LeXFiles and LegalLAMA: Facilitating English Multinational Legal Language Model Development ⚖️ 👩⚖️ 🗺️
@inproceedings{chalkidis-etal-2023-lexfiles,
title = "{L}e{XF}iles and {L}egal{LAMA}: Facilitating {E}nglish Multinational Legal Language Model Development",
author = "Chalkidis, Ilias and
Garneau, Nicolas and
Goanta, Catalina and
Katz, Daniel and
S{\o}gaard, Anders",
booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = jul,
year = "2023",
address = "Toronto, Canada",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.acl-long.865",
pages = "15513--15535"
}
The LeXFiles is a new diverse English multinational legal corpus that we created including 11 distinct sub-corpora that cover legislation and case law from 6 primarily English-speaking legal systems (EU, CoE, Canada, US, UK, India). The corpus contains approx. 19 billion tokens. In comparison, the "Pile of Law" corpus released by Hendersons et al. (2022) comprises 32 billion in total, where the majority (26/30) of sub-corpora come from the United States of America (USA), hence the corpus as a whole is biased towards the US legal system in general, and the federal or state jurisdiction in particular, to a significant extent.
from datasets import load_dataset
dataset = load_dataset('lexlms/lex_files', name='eu-legislation')| Corpus | Corpus alias | Documents | Tokens | Pct. | Sampl. (a=0.5) | Sampl. (a=0.2) |
|---|---|---|---|---|---|---|
| 🇪🇺 EU Legislation | eu-legislation |
93.7K | 233.7M | 1.2% | 5.0% | 8.0% |
| 🇪🇺 EU Court Decisions | eu-court-cases |
29.8K | 178.5M | 0.9% | 4.3% | 7.6% |
| 🇪🇺 ECtHR Decisions | ecthr-cases |
12.5K | 78.5M | 0.4% | 2.9% | 6.5% |
| 🇬🇧 UK Legislation | uk-legislation |
52.5K | 143.6M | 0.7% | 3.9% | 7.3% |
| 🇬🇧 UK Court Decisions | uk-court-cases |
47K | 368.4M | 1.9% | 6.2% | 8.8% |
| 🇮🇳 Indian Court Decisions | indian-court-cases |
34.8K | 111.6M | 0.6% | 3.4% | 6.9% |
| 🇨🇦 Canadian Legislation | canada-legislation |
6K | 33.5M | 0.2% | 1.9% | 5.5% |
| 🇨🇦 Canadian Court Decisions | canadian_decisions |
11.3K | 33.1M | 0.2% | 1.8% | 5.4% |
| 🇺🇸 U.S. Court Decisions [1] | us-court-cases |
4.6M | 11.4B | 59.2% | 34.7% | 17.5% |
| 🇺🇸 U.S. Legislation | us-legislation |
518 | 1.4B | 7.4% | 12.3% | 11.5% |
| 🇺🇸 U.S. Contracts | us-contracts |
622K | 5.3B | 27.3% | 23.6% | 15.0% |
| Total | lexlms/lex_files |
5.8M | 18.8B | 100% | 100% | 100% |
[1] We consider only U.S. Court Decisions from 1965 onwards (cf. post Civil Rights Act), as a hard threshold for cases relying on severely out-dated and in many cases harmful law standards. The rest of the corpora include more recent documents.
[2] Sampling (Sampl.) ratios are computed following the exponential sampling introduced by Lample et al. (2019).
Additional corpora not considered for pre-training, since they do not represent factual legal knowledge.
| Corpus | Corpus alias | Documents | Tokens |
|---|---|---|---|
| 🗺️ Legal web pages from C4 | legal-c4 |
284K | 340M |
LegalLAMA is a diverse probing benchmark suite comprising 8 sub-tasks that aims to assess the acquaintance of legal knowledge that PLMs acquired in pre-training.
from datasets import load_dataset
dataset = load_dataset('lexlms/legal_lama', name='contract_sections')| Corpus | Corpus alias | Examples | Avg. Tokens | Labels |
|---|---|---|---|---|
| 🇨🇦 Criminal Code Sections (Canada) | canadian_sections |
321 | 72 | 144 |
| 🇪🇺 Legal Terminology (EU) | cjeu_term |
2,127 | 164 | 23 |
| 🇺🇸 Contractual Section Titles (US) | contract_sections |
1,527 | 85 | 20 |
| 🇺🇸 Contract Types (US) | contract_types |
1,089 | 150 | 15 |
| 🇪🇺 ECHR Articles (CoE) | ecthr_articles |
5,072 | 69 | 13 |
| 🇪🇺 Legal Terminology (CoE) | ecthr_terms |
6,803 | 97 | 250 |
| 🇺🇸 Crime Charges (US) | us_crimes |
4,518 | 118 | 59 |
| 🇺🇸 Legal Terminology (US) | us_terms |
5,829 | 308 | 92 |
We considered the following PLMs in our experimentations;
- RoBERTa
- LegalBERT
- CaseLawBERT
- PoL-BERT
- LexLM
To evaluate these PLMs on LegalLAMA, run the following script:
sh scripts_lama/run_lama.sh
We release 2 new legal-oriented PLMs, dubbed LexLMs, warm-started from the RoBERTa models, and further pre-trained on the "LeXFiles" corpuss for 1M additional steps.
| Model Name | Alias | Layers | Hidden Units | Attention Heads | Parameters |
|---|---|---|---|---|---|
| Lex-LM (Base) | lexlms/legal-roberta-base |
12 | 768 | 12 | 123.9M |
| Lex-LM (Large) | lexlms/legal-roberta-large |
24 | 1024 | 16 | 354.0M |
You can load any model with the standard HF AutoModel code.
from transformers import AutoModel, AutoTokenizer
model = AutoModel.from_pretrained("lexlms/legal-roberta-base")
tokenizer = AutoTokenizer.from_pretrained("lexlms/legal-roberta-base")Initially, we train a custom BPE tokenizer:
python train_tokenizer.py
Then, we train an LM from scratch with Masked Language Modelling (MLM):
sh scripts_mlm/train_lm.sh
Model type (architecture, size) and other training specifications can be modified in the script:
MODEL_PATH='lexlm-large'
MODEL_MAX_LENGTH=512
TOTAL_STEPS=1000000
BATCH_SIZE=512
For example to fine-tune a PLM on EURLEX, you have to run:
sh lex-glue/scripts/run_eurlex.sh
PLM and other training specifications can be modified in the script:
MODEL_PATH='lexlms/legal-roberta-large'