Closing the acronym gap in educational retrieval with ontology-guided augmentation

Retrieval-augmented generation (RAG) combines information retrieval with language generation to ground model outputs in external evidence, improving factuality, traceability, and freshness, features essential for educational assistants. However, in acronym-dense course repositories, locally defined short forms often sit far from their expansions in embedding space, yielding systematic retrieval failures even when relevant passages exist. Prior acronym work mainly targets identification and disambiguation as standalone NLP tasks, typically evaluated on scientific or biomedical corpora, and rarely quantifies downstream retrieval impact in RAG pipelines. We close this “acronym gap” with an ontology-guided augmentation strategy that preserves authoring workflows. We build a corpus-scale acronym ontology via high-recall candidate mining, LLM-assisted definition extraction with evidence spans, global corpus-level disambiguation, and light expert curation. We then apply Query-Time Augmentation (QTA) to append likely short/long forms to user queries and Ingestion-Time Enrichment (ITE) to inject complementary forms during indexing; optional embedding fine-tuning further aligns acronym semantics. On a ∼10,000-document Catalan university corpus from the Barcelona School of Informatics, we evaluate with nDCG@10, Recall@10/100, and Success@100 using paired-bootstrap significance tests over queries. QTA delivers a 45.5% average overall improvement across models, more than doubles Recall@10 for acronym-to-expansion queries, and QTA + ITE yields gains up to 64.5% over baseline.