关键词: 建筑信息模型, 深化设计, 大语言模型, 规则转译, 思维链, 自动合规性审查

Abstract:

Drawing review records generated during detailed design are often colloquial. Their semantic ambiguity, inconsistent engineering terminology, and strong dependence on drawing context make engineering experience difficult to reuse. The objective was to develop a structured translation method that converts unstructured engineering text into computable drawing-review rules, reduces the cost of manual rule authoring, and improves the efficiency of accumulating reusable engineering experience. Building Information Modeling (BIM) checking requirements were used as the operational constraint. Each candidate rule was represented as a quadruple comprising a review subject, a review attribute, a logical operator, and a threshold or object value. A domain semantic dictionary was constructed to align colloquial and nonstandard construction terms with Industry Foundation Classes (IFC) entities, and attributes. The implementation workflow contained six connected stages. First, raw records were cleaned, irrelevant meeting information was removed, separated questions and replies were merged, and compound entries were divided into minimal issue units. Second, BIM checkability was determined for each unit or requiring human judgment according to the availability of explicit objects, geometric relations, measurable attributes, spatial constraints, or existence requirements. Third, BIM-checkable units were decomposed into the four semantic slots while preserving thresholds, units, reference objects, and uncertainty markers. Fourth, extracted entities and properties were disambiguated using the semantic dictionary and IFC terminology; ambiguous matches were retained as candidates for review. Fifth, the logical operator was matched to available computational functions for attribute, spatial, and existence constraints. Sixth, the completed slots were instantiated as Structured Natural Language (SNL) statements with prerequisite conditions and executable checking clauses. A Large Language Model (LLM) prompt architecture incorporating chain-of-thought reasoning was used for staged reasoning and self-checking of results. Two hundred samples were selected from the company’s historical drawing review records. Five experienced engineers created the expert reference answers. A separate panel of ten senior engineers independently scored anonymized outputs for BIM checkability, component naming, component attributes, logical operators, and SNL conformity. A general-model group using a simple prompt was compared with the proposed chain-of-thought-enhanced group under the same base-model environment. The chain-of-thought-enhanced group obtained score rates of 98.0% for checkability, 91.7% for component names, 91.3% for component attributes, 92.0% for logical operators, and 94.3% for SNL conformity. Compared with the general-model group, the corresponding gains were 27.0, 34.4, 38.6, 33.6, and 43.9 percentage points. The largest improvement occurred in SNL conformity, indicating that the coordinated use of the semantic dictionary, rule quadruple, computational-function constraints, and staged prompts substantially reduced uncontrolled expression. Score rates for simple, routine, and complex samples were 97.3%, 93.9%, and 86.6%, respectively. The improvement was more pronounced for complex samples. Processing all 200 samples required approximately 20 minutes for the general-model group and 1.5 hours for the enhanced group. Processing by the enhanced group remained markedly faster than expert reference construction, which required approximately 50 person-hours, or more than 10 hours with five experts working in parallel. Coordinating controlled semantic slots, a domain semantic dictionary, SNL templates, and staged reasoning can improve the consistency, conformity, and practical efficiency of translating drawing review records into reusable BIM checking rules. The workflow provides an implementable route for converting dispersed engineering experience into structured digital assets while retaining uncertainty and manual-review flags.

Key words: building information modeling, detailed design, large language model, rule translation, chain-of-thought, automated compliance checking