Context Completeness in LLM-Based Extraction of Dental Bond-Strength Data

Source abstract

A dental adhesive bond-strength value is interpretable only together with the conditions that produced it: the bonded substrate, the surface treatment, the mechanical test, and any artificial aging. Large language models (LLMs) are increasingly applied to mine such numbers from the literature, yet current evaluations score almost exclusively the extracted number. A record storing only a material name and a value is therefore potentially misleading, since one adhesive can span a several-fold range across contexts. We contribute a nine-field context-complete schema with an explicit usability criterion, a two-stage validation layer combining source-grounded re-grounding with deterministic rules, and a quantification of the numeric–context gap across a closed model and two open families under zero-shot, few-shot, and low-rank fine-tuning. On 120 annotated open-access articles (70 training, 50 held out), few-shot prompting recovers numeric fields reliably (F1 ≈ 0.85–0.93), whereas context-completeness F1 is far lower (0.62–0.78) and degrades faster with scale. The validation layer raises the precision of retained records by six to seven points at a recall cost of three to four. Completeness, not numeric accuracy, is the binding constraint for trustworthy adhesive databases.