CNN Sentiment Analysis

This repository provides code for "Word Embedding Aware Convolutional Networks for Sentiment Analysis". We evaluate various word embeddings on the performance of convolutional networks in the context of sentiment analysis tasks. The paper is available here.

Datasets

This project uses the Movie Reviews, SST-2, and MPQA datasets. These preprocessed datasets can be obtained from the Academia Sinica NLP Lab. Further, we use several types of word embeddings -- Word2Vec, GloVe, and Numberbatch.

Use python3 create_datasets.py to build the datasets from their raw formats.

Training

Use train.sh to train all the models on each dataset and embedding type. For more fine-grained training, refer to the individual flags provided in train.py. For example, the following command can be used to train a CNN on the Movie Reviews corpus using Word2Vec embeddings:

python3 train.py --name=cnn_mr_rand --dataset=mr \
                 --epochs=20 --batch=32 --lr=1e-4 \
                 --reg=1e-3 --edims=300 --etype=w2v

By default, models will be checkpointed according to their validation accuracies in checkpoints/. The training loss, training accuracy, validation loss, and validation accuracy statistics will be logged in stats/. mkdir checkpoints stats can create these directories if they do not already exist.

Testing

Use eval.py to test the models. MR and MPQA have to be tested with cross-validation, but SST-2 has a dedicated test set. The dictionaries mr_epochs and mpqa_epochs control what number of epochs to train the models until. These are determined from the validation accuracies, so they can be changed depending on the model checkpoints.