Subject Title: Artificial intelligence and legal decisions: comparison
of the performance of artificial intelligence techniques in order to
understand and anticipate judges' reasoning based on evidence
(evidential reasoning)

Supervisors:
Caroline BAZZOLI, Jean Kuntzmann Laboratory,
caroline.bazzoli@univ-grenoble-alpes.fr
Jean-Pierre CHEVALLET, Laboratoire d'Informatique de Grenoble,
jean-pierre.chevallet@univ-grenoble-alpes.fr
Duration: 5-6 months
Keywords : NLP; Machine learning ;Deep learning ; Text mining ;

Introduction

The objective of the project is to test and compare the performance of
artificial intelligence systems derived from two different
methodologies for predicting court decisions and identifying the
criteria (facts and evidence) that influence the reasoning of judges.
The first method consists in creating a mathematical model of judges'
decisions based on a learning process that relies on the detailed
annotation of several hundred judgments. The second method, is based on
deep neural networks trained on the same corpus of judgments annotated
with decision labels.

This project is grounded on a close cooperation between computer
scientists, jurists and statisticians. It involves both academic and
industrial partners. It is the first academic research in France whose
objective is to measure the performance of AI in the field of legal
sciences. It seeks to find out to what extent state-of-the-art
artificial intelligence models are likely to help understanding and
anticipating judges' decisions.

Internship subject

The objective of the internship is to compare the performance of two
scientific methods for analyzing the decision-making process of judges:
on the one hand, a mathematical modeling method based on human
annotation work, which requires both significant human resources and
advanced legal skills; on the other hand, the use of recent advanced in
natural language processing models based on word embeddings, i.e.
representing words as vectors of numbers, based on their context (e.g.
BERT Bidirectional Encoder Representations from Transformers) to build
efficient text classification on legal texts.

In this project, the student must setup an experiment that will test
the capacity of a Neural Network (NN) to be trained to learn the judges
decision. The NN will have as input vectors embedding transformations
using FlauBERT developed in LIG as the sources text are in French.

The internship missions are:

Bibliographic study on the predictive efficiency of supervised
classification applicable to our context by presenting their advantages
and disadvantages.

Programming of the Natural Language pre-processing phase of the text of
court decisions for the use of FlauBERT.

Test if a NN with text embeddings input, can learn and predict judges
decisions only from judgments corpus

Analyze and understand the usefulness on human annotation in the
efficiency of the learning processing

Find a way to automatically highlight in the original text, the
passages that has been evaluated as strongly influential for the NN
decision making.

Candidate profile

Master 2 in Computer science of Applied Mathématics

Knowledge of programming tools in the machine learning domain : R,
Python, pyTorch, etc.

Theoretical knowledge in multivariate statistics, logistic regression
and data analysis (classification, clustering and neural networks,
DeepLearning)

Scientific English. French reading could be important as the text
collection is in French.

Practical informations

Location : Laboratoire Informatique de Grenoble, Bâtiment IMAG,
Université Grenoble Alpes, 700 avenue Centrale, 38401 Domaine
Universitaire de Saint-Martin-d'Hères.

Usual internship gratuity (around 540,00¤ per month)

Duration : 5 or 6 months

Supervisors :
Jean-Pierre CHEVALLET (LIG), Caroline BAZZOLI (Laboratoire Jean
Kuntzmann), Etienne VERGES (Centre de recherche Juridique)