Movie Review Sentiment Classification

The project is done using Jupyter Notebook with Python 3.7, PyTorch 1.0.1, fastai 1.0.52, gensim, ...

Building a sentiment classifier that takes movie review text and output rating from 1 to 10 using Word2Vec, Bidirectional LSTM, AWD LSTM and more.

Directory Structure

project
├─data
│  ├─test_set.ss                    Test dataset
│  ├─training_set.ss                Training dataset
│  └─validation.ss                  Validation dataset
├─images                            Notebook images
├─language
│  └─movie_corpus.txt               Corpus for training Word2Cec model  
├─rating_model_fastai.ipynb         Plan B notebook
├─rating_model.ipynb                Plan A notebook
├─word2vec.ipynb                    Word2Vec model training notebook
│
...

Report

Reports with implementation introduction, code explanation and result analysis are all embedded in the notebooks for better coherence.

Plan A: Bidirectional LSTM with word2vec as embedding

Training Word2Vec model

The corpus I used is a self-made 10M movie review + Harry Potter sentence collection. File at language/movie_corpus.txt. The dimension of Word2Vec model is 100.

Please see word2vec.ipynb

Training Classifier

Mainly use PyTorch

Please see rating_model.ipynb

Plan B: Transfer Learning LSTM using FastAI

Mainly use FastAI - an high-level library for easier working with PyTorch.

Please see rating_model_fastai.ipynb

Predicts of Test Set

Choosing the result of 'Plan B' for its better performance.

Please see senti_output.ss

License

MIT, see the LICENSE file for details.