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. The model was trained on the LLeQA dataset for legal information retrieval in French.

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:

python

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

model = SentenceTransformer('maastrichtlawtech/distilcamembert-lleqa')
embeddings = model.encode(sentences)
print(embeddings)

🤗 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.

python

from transformers import AutoTokenizer, AutoModel
import torch


#Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

# 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('maastrichtlawtech/distilcamembert-lleqa')
model = AutoModel.from_pretrained('maastrichtlawtech/distilcamembert-lleqa')

# 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, mean pooling.
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
print(sentence_embeddings)

Evaluation

We evaluate the model on the test set of LLeQA, which consists of 195 legal questions with a knowlegde corpus of 27.9K candidate articles. We report the mean reciprocal rank (MRR), normalized discounted cumulative gainand (NDCG), mean average precision (MAP), and recall at various cut-offs (R@k).

MRR@10	NDCG@10	MAP@10	R@10	R@100	R@500
36.67	37.24	29.26	52.95	78.07	90.17

Training

Background

We utilize the distilcamembert-base model and fine-tuned it on 9.3K question-article pairs in French. We used a contrastive learning objective: given a short legal question, the model should predict which out of a set of sampled legal articles, was actually paired with it in the dataset. Formally, we compute the cosine similarity from each possible pairs from the batch. We then apply the cross entropy loss with a temperature of 0.05 by comparing with true pairs.

Hyperparameters

We trained the model on a single Tesla V100 GPU with 32GBs of memory during 20 epochs (i.e., 5.4k steps) using a batch size of 32. We used the AdamW optimizer with an initial learning rate of 2e-05, weight decay of 0.01, learning rate warmup over the first 50 steps, and linear decay of the learning rate. The sequence length was limited to 384 tokens.

Data

We use the Long-form Legal Question Answering (LLeQA) dataset to fine-tune the model. LLeQA is a French native dataset for studying legal information retrieval and question answering. It consists of a knowledge corpus of 27,941 statutory articles collected from the Belgian legislation, and 1,868 legal questions posed by Belgian citizens and labeled by experienced jurists with a comprehensive answer rooted in relevant articles from the corpus.

Citation

bibtex

@article{louis2023interpretable,
  author = {Louis, Antoine and van Dijck, Gijs and Spanakis, Gerasimos},
  title = {Interpretable Long-Form Legal Question Answering with Retrieval-Augmented Large Language Models},
  journal = {CoRR},
  volume = {abs/2309.17050},
  year = {2023},
  url = {https://arxiv.org/abs/2309.17050},
  eprinttype = {arXiv},
  eprint = {2309.17050},
}