| Title |
LLM-Based Interpretation and Refinement of Ambiguous Site Instructions - A Support System for Novice Site Managers in Apartment Finishing Works - |
| Authors |
Ju-Yeon Park ; Min-Jung Kim ; JeeHee Lee |
| DOI |
https://dx.doi.org/10.6106/KJCEM.2026.27.3.055 |
| Keywords |
Ambiguous on-site verbal instructions; Apartment finishing works; LLM |
| Abstract |
Ambiguous on-site verbal instructions in apartment finishing works can lead to misconstruction and rework, especially for novice site managers. This study proposes an LLM-based support framework that classifies such instructions into three ambiguity types (A/B/C), diagnoses missing information, and generates refined written instructions and checklists for field use. Using 54 real on-site utterances collected from apartment finishing works, 600 synthetic instructions were generated with GPT-5.1 to train a TF?IDF and logistic regression classifier. On an external set of 54 real instructions, the model achieved an accuracy of 0.65 and an F1-macro of 0.64, with relatively better performance on Type A and systematic confusion between Types B and C. A completeness analysis showed that average coverage improved from 0.35 in the original instructions to 0.88 in the refined outputs. These results suggest that LLM-based refinement can support novice site managers in interpreting and supplementing ambiguous verbal on-site instructions. |