Interdisciplinary research has gradually become one of the main driving forces to promote original innovation of scientific research, and how to measure the interdisciplinarity of science project is becoming an important topic in the science foundation managements. Existing researches mainly using methods, such as academic degree or institutional discipline or discipline category mapping of journals, to measure the interdisciplinarity. This study proposes an approach to mine and capture the different or complementary characteristics of interdisciplinarity of projects by combining text mining and machine learning methods. First, we construct the classification system and extract a raw paper and its discipline matrix according to the discipline category of journals where the references were published in. Second, we cut the matrix to summarise the distribution of key disciplines in each paper and extract the text features in the abstract and title to form a training set. Finally, we compare and analyse the classification effects of Naive Bayesian Model, Support Vector Machine and Bidirectional Encoder Representations from Transformers (BERT) model. Then, the model evaluation indicators show that the best classification effect was achieved by the BERT model. Therefore, the deep pre-trained linguistic model BERT is chosen to predict the discipline distribution of each project. In addition, the different aspects of interdisciplinarity are measured using network coherence and discipline diversity indicators. Besides, experts are invited to evaluate and interpret the results. This proposed approach could be applied to deeply understand the discipline integration from a new perspective.