This repo contains TABSA models for Sentihood dataset. The models implemented here use the auxiliary sentence approach introduced in Utilizing BERT for Aspect-Based Sentiment Analysis via Constructing Auxiliary Sentence
Note: The model predictions are present in predictions directory with filename output.jsonl
Download the Sentihood dataset and place it in the data directory.
run python3 generate_datasets.py to generate the auxiliary sentence dataset.
This codebase uses Hydra for configuration management. You can change the configuration present in conf/config.yaml.
Use dataset argument to use either the NLI dataset or QA dataset.
To use QA_M dataset run
python3 main.py dataset=QA_M
or to use NLI_M dataset run
python3 main.py dataset=NLI_M
| Model | Sent Acc | Sent AUC | Asp Acc | Asp F1 | Asp AUC |
|---|---|---|---|---|---|
| BERT | 0.8964 | 0.9526 | 0.7499 | 0.8558 | 0.9654 |
| RoBERTa | 0.9243 | 0.9645 | 0.7728 | 0.8509 | 0.9686 |
| RoBERTa 4 layers | 0.9276 | 0.9757 | 0.7419 | 0.8829 | 0.9667 |
| RoBERTA+BiLSTM | 0.9104 | 0.9645 | 0.7328 | 0.8678 | 0.9591 |
| Model | Sent Acc | Sent AUC | Asp Acc | Asp F1 | Asp AUC |
|---|---|---|---|---|---|
| BERT | 0.9013 | 0.9525 | 0.7312 | 0.8595 | 0.9631 |
| RoBERTa | 0.9285 | 0.9665 | 0.7387 | 0.8877 | 0.9654 |
| RoBERTa 4 layers | 0.9219 | 0.9589 | 0.6998 | 0.8234 | 0.9575 |
| RoBERTA+BiLSTM | 0.9021 | 0.9608 | 0.7344 | 0.9019 | 0.9495 |
Wandb training logs for experiments on QA M and NLI M data.
The evaluation metrics are from BERT for ABSA