semeval2023

Environment

python=3.8 All the depended packages are listed in requirements.txt

The structure of directory

└── task4
    ├── README.md
    ├── apps
    ├── bin
    ├── configuration
    ├── data
    ├── data_man
    ├── logs
    ├── metric
    ├── modeling
    └── papers

some experiments

At first stage, we used all the fields to encode information with transformers-based models. But we found just precise is vital for the model performance.

The best encoder model is roberta-large, compared with bert-base-uncased, bert-large-uncased, albert, etc.

submission records

└── task4
    ├── apps
    ├── bin
    ├── configuration
    ├── data
    │   ├── kfold
    │   ├── submission            -------- 提交文件所在的目录
    │   │   └── labels_1.tsv        
    │   ├── test_data
    │   ├── training_data
    │   └── validate_data
    ├── data_man
    ├── logs
    ├── metric
    ├── modeling
    └── papers

1. 文件名： labels_1.tsv
2. 实验参数： baseline_argument_data_module,rewrite_argument_dataset,class_balanced_loss_argument_model,roberta-large,baseline_argument_data_module,rewrite_argument_dataset,class_balanced_loss_argument_model,roberta-large,16,35,val_f1 进行8折交叉验证的结果。