SentiX

This is the source code of our COLING 2020 paper <SentiX: A Sentiment-Aware Pre-Trained Model for Cross-domain Sentiment Analysis>

Requirements

• Python 3.6 or higher.
• PyTorch 1.2.0.
• transformers PyTorch 1.2.0.

Pre-trained Models (PyTorch)

The following pre-trained models are available for download from Google Drive:

• SentiX: PyTorch version, same setting with BERT-base，loading model with transformers.

Training

To train SentiX, simply run:

python pretrain_multigpu_final.py --dataset data_all --max_seq_len 256 --batch_size 16 --gradient_accumulation_steps 4 

Evaluation Instructions

To test after setting model path:

run_multidomain_sentiment_analysis.sh

File Structure

• pretrain_multigpu_final.py: main file of pre-trained model
• data_processing: data processing
• models： our models
• data_utils_pretrain_final.py: data_loader