Conf-UNet: A Model for Speculation on Unknown Oracle Bone Characters

Oracle Bone Characters (OBC) are the oldest developed pictographs in China, having been created over 3000 years ago. As a result of their antiquity and the scarcity of relevant historical sources, identifying and interpreting OBC has been an ongoing challenge for scholars of oracle bone characters. Large Seal Script evolved from OBC, and retains many of its features, making the linking of these two scripts an urgent task. The traditional method of textual study and interpretation can be time-consuming, requires a high degree of professionalism, and demands significant material and human resources. To address these issues, we propose the Conf-UNet model, a deep learning approach that incorporates a U-net combined with a multi-head self-attention mechanism. This model was used to link Large Seal Script with unidentified oracle bone characters to reduce the workload and provide linguists with a reliable method to speculate the sealed characters that correlate with unidentified OBC. The proposed model also enables linguists to use deep learning to research the evolution of pictographs. Experiments on the HWOBC-A dataset demonstrate that our model outperforms other models on this task of identifying oracle bone characters.