jonc/cei-benchmark

CEI: A Benchmark for Evaluating Pragmatic Reasoning in Language Models

Dataset Description

CEI (Contextual Emotional Inference) is a benchmark of 300 expert-authored scenarios for evaluating how well language models interpret pragmatically complex utterances in social contexts. Each scenario presents a communicative exchange involving indirect speech (sarcasm, mixed signals, strategic politeness, passive aggression, or deflection) where the speaker's literal words diverge from their actual emotional state.

• Paper: CEI: A Benchmark for Evaluating Pragmatic Reasoning in Language Models (DMLR 2026)
• Repository: https://github.com/jon-chun/cei-tom-dataset-base
• Zenodo: https://doi.org/10.5281/zenodo.18528705
• License: CC-BY-4.0 (data), MIT (code)

Dataset Structure

Scenarios

• 300 scenarios across 5 pragmatic subtypes (60 each)
• 3 independent annotations per scenario (900 total)
• Predefined splits: train (211), validation (48), test (41), stratified by subtype and power relation

Fields

Field	Type	Description
id	int	Scenario ID (unique within subtype)
subtype	string	Pragmatic subtype (sarcasm-irony, mixed-signals, strategic-politeness, passive-aggression, deflection-misdirection)
context	string	Situational context (2-4 sentences)
speaker	string	Speaker's role in the scenario
listener	string	Listener's role in the scenario
utterance	string	The speaker's pragmatically ambiguous utterance
power_relation	string	Power dynamic: peer, high-to-low, or low-to-high
gold_standard	string	Gold-standard emotion (majority vote + expert adjudication)
ann1_emotion	string	Annotator 1's emotion label (Plutchik)
ann2_emotion	string	Annotator 2's emotion label (Plutchik)
ann3_emotion	string	Annotator 3's emotion label (Plutchik)
valence_mean	float	Mean valence rating across annotators (-1.0 to +1.0)
arousal_mean	float	Mean arousal rating across annotators (-1.0 to +1.0)
dominance_mean	float	Mean dominance rating across annotators (-1.0 to +1.0)

Pragmatic Subtypes

Subtype	Description	Fleiss' kappa
Sarcasm/Irony	Speaker says the opposite of what they mean	0.25
Passive Aggression	Hostility expressed through superficial compliance	0.22
Strategic Politeness	Polite language masking negative intent	0.20
Mixed Signals	Contradictory verbal and contextual cues	0.16
Deflection/Misdirection	Speaker redirects to avoid revealing feelings	0.06

Labels

• Primary emotion: One of Plutchik's 8 basic emotions (joy, trust, fear, surprise, sadness, disgust, anger, anticipation)
• VAD ratings: Mean Valence, Arousal, Dominance across 3 annotators, mapped to [-1.0, +1.0]
• Gold standard: Majority vote with expert adjudication for three-way splits

Power Relations

• Peer (72%), High-to-Low authority (20%), Low-to-High authority (7%)

Key Statistics

• Inter-annotator agreement: Overall kappa = 0.21 (fair), ranging from 0.06 (deflection) to 0.25 (sarcasm)
• Human accuracy (vs. gold): 61% mean, 14.3% unanimous, 31.3% three-way split
• Best LLM baseline: 25.0% accuracy (Llama-3.1-70B, zero-shot) vs. 54% human majority agreement
• Random baseline: 12.5% (8-class)

Intended Uses

• Benchmarking LLM pragmatic reasoning capabilities
• Diagnosing model failure modes on indirect speech subtypes
• Research on emotion inference, social AI, Theory of Mind
• Soft-label training using per-annotator distributions

Limitations

• All scenarios are expert-authored (not naturalistic)
• English only
• 15 undergraduate annotators from a single institution
• Small scale (300 scenarios) optimized for annotation quality over quantity

Citation

@article{chun2026cei,
  title={CEI: A Benchmark for Evaluating Pragmatic Reasoning in Language Models},
  author={Chun, Jon and Sussman, Hannah and Pechon-Elkins, Mateo and Mangine, Adrian and Kocaman, Murathan and Sidorko, Kirill and Koirala, Abhigya and McCloud, Andre and Akanwe, Wisdom and Gassama, Moustapha and Enright, Anne-Duncan and Dunson, Peter and Ng, Tiffanie and von Rosenstiel, Anna and Idowu, Godwin},
  journal={Journal of Data-centric Machine Learning Research (DMLR)},
  year={2026}
}