bert-base-chinese fine-tuned on Douban reviews

A sentiment classifier (negative / positive) made for CHI3242.

• Base model: bert-base-chinese, with a new classifier on the [CLS] vector
• Data: Douban reviews from qhchina-data; 1 star = negative, 5 stars = positive
• Training: 5,000 + 5,000 reviews, 1 epoch, learning rate 2e-5
• Test accuracy (1,000 + 1,000 held-out reviews): 0.930