| Venue | EMNLP 2026 |
|---|---|
| Track | Main Conference |
| Status | Published |
| Topic | Social pragmatic inference |
| Publication | ACL Anthology |
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 published in the proceedings of 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
- Shiwei Hong, Junjie Ma, Emma Jiren Wang, Ethan Z. Rong, Siying Hu, Haichang Li, Ziying Wang, and Zhicong Lu. 2026. You Really Didn’t Get That? Benchmarking Social Pragmatic Inference for Indirect and Playful Chinese Online Comments. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026). Association for Computational Linguistics, Budapest, Hungary, 19623–19640. https://doi.org/10.18653/2026.emnlp-main.846.