Datasets:

AiLab-IMCS-UL
/

twitter_emotions-ru

language:
  - ru
  - en
license: apache-2.0
task_categories:
  - text-classification
dataset_info:
  - config_name: simplified_ekman
    features:
      - name: ru_text
        dtype: string
      - name: text
        dtype: string
      - name: labels
        dtype:
          class_label:
            names:
              '0': sadness
              '1': joy
              '2': love
              '3': anger
              '4': fear
              '5': surprise
      - name: labels_ekman
        dtype:
          class_label:
            names:
              '0': anger
              '1': disgust
              '2': fear
              '3': joy
              '4': sadness
              '5': surprise
    splits:
      - name: train
        num_bytes: 103759867.36530161
        num_examples: 333447
      - name: validation
        num_bytes: 12970022.317349194
        num_examples: 41681
      - name: test
        num_bytes: 12970022.317349194
        num_examples: 41681
    download_size: 68831057
    dataset_size: 129699912
configs:
  - config_name: simplified_ekman
    data_files:
      - split: train
        path: simplified_ekman/train-*
      - split: validation
        path: simplified_ekman/validation-*
      - split: test
        path: simplified_ekman/test-*

Twitter Emotions dataset in Russian

The original dataset: Emotions

The derived dataset was machine translated from English into Russian using the free Google Translate API (with deep-translator). It also contains an additional labels_ekman column that maps the original emotion classes to the Paul Ekman's classification.

The translation script:

import pandas as pd
from deep_translator import GoogleTranslator
from deep_translator.exceptions import TranslationNotFound

# Loads the dataset and drops the ID column
df = pd.read_csv("text.csv").iloc[:, 1:]

translator = GoogleTranslator(source="en", target="ru")

def translate_samples(samples):
    texts = samples["text"].tolist()

    while True:
        try:
            translated = translator.translate_batch(texts)
            break
        except TranslationNotFound:
            print(f"Translation failed for '{texts}', retrying...")

    # Replaces None values with the original text if translation was not successful
    translated = [
        t if t is not None else orig
        for t, orig in zip(translated, texts)
    ]

    # Prints replacements
    for t, orig in zip(translated, texts):
        if t == orig:
            print(f"Replaced {orig} with {t}")

    samples["ru_text"] = translated
    return samples

# Apply batch translation
batch_size = 500
translated_dataset = df.groupby(df.index // batch_size, group_keys=False).apply(translate_samples)

The column labels contains the emotion classes of the original dataset:

0: sadness
1: joy
2: love - not distinguished in the Ekman's classification
3: anger
4: fear
5: surprise

The column labels_ekman contains the corresponding Ekman's emotion classes:

0: anger
1: disgust - omitted, since not used in the original dataset
2: fear
3: joy
4: sadness
5: surprise

The mapping from the original to the Ekman's classification is made as follows:

Original	Ekman
sadness (0)	sadness (4)
joy (1)	joy (3)
love (2)	joy (3)
anger (3)	anger (0)
fear (4)	fear (2)
surprise (5)	surprise (5)

Acknowledgements

This work was supported by the EU Recovery and Resilience Facility project Language Technology Initiative (2.3.1.1.i.0/1/22/I/CFLA/002).

Twitter Emotions dataset in Russian

See also

Acknowledgements