Research on intelligent diagnosis method of construction quality problems based on unlabeled short text

During the construction process, the recorded construction quality problems (CQPs) data serves as an objective representation of construction quality status, holding significant value for quality control. However, current subcategory quality issues such as apparent quality and structural bearing capacity typically exhibit characteristics of multi-source origins, short-text format, and unlabeled nature. These characteristics make conventional methods inadequate for achieving automatic fine-grained cause identification, while also resulting in mixed information within quality knowledge bases that proves difficult to utilize. Addressing the challenges in theme recognition for unlabeled short-text quality data - including feature sparsity, polysemy of professional terminology, and high labeling costs - this paper proposes an unsupervised theme mining framework based on an improved K-means algorithm (POS-Word2vec-PCA-L2-Kmeans-WF, PWPL2KW). This framework addresses the challenges of indistinct feature extraction, semantic sparsity, and high dimensionality in CQPs texts—stemming from their domain-specific expertise and short text length—by integrating domain lexicon-guided feature filtering, sentence vector representation based on word vector aggregation, dimensionality reduction and L2 normalization, dynamic clustering optimization, and interpretable thematic mapping. The experimental results demonstrate that compared to other clustering methods, the PWPL2KW model exhibits superior performance in clustering unlabeled short-text data of CQPs. It efficiently generates clear and interpretable thematic categories, enabling automated labeling of quality problems and intelligent diagnosis through identification of high-frequency causative factors. The model facilitates granular thematic analysis of diverse quality defects and contributes to building a systematic quality issue database, thereby supporting sustainable analysis of CQPs.