Document-level sentiment classifier for Russian restaurant reviews. Given the full text of a review, it predicts the overall sentiment: Positive, Negative, or Neutral.

This is a fine-tune of cointegrated/rubert-tiny2 (a compact Russian BERT, 29M parameters) for sequence classification. It complements the aspect extraction model billerjully/BERT-absa-rest-reviews-ate: together they answer "what is the review about" and "how does the reviewer feel overall".

Usage

python

from transformers import pipeline

asc = pipeline(
    "text-classification",
    model="billerjully/BERT-absa-rest-reviews-asc",
    top_k=None,  # return probabilities for all three classes
)

text = "Ужасное место. Еда пришла холодной, официант хамил, цены завышены."
print(asc(text)[0])

[{'label': 'Positive', 'score': 0.71},
 {'label': 'Negative', 'score': 0.15},
 {'label': 'Neutral', 'score': 0.14}]

Training data

Trained on billerjully/ru-absa-restaurant-reviews — 7,253 Russian restaurant reviews:

Label	Count	Share
Positive	6,163	85.0%
Negative	591	8.1%
Neutral	499	6.9%

Limitations

Document-level only; no aspect-level sentiment.
Domain-specific (restaurant reviews) and Russian language only.

Aspect extraction model: billerjully/BERT-absa-rest-reviews-ate
Dataset: billerjully/ru-absa-restaurant-reviews
Base model: cointegrated/rubert-tiny2

billerjully/BERT-absa-rest-reviews-asc

BERT-absa-rest-reviews-asc

Usage

Training data

Limitations

Related

billerjully/
BERT-absa-rest-reviews-asc