, and conservative multiple-testing corrections across 4 diverse languages.
  • Writing Quality: ⭐⭐⭐⭐⭐ Exemplary prose structure, lucid psychometric formulations, and seamless bridging of social psychology with NLP.
  • Value: ⭐⭐⭐⭐⭐ Foundational contribution to data perspectivism, human-annotator bias modeling, and culture-aware LLM alignment.