YUE-PUB-Speech

This repo contains the data for our paper YUE-PUB-Speech: A Speech-based Pragmatic Understanding Benchmark for Cantonese (Interspeech 2026).

Links

• 🤗 Dataset: https://huggingface.co/datasets/Multilingual-NLP/YUE-PUB-Speech

• 💻 Code: https://github.com/swaggy66/Yue-PUB-Speech

• The benchmark consists of 14 tasks collected from 7 existing pragmatic datasets, covering different forms of pragmatic reasoning, including response classification, implicature recovery, figurative language understanding, presupposition inference, and reference resolution.

Task Description

The benchmark consists of 14 pragmatic understanding tasks, which are grouped into three major pragmatic phenomena.

Task ID	Phenomenon	Description	Source
Task 1–2	Implicature	Response Classification	CIRCA
Task 3	Implicature	Response Classification	CIRCA
Task 4	Implicature	Implicature Recovery	GRICE
Task 5–6	Implicature	Agreement / Sarcasm Detection	FigQA
Task 7	Implicature	Figurative Language Understanding	FLUTE
Task 8–9	Implicature	Figurative Language Understanding	FLUTE
Task 10	Implicature	Implicature Natural Language Inference (NLI)	IMPPRES
Task 11	Presupposition	Presupposition Natural Language Inference (NLI)	IMPPRES
Task 12	Presupposition	Presupposition Question Answering	DailyDialog
Task 13	Reference	Deictic Question Answering	GRICE
Task 14	Reference	Reference Resolution via Metonymy	Metonymy

Data Description

This is a dataset suitable for the speech-based pragmatic understanding task with multiple-choice question answering.

All datasets are stored in the data/ folder.

• The benchmark covers 3 major pragmatic phenomena.

phenomena = [
    "implicature",
    "presupposition",
    "reference"
]

• The benchmark consists of 14 tasks collected from 7 existing pragmatic datasets.

sources = [
    "CIRCA",
    "GRICE",
    "FigQA",
    "FLUTE",
    "IMPPRES",
    "DailyDialog",
    "Metonymy"
]

• The benchmark contains 1,680 dialogue instances with approximately 10.87 hours of recorded Cantonese speech.

• The speech recordings are collected from 4 native Cantonese speakers (2 male and 2 female).

• The benchmark supports the following input modalities.

modalities = [
    "text",
    "speech",
    "text+speech"
]

• Each instance contains the dialogue context, question, answer options, ground-truth answer, and the corresponding speech recording. An example is shown below.

{
    "id": "0",
    "context": "X同Y係星期五放工時間差唔多嘅同事。",
    "question": "想唔想我順便車你返屋企？",
    "response": "我想我自己車我自己返屋企。",
    "options": [
        "Direct answer",
        "Indirect answer"
    ],
    "answer": "Direct answer",
    "audio": "audio/0.wav"
}

Dataset Statistics

Property	Value
Language	Cantonese
Modality	Text + Speech
Total Instances	1,680
Audio Duration	10.87 Hours
Speakers	4 Native Speakers (2 Male, 2 Female)
Pragmatic Phenomena	3
Tasks	14
Source Datasets	7

Pragmatic Phenomena

Category	#Tasks	#Instances
Implicature	10	1,200
Presupposition	2	240
Reference	2	240

Evaluation

The primary evaluation metric is:

• Accuracy

Citation

@inproceedings{wen2026yuepubspeech,
  title     = {{YUE-PUB-Speech}: A Speech-based Pragmatic Understanding Benchmark for Cantonese},
  author    = {Wen, Yajie and Gong, Ziwei and Wu, Chengyan and Gong, Xiyun and Xue, Yun and Hirschberg, Julia and Ma, Bolei},
  booktitle = {Proceedings of Interspeech 2026},
  year      = {2026},
  address   = {Sydney, Australia},
  publisher = {International Speech Communication Association (ISCA)}
}