seebeyond / lstm_sentiment

Sentiment analysis on the yelp review dataset using LSTMs

LSTM sentiment analysis on the Yelp-dataset

This project does sentiment analysis on the yelp-review dataset. This is done by modeling the star rating and the number of votes for the review being cool, funny and helpful as a function of the written text in the review.

The models evaluated are different combinations of recurrent neural networks, embeddings and convolutional neural networks. All the models are implemented in PyTorch.

Visualization

For visualization of training and other data Visdom is used. This means that you have to start a Visdom server before being able to visualize anything.

python -m visdom.server

Settings

A python script called 'settings.py' contains all the parameters that are meant to be modified for different test. This includes the number of epochs of training that should be performed, the learning rate of the optimizer, the model to be used (defined by a dictionary with the model to be used and the parameters for this model), the dataset to be used, if there should be visualization or not etc. The settings script also takes care of parsing of command line arguments for the other scripts and to see what can be set using these just run for example:

python train.py --help

For a list of supported arguments.

Training a model

When the setting have been set properly training a model should be as easy as

python train.py --data-path PATH_TO_TRAIN_DATA

Evaluating model on new data

To evaluate a trained model on new data start by making sure that the model and dataset type in the settings script are the same as the one you want to evaluate.

Then evaluation should be performed through:

python evaluate.py --load-path PATH_TO_CHECKPOINT_OF_MODEL --data-path PATH_TO_EVAL_DATA

Additional potentially useful resources

Pre-trained word embeddings for PyTorch: https://github.com/iamalbert/pytorch-wordemb