nimuezorro/bilingual_children_speech
Bilingual Children Speech Dataset Dataset Details This dataset was created with data taken from the Kaggle Dataset Corpus of bilingual children's speech. The original dataset includes much more data but for the purpose of this dataset only child utterances, l1, child_id, and age were extracted. The original dataset also includes much more free flowing dialogue and shorter utterances. Therefore, a script was used to extract the target child's English utterances… See the full description on the dataset page: https://huggingface.co/datasets/nimuezorro/bilingual_children_speech.
Bilingual Children Speech Dataset
Dataset Details
This dataset was created with data taken from the Kaggle Dataset `Corpus of bilingual children's speech`.
The original dataset includes much more data but for the purpose of this dataset only child utterances, l1, child_id, and age were extracted.
The original dataset also includes much more free flowing dialogue and shorter utterances. Therefore, a script was used to extract the target child's English utterances and turn them into medium-length text chunks for supervised machine learning.
The original data is rather unbalanced so I used ChatGPT to create synthetic utterances of l1's that have less corresponding data. These utterances are tagged with 'synthetic' and are only added in the file childes_l1_dataset_balanced.jsonl
Available Configurations
original
Contains only utterances derived from the original CHILDES/Paradis corpus.
balanced
Contains the original data plus synthetic ChatGPT-generated samples for underrepresented L1 groups.
Content
The data was collected in 2002 in Edmonton, Canada where 25 children learning English as a second language participated in an interview. Children were asked about varying topics such as their age, favorite food, school attendance, etc. For a full list of interview questions, please visit the official Paradis website.
Included L1
- Farsi
- Arabic
- Cantonese
- Spanish
- Mandarin
- Korean
- Ukrainian
- Japanese
- Romanian
Uses
This dataset was created for an academic text classification assignment focused on predicting a child’s first language from English speech patterns.
The dataset is relatively small and should primarily be used for research and educational purposes rather than production systems. Please note the use of OpenAI to create synthetic data in childes_l1_dataset_balanced.jsonl.
Dataset Structure Example
{ "text": "my birthday's finished ah um it i don't know when i go to grade one it's my birthday um mrs em and mrs bee um two of them yes i like everything but i just i don't like blocks i don't like it everybody fight with blocks nobody giving that to me nobody giving that to her and they say doesn't want to give me blocks um paint and house and school and paint and draw", "l1": "Farsi", "child_id": "BNFS1", "age": "73", "source_file": "bnfs1.cha" }
Again for synthetic texts we tag child_id, age, and source_file with 'synthetic'