We describe a research project on resolving semantic differences for multi-agent systems (MAS) in electronic commerce. The approach can be characterized as follows: (1) agents in a MAS may have their own specific ontologies defined on top of a shared base ontology; (2) concepts in these ontologies are represented as frame-like structures based on DAML+OIL language; (3) the semantic differences between agents are resolved at runtime through inter-agent communication; and (4) the resolution is viewed as an abductive inference process, and thus necessarily involves approximate reasoning.