requirements_ambiguity_v2

This model is a fine-tuned version of GroNLP/bert-base-dutch-cased on a private dataset with 2,523 labeled software requirements for ambiguity detection in Dutch.

Please contact me via LinkedIn if you have any questions about this model or the dataset used.

The dataset and this model were created as part of the final project assignment of the Natural Language Understanding course (XCS224U) from the Professional AI Program of the Stanford School of Engineering.

It achieves the following results on the evaluation set:

• Loss: 0.7485
• Accuracy: 0.8458
• F1: 0.8442
• Recall: 0.7474

Intended uses & limitations

The model performs automated ambiguity detection through binary text classification. Its intended use is as a tool voor requirements engineers to detect spurious and ambiguous formulations.

Training and evaluation data

The model was trained on ReqAmbi dataset. This dataset is private and contains 2,523 requirement formulations. Each requirement is manually labeled 0 (unambiguous) or 1 (ambiguous). The dataset is split 2,019/253/253 into train, validation and test. The reported metrics are from the evaluation on the test set. The validation set was used for cross-validation during training.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 64
• eval_batch_size: 64
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1	Recall
0.5268	1.0	36	0.5424	0.8063	0.8057	0.7263
0.318	2.0	72	0.4688	0.8182	0.8182	0.7579
0.1244	3.0	108	0.6019	0.8379	0.8366	0.7474
0.0308	4.0	144	0.7485	0.8458	0.8442	0.7474

Framework versions

• Transformers 4.24.0
• Pytorch 2.0.0
• Datasets 2.9.0
• Tokenizers 0.11.0