Exploration of the cultural attributes of Chinese character sculpture using machine learning technology

Abstract

The article employs machine learning, specifically the CLIP (Contrastive Language-Image Pretraining) model, to analyze Chinese character sculptures’ cultural attributes. It overcomes challenges in multi-dimensional data processing and high digitization costs. The process involves normalizing sculpture images, using FastText for vector representations of Chinese characters, and mapping text to the same embedding space as images for word embedding. The CLIP model, through unsupervised training, minimizes the negative logarithmic likelihood loss between image and text embeddings to establish cultural attribute representations. Key findings include the CLIP model’s improved performance over the M3 model, with a 5.4% higher average AUC. The model demonstrates high efficiency and accuracy, evident in its low RMSE (0.034) and MAE (0.025) and fast analysis time of 182 ms. This approach effectively and accurately analyzes the cultural attributes of Chinese character sculptures, addressing existing research gaps.