README.md

metadata

language: zh
tags:
  - sentiment-analysis
  - pytorch
widget:
  - text: 房间非常非常小，内窗，特别不透气，因为夜里走廊灯光是亮的，内窗对着走廊，窗帘又不能完全拉死，怎么都会有一道光射进来。
  - text: 尽快有洗衣房就好了。
  - text: 很好，干净整洁，交通方便。
  - text: 干净整洁很好

Note

BERT based sentiment analysis, finetune based on https://huggingface.co/IDEA-CCNL/Erlangshen-Roberta-330M-Sentiment .

The model trained on hotel human review chinese dataset.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification, TextClassificationPipeline

MODEL = "tezign/Erlangshen-Sentiment-FineTune"

tokenizer = AutoTokenizer.from_pretrained(MODEL)

model = AutoModelForSequenceClassification.from_pretrained(MODEL, trust_remote_code=True)

classifier = TextClassificationPipeline(model=model, tokenizer=tokenizer)

result = classifier("很好，干净整洁，交通方便。")

print(result)

"""
print result
>> [{'label': 'Positive', 'score': 0.989660382270813}]
"""

Evaluate

We compared and evaluated the performance of Our finetune model and the Original Erlangshen model on the hotel human review test dataset(5429 negative reviews and 1251 positive reviews).

The results showed that our model substantial improved the precision and recall of positive reviews:

Our finetune model:
              precision    recall  f1-score   support

    Negative       0.99      0.98      0.98      5429
    Positive       0.92      0.95      0.93      1251

    accuracy                           0.97      6680
   macro avg       0.95      0.96      0.96      6680
weighted avg       0.97      0.97      0.97      6680

======================================================

Original Erlangshen model:
              precision    recall  f1-score   support

    Negative       0.81      1.00      0.90      5429
    Positive       0.00      0.00      0.00      1251

    accuracy                           0.81      6680
   macro avg       0.41      0.50      0.45      6680
weighted avg       0.66      0.81      0.73      6680