Domain-adapted mBERT for French Legal Practice

This is a sentence-transformers model: It maps sentences & paragraphs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search.

Pretrained transformers model on the top 102 languages with the largest Wikipedia using a masked language modeling (MLM) objective, fitted using Transformer-based Sequential Denoising Auto-Encoder for unsupervised sentence embedding learning with one objective : french legal domain adaptation.

This way, the model learns an inner representation of the french legal language in the training set that can then be used to extract features useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a standard classifier using the features produced by the model as inputs.

Usage (Sentence-Transformers)

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence-transformers

Then you can use the model like this:

from sentence_transformers import SentenceTransformer
sentences = ["This is an example sentence", "Each sentence is converted"]

model = SentenceTransformer("louisbrulenaudet/tsdae-lemone-mbert-base")
embeddings = model.encode(sentences)
print(embeddings)

Usage (HuggingFace Transformers)

Without sentence-transformers, you can use the model like this: First, you pass your input through the transformer model, then you have to apply the right pooling-operation on-top of the contextualized word embeddings.

from transformers import AutoTokenizer, AutoModel
import torch


def cls_pooling(model_output, attention_mask):
    return model_output[0][:,0]


# Sentences we want sentence embeddings for
sentences = ['This is an example sentence', 'Each sentence is converted']

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained("louisbrulenaudet/tsdae-lemone-mbert-base")
model = AutoModel.from_pretrained("louisbrulenaudet/tsdae-lemone-mbert-base")

# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors="pt")

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)

# Perform pooling. In this case, cls pooling.
sentence_embeddings = cls_pooling(model_output, encoded_input["attention_mask"])

print("Sentence embeddings:")
print(sentence_embeddings)

Training

The model was trained with the parameters:

DataLoader:

torch.utils.data.dataloader.DataLoader of length 25000 with parameters:

{'batch_size': 4, 'sampler': 'torch.utils.data.sampler.RandomSampler', 'batch_sampler': 'torch.utils.data.sampler.BatchSampler'}

Loss:

sentence_transformers.losses.DenoisingAutoEncoderLoss.DenoisingAutoEncoderLoss

Parameters of the fit()-Method:

{
    "epochs": 1,
    "evaluation_steps": 0,
    "max_grad_norm": 1,
    "optimizer_class": "<class 'torch.optim.adamw.AdamW'>",
    "optimizer_params": {
        "lr": 3e-05
    },
    "scheduler": "constantlr",
    "steps_per_epoch": null,
    "warmup_steps": 10000,
    "weight_decay": 0
}

Training data

The training database consisted of 100,000 random sentences, each over 40 characters long, from the :

• French Intellectual Property Code (Code de la propriété intellectuelle)
• French Civil Code (Code civil)
• French Labor Code (Code du travail)
• French Monetary and Financial Code (Code monétaire et financier)
• French Commercial Code (Code de commerce)
• French Penal Code (Code pénal)
• French Consumer Code (Code de la consommation)
• French Environment Code (Code de l'environnement)
• French General Tax Code (Code général des Impôts)
• French Code of civil procedure (Code de procédure civile)

The number of sentences per code may not exceed 15,000.

The DenoisingAutoEncoderDataset is crafted to provide pairs of noisy and clean data instances. This arrangement allows the denoising autoencoder model to learn and enhance its ability to reconstruct or generate clean data from the corrupted versions provided in the dataset.

By providing pairs of noisy and clean data instances from legal texts, the denoising autoencoder can learn to reconstruct or denoise the noisy, domain-specific text, effectively capturing the intricate linguistic nuances and domain-specific features. This learning process assists in building a model that can generalize better to the legal domain, even when initially trained on more general or diverse datasets.

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False})
)

Citing & Authors

If you use this code in your research, please use the following BibTeX entry.

@misc{louisbrulenaudet2023,
  author =       {Louis Brulé Naudet},
  title =        {Domain-adapted mBERT for French Legal Practice},
  year =         {2023}
  howpublished = {\url{https://huggingface.co/louisbrulenaudet/tsdae-lemone-mbert-base}},
}

Feedback

If you have any feedback, please reach out at louisbrulenaudet@icloud.com.