[简体中文 | English]

Use GloVe/Word2Vec embedding, LSTM and Attention to do SST sentiment classification.

Word to idx is ready in /output_data directory

We use 2 word embedding technique: GloVe & Word2Vec

GLoVe: We use glove.840B.300d.txt. You can download it from stanfordnlp

Word2Vec: We use GoogleNews-vectors-negative300.bin. You can download .gz file from google drive

• It's obvious that word embedding play an important roll in training
• In this project, we set large dropout for better test performance
• Models with Attention outperform those without 0.5% approximately

python main.py run

If you have a test.tsv different from the one in this project's ./data/SST directory, you should rename your test.tsv to test_pj.tsv, put it in the ./data/SST directory, and simply run:

python preprocess.py

If you encounter any trouble, here may have some details.

This will use the saved word map to interpret your test.tsv to tokens.

For testing and getting predictions, simply run:

python main.py run --status='test' --best_model="checkpoints/BEST_checkpoint_SST-2_TextAttnBiLSTM_SST.pth"

The prediction for test dataset will be saved in ./prediction.tsv

Refer to class ModelConfig in TextAttnBiLSTM.py and configurations in config.py

• You don't need to download pre-trained word-embedding model if you are using SST dataset

• You can simply run main.py and get a model

• More information for model analysis metrics

Here for implementation of SOTA model

This project is based on this repo@Doragd