Abstract: When designers utilize text prompts to retrieve resources in a design repository, the keyword-matching approach usually fails to capture the real design semantics in the text and recommends irrelevant design resources. To address this problem, this paper proposes a design material recommendation system exploiting word vectors and self-attention mechanisms. First, the model segments the text into tokens and takes the token word vectors from the pre-trained word vector model as input. Then, we exploit the self-attention mechanism to learn the weight of tokens and the mapping function between tokens and design semantics. Finally, we retrieve and sort the design resources by computing the vector similarity between the model output and annotated resource data. The system can retrieve other design resources with different semantic tags as well by unifying the similarity between the different semantics into weight vectors and converting the data dimension to be consistent with the model output. With 1 300 e-commerce design cases and 200 fonts, the efficiency of user-assessed recommendation results is 55% and 57.3%, respectively, which are above average for similar works and validate the feasibility and effectiveness of the proposed system.