PaperlayerView exact revision on arXiv

Source-linked AI summary

Less Is Moral: A CHARMing Framework for Moral Foundations Detection in Endorsement Behaviour

Huixiang Fu, Marian-Andrei Rizoiu

arXiv:2609.03330v2cs.CLcs.CYcs.SI

TL;DR

Moral foundation detectors often generalize poorly across domains, provide weak rationale grounding, and make prompting-based LLM analysis costly. CHARM addresses these gaps with a lightweight, theory-integrated framework combining MAC grounding, rationale alignment, and hate-speech modulation, and reports stronger supervised-baseline performance across out-of-domain datasets. Its COVID-19 application associates producer-side moral framing with endorsement, while the authors identify limited polarity-level data and weaker transfer to abstract social-norm datasets as boundaries.

  • Problem

    Existing moral foundation detectors have limited cross-domain generalization and rationale grounding, while prompting-based LLM detectors are costly and often decoupled from psychological theory.

  • Method

    CHARM uses a LoRA-adapted LLM with MAC cross-attention, rationale-aligned pooling, and hate-speech-aware modulation to predict polarity-aware moral dimensions.

  • Results

    CHARM surpasses supervised baselines on every out-of-domain dataset in both AUC and F1, improves AUC by up to 29.5% in-domain and 5.1% out-of-domain, and predicts endorsement from producer-side moral framing.

  • Takeaways & Limitations

    CHARM provides a scalable framework for measuring moral framing and studying its association with online endorsement and network homophily.

  • Takeaways & Limitations

    Polarity-level annotations remain limited, and weaker results on VIG and MIC indicate constrained generalization to more abstract social-norm content.

Abstract

from arXiv · show

Moral language plays a central role in shaping online endorsement and the diffusion of information, yet existing moral foundation detection systems often suffer from poor cross-domain generalization, weak rationale grounding, and reliance on costly prompting-based large language models (LLMs). We introduce CHARM, a MAC- and Hate-speech-Aware Rationalealigned Moral foundation detection framework built on a lightweight fine-tuned LLM, which integrates complementary moral grounding, rationale alignment, and polarity-aware hate speech signals to support more robust and faithful moral prediction. Unlike prior dictionary-, fine-tune-, or prompt-based detectors, which decouple computation from psychological theory, CHARM is built so that each component -- MAC cross-attention, rationale alignment, and hate-speech modulation -- operationalizes a distinct psychological construct. Using a 30\% subsample of the MFTC, MFRC, and News training pools together with the richer supervision in MFTCXplain, CHARM improves AUC by up to 15.3\% in-domain, surpasses the supervised baselines on every out-of-domain dataset in both AUC and F1, and offers a scalable, low-cost alternative to prompting-based LLM detectors. We further apply CHARM to large-scale COVID-19 discourse on Twitter and show that moral value alignment is strongly associated with online endorsement behavior. By making moral framing measurable at scale, CHARM offers a practical tool for studying the spread of morally charged misinformation. Code and additional materials: https://github.com/HuixiangF/CHARM/.

1 Introduction

CHARM addresses weak theory grounding, rationale faithfulness, cross-domain robustness, and efficiency in moral foundation detection. It integrates psychological structure and richer supervision, then links moral framing to COVID-19 endorsement and network structure.

  • Existing detectors often decouple moral classification from MFT and MAC, collapse virtue and vice, and treat hate-speech detection separately.
  • CHARM uses MAC cross-attention, rationale alignment, and hate-speech modulation so each component operationalizes a distinct psychological construct.
  • MAC grounding produces the largest cross-domain ablation drops, including ARG F1 .54→.43 and VIG F1 .67→.50.
  • Up to 29.5% in-domain and 5.1% out-of-domain AUC improvements are reported, with higher AUC and F1 than supervised baselines on every OOD dataset.The system uses 30% of several training pools plus MFTCXplain supervision and remains competitive with prompting-based LLMs at lower inference cost.
  • COVID-19 analysis tests whether creators’ moral framing predicts endorsement and whether endorsement networks show moral homophily.The reported network is morally assortative across foundations, strongest along loyalty at r = .46.

2 Related Work

Moral foundation detection combines competing theoretical frameworks with increasingly contextual computational methods. Existing approaches improve sensitivity or transferability but retain limitations in theory integration, cross-domain generalization, and auditable evidence.

  • MFT describes five foundations—care/harm, fairness/cheating, loyalty/betrayal, authority/subversion, and purity/degradation—with liberty/oppression added later.
  • MAC identifies seven cooperation-based moral domains, complementing MFT through mappings such as group loyalty to loyalty/betrayal and deference to authority/subversion.
  • Lexicon methods scale but struggle with contextual and implicit expressions, while fine-tuned models improve contextual sensitivity yet generalize poorly across domains.
  • Prompting-based LLMs improve transferability but depend on closed proprietary APIs, leaving moral detection weakly grounded in theory and auditable evidence.
  • CHARM uses MFTCXplain for rationale-rich training and cross-lingual evaluation, and adds MAC grounding for cross-domain MFT detection.

3 CHARM Framework

CHARM is a two-stage, theory-integrated framework that combines foundation encoding with rationale-guided, MAC-grounded, and hate-speech-aware polarity prediction. Its modules transform complementary psychological and annotation signals into auditable moral representations.

  • Two-stage architecture: Stage 1 adapts a LoRA encoder to learn foundation-level moral representations; Stage 2 builds polarity-aware predictions on the resulting encoder.
  • Rationale alignment: The rationale selector predicts token-level rationale probabilities from encoder representations and is pretrained with human span supervision.
  • Rationale alignment: A total variation regularizer encourages contiguous rationale predictions across adjacent valid token positions.
  • Rationale alignment: Rationale-steered pooling interpolates encoder attention with predicted rationale attention, focusing representations on human-identified moral evidence while retaining context.
  • MAC-theory grounding: MAC scores become seven projected tokens, and ten learnable virtue–vice queries cross-attend to them before attention pooling produces a MAC representation.
  • MAC-theory grounding: The pooled MAC representation is fused with the rationale-guided text representation to inject cooperation-domain structure into moral representations.
  • Hate-speech modulation: A hate-speech classifier converts predicted hate probability into FiLM scaling and shifting, modulating the MAC-enhanced representation before 10-dimensional moral prediction.

4 Experiments

CHARM is evaluated across diverse moral-detection datasets using limited training data, richer MFTCXplain supervision, and comparisons with fine-tuned and prompting-based baselines. It generally improves in-domain and out-of-domain performance while providing polarity-aware, rationale-aligned, and cost-efficient prediction.

  • Experimental setup: CHARM uses nine datasets and trains in two stages, combining foundation supervision with MFTCXplain polarity, rationale, and hate-speech annotations.The main experiments use 30% of the MFTC, MFRC, and News training data while retaining original train/test splits.
  • In-domain performance: CHARM improves over MFORMER on MFRC and News by .07 and .11 AUC, respectively, but MFORMER remains stronger on MFTC F1 (.77 vs. .72).CHARM remains competitive on MFTC, where its AUC is .87 versus MFORMER’s .89.
  • Generalization under distribution shift: CHARM outperforms MFORMER on all five OOD datasets in both AUC and F1, and matches or exceeds Tuning-GPT4o-mini on four of five OOD AUC comparisons.Against zero-shot models, performance is more mixed: CHARM beats Qwen-32B-Instruct in AUC on all nine datasets and in F1 on eight.
  • Moral polarity classification: CHARM improves MORALBERT’s AUC by .33 on MFTCXplain and .21 on HateBR for 10-dimensional moral polarity classification.The model also performs well on Portuguese HateBR despite not being designed specifically for multilingual learning.
  • Inference cost: CHARM requires no per-example inference-time API calls, unlike prompting-based approaches that incur costs for each inference pass.MOVA costs approximately $.066 and $.065 per 1K input tokens on SC and MIC, respectively.
  • Ablation studies: Removing MAC causes the largest OOD degradation, including F1 drops from .54 to .43 on ARG and from .67 to .50 on VIG.The results indicate that MAC provides structural grounding for implicit moral interpretation under distribution shift.
  • Ablation studies: Rationale supervision improves performance and rationale quality, especially on longer or compositional inputs such as MFRC and VIG.Removing it lowers F1 from .69 to .65 on MFRC and from .67 to .60 on VIG, while the full model improves all rationale metrics.
  • Ablation studies: Removing hate-speech supervision reduces performance on MFTCXplain10d, SC, and VIG, while the complete system achieves the strongest hate-speech detection performance.The three supervision signals provide complementary gains, and the full configuration reaches average AUC=.83 and F1=.57 in leave-one-corpus-out analysis.

5 Moral Alignment and Endorsement Behaviour

The COVID-19 case study uses CHARM-derived moral profiles to test whether producer-side moral framing predicts endorsement and whether endorsement ties exhibit moral homophily. Moral features add predictive signal beyond engagement covariates, while endorsement networks show positive alignment across foundations.

  • Study design: The case study aggregates CHARM-inferred tweet-level moral scores into user profiles and models repeated retweet interactions as directed-pair endorsement classification.Positive pairs are repeated retweet interactions occurring more than five times; negatives use a two-hop non-endorsement constraint.
  • Moral alignment and endorsement: Endorsed pairs have higher moral cosine similarity than non-endorsed pairs, with a reported difference of .042 and significant Mann–Whitney and KS tests at p < 10^-119.The separation is consistent across all five foundations, with loyalty and fairness showing the sharpest distributional gap.
  • Endorsement prediction: Moral features add predictive signal beyond network and behavioral covariates, while behavioral and network features still dominate the full model’s SHAP importance.Author purity ranks third and moral Euclidean distance fifth among the full model’s top ten features.
  • Endorsement prediction: Author-side features occupy seven of the ten highest positions in the moral-only model, with author purity and moral cosine similarity ranked first and second.Author purity is the strongest individual moral predictor despite having the lowest network assortativity, r = .27.
  • Moral homophily: All five moral foundations show significantly positive assortativity in the endorsement network, indicating robust moral homophily between connected users.The figure compares foundation-specific assortativity with degree-preserving random networks and reports bootstrap confidence intervals.
  • Human validation: Human validation on 100 stratified tweets yields observed agreement .83, Gwet’s AC1 .76, macro AUC .84, and macro average precision .56.These results support using CHARM-derived moral scores in the COVID-19 case study, with caution for low-prevalence categories.

6 Conclusions

CHARM is presented as a lightweight moral foundation detector that combines rationale alignment, MAC-based cooperation signals, and hate-speech-aware polarity modeling. Across datasets, it shows robustness, faithfulness, and sample efficiency, while linking endorsement behavior to latent moral alignment.

  • CHARM unifies rationale alignment, MAC-based cooperative signals, and hate-speech-aware polarity modeling within a fine-tuned LLM architecture.
  • Across multiple datasets, CHARM demonstrates strong robustness, faithfulness, and sample efficiency while remaining substantially more efficient than prompt-based LLM approaches.
  • Author purity is the strongest individual predictor of endorsement but has the weakest network assortativity (r = .27), suggesting endorsement can extend beyond moral congruence.

Limitations

The authors identify limits in polarity-level supervision, cross-domain generalization to abstract moral reasoning, and automatically generated MAC labels. These constraints bound the coverage, transferability, and reliability of the framework’s supervision.

  • Polarity-level moral annotations are limited, and most existing datasets do not cover liberty/oppression, restricting the 10-dimensional setting largely to MFTCXplain and HateBRMoralXplain.
  • CHARM performs more weakly than prompt-based approaches on VIG and MIC, where abstract social norms differ from the concrete discourse dominating its training data.
  • MAC supervision uses automatically generated eMACDscore labels because no publicly available human-annotated MAC dataset exists, potentially introducing downstream training noise.

Ethics Statement

The paper frames CHARM as a tool for analyzing moral framing and conflict rather than judging moral rightness. It cautions that predictions may reflect annotation and training-data biases, especially in sensitive contexts, and should not independently guide high-stakes decisions.

  • The study analyzes moral reasoning discourse using publicly available social media, news, and online discussion datasets.
  • Because moral annotations can reflect cultural and annotator-specific biases, model predictions should not be interpreted as objective judgments of morality.
  • CHARM may inherit training-data biases in politically or culturally sensitive contexts and should not be deployed alone for high-stakes moderation or decision-making.

A.7 Ablation Results on AUC

Ablation results indicate that rationale supervision, MAC representations, and hate-speech supervision provide complementary benefits. Their effects vary by evaluation setting, with rationale alignment supporting robustness, MAC supporting social generalization, and hate-speech supervision helping on toxicity-related data.

  • Removing rationale supervision consistently degrades performance on most datasets, particularly under cross-domain evaluation.
  • Removing MAC-enhanced representations causes the largest drops on socially grounded datasets such as SC and MIC.
  • The hate-speech auxiliary objective primarily improves performance on politically charged or toxicity-related datasets, especially MFTC and MFRC.

A.8 Training-Corpus Contribution

The leave-one-corpus-out analysis evaluates how MFRC, MFTC, News, and MFTCXplain contribute to CHARM under a shared configuration and protocol. Results indicate complementary value from corpus diversity and richer supervision, with MFTCXplain especially important for polarity, rationale, and hate-speech signals.

  • Analysis design: The leave-one-corpus-out analysis measures each corpus contribution using the same model configuration and evaluation protocol as the full model.It is not a controlled architecture-only comparison with the baselines.
  • Corpus contributions: The full configuration achieves the strongest average performance, with MFRC making the largest contribution and MFTC and News providing complementary domain coverage.The average is computed across nine evaluation datasets.
  • Corpus contributions: Removing MFTCXplain lowers average performance and produces the clearest declines on MFTCXplain and HateBR.The analysis attributes this value to MFTCXplain’s polarity, rationale, and hate-speech supervision.
  • Interpretation: Variation across individual datasets suggests corpus-specific interactions rather than uniform gains from every training source.Overall, the results reflect complementary value from corpus diversity and richer supervision.
  • Rationale evidence: Rationale-supervised examples show CHARM retaining label-relevant evidence that the model without rationale supervision misses.Examples include loyalty cues from teamwork and care cues from suffering and assistance.

A.10 Per-Foundation Performance Analysis

Per-foundation analysis examines robustness under distribution shift and the moral representations used in endorsement analysis. CHARM’s ranking remains comparatively stable when F1 declines, while user profiles collapse sparse virtue and vice dimensions into five foundation-level representations.

  • Cross-domain performance: CHARM maintains relatively stable degradation under distribution shift, with larger robustness gaps mainly on short-text and socially ambiguous datasets such as SC and MIC.Figure 6 presents dataset-level Macro-F1 and Figure 7 presents AUC comparisons.
  • Cross-domain performance: Out-of-domain performance decreases are generally moderate, suggesting transferable moral representations rather than reliance solely on dataset-specific lexical cues.AUC remains comparatively stable even when F1 declines, indicating preserved ranking consistency under domain shift.
  • Endorsement setup: Endorsement analysis uses directed retweet interactions occurring more than five times as positive samples and two-hop non-endorsement constraints for negative samples.Mentions and quote tweets are excluded because they may reflect disagreement or contextual commentary.
  • Moral dimension distributions: Positive moral poles are substantially denser and more stable than negative poles, while several vice dimensions remain highly sparse in large-scale Twitter discourse.The distribution is shown across ten polarity-aware moral dimensions.
  • Moral dimension distributions: Virtue and vice polarity scores are collapsed into five foundation-level moral representations to reduce sparsity and improve robustness in user-level endorsement profiles.This transformation is applied when constructing user-level moral profiles for endorsement analysis.

B.2 Manual Validation of the COVID-19 Case Study

Manual validation assesses CHARM’s COVID-19 moral predictions against blinded human annotations and supports their use in downstream endorsement analyses. Aggregate alignment is strong, but rare-category estimates remain unstable because several foundations have very few positive examples.

  • Validation design: Three blinded annotators label a stratified sample of 100 COVID-19 tweets, with majority voting defining the human reference.Tweets are stratified by predicted moral foundation and intensity to cover diverse predicted profiles.
  • Annotation reliability: Observed agreement is .830, Fleiss’ κ is .415, PABAK is .660, and Gwet’s AC1 is .760 across tweet–label decisions.Prevalence-adjusted measures complement raw agreement because low-prevalence labels can inflate agreement-based impressions.
  • Human–model alignment: CHARM achieves a polarity-level macro AUC of .842 and macro average precision of .555 against majority-voted human labels.The 95% bootstrap confidence interval for macro AUC is [.801, .876].
  • Human–model alignment: The validation results support using CHARM’s continuous moral predictions in downstream endorsement analyses.This conclusion concerns aggregate alignment with human judgments in the COVID-19 domain.
  • Limitations: Fairness, Loyalty, Purity, and Betrayal have only 2, 5, 5, and 6 majority-voted positive examples, respectively.High category-level AUC or agreement can therefore be unstable and partly driven by the large number of negative instances.
  • Limitations: The study emphasizes aggregate validation results rather than strong conclusions from individual rare-category estimates.Label-level estimates for low-prevalence categories require caution.
Loading 2609.03330v2…