For the fragmentary management needs of multi-modal, intelligent, refined, knowledgeable and reorganized professional content resources in the knowledge service industry, how to efficiently generate and apply professional knowledge to promote the innovation and development of the real economy has become a common strategic choice and a challenging problem. Therefore, this paper studies the multi-relation classification system and knowledge relationship labeling standard of content resources in eight strategic emerging industries, and develops a consistent multi-relation classification description system for service-oriented professional content resources. Then an extraction model of multiple semantic relations based on domain knowledge is proposed, which can distinguish various entities (e.g., knowledge, information, resources, services, objects), and meet the requirements of automatic deframe, aggregation and intelligent extraction among entities. A multi-level attention mechanism is used to highlight representational details. At the same time, it designs and optimizes meta-attributes between the core source and the target in heterogeneous contexts through the knowledge graph. Unlike previous baseline models, the proposed model structure supports end-to-end learning in the specific domain without explicit dependence on external knowledge. The experiments show that the proposed method can effectively improve the accuracy of the multi-relation classification and the efficiency of professional knowledge service.