Sentiment Analysis SST-5

This repository contains an LSTM model implemented by PyTorch to perform sentiment classification on the Stanford Sentiment Treebank (SST-5) dataset. We train the model with/without pretrained embeddings and conduct several experiments on different hyperparameters. Since the SST-5 dataset contains sentiment labels for each token in sentences, we develop a modified model to utilze this information.

We use the 300-dim GloVe embeddings from 6B tokens and provide a report to introduce the implementation details and evaluation results.

Project Structure and Environments

The required environments are as follows.

• torch-1.11.0
• torchtext-0.12.0
• pytreebank-0.2.7: used to load the datasets.

The structure of our projects is as follows.

• sentiment analysis: basic version.
  • /codes: contain all codes.
    • test.py: the entrance of our project.
    • train.py: define the class to train the model.
    • model.py: define the model.
    • utils.py: load the data for training and evaluations.
    • config.yaml: store the configurations for model trianing.
  • /weight: the directory to save models, training logs and results.
  • /data: the directory of the datasets and pretrained embeddings.
• improved: improved version.
  • /codes: contain all codes.
    • test_improved.py: the entrance of the improved model.
    • train_improved.py: define the class to train the model.
    • model_improved.py: define the improved model.
    • utils_improved.py: load the data for training and evaluations.
    • config.yaml: store the configurations for model trianing.
  • /weight: the directory to save models and training logs.
  • /data: the directory of corpus used in training and evaluation.
• report.pdf: a brief introduction of our implementation details, the improved model and evaluation results.

Usage

First download the Stanford Sentiment Treebank (SST-5) dataset and the pretrained embeddings into the /data directory. Then unzip them.

After that, you can use the following command to run our codes in the /codes directory.

python test.py --config=config.yaml

The meaning of each configuration can be found in our report.