| --- |
| tags: |
| - model_hub_mixin |
| - pytorch_model_hub_mixin |
| - address-parsing |
| - deepparse |
| license: mit |
| datasets: |
| - deepparse/worldwide-addresses |
| library_name: deepparse |
| --- |
| |
| # deepparse model2vec-8m-attention |
|
|
| A [deepparse](https://deepparse.org) address parser (Seq2Seq, with attention mechanism) using the |
| [minishlab/potion-base-8M](https://huggingface.co/minishlab/potion-base-8M) Model2Vec static embeddings (256 |
| dimensions). |
|
|
| ```python |
| from deepparse.parser import AddressParser |
| |
| address_parser = AddressParser(model_type="model2vec-8m", attention_mechanism=True) |
| address_parser("350 rue des Lilas Ouest Québec Québec G1L 1B6") |
| ``` |
|
|
| Install it with `pip install deepparse[model2vec]`. |
|
|
| ## Training |
|
|
| - Data: the [deepparse/worldwide-addresses](https://huggingface.co/datasets/deepparse/worldwide-addresses) dataset |
| (20 countries, 100,000 clean and 50,000 incomplete addresses per country), 90% for training and 10% for validation. |
| `Suburb` and `District` are mapped to `Municipality`, and the `County` and `Country` words are removed. |
| - Architecture: encoder and decoder hidden size 256, one layer each, with an attention mechanism. |
| - Selection: the hidden size (128, 256, 512 or 1024) with the lowest mean validation loss over 5 seeds, then the |
| seed with the lowest validation loss (seed 42, best epoch 26). |
|
|
| ## Results |
|
|
| | | Accuracy (%) | |
| |---|---| |
| | Test, this model | 98.17 | |
| | Test, mean ± std over the 5 seeds of this configuration | 97.99 ± 0.29 | |
| | Validation, this model | 98.41 | |
|
|
| The accuracy is the per-word tag accuracy on the test split of the rebuilt worldwide-addresses data. |
|
|
| ## Limitations |
|
|
| The source data has no `Orientation` nor `GeneralDelivery` tags: the model has these outputs but never predicts them |
| (the street orientation, e.g. `Ouest` or `NW`, is part of the `StreetName`). South Korea has almost only incomplete |
| addresses. |
|
|
