---
configs:
- config_name: all
  data_files: All.csv.gz
  default: true
- config_name: quality
  data_files: Quality.csv.gz
- config_name: readability
  data_files: Readability.csv.gz
- config_name: sentiment
  data_files: Sentiment.csv.gz
license: cc-by-sa-4.0
task_categories:
- text-classification
- feature-extraction
- translation
language:
- en
tags:
- tatoeba
- multilingual
---

# Tatoeba English Translation Dataset

## Dataset Summary

This dataset is derived from the [Tatoeba](https://tatoeba.org/en) database, focusing on English sentences and their translations. It includes assessments of English sentences using text quality, sentiment, and readability models. The dataset is designed for tasks related to multilingual text quality, readability, and sentiment analysis.

## Supported Tasks and Leaderboards

- Quality Assessment
- Readability Prediction
- Sentiment Analysis

## Dataset Structure

### Data Fields

- `ID`: Unique identifier for the English sentence
- `English`: The English sentence
- `TranslationID`: Unique identifier for the translation
- `Language`: ISO code of the translation language
- `Translation`: The translated sentence
- `Quality`: Text quality score of the English sentence
- `Readability`: Readability score of the English sentence
- `Sentiment`: Sentiment score of the English sentence

### Data Splits

The dataset is divided into the following CSV files:
- `All.csv`: Focused on the English and translated text
- `Quality.csv`: Focused on text quality scores
- `Readability.csv`: Focused on readability scores
- `Sentiment.csv`: Focused on sentiment scores

Each of the quality, readability, sentiment files contains a maximum of 50 000 rows.

Those rows were from 25 000 pairs (2 500 from each of 10 bins) split into English and non-English entries.

## Dataset Creation

### Source Data

The source data is the Tatoeba database, which contains user-contributed translations.

### Annotations

English sentences were assessed using:
- [agentlans/deberta-v3-base-quality-v2](https://huggingface.co./agentlans/deberta-v3-base-quality-v2)
- [agentlans/deberta-v3-base-readability-v2](https://huggingface.co./agentlans/deberta-v3-base-readability-v2)
- [agentlans/deberta-v3-base-tweet-sentiment](https://huggingface.co./agentlans/deberta-v3-base-tweet-sentiment)

### Personal and Sensitive Information

The dataset may contain personal names and locations as part of the sentences.

## Considerations for Using the Data

### Social Impact of Dataset

This dataset can be used to improve machine translation systems and text analysis tools, potentially bridging language barriers and enhancing cross-cultural communication.

### Discussion of Biases

The dataset may reflect biases present in the original Tatoeba database and in the assessment models used.

### Other Known Limitations

The quality, readability, and sentiment scores are model-generated and may not always accurately reflect human judgments.

## Contributions

- Thanks to the Tatoeba Project contributors for the original sentence pairs.
- The rest of the analyses are mine.