This repo is an implementation of Google's paper on Bidirectional Encoder Representations from Transformers(BERT) for sentiment analysis on articles, using huggingface's Pytorch NLP wrapper: https://github.com/huggingface/pytorch-transformers.
The BERT model was placed in a RESTFul API which can run on any server to output the overall sentiment of the article.
Model weights: https://drive.google.com/file/d/1pBvpez7OTFjE7UAdeyDS0pqhA0PkXjvv/view?usp=sharing
- Clone/download this repository
- Install dependencies
pip install -r requirements.txt- Download the model weights and add it to the folder
- Launch the API
python bert_inference.py Use curl or POSTMAN to post your article to the API. A sentiment(positive/negative) along with the confidence will be returned.
e.g. curl (IP address):8000 -F input=@path_to_your_article.txt
The model was fine-tuned using the training data of the IMDB movie review dataset which consist of 50k labelled training/testing data (evenly split). Download link for dataset: https://www.kaggle.com/iarunava/imdb-movie-reviews-dataset
The model achieved an accuracy of 92.6% when tested on the 25k instances of test data in the IMDB dataset. Model was trained on a GeForce GTX 1660 with sentence length of 256 and batch size of 6 for 4 epochs with a learning rate of 2e-5.
The model was subsequently tested on 3 random news articles (in the articles folder), results displayed shows the overall sentiment of the article being positive/negative along with the confidence level:
- Banks in Hong Kong condemn violence, urge restoration of 'harmony' Result: Negative (0.72)
- Pompeo praises ‘US ally’ Denmark after Trump cancels visit Result: Positive (0.95)
- Evidence suggests microplastics in water pose ‘minimal health risk’ Result: Positive (0.89)