acceptodds
Desk rejected at ICLR 2027

COERCION:Benchmarking Moral Coercion in Dialogues via Psychological Dimension Decoupling

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Moral coercion is a form of implicit manipulation that uses shared ethical norms to exert interpersonal pressure. While existing NLP models detect explicit toxicity, they struggle with moral coercion due to its polite phrasing and contextual dependency, Current datasets on moral coercion are insufficient, often relying on movie scripts that lack the pragmatic complexity of actual social interactions. To address this gap, we build COERCION, a dataset of 4,700 multi-turn dialogues sourced from social media. To model the latent manipulative mechanism of moral coercion, we design a taxonomy decomposing moral coercion into four quantifiable psychological dimensions: Obligation, Constraint, Value Judgement, and Toxicity. We further propose a two-stage dimension-aware pipeline that extracts these dimensions as structural anchors, guiding downstream Large Language Models (LLMs) toward deeper pragmatic reasoning rather than shortcut learning. Statistical analysis indicates that our dataset reflects moral coercion characteristics. Experimental results demonstrate that grounding LLMs in these dimensions improves detection accuracy. Furthermore, ablation results indicate that the joint integration of ValueJudgement and Constraint is sufficient to sustain near-optimal performance across models, demonstrating their critical contribution to the exploration of moral coercion.

Primary area:
datasets and benchmarks
Keywords:
Moral Coercion Detection, Multi-turn Dialogue, Pragmatic Reasoning, Implicit Manipulation, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Mp8NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the detection of implicit moral coercion in multi-turn dialogues. The authors introduce COERCION, a dataset of 4,700 dialogues constructed from social media interactions, supplemented by GPT-4o-generated examples and human annotations. Each dialogue is labeled for the presence of moral coercion and its intensity on a 0–5 scale. The paper further decomposes moral coercion into four psychological dimensions: Obligation, Constraint, Value Judgement, and Toxicity. Based on these dimensions, the authors propose a two-stage framework that first extracts dimension-specific features and then incorporates them alongside the original dialogue for coercion detection and intensity prediction.

The experiments evaluate cross-domain generalization, in-domain fine-tuning, and coercion intensity prediction, together with ablation studies on the proposed psychological dimensions. Fine-tuning on COERCION achieves binary classification F1 scores of approximately 0.84-0.85 and intensity prediction accuracy of around 76%. Incorporating dimension-specific features also improves detection performance over direct classification baselines. The ablation results suggest that Constraint and Value Judgement are the most informative dimensions, with their combination achieving comparable or better performance than the full four-dimensional feature set across multiple models.

Critical strengths

A targeted benchmark for moral coercion detection. The paper introduces COERCION, a dataset of 4,700 multi-turn dialogues annotated with both binary coercion labels and six-level intensity scores. Unlike existing datasets primarily focused on explicit toxicity or general emotion recognition, this benchmark addresses the underexplored problem of implicit moral coercion, providing a useful resource for future research.

Systematic empirical evaluation and feature analysis. Through cross-domain comparisons, in-domain fine-tuning, and ablation studies, the paper demonstrates the benefits of task-specific training for moral coercion detection. In particular, Constraint and Value Judgement show relatively consistent contributions across different models, providing preliminary empirical insights into the linguistic mechanisms underlying moral coercion.

Critical weaknesses

Insufficient positioning against closely related methods. Although the paper discusses the MENTALMANIP dataset, it does not adequately engage with subsequent research on intent-aware and multi-stage prompting approaches for conversational manipulation detection. Without sufficient comparisons against these closely related methods, it remains unclear whether the proposed psychological dimensions offer meaningful methodological advances beyond existing structured prompting strategies.

Concerns about dataset reliability and submission compliance. The paper uses GPT-4o to generate synthetic dialogue examples but does not clearly report their proportion in the dataset or provide sufficient validation on independent real-world dialogues. This raises concerns about dataset representativeness and the generalizability of the experimental findings. Additionally, the paper omits the mandatory AI Use Statement required by ICLR 2027, despite explicitly using generative AI, constituting a violation of the conference's disclosure requirements.

Questions for authors

Could the authors clarify what methodological advantages the proposed dimension-aware pipeline offers over existing approaches that explicitly model conversational intent or intermediate reasoning signals for manipulation detection? In particular, what evidence demonstrates that the proposed psychological dimensions provide benefits beyond those achievable through existing structured prompting strategies?

Reviewer_oiB9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces COERCION, a dataset of 4,700 multi-turn dialogues for detecting moral coercion and predicting its intensity. The authors define four psychological dimensions—Obligation, Constraint, Value Judgement, and Toxicity—and propose a two-stage LLM pipeline that extracts these features to assist downstream classification. Experiments across several language models evaluate the dataset and the proposed pipeline, with ablation studies suggesting that Constraint and Value Judgement are particularly useful for detecting moral coercion.

Critical strengths

The paper addresses a meaningful and relatively underexplored problem of detecting implicit moral coercion in multi-turn dialogues. The inclusion of human-annotated coercion intensity labels could provide a useful resource for future research.

Critical weaknesses

1. Insufficient transparency and validation of dataset construction.

As a dataset paper, this work fails to provide essential information about data collection, including the source platforms, crawling keywords, and sampling criteria. More importantly, the authors use GPT-4o to generate synthetic dialogues but do not disclose the proportion of synthetic data in the dataset or in each train/validation/test split. It is therefore unclear how much of the evaluation is conducted on authentic human conversations. Given that the paper's primary contribution is a benchmark intended to reflect real-world moral coercion, these omissions raise fundamental concerns about the dataset's validity, representativeness, and reproducibility.

2. Lack of essential baselines and insufficient methodological novelty.

The proposed method relies on DeepSeek-V3.2 to extract psychological dimensions, which are then provided to smaller downstream LLMs for classification. However, the authors do not evaluate DeepSeek-V3.2 as a direct classification baseline, nor do they test whether the downstream models can achieve similar improvements by extracting these dimensions themselves through prompting.

Consequently, the experiments cannot establish whether the reported improvements come from the proposed psychological dimension decomposition, the additional reasoning steps, or simply the use of a stronger upstream model. The four dimensions themselves also lack independent human validation, as acknowledged in the limitations.

Given that the proposed methodology primarily consists of prompting existing LLMs without substantial technical innovation, these missing controls significantly weaken the paper's methodological contribution and its claims of effectiveness.

Questions for authors
  1. Dataset construction: Which social media platforms were used, and what keywords and sampling criteria were applied during data collection? What percentage of the dataset was generated by GPT-4o, particularly in the test set? Can the authors report performance separately on naturally collected and synthetic dialogues?

  2. Missing baselines: Why was DeepSeek-V3.2, which serves as the upstream feature extractor, not evaluated as a direct classification baseline? Additionally, how would the downstream models perform if they were prompted to extract the four psychological dimensions themselves before classification, without relying on DeepSeek-V3.2?

  3. Validity of psychological dimensions: How were the four psychological dimensions and their numerical scores validated? Since these scores are generated by DeepSeek-V3.2 without independently annotated ground truth, what evidence supports their reliability and the claim that the observed improvements result from psychological dimension decomposition rather than the upstream model's reasoning capabilities?

Reviewer_y61zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces a dataset of 4,700 multi-turn dialogues for moral coercion detection. The dataset includes both collected and synthetic dialogues, with each dialogue assessed by three human annotators. The authors use DeepSeek to assign scores to four psychological dimensions. These scores are then provided alongside the original dialogue as features to help a downstream model predict whether moral coercion is present.

Critical strengths

The paper addresses an interesting and important problem. Understanding implicit intentions and interpersonal pressure in conversations requires looking beyond explicitly hostile language. A dataset focused on moral coercion could provide a useful resource for studying how models interpret these interactions.

Critical weaknesses

The proportions of collected and synthetic dialogues are not reported, and the specific source platforms are unclear. Although the authors describe manual filtering, more detail is needed on how they assessed whether generated examples reflect realistic interactions. These omissions make it difficult to assess what kinds of conversations the benchmark represents.

Judgments of moral coercion may depend on prior obligations, relationships between speakers, and cultural expectations. It is unclear whether the available dialogue provides enough information to make these judgments consistently. Having three annotators per dialogue helps, but does not fully resolve contextual ambiguity. The paper should clarify what information annotators received, how they handled ambiguous cases, and how they distinguished coercion from ordinary persuasion or legitimate reminders of responsibility. Also, while the paper reports educational backgrounds, annotator recruitment and specific cultural backgrounds remain unclear.

Although the paper cites relevant literature for each dimension, it is unclear how these sources motivate this particular combination of four dimensions. The paper also does not report a comparison with a standard chain-of-thought baseline. Such a comparison under a comparable inference budget would help establish whether the proposed dimension scores offer additional value over simply analyzing the dialogue before making a prediction.

Questions for authors

What proportions of the dataset were collected versus synthetic, which platforms were used, and how was the realism of synthetic examples assessed?

What context and guidelines did annotators use to distinguish coercion from ordinary persuasion or legitimate responsibility? How were annotators recruited, and what cultural backgrounds were represented?

How does the cited literature justify selecting these four dimensions together?

Do you have results comparing the proposed method with a standard CoT baseline under a comparable inference budget?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.