This study aimed to explore young children’s vocabulary development using a machine learning technique, variational autoencoder (VAE). The VAE is an unsupervised neural network that maps high-dimensional input data onto a dimension-reduced latent space and then regenerates the data. The complex input features could be visualized in a low-dimensional latent space while maintaining its interpretability. We used parent-reported questionnaire data extracted from a publicly available database, involving American young children (𝑁= 5,520) and applied VAE. The two-dimensional latent space in the adopted model demonstrated that vocabulary development had a quasi-one-dimensional structure shaped by an arc. Its rotation and radial directions represented changes in total vocabulary size and individual differences, respectively. We found that some categories in the questionnaire (e.g., Sounds, Animals) were more likely to develop earlier in the outer path of the arc, whereas others (e.g., Action words, Pronouns) tended to develop predominantly in the inner path of the arc. Furthermore, a simulation case study using longitudinal data suggested that some specific lexical items were crucial in characterizing the universality and diversity of different developmental trajectories in the latent space. Our approach will contribute to quantitatively depicting the development of children’s vocabulary in a more fine-grained and nuanced manner, providing a synergetic bridge between machine learning and developmental science.