BERT (post-)trained from review corpus to understand sentiment, options and various e-commence aspects.

BERT_Review is cross-domain (beyond just laptop and restaurant) language model with one example from randomly mixed domains, post-trained (fine-tuned) on a combination of 5-core Amazon reviews and all Yelp data, expected to be 22 G in total. It is trained for 4 epochs on bert-base-uncased. The preprocessing code here.

Model Description

The original model is from BERT-base-uncased trained from Wikipedia+BookCorpus.
Models are post-trained from Amazon Dataset and Yelp Dataset.


Loading the post-trained weights are as simple as, e.g.,

import torch
from transformers import AutoModel, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("activebus/BERT_Review")
model = AutoModel.from_pretrained("activebus/BERT_Review")

Evaluation Results

Check our NAACL paper BERT_Review is expected to have similar performance on domain-specific tasks (such as aspect extraction) as BERT-DK, but much better on general tasks such as aspect sentiment classification (different domains mostly share similar sentiment words).


If you find this work useful, please cite as following.

    title = "BERT Post-Training for Review Reading Comprehension and Aspect-based Sentiment Analysis",
    author = "Xu, Hu and Liu, Bing and Shu, Lei and Yu, Philip S.",
    booktitle = "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics",
    month = "jun",
    year = "2019",
Downloads last month
Hosted inference API
Mask token: [MASK]
This model can be loaded on the Inference API on-demand.