SENTA

This repository is the official implementation of the paper Interventional Aspect-Based Sentiment Analysis.

Requirements

python3 / pytorch 1.7 / transformers 4.5.1

Dataset

Dataset is included in file. Pre-trained models such as bert-base-uncased can be download from Hugging Face models

The expected structure of files is:

ATLOP
 |-- dataset
 |    |-- asc
 |    |    |-- laptop 
 |    |    |    |-- train.json
 |    |    |    |-- dev.json
 |    |    |    |-- test.json
 |    |    |    |-- test_enriched.json
 |    |    |    |-- build_data.py
 |    |    |-- rest 
 |    |    |    |-- train.json
 |    |    |    |-- dev.json
 |    |    |    |-- test.json
 |    |    |    |-- test_enriched.json
 |    |    |    |-- build_data.py
 |-- run_scm.py
 |-- model.py
 |-- utils.py
 |-- README.md

Training

You can set the --num_train_epochs and --model_name_or_path argument in training scripts.

STEP1: Train a Confounding Model

Train the confounding model on Laptop domain with the following command:

# example for bert-base-uncased
>>  python run_scm.py 
\--model_type bert  --data_dir dataset/asc 
\--per_gpu_train_batch_size 128 
\--per_gpu_eval_batch_size 128 
\--learning_rate 1e-5 
\--num_train_epochs ${num_train_epochs}
\--max_seq_length 64
\--doc_stride 128 
\--threads 10 
\--do_lower_case 
\--evaluate_during_training 
\--logging_steps 500 
\--save_steps 500 
\--dataset eaptop 
\--output_dir ./output_laptop_base 
\--model_name_or_path ${pre_trained_models_path}/bert-base-uncased 
\--do_train 
\--do_eval

STEP2: Train a Interventional Model

Train the interventional model on Laptop domain with the following command:

# example for bert-base-uncased
>> python run_scm.py 
\--model_type bert  --data_dir dataset/asc 
\--per_gpu_train_batch_size 128 
\--per_gpu_eval_batch_size 128 
\--learning_rate 1e-5 
\--num_train_epochs ${num_train_epochs}
\--max_seq_length 64
\--doc_stride 128 
\--threads 10 
\--do_lower_case 
\--evaluate_during_training 
\--logging_steps 500 
\--save_steps 500 
\--dataset laptop 
\--output_dir ./output_laptop_base 
\--model_name_or_path ${pre_trained_models_path}/bert-base-uncased 
\--do_train 
\--do_eval 
\--do_intervention  

Evaluation

To evaluate the trained interventional model in Ori test set,

# example for bert-base-uncased
>>  python run_scm.py 
\--model_type bert  
\--data_dir dataset/asc 
\--do_lower_case 
\--dataset laptop 
\--output_dir ./output_laptop_base 
\--model_name_or_path ${pre_trained_models_path}/bert-base-uncased 
\--do_test  
\--do_intervention

To evaluate the trained interventional model in Change test set(ARTS),

# example for bert-base-uncased
>>  python run_scm.py 
\--model_type bert  --data_dir dataset/asc 
\--do_lower_case 
\--evaluate_during_training 
\--dataset laptop 
\--output_dir ./output_laptop_base 
\--model_name_or_path ${pre_trained_models_path}/bert-base-uncased 
\--do_test 
\--do_test_enriched  
\--do_intervention

How to Cite

@article{bi2021interventional,
  title={Interventional Aspect-Based Sentiment Analysis},
  author={Bi, Zhen and Zhang, Ningyu and Ye, Ganqiang and Yu, Haiyang and Chen, Xi and Chen, Huajun},
  journal={arXiv preprint arXiv:2104.11681},
  year={2021}
}