Búsqueda semántica de sentencias judiciales usando técnicas actuales del NLP y métodos de recuperación de información

Abstract

The inefficiency of the Colombian justice system is a dilatory phenomenon; the Colombian justice system has 11 judicial officials for every 100.000 inhabitants, significant congestion in the number of judicial processes in inventory, and consequently faces acts of corruption[5]. The latest developments in deep learning algorithms and techniques such as transfer learning have filled with new possibilities for NLP (Natural Language Processing), although it is increasingly projects of digitization, especially in the jurisdictional field; this constant torrent of data information that has been created makes it possible to train Machine Learning models with data that allow the incursion of research in AI models for a judicial decision in Colombia. We propose a method to improve the quality of the searches and ranking-based text classification models for case law texts using transfer learning and BERT[1]. The search is based on the conversion of 23.750 law text of the "Corte Constitutional" into multidimensional vectors based on pre-trained models and the search of these vectors using a similarity search using FAIIS [1], Euclidean distance, and KNN for text categorization field [2]. The results of this model were connected to a web interface and were evaluated based on the methodology proposed by Knijnenburg where a pragmatic user-based method is sought to measure the proposed recovery system[33], the questions were posed to a group of lawyers and law students, the results were compared with the traditional search engines and has a higher perceived quality of recommendation according to the results.