Abstract: 【Purpose/significance】The textual structure and semantic content of academic text is reflected by its structure function.
The current researches mainly focus on its multi-disciplinary and multi-level recognition. There is a lack of the comparative research on the relations among model, discipline and level.【Method/process】This paper built a multi-disciplinary experimental corpus, in⁃cluding the disciplines such as traditional Chinese medicine, library and information science, computer, environmental science and phytology. The structure function of academic texts was recognized from sentence, paragraph and chapter content based on the deep learning models of CNN, LSTM, BERT.【Result/conclusion】The experimental results show that BERT performs best in the multi-dis⁃ciplinary and multi-level environment, the deep learning models perform well in the multi-disciplinary and chapter-content level en⁃vironment, the recognition effectiveness of academic text in traditional Chinese medicine was much better than those of other disci⁃plines. In addition, the confusion matrix is used to reveal the misidentification errors, and the causes of the errors associated with the identification of academic texts were analyzed simultaneously.【Innovation/limitation】This study provides first-hand empirical data for
the research of academic text structure function recognition.