Sentiment Classification using Self-Attention Model and POS Embeddings

This is an extension repo of the paper: ``A Structured Self-Attentive Sentence Embedding'' published by IBM and MILA. https://arxiv.org/abs/1703.03130

The repo is forked from https://github.com/ExplorerFreda/Structured-Self-Attentive-Sentence-Embedding

Usage

get_data.py

Split the official Yelp dataset review.json to training, dev, and testing. Tokenize sentences. Generate the vocabulary.

get_tensors.py

Transform tokens/POS tags to indices. Train POS2vec using word2vec Python library. Zero-pad word and POS sequences.

feature_generator.py

Use PyTorch Dataloader class to generate a batch of features and labels. This will speed up training process.

model.py

Model for using word2vec feature only.

model_pos.py

Model for using word2vec and POS2vec featurs.

model_pos_attention.py

Model for separate attention layers for the two features.

train*.py

Training codes for all combinations of parameters. Need to refactorize them to be one file and accept arguments.

The best accuracy is 73.05% on the testing data.