Information retrieval for question-answering remains underexplored in specialised domains and under-resourced language variants such as European Portuguese. Existing benchmarks largely target general-domain English data and document-centric retrieval, failing to capture the semantic alignment required for linguistic consultation tasks over curated question–answer (QA) pairs. We address this gap by introducing a controlled evaluation framework for retrieval over the "Ciberdúvidas da Língua Portuguesa" corpus, comprising 29,145 expert-validated QA entries. Our approach systematically analyses the interaction between indexing strategies, encoder models, and retrieval paradigms, while modelling real-world query variability through a paraphrase-based benchmark of 600 queries across five user profiles, manually validated by a professional linguist to ensure semantic fidelity. Experiments show that dense retrieval with an IR-optimised monolingual encoder significantly outperforms both sparse (BM25) and hybrid methods, achieving a Mean Reciprocal Rank (MRR) of 0.93. Notably, hybrid retrieval underperforms due to lexical mismatch interference, challenging prevailing assumptions in the literature. Our contributions include a novel benchmark framework for linguistic QA retrieval, empirical evidence supporting monolingual IR-specialised models, and insights into retrieval robustness under paraphrastic variation, enabling improved QA systems for specialised and low-resource environments.

--

Information Retrieval,Question-Answering,European Portuguese,Sentence Encoders,Natural Language Processing

Best paper award