Gas sensing materials based on transition metal perovskite oxides (TMPOs) have garnered extensive attention across various fields such as air quality control, environmental monitoring, healthcare, and national defense security. To overcome challenges encountered in traditional research, a deep learning framework combining natural language processing technology (Word2Vec) and crystal graph convolutional neural network (CGCNN) was adopted in this study, proposing a predictive method that incorporates a comprehensive data set consisting of 1.2 million literature abstracts and 110,000 crystal structure data entries. This method assessed the optimal combination of zinc-cobalt bimetallic ions complexed with ligands as precursors for perovskite oxides. The regulatory function of ligand concentration on the p-n transformation of zinc-cobalt oxide sensing performance was identified, and optimization strategies were provided. The Zn(II)/Co(III)/1-methyl-1