Along with the large and increasing number of scientific workflows publicly accessible on Web repositories, the discovery of workflow fragments is significant to promote the reuse or repurposing of best-practices evidenced in legacy workflows, when novel scientific experiments are to be conducted. This paper proposes a novel crossing-workflow fragment discovery mechanism, where an activity knowledge graph is constructed to capture flat invocation relations between activities, and hierarchical parent–child relations specified upon sub-workflows and their corresponding activities. Semantic relevance of activities and sub-workflows is calculated based on their representative topics, where these topics are generated by applying the biterm topic model. Given a requirement specified in terms of a workflow fragment template, individual candidate activities or sub-workflows are discovered when considering their semantic relevance and short-document descriptions. Candidate fragments are constructed through discovering the relations in activity knowledge graph specified upon candidate activities or sub-workflows. These fragments are evaluated by balancing their structural and semantic similarities. Evaluation results show that our approach is accurate in discovering appropriate crossing-workflow fragments in comparison with the state of art’s techniques.