Case study · Applied AI · Information retrieval

JurisPortail — semantic search for Moroccan law

A retrieval system and mobile app that let people search French-language Moroccan legal articles by meaning rather than exact keywords.

Context: M.S. Software Engineering thesis, Al Akhawayn University in Ifrane, Summer 2023. Supervised by Dr. Asmae Mourhir.

Problem

Moroccan legal texts are published online as PDF documents, so search is limited to keyword matching.
Keyword matching misses the context and phrasing of legal language, which makes finding the right article slow for professionals, researchers and citizens.

Sole author of the thesis: data extraction, pre-processing, retrieval models, evaluation and the mobile application design.

No large labelled set of query–article pairs, which rules out supervised fine-tuning.
Source PDFs vary in structure; documents with tables inside articles were excluded to keep extraction reliable.
French legal vocabulary, where exact terms and their context both matter.

Semi-automated extraction of 4,048 articles from six legal codes, including the Commercial Code, Labour Code, Penal Code and Code of Obligations and Contracts.
French pre-processing pipeline (tokenisation, punctuation and stop-word removal, lemmatisation, accent cleaning) that roughly halved the token count.
Three search approaches: BM25 and two variants (BM25L, BM25+), TF-IDF with Latent Semantic Indexing, and a hybrid combining BM25+ with TF-IDF/LSI.
Similar-article recommendations from Sentence-CamemBERT embeddings indexed with ANNOY for approximate nearest-neighbour lookup.
A React Native application on a Laravel backend for search, similar articles and relevance marking.

Lexical and latent-semantic models over fine-tuned transformers for search, because they need no labelled training pairs.
Transformer embeddings kept for recommendations, where similarity between whole articles matters more than exact terms.
Evaluation combined Recall@8 and mean average precision with review by a legal domain expert and relevance marking by ten test users.

On the thesis evaluation set, the hybrid ranker reached the highest Recall@8 (0.92) and BM25+ the highest mean average precision (0.70).
The domain expert approved BM25, BM25+, TF-IDF/LSI and the recommendation model, and rejected BM25L.
These are offline research results, not production usage figures.