Jurisprudence search based on facts similarity using NLP and ML techniques.

Part of a lawyers job is to understand the clients problem, to textually describe its facts and to apply the sources of law. To support a new legal case, a handful of past judgments on similar cases are typically employed by the lawyers, but finding them is currently a time-consuming procedure. To a...

ver descrição completa

Detalhes bibliográficos
Autor: Ruiz, Rodrigo Amorim
Formato: tesis de maestría
Estado:Versión publicada
Fecha de publicación:2021
País:Brasil
Recursos:Universidade de São Paulo (USP)
Repositorio:Biblioteca Digital de Teses e Dissertações da USP
Idioma:inglés
OAI Identifier:oai:teses.usp.br:tde-14022022-122906
Acesso em linha:https://www.teses.usp.br/teses/disponiveis/3/3141/tde-14022022-122906/
Access Level:acceso abierto
Palavra-chave:Aprendizado computacional
Artificial intelligence
Bag-of words
Cosine similarity
Deep learning
FastText
GloVe
Inteligência artificial
Jurisprudence
Jurisprudência
Linguagem Natural
Logistic regression
Long short-term memory
Machine learning
Multilayer perceptron
Naive bayes
Natural language processing
Neural network
Redes neurais
Siamese neural network
TF-IDF
Transfer learning
Transformer
Word embedding
Word2Vec
Descrição
Resumo:Part of a lawyers job is to understand the clients problem, to textually describe its facts and to apply the sources of law. To support a new legal case, a handful of past judgments on similar cases are typically employed by the lawyers, but finding them is currently a time-consuming procedure. To address this problem, we built a machine learning model responsible for classifying similarity between two facts descriptions. This similarity metric measures how much (from 0 to 1) a legal decision may be used to support another. We trained different model architectures combining several state-of-the-art natural language processing and machine learning techniques using an extracted dataset from the Superior Court of Justice website of past judgments, which enabled the dynamic construction of facts description pairs when one case cites another as a reference. The final best architecture employs TF-IDF for encoding and reducing dimensionality of our input documents, a Siamese Neural Network (SNN) with a Multilayer Perceptron (MLP) for feature extraction and a final layer, another MLP, responsible for concatenating and classifying the features into the similarity metric, achieving 85.98% accuracy, 83.89% precision and 89.06% recall. Such a model would enable the lawyer to compare a case facts description with several judgments of the jurisprudence and start their search on the most similar ones.