RoBERTaABSA

Implementation for paper Does syntax matter? A strong baseline for Aspect-based Sentiment Analysis with RoBERTa [NAACL 2021], a work focusing on Aspect-level Sentiment Classification (ALSC). It conducts a detailed study on the performance gain of dependency tree in ALSC, and provides a strong baseline using RoBERTa.

You can find more information here:

• Paper
• Code
• Paperwithcode

For any questions about code or paper, feel free to create issues or email me via [email protected].

If you are interested in on the whole ABSA task, please have a look at our ACL 2021 paper A Unified Generative Framework for Aspect-Based Sentiment Analysis.

Dependencies

We recommend to create a virtual environment.

conda create -n absa
conda activate absa

packages:

• python 3.7
• pytorch 1.5.1
• transformers 2.11.0
• fastNLP < 1.0
  • pip install git+https://github.com/fastnlp/fastNLP@dev
• fitlog
  • pip install git+https://github.com/fastnlp/fitlog

All codes are tested on linux only.

Data

English Datasets are released in Dataset folder for reproduction. If you want to process with your own data, please refer to scripts in Dataset folder.

Usage

To get ALSC result:

To get ALSC results (see Paperwithcode), run finetune.py in Train folder. Before code running, remember to check Notes below, and make sure --data_dir and --dataset are filled with correct dataset filepath and dataset name.

We also provide detailed arguments and logs here:

It is worth noting that above results are only tested by one run rather than averaged runs reported in the paper. Also remember to check Notes below.

To reproduce the whole experiment in paper:

It includes four steps to reproduce the whole experiment in our paper:

1. Fine-tuning models on ALSC datasets using codes in Train folder, which will save fine-tuned models after fine-tuning.

   python finetune.py --data_dir {/your/dataset_filepath/} --dataset {dataset_name}

2. Generate induced trees using code in Perturbed-Masking folder, which will output datasets serving as input for different models.

   python generate_matrix.py --model_path bert --data_dir /user/project/dataset/ --dataset Restaurant

• model_path can be either bert/roberta/xlmroberta/xlmbert, or the model path where the above fine-tuned model is put.

3. Generate data with different input format corrsponding to specific model.

• ASGCN Input Data:

python generate_asgcn.py --layers 11

• PWCN Input Data:

python generate_pwcn.py --layers 11

• RGAT Input Data:

python generate_rgat.py --layers 11

4. Run code in ASGCN, PWCN and RGAT.

Disclaimer

• We made necessary changes to the original code of ASGCN, PWCN , RGAT and Perturbed-Masking. All the changes we have made are opensourced. We believe all the changes are under MIT License permission.
• Errors may be raised if run above codes following their original steps. We recommand to run them (ASGCN, PWCN , RGAT and Perturbed-Masking) following the readme description in corresponding folders.

Notes

• The learning rate in the paper was written incorrectly and should be corrected to 2e-5 for RoBERTa.
• Remember to split validation set on your own data. The "dev" argument should be filled with corresponding validation filepath in the trainer of finetune.py. We did not provide a validation set partition here, which was an issue that we previously overlooked. Yet in the implementation of our experiment, we use validation set to evaluate performance of different induced trees.

Reference

If you find this work useful, feel free to cite：

@inproceedings{DBLP:conf/naacl/DaiYSLQ21,
  author    = {Junqi Dai and
               Hang Yan and
               Tianxiang Sun and
               Pengfei Liu and
               Xipeng Qiu},
  editor    = {Kristina Toutanova and
               Anna Rumshisky and
               Luke Zettlemoyer and
               Dilek Hakkani{-}T{\"{u}}r and
               Iz Beltagy and
               Steven Bethard and
               Ryan Cotterell and
               Tanmoy Chakraborty and
               Yichao Zhou},
  title     = {Does syntax matter? {A} strong baseline for Aspect-based Sentiment
               Analysis with RoBERTa},
  booktitle = {Proceedings of the 2021 Conference of the North American Chapter of
               the Association for Computational Linguistics: Human Language Technologies,
               {NAACL-HLT} 2021, Online, June 6-11, 2021},
  pages     = {1816--1829},
  publisher = {Association for Computational Linguistics},
  year      = {2021},
  url       = {https://doi.org/10.18653/v1/2021.naacl-main.146},
  doi       = {10.18653/v1/2021.naacl-main.146},
}