|
--- |
|
license: cc-by-4.0 |
|
task_categories: |
|
- token-classification |
|
language: |
|
- en |
|
size_categories: |
|
- 100K<n<1M |
|
tags: |
|
- education |
|
dataset_info: |
|
features: |
|
- name: audio_path |
|
dtype: string |
|
- name: asr_transcript |
|
dtype: string |
|
- name: original_text |
|
dtype: string |
|
- name: mutated_text |
|
dtype: string |
|
- name: index_tags |
|
dtype: string |
|
- name: mutated_tags |
|
dtype: string |
|
splits: |
|
- name: DEL |
|
num_bytes: 208676326 |
|
num_examples: 351867 |
|
- name: SUB |
|
num_bytes: 243003228 |
|
num_examples: 351867 |
|
- name: REP |
|
num_bytes: 303304320 |
|
num_examples: 351867 |
|
download_size: 0 |
|
dataset_size: 754983874 |
|
--- |
|
# Dataset Card for Running Records Errors Dataset |
|
|
|
## Dataset Description |
|
|
|
- **Repository:** |
|
- **Paper:** |
|
- **Leaderboard:** |
|
- **Point of Contact:** |
|
|
|
### Dataset Summary |
|
|
|
The Running Records Errors dataset is an English-language dataset containing 1,055,601 sentences based on the Europarl corpus. As described in our paper, |
|
we take the sentences from the English version of the Europarl corpus and randomly inject three types of errors into the sentences: *repetitions*, where |
|
certain words or phrases are repeated, *substitutions*, where certain words are replaced with a different word, and *deletions*, where the word is completely |
|
omitted. The sentences are then passed into a TTS pipeline consisting of TacoTron2 and HifiGAN model to produce audio recordings of those mutated sentences. Lastly, |
|
the data is passed into a Quartznet 15x5 model which produces a transcript of the spoken audio. |
|
|
|
### Supported Tasks and Leaderboards |
|
|
|
The original purpose of this dataset was to construct a model pipeline that could score running records assesments given a transcript of a child's speech along with |
|
the true text for that assesment. However, we provide this dataset to support other tasks involving error detection in text. |
|
|
|
### Languages |
|
|
|
All of the data in the dataset is in English. |
|
|
|
## Dataset Structure |
|
|
|
### Data Instances |
|
|
|
For each instance, there is a string for the audio transcript, a string for the original text before we added any errors, as well as a string of the sentence with the errors we generated. |
|
In addition, we provide two lists. One list denotes the original position of each word in the mutated text, and the second list denotes the error applied to that word. |
|
|
|
### Data Fields |
|
|
|
- asr_transcript: The transcript of the audio processed by our Quartznet 15x5 model. |
|
- original_text: The original text that was in the Europarl corupus. This text contains no artificial errors. |
|
- mutated_text: This text contains the errors we injected. |
|
- index_tags: This list denotes the original position of each word in `mutated_text.` |
|
- mutated_tags: This list denotes the error applied to each word in `mutated_text.` |
|
|
|
### Data Splits |
|
|
|
- DEL: Sentences that have had random words removed. |
|
- REP: Sentences that have had repetitions inserted. |
|
- SUB: Sentences that have had words randomly substituted. |
|
|
|
## Dataset Creation |
|
|
|
### Curation Rationale |
|
|
|
[More Information Needed] |
|
|
|
### Source Data |
|
|
|
#### Initial Data Collection and Normalization |
|
|
|
[More Information Needed] |
|
|
|
#### Who are the source language producers? |
|
|
|
[More Information Needed] |
|
|
|
## Additional Information |
|
|
|
### Dataset Curators |
|
|
|
This dataset was generated with the guidance of Carl Ehrett. |
