BERT-absa-rest-reviews-ate
Aspect Term Extraction (ATE) model for Russian restaurant reviews. Given a review, it finds aspect terms (words the reviewer is talking about) and assigns each to one of five categories: Food, Service, Price, Interior, Delivery.
This is a fine-tune of cointegrated/rubert-tiny2 (a compact Russian BERT, 29M parameters) for token classification with BIO tagging. It is the first stage of an Aspect-Based Sentiment Analysis (ABSA) pipeline; see the companion sentiment model billerjully/BERT-absa-rest-reviews-asc.
Labels
| Tag | Category | Examples of terms |
|---|---|---|
B-Food / I-Food | Food | паста, стейк, порции, суп |
B-Service / I-Service | Service | официант, персонал, обслуживание |
B-Price / I-Price | Price | цены, счёт, стоимость |
B-Interior / I-Interior | Interior | интерьер, музыка, атмосфера |
B-Delivery / I-Delivery | Delivery | доставка, курьер, заказ |
O | — | everything else |
Usage
from transformers import pipeline
ate = pipeline(
"token-classification",
model="billerjully/BERT-absa-rest-reviews-ate",
aggregation_strategy="simple", # merges B-/I- tokens into whole terms
)
text = "Паста была превосходной, но официант нас совсем забыл, а цены завышены."
for aspect in ate(text):
print(aspect["entity_group"], text[aspect["start"]:aspect["end"]], round(aspect["score"], 3))
Food паста 0.98
Service официант 0.98
Price цены 0.92
Tip: offsets (
start/end) are Python code-point indices. If you highlight spans in JavaScript, slice the text as an array of code points (Array.from(text)) — emoji in reviews shift UTF-16 indices otherwise.
Training data
Trained on billerjully/ru-absa-restaurant-reviews — 7,253 Russian restaurant reviews with token-level BIO annotation. Aspect mention counts:
| Category | Mentions |
|---|---|
| Food | 9,481 |
| Service | 4,985 |
| Interior | 2,892 |
| Price | 1,087 |
| Delivery | 849 |
Limitations
- Domain-specific: trained only on restaurant reviews; quality degrades on other domains.
- Mostly single-token aspects:
I-tags are very rare in the training data (< 60 occurrences), so multi-word aspect terms are seldom produced. - Recall is imperfect: the model is small (29M params) and may miss some aspects, especially rare Price/Delivery mentions.
- Russian language only.
Related
- Sentiment model: billerjully/BERT-absa-rest-reviews-asc
- Dataset: billerjully/ru-absa-restaurant-reviews
- Base model: cointegrated/rubert-tiny2