The Group Steiner Tree Problem (GSTP) has been popularly used to formulate graph-based tasks on the Web where the number of groups (i.e., g), e.g., representing the number of keywords in knowledge graph search, is assumed to be small. This assumption never holds in emerging tasks such as the summarization of knowledge graphs, where g represents the number of distinct entity description patterns that increases with graph order (i.e., n) and reaches 105 in practice. Existing algorithms become impractical, since they are optimized for large n but not for large g. In this paper, we devise novel approximation algorithms for GSTP that exhibit scalability in relation to both n and g. When g is large, our algorithms outperform existing algorithms that have a comparable approximation ratio by orders of magnitude in running time, showing their unique ability to practically support such challenging Web applications.