Propicto/propicto-commonvoice

---
license: cc0-1.0
task_categories:
- translation
language:
- fr
tags:
- pictograms
- AAC
pretty_name: Propicto-commonvoice
---

# Propicto-commonvoice

## 📝 Dataset Description

- **Public:** True
- **Tasks:** MT

Propicto-commonvoice is a dataset of aligned speech-id/transcription/pictograms (the pictograms correspond to the identifier associated with an ARASAAC pictogram) in French.
It was created from the CommonVoice-15.0 French corpus.

Propicto-commonvoice contains three CSV files: `train`, `valid`, and `test`, with the following statistics:
| **Split** | **Number of utterances** |
|:-----------:|:-----------------------:|
| train | 527,544 |
| valid | 16,130 |
| test | 16,132 |

## ⚒️ Dataset Structure

Each file contains the following information :
```csv
clips : the unique identifier of the utterance, which corresponds to a unique audio clip file (in mp3) from the commonvoice dataset
text : the transcription of the audio clip
pictos : the sequence of id pictograms from ARASAAC
tokens : the sequence of tokens, each of them is the keyword associated to the ARASAAC id pictogram
```

## 💡 Dataset example

For the given sample :
```csv
clips : common_voice_fr_24683664.mp3
text : l'auteur est connu comme auteur de romans policiers
pictos : [8476, 11258, 8456, 12313, 11258, 7074, 2450, 5547]
tokens : le écrivain connaître comme écrivain de livre agent_de_police_municipale
```

- The `text` is the associated transcription, in en : “the author is known as a writer of detective novels”.<br />
- `pictos` is the sequence of pictogram IDs, each of them can be retrieved from here : 8476 = https://static.arasaac.org/pictograms/8476/8476_2500.png<br />
- `tokens` are retrieved from a specific lexicon and can be used to train translation models.


## ℹ️ Dataset Sources

- **Repository:** [cv-corpus-15.0-2023-09-08-fr](https://commonvoice.mozilla.org/fr/datasets)
- **Papers :**
- [Common Voice: A Massively-Multilingual Speech Corpus](https://aclanthology.org/2020.lrec-1.520/) (Ardila et al., LREC 2020)

## 💻 Uses

Propicto-CommonVoice is intended for training Speech-to-Pictogram and Text-to-Pictogram translation models.
It can also be used to fine-tune large language models for translation into pictograms.

## ⚙️ Dataset Creation

The dataset was created using a specific formalism that converts French oral transcriptions into corresponding sequences of pictograms.
This formalism incorporates grammatical rules to handle specific phenomena (e.g., negation, named entities, pronominal forms, plural forms) in French, as well as a dictionary associating each ARASAAC pictogram ID with a set of keywords (tokens).
It was presented in:
[A Multimodal French Corpus of Aligned Speech, Text, and Pictogram Sequences for Speech-to-Pictogram Machine Translation](https://aclanthology.org/2024.lrec-main.76/) (Macaire et al., LREC-COLING 2024).

**Source Data:** Read speech (oral transcriptions).

## ⁉️ Limitations

Translations may contain inaccuracies due to incorrect or missing mappings of words to pictograms.

## 💡 Information

- **Curated by:** Cécile MACAIRE
- **Funded by :** [PROPICTO ANR-20-CE93-0005](https://anr.fr/Projet-ANR-20-CE93-0005)
- **Language(s) (NLP):** French
- **License:** cc0-1.0

## 📌 Citation

```bibtex
@inproceedings{macaire24_interspeech,
title = {Towards Speech-to-Pictograms Translation},
author = {Cécile Macaire and Chloé Dion and Didier Schwab and Benjamin Lecouteux and Emmanuelle Esperança-Rodier},
year = {2024},
booktitle = {Interspeech 2024},
pages = {857--861},
doi = {10.21437/Interspeech.2024-490},
issn = {2958-1796},
}
```

## 👩‍🏫 Dataset Card Authors

**Cécile MACAIRE, Chloé DION, Emmanuelle ESPÉRANÇA-RODIER, Benjamin LECOUTEUX, Didier SCHWAB**