YUE-PUB-Speech: A Speech-Based Pragmatic Understanding Benchmark for Cantonese

MCML Authors

Bolei Ma

→ Group Frauke Kreuter
 Social Data Science and AI

Abstract

Pragmatic understanding is essential for natural human communication, yet most existing studies focus on text-only settings. In this paper we introduce YUE-PUB-Speech, the first multimodal pragmatic dataset for Cantonese, a low-resource<br>language, constructed by sampling pragmatically rich texts and recording corresponding speech by trained annotators. The corpus provides paired text–speech instances that capture pragmatic meaning in spoken language without requiring fine-grained phonological or word-level alignment. Baseline experiments show that incorporating speech signals improves pragmatic understanding compared to text-only models, highlighting the importance of multimodal resources for pragmatic research in speech and language processing.

inproceedings WGW+26

---

Interspeech 2026

27th Annual Conference of the International Speech Communication Association. Sydney, Australia, Sep 27-Oct 01, 2026. To be published. Preprint available.

---

Authors

Y. Wen • Z. Gong • C. Wu • X. Gong • Y. Xue • J. Hirschberg • B. Ma

---

Links

PDF URL

---

Research Area

 C4 | Computational Social Sciences