A sememe tree constructs a hierarchical structure capable of expressing lexical meaning by combining sememes (the smallest semantic units) and relations. This paper finds that the term exhibits systematicity and that the sememe trees of terms within the same system are similar. Based on this phenomenon, this paper proposes an Adaptive Nearest Neighbor Sememe Tree Generation Model (ANN-STGM). ANN-STGM addresses the insufficient training problem of the decoder on low-frequency samples. This model constructs the term sememe tree in the training set as explicit memory sets. When generating (first) sememes or relations, the decoder retrieves the core concepts, important attributes, or relation types contained in terms with the same or similar conceptual structures to help the model generate sememe tree. Experimental evidence shows that the F1 of sememe-relation-sememe triple reaches 30.74%, the relation F1 reaches 52.45%, the sememe F1 reaches 51.47% and the first sememe F1 reaches 68.83%. These values show an improvement of 2.28%, 0.79%, 1.00% and 0.78%, respectively, compared to the baseline.