AIFI

This repo contains the data and code for our paper Modeling Adaptive Inter-Task Feature Interactions via Sentiment-Aware Contrastive Learning for Joint Aspect-Sentiment Prediction in AAAI 2024.

Requirements

Please note that some packages (such as transformers) are under highly active development, so we highly recommend you to install the specified version of the following packages:

• numpy==1.24.3
• six==1.16.0
• torch==2.1.2
• transformers==4.36.2

Or you can install it directly with the command:

pip install -r requirements.txt

Quick Start

• Set up the environment as described in the above section

• Download the pre-trained bert model (you can also use larger versions for better performance depending on the availability of the computation resource), put it under the folder

  cache_bert_large

  The catalog of downloaded models is as follows：

  └─cache_bert_large
    ├─config.json
    ├─pytorch_model.bin
    ├─tokenizer_config.json
    ├─tokenizer.json
    ├─vocab.txt

• Run command

  bash script/run.sh

  which runs the **ABSA** task on the **laptop14** / **rest16** / **twitter** dataset.

• The final model, metrics results, and experimental parameters, will be in the generated OUT folder

Detailed Usage

time=$(date "+%m_%d_%H:%M:%S")
BERT_DIR="cache_bert_large"
DATA_DIR="data/absa"
DATA_Type="laptop14"
OUT="OUT"
SAVE=$OUT/$DATA_Type/$time

CUDA_VISIBLE_DEVICES=1 \
python -m main.run_joint_span \
  --max_temperature 1 \
  --min_temperature 0.05 \
  --use_si \
  --weight_si 0.15 \
  --weight_temp 0.1 \
  --weight_kl 0.1 \
  --shared_weight 0.1 \
  --num_train_epochs 80 \
  --vocab_file $BERT_DIR/vocab.txt \
  --bert_config_file $BERT_DIR/config.json \
  --init_checkpoint $BERT_DIR/pytorch_model.bin \
  --data_dir $DATA_DIR \
  --output_dir $SAVE\
  --train_file ${DATA_Type}_train.txt \
  --predict_file ${DATA_Type}_test.txt \
  --train_batch_size 64 \
  --predict_batch_size 64 \
  --learning_rate 3e-5 \

• $DATA_DIR represents the path to the dataset. data/absa
• $DATA_Type refers to one of the four datasets in [laptop14, rest_total, twitter1, twitter2, twitter3, twitter4, twitter5, twitter6, twitter7, twitter8, twitter9, twitter10]

Note that 10 cross-validations are required on the twitter dataset here, and all the final results need to be averaged.

More details can be found in the paper and the help info in the main/run_joint_span.py.

Citation

If the code is used in your research, please star our repo and cite our paper as follows:

@inproceedings{chen2024modeling,
  title={Modeling Adaptive Inter-Task Feature Interactions via Sentiment-Aware Contrastive Learning for Joint Aspect-Sentiment Prediction},
  author={Chen, Wei and Liu, Yuxuan and Zhang, Zhao and Zhuang, Fuzhen and Zhong, Jiang},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={16},
  pages={17781--17789},
  year={2024}
}