Work

Preregistration: Answer quality of Norwegian extractive question answering under fixed context-token budgets, by reduction method and Claude model (confirmatory study)

Year: 2026
Type: article
Abstract: Preregistration of a confirmatory study of how much answer quality is lost in Norwegian Bokmål extractive question answering when the context is cut to a fixed share of its Claude tokens. Three reduction methods (truncation, BM25 sentence selection, LLMLingua-2) at budgets of 70, 50 and 30 % are compared with full context and with no context on the 472 questions of the NorQuAD test split, for Claude Haiku 5.5, Claude Sonnet 5.5 and Claude Opus 5.5. Nine hypotheses (BM25 and LLMLingua-2 at 50 % non-inferior to full context with a 5-point F1 margin, and BM25 against LLMLingua-2, per model) are decided with a paired cluster bootstrap and Bonferroni-adjusted intervals (k = 9); truncation at 50 % is reported descriptively. Part 2 of tokenpris (https://doi.org/10.5281/zenodo.24193112). Files: (1) a .tar.gz git archive of the signed tag prereg-v1 (preregistration, plan, decision records, locked analysis code); (2) PREREG.md; (3) and (4) OpenTimestamps proofs of the tarball and of PREREG.md, stamped after the tag and before this deposit (pending proofs are normal and are upgraded later). The DOI is written to the repository only after publication; the test split is opened only then. Licences: documents CC BY 4.0; code in the archive MIT; third-party material keeps its own licence (NorQuAD CC0-1.0; see LICENSE in the archive).
Language: English
Open Access status: green
OpenAlex ID: W7226016478
FWCI: 1.397
Cites:
Cited by: 1
Related to: 10