SentiTurca - A Sentiment Analysis Benchmark for Turkish

Dataset Card for SentiTurca

SentiTurca is a sentiment analysis benchmarking dataset including movie reviews, hate speech and e-commerce reviews classification.

Datasets

e-commerce: The e-commerce reviews are scraped from e-commerce websites Trendyol.com and Hepsiburada.com, including review for many product types such as cloths, toys, books, electronics and more.
E-commerce reviews has their stand alone HF repo as well.

movies The movie reviews are scraped from two movie review websites, Sinefil.com and Beyazperde.com. Here, we used 2 labels but for a total challenge of 10 label classification can be found under this dataset's stand alone HF repo. This dataset is also a part of TrGLUE benchmark under the task name sst2.

hate This dataset is the Turkish Hate Map, scraped from Eksisozluk.com and including 4 labels: offense, hate, neutral and civilized.

Dataset statistics

Here are the dataset sizes and number of labels:

Subset	size	num labels
e-commerce	103K	5
movies	78K	2
hate	52K	4

Benchmarking

We benchmarked BERTurk on all of our datasets. All benchmarking scripts can be found under the dedicated SentiTurca Github repo.

Subset	metrics	success
movies	acc./F1	87.4/91.3
e-commerce	acc./F1	0.66/0.64
hate	acc./F1	0.61/0.58

As one sees, hate dataset is quite challenging. For a full critique of the benchmark please visit our research paper.

Citation

Coming soon!