GiulioZh/ProsodyEmoji

The Prosody of Emojis (ACL 2026)

Dataset Summary

Prosodic features such as pitch, timing, and intonation are central to spoken communication, conveying emotion, intent, and discourse structure. In text-based settings, emojis act as visual surrogates that add affective and pragmatic nuance.

This dataset examines how emojis influence prosodic realisation in speech and how listeners interpret prosodic cues to recover emoji meanings. It contains human speech data collected through a controlled elicited production task. The data demonstrates that speakers systematically adapt their prosody based on emoji cues, and that listeners can recover intended meanings significantly above chance.

This repository serves two purposes:

1. Machine Learning/Speech Resources: A clean corpus of audio mappings linking text, emojis, and spoken prosody.
2. Research Reproducibility: A complete replication package containing the original data and R scripts used to generate the findings in the ACL 2026 paper.

---

Dataset Structure

The repository is organized into two distinct components to serve both NLP practitioners and linguistics researchers.

1. The Core Audio Corpus (For TTS & Audio Classification)

Contains the raw .wav files and clean metadata linking the audio to transcripts, emojis, and intended meanings.

• wav_files/: Directory containing all raw audio recordings.
• metadata.csv: The core mapping file required by the Hugging Face datasets library. Contains:
  • file_name: Path to the audio file.
  • transcription: The text of the spoken sentence.
  • emoji: The emoji prompt used to elicit the prosody.
  • ex1_intent: The intended meaning of the emoji in context.
  • ex2_interpretation: How listeners interpreted the tone of the recording.
• speakers_metadata.csv: Demographics (Age, Gender, Hardware) and calculated expressivity scores for the 124 unique British English speakers.
• listeners_metadata.csv: Demographics for the 112 listeners who participated in the evaluation experiments.

2. Analysis (For Research Reproducibility)

Contains the exact data files and R scripts necessary to reproduce the statistical analysis and Bayesian multilevel modelling from the paper.

• Analysis Data: RQ1_data.csv, RQ2_data.csv, RQ3_data.csv, RQ4_data.csv, RQ4_data_stretch.csv
• R Scripts: rq1.R, rq2.R, rq3.r, rq4_v3.R

Note for reproducing: To run the R scripts, download the repository and ensure your R working directory is set to the folder containing both the scripts and the CSV files.

---

Participant Demographics

Speakers

• Total: 124 unique speakers
• Language/Accent: British English
• Gender: 89 Female, 35 Male/Other
• Age: Mean 42.4 years (Range: 19 - 84)

Listeners

• Total: 112 unique evaluators across Experiments 3 and 4
• Task: Evaluated the communicative success and prosodic similarity of the elicited recordings.

---

Citation

If you use this dataset or the accompanying analysis scripts in your research, please cite our ACL 2026 paper:

@inproceedings{zhou-etal-2026-prosody-emojis,
    title = "The Prosody of Emojis",
    author = "Zhou, Giulio and Lam, Tsz Kin and Birch, Alexandra and Haddow, Barry",
    booktitle = "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026)",
    year = "2026",
    publisher = "Association for Computational Linguistics",
}