bert-base-uncased-finetuned-advanced-srl_arg

This model is a fine-tuned version of bert-base-uncased on the English Universal Propbank dataset for the Semantics Role Labeling (SRL) task. It achieves the following results on the evaluation set:

• Loss: 0.0914
• Precision: 0.8664
• Recall: 0.8673
• F1: 0.8669
• Accuracy: 0.9812

Model description

This more advanced SRL model uses similar apporach as the Augment method described in NegBERT (Khandelwal, et al. 2020). That is, adding a special token ([V]) immediately before the predicate:

This [V] is a sentence.

Note that the special token and the predicate is considered a whole. That is, the actual sentence is like

'This' '[V] is' 'a' 'sentence' '.'

Usages

The model labels semantics roles given input sentences. See usage examples at https://github.com/dannashao/bertsrl/blob/main/Evaluation.ipynb

Training and evaluation data

The English Universal Proposition Bank v1.0 data. See details at https://github.com/UniversalPropositions/UP-1.0

Training procedure

See details at https://github.com/chuqiaog/Advanced_NLP_group_1/blob/main/A3/A3_main.ipynb

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
0.0457	1.0	2655	0.0849	0.8447	0.8644	0.8544	0.9792
0.0322	2.0	5310	0.0883	0.8586	0.8679	0.8632	0.9806
0.0234	3.0	7965	0.0914	0.8664	0.8673	0.8669	0.9812

Framework versions

• Transformers 4.37.0
• Pytorch 2.0.1+cu117
• Datasets 2.16.1
• Tokenizers 0.15.1