Datasets:

Modalities:
Text
Formats:
csv
Languages:
Kazakh
ArXiv:
Libraries:
Datasets
pandas
License:

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

KazSAnDRA

Kazakh Sentiment Analysis Dataset of Reviews and Attitudes, or KazSAnDRA, is a dataset developed for Kazakh sentiment analysis. KazSAnDRA comprises a collection of 180,064 reviews obtained from various sources and includes numerical ratings ranging from 1 to 5, providing a quantitative representation of customer attitudes.

In the original study, KazSAnDRA was utilised for two distinct tasks:

  1. polarity classification (PC), involving the prediction of whether a review is positive or negative:
    • reviews with original scores of 1 or 2 were classified as negative and assigned a new score of 0,
    • reviews with original scores of 4 or 5 were classified as positive and assigned a new score of 1,
    • reviews with an original score of 3 were categorised as neutral and were excluded from the task.
  2. score classification (SC), where the objective was to predict the score of a review on a scale ranging from 1 to 5. To align with the enumeration used for labelling in the classifier, which starts from 0 rather than 1, labels 1–5 were transformed into 0–4.

KazSAnDRA consists of seven CSV files. File full.csv contains all the 180,064 reviews and ratings from 1 to 5. Files train_pc.csv, valid_pc.csv, and test_pc.csv are the training, validation, and testing sets for the polarity classification task, respectively. Files train_sc.csv, valid_sc.csv, and test_sc.csv are the training, validation, and testing sets for the score classification task, in turn.

All files, except for full.csv, include records containing a custom review identifier (custom_id), the original review text (text), the pre-processed review text (text_cleaned), the corresponding review score (label), and the domain information (domain). File full.csv includes records containing a custom review identifier (custom_id), the original review text (text), the corresponding review score (label), and the domain information (domain).

Dataset Statistics

For the sake of maintaining consistency and facilitating reproducibility of our experimental outcomes among different research groups, we partitioned KaZSAnDRA into three distinct sets: training (train), validation (valid), and testing (test) sets, following an 80/10/10 ratio.

Task Train Valid Test Total
# % # % # % # %
PC 134,368 80 16,796 10 16,797 10 167,961 100
SC 140,126 80 17,516 10 17,516 10 175,158 100

The distribution of reviews across the three sets based on their domains and scores for the PC task:

Domain Train Valid Test
# % # % # %
Appstore 101,477 75.52 12,685 75.52 12,685 75.52
Market 22,561 16.79 2,820 16.79 2,820 16.79
Mapping 6,509 4.84 813 4.84 814 4.85
Bookstore 3,821 2.84 478 2.85 478 2.85
Total 134,368 100 16,796 100 16,797 100
Score Train Valid Test
# % # % # %
1 110,417 82.18 13,801 82.17 13,804 82.18
0 23,951 17.82 2,995 17.83 2,993 17.82
Total 134,368 100 16,796 100 16,797 100

The distribution of reviews across the three sets based on their domains and scores for the SC task:

Domain Train Valid Test
# % # % # %
Appstore 106,058 75.69 13,258 75.69 13,257 75.69
Market 23,278 16.61 2,909 16.61 2,910 16.61
Mapping 6,794 4.85 849 4.85 849 4.85
Bookstore 3,996 2.85 500 2.85 500 2.85
Total 140,126 100 17,516 100 17,516 100
Score Train Valid Test
# % # % # %
5 101,302 72.29 12,663 72.29 12,663 72.29
1 20,031 14.29 2,504 14.30 2,504 14.30
4 9,115 6.50 1,140 6.51 1,139 6.50
3 5,758 4.11 719 4.10 720 4.11
2 3,920 2.80 490 2.80 490 2.80
Total 140,126 100 17,516 100 17,517 100

How to Use

To load the subsets of KazSAnDRA separately:

from datasets import load_dataset

full = load_dataset("issai/kazsandra", "full")
pc = load_dataset("issai/kazsandra", "polarity_classification")
sc = load_dataset("issai/kazsandra", "score_classification")
Downloads last month
34

Models trained or fine-tuned on issai/kazsandra