Abstract: 【Purpose/significance】Scientific data, as one of the expressions of scientific research achievements, is often hidden in aca⁃ demic papers in the form of informal citations. It provides a new idea for organizing scientific data elements by identifying data refer⁃ ence information from academic papers.【Method/process】To improve the proportion of positive text and enhance the identification ef⁃ fect of data citation sentences, based on the full-text data of papers in the field of bioinformatics, three methods for sampling unbal⁃ anced corpora were used: chapter structure recognition and data augmentation, random undersampling, and feature word filtering. Then, five text classification models were combined to build a data citation recognition process.【Result/conclusion】It is found that identifying data citations from academic papers is an effective link to refine the organization of data elements. Text structure recogni⁃ tion and unbalanced corpus sampling can effectively improve the performance of data reference recognition. Compared with traditional machine learning models, BERT-like deep learning model has better performance in data reference text classification.【Innovation/ limitation】Identifying informal data quotes from academic papers brings a new perspective to the organization of data elements and is an efficient method to collect high-value data elements. However, because the data cited in the paper is irregular and sparse, there is still room for improvement in the classification accuracy rate.