This paper presents a bidirectional neural machine translation system between French and Fon, a major language spoken in Benin and belonging to the Gbe family. Unlike existing generic translation tools, our system is specifically designed to address the linguistic, cultural, and computational challenges of low-resource African languages. The proposed architecture builds upon Facebook AI’s NLLB-200 model, which we adapt through cross-linguistic transfer from Ewe to Fon, taking advantage of structural similarities within the Gbe languages. To further enhance performance, we employ the T-Projection method for more reliable annotation and integrate a continuous improvement framework driven by real-time user feedback. Evaluation was conducted on a 73MB French–Fon parallel corpus. The results indicate a 25% improvement in the translation of idiomatic expressions, as well as a 40% reduction in inference time through knowledge distillation. Beyond linguistic accuracy, the system introduces a cultural evaluation module, enabling context-aware translation in domains such as Vodoun practices, royal discourse, and traditional expressions. This ensures not only linguistic fidelity but also cultural adequacy. The system achieves a BLEU score of 0.94 and a user satisfaction rate of 0.93, confirming its effectiveness and relevance for real-world use in Beninese contexts.

The analysis and exploration of traces of mobility produced by various mobile objects is a research topic that has attracted great interest in recent years. In this article, we present a classification (or clustering) approach adapted to the data of people moving under the constraints of a road network. A similarity measure is proposed to compare the trajectories studied with each other, taking into account the displacement constraints imposed by the network. This measurement is exploited to build a graph translating the different similarity relations maintained by the trajectories between them. We partition this graph using an algorithm using the notion of modularity as a quality criterion in order to discover communities (or clusters) of trajectories which are strongly linked and which exhibit a common behavior. We have implemented and tested the proposed approach on several synthetic datasets through which we show its operation.