09:30
12:00

This thesis deals with the recommendation of books in Arabic from heterogeneous sources, reconciling personalization, explainability, and temporal adaptation. Using a Goodreads corpus, we set up a traceable pipeline: normalization of reviews (MSA/dialects), deduplication and cleaning, harmonization of metadata (author, title, description), and parsimonious semantic enrichment. We propose D-MARS, which encodes each modality according to its nature (reviews via a language model adapted to Arabic dialects, ratings via an ordinal perceptron, metadata via dedicated representations), then aggregates by hierarchical attention combining intra-view (metadata) and inter-view (source calibration) weightings before modeling interactions. The D-MARS+ extension introduces incremental learning by sessions with controlled replay, limiting forgetting and stabilizing quality over time. Evaluation (strict temporal split, ablations, sensitivity) is based on an interpretable reading of attention weights, showing how information is distributed between opinions, metadata, and notes according to the richness of the inputs.
The contributions focus on (i) a cleaned corpus and a reproducible pipeline adapted to Arabic specificities, (ii) an explainable fusion exploiting the complementarity of views without unnecessary complexity, and (iii) a simple and effective incremental strategy improving the temporal consistency of ranks. Limitations include dialectal noise, uneven metadata completeness, and encoding costs. Future prospects include objectives better aligned with rank, distillation/quantification for inference, backing by a semantic graph (authors/themes), and in vivo evaluations integrating diversity, fairness, and robustness. 

? ailleurs qu'au LaBRI