Add the model card
Browse files
README.md
CHANGED
|
@@ -2,10 +2,62 @@
|
|
| 2 |
tags:
|
| 3 |
- model_hub_mixin
|
| 4 |
- pytorch_model_hub_mixin
|
|
|
|
|
|
|
| 5 |
license: mit
|
|
|
|
| 6 |
---
|
| 7 |
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
tags:
|
| 3 |
- model_hub_mixin
|
| 4 |
- pytorch_model_hub_mixin
|
| 5 |
+
- address-parsing
|
| 6 |
+
- deepparse
|
| 7 |
license: mit
|
| 8 |
+
library_name: deepparse
|
| 9 |
---
|
| 10 |
|
| 11 |
+
# deepparse BPEmb Attention
|
| 12 |
+
|
| 13 |
+
A [deepparse](https://deepparse.org) address parser (Seq2Seq, with attention mechanism) using the multilingual [BPEmb](https://nlp.h-its.org/bpemb/) subword embeddings (100,000 merges, 300 dimensions), embedded per word by a bidirectional LSTM.
|
| 14 |
+
|
| 15 |
+
```python
|
| 16 |
+
from deepparse.parser import AddressParser
|
| 17 |
+
|
| 18 |
+
address_parser = AddressParser(model_type="bpemb", attention_mechanism=True)
|
| 19 |
+
address_parser("350 rue des Lilas Ouest Québec Québec G1L 1B6")
|
| 20 |
+
```
|
| 21 |
+
|
| 22 |
+
## Architecture
|
| 23 |
+
|
| 24 |
+
One-layer LSTM encoder and decoder, hidden size 1,024, BPEmb embeddings of 300 dimensions, with an attention mechanism. The
|
| 25 |
+
model predicts 8 tags (`StreetNumber`, `StreetName`, `Unit`, `Municipality`, `Province`, `PostalCode`, `Orientation`
|
| 26 |
+
and `GeneralDelivery`) and the end-of-sequence tag.
|
| 27 |
+
|
| 28 |
+
## Training
|
| 29 |
+
|
| 30 |
+
Trained in 2020 on the dataset described in the
|
| 31 |
+
[deepparse-address-data](https://github.com/GRAAL-Research/deepparse-address-data) repository (addresses of 20
|
| 32 |
+
countries, built from the [libpostal](https://github.com/openvenues/libpostal) data). See
|
| 33 |
+
[Leveraging Subword Embeddings for Multinational Address Parsing](https://arxiv.org/abs/2006.16152) and
|
| 34 |
+
[Multinational Address Parsing: A Zero-Shot Evaluation](https://arxiv.org/abs/2112.04008).
|
| 35 |
+
|
| 36 |
+
## Results
|
| 37 |
+
|
| 38 |
+
- 41 zero-shot countries: 86.00% mean per-country accuracy (24.48% to 99.77%).
|
| 39 |
+
- On the 20 training countries, the attention mechanism improves the accuracy by around 0.5%.
|
| 40 |
+
|
| 41 |
+
Per-country tables: the [Results page](https://deepparse.org/results.html) of the documentation.
|
| 42 |
+
|
| 43 |
+
## Limitations
|
| 44 |
+
|
| 45 |
+
- On the test split of the newer
|
| 46 |
+
[worldwide-addresses](https://huggingface.co/datasets/deepparse/worldwide-addresses) data (same 20 countries,
|
| 47 |
+
clean and incomplete addresses, 260,000 addresses), this model reaches 84.38% per-word accuracy (measured in
|
| 48 |
+
November 2026). It is weaker on incomplete addresses (e.g. a postal code or a municipality alone) than on complete
|
| 49 |
+
ones.
|
| 50 |
+
- In practice, it rarely predicts the `Orientation` and `GeneralDelivery` tags: in November 2026, it predicted
|
| 51 |
+
neither on 12,000 addresses of the worldwide-addresses test split.
|
| 52 |
+
|
| 53 |
+
## Citation
|
| 54 |
+
|
| 55 |
+
```bibtex
|
| 56 |
+
@misc{yassine2020leveraging,
|
| 57 |
+
title={{Leveraging Subword Embeddings for Multinational Address Parsing}},
|
| 58 |
+
author={Marouane Yassine and David Beauchemin and François Laviolette and Luc Lamontagne},
|
| 59 |
+
year={2020},
|
| 60 |
+
eprint={2006.16152},
|
| 61 |
+
archivePrefix={arXiv}
|
| 62 |
+
}
|
| 63 |
+
```
|