Named entity recognition for construction documents based on fine-tuning of large language models with low-quality datasets

Abstract

Named Entity Recognition (NER) is a fundamental task for automatically processing and reusing documents. In traditional methods, machine learning has been used relying on costly high-quality datasets. This paper proposed an NER method based on fine-tuning Large Language Models (LLMs) with low-quality datasets for construction documents. Firstly, low-quality datasets were semi-automatically generated from national standards, qualification textbooks, and lexicons, including datasets of generation-type, tagging-type and question-answering type. Then, they were used to fine-tune an LLM for NER of structural elements to obtain optimal parametric fine-tuning conditions. Next, the results of optimally fine-tuned LLM were used to iterate the low-quality dataset to improve the performance. The F1 finally reached 0.756. Similar results were obtained on two other types of named entities, illustrating the generalizability. This paper provided a more effective and efficient method for the construction documents reuse. Future research should explore how to achieve better results by using other methods.