【Purpose/significance】 In order to clarify the semantic role of scientific entities in academic texts, so as to establish the

knowledge structure of terminologies in specific fields, this paper proposes a classification method of scientific entities that takes the relationship between entities as a feature engineering. From the perspective of semantic attribute of academic research, the keywords of academic texts are divided into "research field", "research problem", "research method", "research tool" and "other".【Method/pro⁃cess】 With the method of dependency syntax analysis, for sentences with two or more entities in the academic text, the shortest depen⁃dency path between entities is mined, the predicate components on the shortest dependency path are extracted as the relationship be⁃tween entities, the relationship between words is taken as the feature, and a 2D matrix is constructed as the input of the convolutional neural network to complete the classification of entities.【Result/conclusion】 The model was verified in the academic literature in the field of "artificial intelligence" downloaded from Web of Science, with an accuracy rate of 89.38%, a recall rate of 92.46%, and an F1value of 0.9089.【Innovation/limitation】 The matrix formed by the relationship between entities is a sparse matrix, which will ad⁃versely affect the calculation speed in the calculation process.