This is the original source cod of KESA: A Knowledge Enhanced Approach For Sentiment Analysis. AACL, 2022.

Requirements

install Python 3.6

If you are using Windows, create python virtual environment by platform, e.g., Pycharm, and click install requirements.txt.

If your are using Ubuntu, create python virtual environment by python3 -m venv kesa, and next activate the virtual environment by source kesa/bin/activate, and then install python packages by pip install -r requirements.

Dataset and Rosources

For dataset MR, SST2, SST5 and IMDB, and sentiment lexicon SentiWordNet 3.0, your can download directly from here. For checkpoints, you can download from here.

Preprocess

Preprocess SentimentWordNet 3.0 to get word-level polarity by preprocess_lexicon/gen_word_level_polarity.py.

Fine-tune

You can fine-tune KESA on SST2 dataset based on checkpoints released by SentiLARE, with label combination is CC (conditional combination) using this command: python run_sentiment_classifier.py --do_train --do_eval --model_type roberta_2task --model_name_or_path ../models/SentiLARE_pretrain_roberta --task_name sst-2 --data_dir ../dataset/SST_2/ --all_data_file ../dataset/SST_2/sst.binary.all --lexicon_file ../dataset/lexicon/SWN.word.polarity --num_train_epochs 3.0 --per_gpu_eval_batch_size 1000 --per_gpu_train_batch_size 32 --max_seq_length 128 --learning_rate 2e-5 --loss_type aggregation --loss_balance_type add_vec --c 0.01 --seed 11

For more commands, please refer to commands.txt