Automated decision classification model for tax appeals commission in Morocco using latent dirichlet allocation

Abstract

This research paper focuses on extracting and classifying information from the Moroccan National Tax Appeals Commission, which is presently nonexistent in the country's legal and tax landscape. This study examines 201 decisions selected from a pool of 562, released between 1999 and 2018, pertaining to corporate tax and involving 550 disputes spanning various corporate tax classifications. The paper aims to propose latent dirichlet allocation (LDA) for topic modeling and compare it with our previous results obtained from the bidirectional encoder representations from transformers (BERT) model. The findings suggest that the rulings can be classified into two primary classifications: those that uphold or reject the tax administration's position. The proposed model shows a good performance, achieving a precision of 9.25% and an accuracy of 9.51%. This highlights the effectiveness of both LDA and BERT models for understanding and classifying topics in tax decision analysis.