When Normalization Hurts: Quantifying Morphological Asymmetry Effects in Arabic Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) pipelines are increasingly used for Arabic question answering, yet applying morphological normalization to only one side of the query–passage pair has received little systematic study. We examine this normalization asymmetry on Mr. TyDi Arabic (3115 development queries; 2,106,586 passages) with a 2×2 design that applies Farasa segmentation to queries, passages, both, or neither under BM25, alongside dense and hybrid baselines and a fixed Qwen2.5-7B-Instruct reader. Asymmetric normalization reduces nDCG@10 by 10–39%; segmenting only the query costs 3.9× more than segmenting only the passages, and this retrieval pattern replicates on the held-out test split. An analyzer ablation supports the interpretation that segmentation creates the penalty, whereas the directional imbalance arises from its interaction with the Lucene ArabicAnalyzer. In the more damaging direction, the recall deficit persists through k=100, and a score-level diagnostic attributes collapses mainly to displacement by detached affix tokens. Generation metrics also decline significantly; the decline is much smaller than that of retrieval against a passage-prefix reference but roughly two to three times larger when answers are scored against TyDi QA minimal answers. An English-trained dense encoder performs no better than our BM25 configuration, so the large mE5 gain should not be generalized. Within this setting, consistency between query and passage normalization matters more than the choice of symmetric scheme, as the two symmetric schemes do not differ significantly.