텍스트 임베딩 기반 건설 내역서 공종코드 자동 매칭 성능 분석

초록

Construction bills of quantities (BoQs) continuously incorporate new knowledge such as emerging construction methods and materials;however, systematic updating and management of these items remain limited. As a result, BoQ items are often recorded as unstructured text,creating difficulties in collecting and standardizing construction data. To address this issue, this study explores text embedding techniques thatcapture the semantic meaning of unstructured BoQ item descriptions. An automatic construction work code matching framework is developedto identify correspondences between BoQ item texts and standard construction work codes. Using the same matching algorithm environment,different text embedding methods are compared to evaluate performance. Four embedding methods are assessed under identical experimentalconditions: TF-IDF, Word2Vec, FastText, and Sentence-BERT. The experimental dataset consists of item descriptions collected from realconstruction BoQs, and performance is evaluated in terms of accuracy and runtime. The results show that TF-IDF achieves the highestmatching accuracy across all code levels, while Word2Vec demonstrates stable performance with faster execution time. In contrast, FastTextand Sentence-BERT show lower accuracy overall. The proposed framework and analysis provide insights for improving the standardization ofBoQ data and enhancing the utilization of construction work codes.

키워드