Lexicalization Is All You Need: Examining the Impact of Lexical Knowledge in a Compositional QALD System

David Maria Schmidt, Mohammad Fazleh Elahi, Philipp Cimiano

Type: other
Venue: arXiv / Bielefeld University

Summary

Compositional QALD over DBpedia 2016-10: multi-parser dependency trees, Lemon lexicon matching, DUDES bottom-up composition, then a flan-t5-small pairwise SPARQL selector. QALD-9 English: multi-model selector micro F1 0.72 (P 0.77 / R 0.67) vs prior SOTA GenRL 0.53. GPT-4 with gold lexical entries in-prompt tops out ~0.35 micro F1; without lexicon lower. Manual lexicon 599 entries (~16h). Artifact https://doi.org/10.5281/zenodo.12610054. EKAW 2024.

Keywords

qald · lexicalization · dudes · sparql · dbpedia · lemon · compositionality · ekaw · bielefeld

Topics

question answering over linked data, lexicalization, compositionality

Research notes

Primary: arxiv abs (cs.AI; also cs.CL, cs.IR). License CC BY 4.0 on HTML at check. Comment: EKAW 2024; LNCS 15370; DOI 10.1007/978-3-031-77792-9_7. Semantic Computing Group, CITEC, Bielefeld University. Correspondence {daschmidt,melahi,cimiano}@techfak.uni-bielefeld.de. Software artifact https://doi.org/10.5281/zenodo.12610054. HF has no paper page (API 404). Discord posted abs. Uses public QALD-9 / DBpedia 2016-10 plus a 599-entry manual Lemon lexicon, not a substantial new hosted corpus, so no datasets_local row.