RimuruChan67/indonesian-googleplay-sentiment

Indonesian Google Play Review Sentiment Dataset

A sentiment classification dataset of 5,336 Indonesian-language Google Play reviews collected from three widely used Indonesian applications: Gojek (on-demand transport and delivery), Tokopedia, and Shopee (e-commerce).

The dataset accompanies the paper Sentiment Classification of Indonesian Google Play Reviews: A Hybrid Evaluation of Random Forest, CNN, and IndoBERT (ICORIS 2026).

Collection

Reviews were retrieved between 23 February and 25 March 2026 with the google-play-scraper library, restricted to the Indonesian language and the Indonesian store region (lang="id", country="id"), sorted by newest, targeting 2,500 reviews per application (7,500 raw reviews in total).

Cleaning steps applied, in order:

1. Removal of exact duplicates (identical application, text, rating and date) — 675 removed
2. Removal of empty texts and missing ratings
3. Exclusion of 3-star reviews, whose polarity is ambiguous — 351 removed
4. Removal of reviews shorter than two words

5,336 reviews remain: Tokopedia 1,834 (34.4%), Gojek 1,782 (33.4%), Shopee 1,720 (32.2%).

Labels

Labels are obtained through weak supervision from star ratings: 1–2 stars are mapped to negative, 4–5 stars to positive. The result is 2,971 negative and 2,365 positive reviews.

Files

googleplay_sentiment_id.csv

column	description
app	Gojek, Tokopedia, or Shopee
review_text	raw review text, unmodified
rating	star rating (1, 2, 4, or 5)
review_date	date of the review
label	positive or negative, derived from the rating
split	train, validation, or test — the stratified 68/12/20 split used in the paper

manual_annotation_496.csv

A validation subset in which 496 reviews were annotated manually from the review text alone, without access to the star rating. This subset quantifies the noise introduced by weak supervision.

column	description
review_text	raw review text
app	source application
rating	star rating
star_label	label derived from the star rating
human_label	label assigned by a human annotator from the text only
agreement	agree or disagree

Agreement: 94.2%, Cohen's kappa 0.88 (almost perfect). Disagreements are strongly asymmetric: of 29 differing cases, 28 carried a positive star rating with critical text. Agreement by rating: 1-star 100.0%, 2-star 96.2%, 5-star 90.8%, and 4-star only 69.2%, which makes the 4-star band the least reliable.

Corpus characteristics

Mean review length 14.9 words (median 9); 34.7% of reviews contain five words or fewer. About 32.5% contain a negation marker and 34.5% contain common slang or abbreviated forms, which makes the corpus a realistic test of informal Indonesian text.

Usage

The dataset has two configurations.

from datasets import load_dataset

# main corpus: 5,336 weakly labeled reviews
reviews = load_dataset("RimuruChan67/indonesian-googleplay-sentiment", "reviews")["train"]
print(reviews[0])

# validation subset: 496 reviews annotated by a human from the text alone
manual = load_dataset("RimuruChan67/indonesian-googleplay-sentiment",
                      "manual_annotation")["train"]
print(manual[0])

To reproduce the exact split used in the paper, filter the split column:

train = reviews.filter(lambda r: r["split"] == "train")            # 3,627
validation = reviews.filter(lambda r: r["split"] == "validation")  #   641
test = reviews.filter(lambda r: r["split"] == "test")              # 1,068

Benchmark results

Reported in the accompanying paper, on the identical held-out test set of 1,068 reviews:

Model	Accuracy	Macro-F1
Random Forest + TF-IDF	0.9157	0.9145
CNN + fastText	0.9260	0.9248
IndoBERT (fine-tuned)	0.9391	0.9382

Limitations

• Labels come from star ratings, not from human annotation; see the validation subset above.
• Only two classes; neutral sentiment is not represented because 3-star reviews were excluded.
• Three applications from two domains only, so findings may not transfer to other domains.
• Reviews reflect a single collection window and may not capture later language change.

Citation

@inproceedings{cahyadi2026sentiment,
  title     = {Sentiment Classification of Indonesian Google Play Reviews:
               A Hybrid Evaluation of Random Forest, CNN, and IndoBERT},
  author    = {Cahyadi, David and Prasetyo, Anang},
  booktitle = {8th International Conference on Cybernetics and Intelligent Systems (ICORIS)},
  year      = {2026}
}

Ethical note

All reviews are publicly visible user content from the Google Play Store. No usernames, account identifiers, or contact details are included. The data was checked for phone numbers, e-mail addresses and URLs, and none were found.