tmutton/wcag-accessibility-issues

WCAG Accessibility Issues

A small synthetic dataset for experimenting with machine learning classification of accessibility defects against WCAG success criteria.

This project was created as a learning exercise to explore the complete Hugging Face text-classification workflow, including dataset preparation, tokenisation, fine-tuning a pretrained Transformer model, evaluation and inference.

Dataset

The dataset contains 100 synthetic accessibility defect descriptions across five WCAG success criteria.

Each class contains 20 examples.

Label	WCAG success criterion
0	1.3.1 Info and Relationships
1	2.1.1 Keyboard
2	2.4.3 Focus Order
3	2.4.7 Focus Visible
4	4.1.2 Name, Role, Value

Example:

Custom radio control does not expose its selected state

Label:

4.1.2 Name, Role, Value

All examples are synthetic and were created specifically for this project. The dataset does not contain real audit findings or customer data.

Dataset Splits

The 100 examples are divided using a stratified split:

Split	Examples	Examples per class
Training	80	16
Validation	10	2
Test	10	2

The original complete dataset is available in data.csv.

split_data.py can be used to reproduce the train, validation and test datasets.

Model Training

The project uses distilbert-base-uncased as the pretrained base model.

DistilBERT is fine-tuned for multiclass sequence classification with five output classes corresponding to the five WCAG success criteria.

The training configuration currently uses:

• 5 epochs
• Batch size of 8
• Learning rate of 2e-5
• Validation after each epoch
• Weighted F1 as the model-selection metric

The training process is implemented in train_model.py.

Initial Results

The initial model achieved:

Metric	Result
Test accuracy	80%
Weighted precision	0.867
Weighted recall	0.800
Weighted F1	0.787

These results are based on a test set containing only 10 examples and should not be interpreted as evidence that the model will achieve 80% accuracy on real-world accessibility defects.

The results demonstrate that the end-to-end classification pipeline is working.

Example Predictions

After fine-tuning, the model correctly classified previously unseen examples such as:

no focus indicator

as:

2.4.7 Focus Visible

and:

can't open menu with keyboard

as:

2.1.1 Keyboard

Repository Files

• data.csv — complete synthetic dataset
• train.csv — training split
• validation.csv — validation split
• test.csv — held-out test split
• split_data.py — creates the stratified dataset splits
• load_dataset.py — loads the CSV files using Hugging Face Datasets
• tokenize_dataset.py — tokenises examples and encodes labels
• load_model.py — loads DistilBERT for sequence classification
• train_model.py — fine-tunes and evaluates the classifier
• predict.py — performs inference using the trained model

Limitations

This is a small educational dataset rather than a production accessibility dataset.

In particular:

• Only five WCAG success criteria are represented.
• The examples are synthetic.
• There are only 20 examples per class.
• The test set contains only 10 examples.
• Real accessibility defects may be substantially more complex or ambiguous.
• Some accessibility defects can relate to more than one WCAG success criterion.

The classifier is a closed-set classifier. It must select one of its five known classes even when an issue actually belongs to a WCAG criterion that is not represented in the dataset.

For example, an issue concerning missing alternative text would normally relate to 1.1.1 Non-text Content, but the current model has no 1.1.1 class available.

Purpose

The purpose of this project is to learn and demonstrate an end-to-end Hugging Face machine learning workflow:

Dataset
   ↓
Train / Validation / Test Split
   ↓
Tokenisation
   ↓
Pretrained DistilBERT
   ↓
Fine-tuning
   ↓
Evaluation
   ↓
Inference

The project can be extended by adding more examples, introducing additional WCAG success criteria and evaluating the resulting classifier against a larger and more representative test dataset.