kelingwang/bert-causation-rating-dr1

Model description

This bert-causation-rating-dr1 model is a fine-tuned biobert-base-cased-v1.2 model on a small set of manually annotated texts with causation labels. This model is tasked with classifying a sentence into different levels of strength of causation expressed in this sentence. Before tuning on this dataset, the biobert-base-cased-v1.2 model is fine-tuned on a dataset containing causation labels from a published paper. This model starts from pre-trained kelingwang/bert-causation-rating-pubmed. For more information please view the link and my GitHub page. The sentences in the dataset were rated independently by two researchers. This dr1 version is tuned on the set of sentences with labels rated by Rater 1.

Intended use and limitations

This model is primarily used to rate for the strength of expressed causation in a sentence extracted from a clinical guideline in the field of diabetes mellitus management. This model predicts strength of causation (SoC) labels based on the text inputs as:

• -1: No correlation or variable relationships mentioned in the sentence.
• 0: There is correlational relationships but not causation in the sentence.
• 1: The sentence expresses weak causation.
• 2: The sentence expresses moderate causation.
• 3: The sentence expresses strong causation. NOTE: The model output is five one-hot logits and will be 0-index based, and the labels will be 0 to 4. It is good to use this python module if one wants to make predictions.

Performance and hyperparameters

Test metrics

This model achieves the following results on the test dataset. The test dataset is a 25% held-out stratified split of the entire dataset with SEED=114514.

• Loss: 5.2014
• Off-by-1 accuracy: 71.1864
• Off-by-2 accuracy: 90.6780
• MSE for ordinal data: 0.7797
• Weighted F1: 0.7164
• Kendall's Tau: 0.8014

This performance is achieved with the following hyperparameters:

• Learning rate: 7.94278e-05
• Weight decay: 0.111616
• Warmup ratio: 0.301057
• Power of polynomial learning rate scheduler: 2.619975
• Power to the distance measure used in the loss function \alpha: 2.0

Hyperparameter tuning metrics

During the Bayesian optimization procedure for hyperparameter tuning, this model achieves the best target metric (Off-by-1 accuracy) of 99.1147, as the result from 4-fold cross-validation procedure based on best hyperparameters.

Training settings

The following training configurations apply:

• Pre-trained model: kelingwang/bert-causation-rating-pubmed
• seed: 114514
• batch_size: 128
• epoch: 8
• max_length in torch.utils.data.Dataset: 128
• Loss function: the OLL loss with a tunable hyperparameter \alpha (Power to the distance measure used in the loss function).
• lr: 7.94278e-05
• weight_decay: 0.111616
• warmup_ratio: 0.301057
• lr_scheduler_type: polynomial
• lr_scheduler_kwargs: {"power": 2.619975, "lr_end": 1e-8}
• Power to the distance measure used in the loss function \alpha: 2.0

Framework versions and devices

This model is run on a NVIDIA P100 CPU provided by Kaggle. Framework versions are:

• python==3.10.14
• cuda==12.4
• NVIDIA-SMI==550.90.07
• torch=2.4.0
• transformers==4.45.1
• scikit-learn==1.2.2
• optuna==4.0.0
• nlpaug==1.1.11