|
Welcome to **RoBERTArg**! |
|
|
|
π€ **Model description**: |
|
This model was trained on ~40k heterogeneous manually annotated sentences (π Stab et al. 2018) of controversial topics (abortion etc.) to classify text into one of two labels: π· **NON-ARGUMENT** (0) and **ARGUMENT** (1). |
|
|
|
**Dataset** |
|
|
|
Please note that the label distribution in the dataset is imbalanced: |
|
* NON-ARGUMENTS: |
|
* ARGUMENTS: |
|
|
|
**Model training** |
|
**RoBERTArg** was fine-tuned on a $RoBERTA_{base}$ pre-trained model using the HuggingFace trainer with the following hyperparameters. The hyperparameters were determined using a hyperparameter search on a 20% validation set. |
|
|
|
``` |
|
training_args = TrainingArguments( |
|
num_train_epochs=2, |
|
learning_rate=2.3102e-06, |
|
seed=8, |
|
per_device_train_batch_size=64, |
|
per_device_eval_batch_size=64, |
|
) |
|
``` |
|
|
|
**Evaluation** |
|
The model was evaluated using 20% of the sentences (80-20 train-test split). |
|
|
|
| Model | Acc | F1 | R arg | R non | P arg | P non | |
|
|----|----|----|----|----|----|----| |
|
| RoBERTArg | 0.8193 | 0.8021 | 0.8463 | 0.7986 | 0.7623 | 0.8719 | |
|
|
|
Showing the **confusion matrix** using the 20% of the sentences as an evaluation set: |
|
|
|
| | ARGUMENT | NON-ARGUMENT | |
|
|----|----|----| |
|
| ARGUMENT | 2213 | 558 | |
|
| NON-ARGUMENT | 325 | 1790 | |
|
|
|
**Intended Uses & Potential Limitations** |
|
The model can be a practical starting point to the complex topic **Argument Mining**. It is a quite challenging task due to the different conceptions of an argument. |
|
|
|
This model is a part of an open-source project providing several models to detect arguments in text. |
|
ππΎ Check out _chkla/argument-analyzer/_ for more details. |
|
|
|
Enjoy and stay tuned! π |
|
|
|
πStab et al. (2018) https://public.ukp.informatik.tu-darmstadt.de/UKP_Webpage/publications/2018/2018_EMNLP_CS_Cross-topicArgumentMining.pdf |