LEGAL-BERT: The Muppets straight out of Law School

LEGAL-BERT is a family of BERT models for the legal domain, intended to assist legal NLP research, computational law, and legal technology applications. To pre-train the different variations of LEGAL-BERT, we collected 12 GB of diverse English legal text from several fields (e.g., legislation, court cases, contracts) scraped from publicly available resources. Sub-domain variants (CONTRACTS-, EURLEX-, ECHR-) and/or general LEGAL-BERT perform better than using BERT out of the box for domain-specific tasks. This is the sub-domain variant pre-trained on EU legislation.

I. Chalkidis, M. Fergadiotis, P. Malakasiotis, N. Aletras and I. Androutsopoulos. "LEGAL-BERT: The Muppets straight out of Law School". In Findings of Empirical Methods in Natural Language Processing (EMNLP 2020) (Short Papers), to be held online, 2020. ( https://aclanthology.org/2020.findings-emnlp.261 )

Pre-training corpora

The pre-training corpora of LEGAL-BERT include:

• 116,062 documents of EU legislation, publicly available from EURLEX ( http://eur-lex.europa.eu ), the repository of EU Law running under the EU Publication Office.

• 61,826 documents of UK legislation, publicly available from the UK legislation portal ( http://www.legislation.gov.uk ).

• 19,867 cases from the European Court of Justice (ECJ), also available from EURLEX.

• 12,554 cases from HUDOC, the repository of the European Court of Human Rights (ECHR) ( http://hudoc.echr.coe.int/eng ).

• 164,141 cases from various courts across the USA, hosted in the Case Law Access Project portal ( https://case.law ).

• 76,366 US contracts from EDGAR, the database of US Securities and Exchange Commission (SECOM) ( https://www.sec.gov/edgar.shtml ).

Pre-training details

• We trained BERT using the official code provided in Google BERT's GitHub repository ( https://github.com/google-research/bert ).
• We released a model similar to the English BERT-BASE model (12-layer, 768-hidden, 12-heads, 110M parameters).
• We chose to follow the same training set-up: 1 million training steps with batches of 256 sequences of length 512 with an initial learning rate 1e-4.
• We were able to use a single Google Cloud TPU v3-8 provided for free from TensorFlow Research Cloud (TFRC) , while also utilizing GCP research credits . Huge thanks to both Google programs for supporting us!

Models list

* LEGAL-BERT-BASE is the model referred to as LEGAL-BERT-SC in Chalkidis et al. (2020); a model trained from scratch in the legal corpora mentioned below using a newly created vocabulary by a sentence-piece tokenizer trained on the very same corpora.

** As many of you expressed interest in the LEGAL-BERT-FP models (those relying on the original BERT-BASE checkpoint), they have been released in Archive.org ( https://archive.org/details/legal_bert_fp ), as these models are secondary and possibly only interesting for those who aim to dig deeper in the open questions of Chalkidis et al. (2020).

Load Pretrained Model

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("nlpaueb/bert-base-uncased-eurlex")
model = AutoModel.from_pretrained("nlpaueb/bert-base-uncased-eurlex")

Use LEGAL-BERT variants as Language Models

Evaluation on downstream tasks

Consider the experiments in the article "LEGAL-BERT: The Muppets straight out of Law School". Chalkidis et al., 2020, ( https://aclanthology.org/2020.findings-emnlp.261 )

Author - Publication

@inproceedings{chalkidis-etal-2020-legal,
    title = "{LEGAL}-{BERT}: The Muppets straight out of Law School",
    author = "Chalkidis, Ilias  and
      Fergadiotis, Manos  and
      Malakasiotis, Prodromos  and
      Aletras, Nikolaos  and
      Androutsopoulos, Ion",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2020",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    doi = "10.18653/v1/2020.findings-emnlp.261",
    pages = "2898--2904"
}

About Us

AUEB's Natural Language Processing Group develops algorithms, models, and systems that allow computers to process and generate natural language texts.

The group's current research interests include:

• question answering systems for databases, ontologies, document collections, and the Web, especially biomedical question answering,
• natural language generation from databases and ontologies, especially Semantic Web ontologies, text classification, including filtering spam and abusive content,
• information extraction and opinion mining, including legal text analytics and sentiment analysis,
• natural language processing tools for Greek, for example parsers and named-entity recognizers, machine learning in natural language processing, especially deep learning.

The group is part of the Information Processing Laboratory of the Department of Informatics of the Athens University of Economics and Business.

Ilias Chalkidis on behalf of AUEB's Natural Language Processing Group

| Github: @ilias.chalkidis | Twitter: @KiddoThe2B |