Issue #103 - LEGAL-BERT: The Muppets straight out of Law School

Akshai Ramesh 16 Oct 2020
The topic of this blog post is language modeling.


BERT (Bidirectional Encoder Representations from Transformers) is a large-scale pre-trained autoencoding language model that has made a substantial contribution to natural language processing (NLP) and has been studied as a potentially promising way to further improve neural machine translation (NMT). 

Given that BERT is based on a similar approach to neural MT in Transformers, there’s considerable interest and research into how the two can be combined” — Dr. John Tinsley, Co-founder and CEO, Iconic Translation Machines 

But, there has been limited investigation on its adaptation guidelines in specialised domains. In this post, we will discuss the systematic investigation of the available strategies when applying BERT in specialised domains as proposed by Chalkidis et. al., (2020).

About Legal Domain

Legal text (e.g., laws, court pleadings, contracts) has distinct characteristics compared to generic corpora. The recognizable characteristics include specialised vocabulary, particularly formal syntax, semantics based on extensive domain-specific knowledge etc., to the extent that legal language is often classified as a ‘sublanguage’. NMT 103 Legal BERT diagram BERT can be pre-trained on a massive corpus of data, and then fine-tuned to a task for which you have a limited amount of data. This allows BERT to provide a significantly higher performance than models that are only able to leverage a small task-specific dataset. Chalkidis et. al., (2020) systematically explore the following strategies for BERT adaptation in the legal domain :

(a) use out of the box pre-trained BERT (BERT-BASE) (b) further pre-train (FP) BERT-BASE on domain-specific corpora (LEGAL-BERT-FP) (c) pre-train BERT from scratch (SC) on domain specific corpora with a new vocabulary of sub-word units (LEGAL-BERT-SC) (d) pre-train a smaller light-weight version of LEGAL-BERT-SC with fewer parameters (LEGAL-BERT-SMALL)

Key Findings

  1. Further pre-training (FP) or pre-training BERT from scratch (SC) on in-domain data produces better results than BERT-BASE for domain-specific tasks.
  2. This paper presents an expanded grid search compared to the guidelines of Devlin et. al., (2019) for fine-tuning BERT, that has a significant impact on performance.
  3. From the results, LEGAL-BERT-SMALL is comparable to LEGAL-BERT (SC and FP) across most datasets while the training time for small models is 4 times faster and requires fewer hardware resources than the LEGAL-BERT.
For detailed information on the training corpora and experimental setup, you can read the complete paper here .

In summary

This paper explores the different strategies for BERT adaptation in domain-specific NLP tasks and presents LEGAL-BERT, a family of BERT models achieving state-of-art results for the legal domain. The authors also present an expanded grid search for fine-tuning BERT. It is interesting to find that the full capacity of larger and computationally more expensive models is not always a necessity in specialised domains. A light-weight BERT model (LEGAL-BERT-SMALL) with fewer parameters can produce comparable results to the LEGAL-BERT-SC. This piece of information is useful for researchers and practitioners who have limited access to large computational resources.
Akshai Ramesh

Akshai Ramesh

Machine Translation Scientist
Akshai is responsible for developing end-to-end enterprise solutions for real-world MT problems and contributes to the wide spectrum of innovations that serve our clients. Akshai has experience working in the area of Natural Language Processing and Machine Learning and has completed his Master's in Computing from Dublin City University, Ireland. During his Master's, he has worked with ADAPT Research Centre and contributed to the low-resource scenario of Indian languages in Neural Machine Translation. His research interests include low resource scenario, model improvement and domain-adaptation of MT.
All from Akshai Ramesh