Knowledge Graph and LoRA Fine-Tuning Large Language Model Fusion Method for Bridge Condition Evaluation

Le résumé fourni par la source

The evaluation of the technical condition of bridges is crucial in ensuring the safety and prolonging the service life of bridges. By conducting regular technical condition evaluation of bridges, potential structural problems can be detected in advance, avoiding sudden accidents and ensuring the safety of transportation. In the early days, bridge technical condition assessment mainly relied on manual work, relying on the experience and judgment of inspectors. With the emergence of various types of bridge assessment specifications, expert systems were introduced into bridge technical condition evaluation, but since expert systems tend to reason based on rules and logic, and must select the options provided by the system or enter the problem in a fixed format, expert systems may not perform well when dealing with complex and ambiguous inputs. In this paper, a large language modeling(LLM) approach combined with knowledge graphs(KG) is proposed, which can address the problems of expert systems' inability to deal with uncertainty and ambiguity, the insufficient semantic parsing ability of traditional knowledge graphs, and the insufficient knowledge learning depth of current general-purpose large language models in vertical domains. Firstly, the knowledge map of bridge technical condition evaluation is constructed through the relevant evaluation standards, and secondly, adopts the LoRA method to fine-tune the large language model to improve the model's ability to extract keywords in the field of bridge technical condition evaluation. Finally, a question-answering test is conducted to determine whether the answers returned by the developed system match the user's intention. The question-answering test shows that it outperforms the expert system for complex and ambiguous questions, and that using a fine-tuned large language model combined with a knowledge graph is more accurate than using the original large language model directly or using the fine-tuned large language model directly.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.