EMNLP 2026

Surfacing the Unsaid: CUE-Bench for Affective Stance in Chinese Discourse

A Chinese discourse benchmark for understanding unsaid emotion through the Explicit-Implicit Stance Matrix.

Zhenyan Zheng^*, Yunyao Zhang^*, Junxi Sheng, Junqing Yu, Zikai Song^†

Huazhong University of Science and Technology
^*Equal contribution ^†Corresponding author

Overview of CUE-Bench and the Explicit-Implicit Stance Matrix

CUE-Bench links what is expressed with what remains unsaid. The Explicit-Implicit Stance Matrix maps surface affect and implied affect into nine interpretable Affective Stances, which guide pragmatic intent and fine-grained emotion inference.

Abstract

Emotion understanding in discourse requires reasoning beyond surface sentiment, since speakers often convey affect through indirect, implicit, polite, ironic, or deliberately mismatched expressions. Existing emotion benchmarks mainly annotate surface polarity or final emotion categories, while lacking a structured account of how explicit expression, implicit affect, pragmatic intent, and fine-grained emotion interact.

To address this gap, we introduce CUE-Bench, a Chinese Unsaid Emotion benchmark centered on Affective Stance. CUE-Bench constructs nine human-interpretable affective stances from Explicit-Implicit polarity interaction and further provides intent and fine-grained emotion annotations for structured affective inference.

Highlights

Unsaid Emotion

Targets cases where literal wording and intended affect diverge, including irony, politeness, understatement, veiled negativity, and formulaic expression.

Explicit-Implicit Stance Matrix

Models affective stance as the interaction between surface expression and implied affect, producing nine interpretable stance categories.

Matrix-Guided Reasoning

Uses Affective Stance as an intermediate scaffold for pragmatic intent understanding and fine-grained emotion classification.

Explicit-Implicit Stance Matrix

The matrix organizes the relation between surface expression and implicit affect. It is not a replacement for fine-grained emotion labels, but an interpretable scaffold for reasoning from what is expressed to what is meant.

Explicit \ Implicit	Positive	Neutral	Negative
Positive	Aligned Positive	Formulaic Positive	Sarcastic Negative
Neutral	Understated Positive	Neutral	Veiled Negative
Negative	Affiliative Positive	Reportive Negative	Aligned Negative

A simple example of Affective Stance: the same utterance can express one affective signal on the surface while implying another affective meaning in context.

Dataset

CUE-Bench provides multi-level supervision for Chinese discourse emotion understanding, spanning explicit affect, implicit affect, Affective Stance, Pragmatic Intent, and Fine-grained Emotion.

52,882instances in the refined full release

9Affective Stance categories

3connected evaluation tasks

Release	Description	Samples
Paper Experimental Split	Train/dev/test split used for the paper experiments.	47,010
Refined Full Release	Final public release with broader coverage of rare stance phenomena.	52,882

Task Suite

Task	Target	What It Evaluates
Affective Stance Recognition	9 stance labels	Whether models can identify the relation between surface affect and implied affect.
Pragmatic Intent Understanding	8 intent labels	Whether models understand communicative strategies such as irony, politeness, suppression, and resistance.
Fine-grained Emotion Classification	25 emotion labels	Whether models can recover the specific affective state conveyed in context.

Experimental Results

Across strong LLM baselines, incorporating Affective Stance improves downstream affective inference, especially for pragmatic intent understanding.

Task	Best Baseline	Matrix-Guided CoT	Overall Gain
Pragmatic Intent Understanding	Direct / Few-shot / CoT	Explicit-Implicit stance guided reasoning	+8.1 percentage points
Fine-grained Emotion Classification	Direct / Few-shot / CoT	Explicit-Implicit stance guided reasoning	+3.1 percentage points

Representative Model Results

Model	Affective Stance W-F1	Pragmatic Intent W-F1	Fine-grained Emotion W-F1	Average
DeepSeek-V4-Flash	0.527	0.465	0.327	0.402
LLaMA-4-Maverick	0.511	0.452	0.331	0.393
GLM-5.1	0.552	0.450	0.333	0.405

Gains summarize improvements over strong prompting baselines in the paper. Full model-level results are reported in the paper.

Community

Discussions are open for questions about CUE-Bench, suggestions for benchmark usage, reports on reproduced results, and feedback on the label schema or release files.

Channel	Use It For	Link
GitHub Discussions	Ask questions, share evaluation results, suggest improvements, or discuss annotation and task definitions.	Open discussions
GitHub Issues	Report reproducibility problems, data loading issues, or bugs in evaluation scripts.	Open issues

Usage

Load the Dataset

git clone https://github.com/AI4SS/CUE-Bench.git
cd CUE-Bench

Run Evaluation

pip install -r requirements.txt
python code/llm_evaluate.py \
  --model gpt-4o-mini \
  --task stance \
  --method direct \
  --shot few

Data Format

{
  "text": "你可真会安排，周六开会最让人开心了。",
  "context": "[工作群聊] 团队连续加班多天，负责人临时通知周六继续开会。",
  "explicit_polarity": "Positive",
  "implicit_polarity": "Negative",
  "affective_stance": "Sarcastic Negative",
  "pragmatic_intent": "Irony",
  "fine_grained_emotion": "Contempt"
}

BibTeX

@misc{zheng2026surfacingunsaidcuebenchaffective,
  title = {Surfacing the Unsaid: CUE-Bench for Affective Stance in Chinese Discourse},
  author = {Zhenyan Zheng and Yunyao Zhang and Junxi Sheng and Junqing Yu and Zikai Song},
  year = {2026},
  eprint = {2608.10810},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url = {https://arxiv.org/abs/2608.10810}
}