bert-large-legal-sentence-classification

bert-large-legal-sentence-classification

Model Description

bert-large-legal-sentence-classification is a finetuned bert-large-cased model that is ready to use for legal sentence classification into rhetorical roles. It was fine-tuned using a repository of analyzed disability-claim decisions issued by the Board of Veterans' Appeals ("BVA") of the U.S. Department of Veterans Affairs.

How to use

With HuggingFace pipeline

from transformers import AutoTokenizer, BertForSequenceClassification
from transformers import pipeline

tokenizer = AutoTokenizer.from_pretrained("samkas125/bert-large-legal-sentence-classification")
model = BertForSequenceClassification.from_pretrained('samkas125/bert-large-legal-sentence-classification')

nlp = pipeline("text-classification", model=model, tokenizer=tokenizer)

example = "The Veteran did not have a psychiatric disorder in service that was unrelated to the use of drugs." 

results = nlp(example)
print(results)

Without HuggingFace pipeline

from transformers import AutoTokenizer, BertForSequenceClassification
from torch import softmax, argmax


tokenizer = AutoTokenizer.from_pretrained("samkas125/bert-large-legal-sentence-classification")
model = BertForSequenceClassification.from_pretrained('samkas125/bert-large-legal-sentence-classification')

sentence = "The Veteran did not have a psychiatric disorder in service that was unrelated to the use of drugs."
encoded_input = tokenizer(sentence, return_tensors='pt')
output = model(**encoded_input)
logits = output.logits
probs = softmax(logits, dim=1)
predicted_class = argmax(probs, dim=1).item()
print(predicted_class)

Check the config.json file to map the index of a class to its name.

Limitations and bias

This model is limited by the size of its training set of 6153 sentences. Additionally, it was only trained on disability claim decisions issued by the U.S. Department of Veterans Affairs. It may not generalize well to non-legal sentences, or sentences from other types of cases within the legal domain.

Training data

A repository of analyzed disability-claim decisions issued by the Board of Veterans' Appeals ("BVA") of the U.S. Department of Veterans Affairs was used to finetune the bert-large-cased model for sequence classification. It consisted of 6153 sentences from 50 cases. The sentences were classified by the Research Laboratory for Law, Logic and Technology (LLT Lab) at the Maurice A. Deane School of Law at Hofstra University, in New York. The analysis consists of classifying sentences in those decisions according to their rhetorical role (as defined within the same repository).