Abstract: To extract information and communications technology (ICT) supply relationships from public text, it is necessary to achieve the matching of ICT bidding projects and supplier products accurately. However, the information on the supplier's official website related to the bidding project is distributed in the product name (entity) and product introduction (document), being difficult to establish its mapping association directly. Therefore, it is necessary to analyze the semantic of the text and to establish an accurate matching. Due to multi types and complex forms of ICT projects and product data, it is difficult to extract their matching semantic information. In addition, the existing text matching models can not differentiate the indiscrimination encoding in different levels of text, causing noise introduction and poor matching performance. To solve the problems, an entity-document level joint matching model was proposed. The model was arranged to use the TextCNN encoder to extract the local semantic information from the product name and the bidding project to eliminate the interference of irrelevant information in the product introduction. Then the CNN-SA encoder was used to extract the local and global information of the product introduction. Finally, the entity-level and document-level matching information was combined to make decisions. Experiment results show that the accuracy of matching mapping between bidding projects and supplier products can reach up to 92%. The method can be provided directly to practice application.