| Venue | EMNLP 2026 |
|---|---|
| Track | Main Conference |
| Status | Accepted |
| Topic | Social pragmatic inference |
| Preprint | arXiv:2609.04384 |
You Really Didn't Get That? Benchmarking Social Pragmatic Inference for Indirect and Playful Chinese Online Comments is a research paper co-authored by Junjie Ma and accepted to the EMNLP 2026 Main Conference. It investigates whether large language models can infer what a speaker means through indirect or playful comments in Chinese online conversations.¹
Background
Recognizing sarcasm or humor does not necessarily establish whom a comment targets or what social purpose it serves. The study examines these distinctions in context, including expressions associated with yinyang guaiqi (阴阳怪气).¹
Benchmark
The authors constructed 4,735 human-validated questions from public Chinese social media interactions. Eight language models proposed questions and answered those written by other models; human reviewers checked the interpretations and answer choices.¹
Findings
The strongest model achieved 81.42% accuracy on questions written by other models. Human accuracy on a 300-item sample was 90.8%. The analysis found that models could recognize a playful tone while misinterpreting its target or function. These results concern the study's Chinese-language benchmark.¹
See Also
- Research — Cultural AI and related research interests
- Publications — Accepted papers
- Home — Junjie Ma's biography
References
- Hong, Shiwei; Ma, Junjie; Wang, Emma Jiren; Rong, Ethan Z.; Hu, Siying; Li, Haichang; Wang, Ziying; and Lu, Zhicong (2026). You Really Didn't Get That? Benchmarking Social Pragmatic Inference for Indirect and Playful Chinese Online Comments. arXiv:2609.04384. Accepted to EMNLP 2026, Main Conference.