More detailed documentation will be uploaded promptly & current codebase will be cleaned up to provide a more consistent interface. Stay tuned!
EqualDefault/, MajDefault/, MinDefault/ provides the training and testing stimuli for the frequency conditions.
Feature-files/ contains the phonological features used.
lr.py: suite to tools to train and obtain heatmaps and accuracy curves
lr_evaluation.py: suite of tools to analyze failure cases on different test stimuli
Transformer_BCE_train.ipynb trains the Transformer Encoder, Transformer_BCE_test.ipynb test the Transformer Encoder on different test stimuli. Transformer_BCE_visualization.ipynb creates saliency maps to analyze feature sensitivity.
LSTM.ipynb trains and test the LSTM Encoder-Decoder model.
| training_cat1/2 | training_cat3 | testing_cat1/2 | testing_cat3 | |
|---|---|---|---|---|
| cvc | 132 | 99 | 32 | 24 |
| cvcvc | 132 | 99 | 32 | 24 |
| vcvc | 132 | 99 | 32 | 24 |
| cvcv | 0 | 99 | 0 | 24 |
| total | 396 + 396 | 396 | 96 | 96 |
| training_cat1/2 | training_cat3 | testing_cat1/2 | testing_cat3 | |
|---|---|---|---|---|
| cvc | 177 | 31 | 32 | 24 |
| cvcvc | 177 | 31 | 32 | 24 |
| vcvc | 177 | 31 | 32 | 24 |
| cvcv | 0 | 31 | 0 | 24 |
| total | 531 + 531 | 124 | 96 | 96 |
| training_cat1/2 | training_cat3 | testing_cat1/2 | testing_cat3 | |
|---|---|---|---|---|
| cvc | 39 | 238 | 32 | 24 |
| cvcvc | 39 | 238 | 32 | 24 |
| vcvc | 39 | 238 | 32 | 24 |
| cvcv | 0 | 238 | 0 | 24 |
| total | 117 + 117 | 952 | 96 | 96 |
| Template | Amount |
|---|---|
| cvc | 32 |
| cvcvc | 32 |
| vcvc | 32 |
| cvcv | 0 |
Mutants: words from "regular" training with the last consonant C changed as follows: If C belongs to class 1: randomly change it either to class 2 or class 3 If C belongs to class 2: randomly change it either to class 3 or class 1 If C belongs to class 3: randomly change it either to class 2 or class 1
| Template | Amount | |
|---|---|---|
| vc | 20 | 20 |
| cvcc | 20 | 20 |
| ccv | 20 |
-h final words are "mutants" of templates cvc and cvcvc