AbstractIn an attempt to propose a robust method for understanding natural language (NL) interface commands, a scheme is proposed that infers intentions from an indirect speech‐act that does not express users' real intentions explicitly. This method classifies the real intentions of the indirect speech‐act into: 1) refusal; 2) reversal; 3) restriction; 4) benefit; and 5) disability. Further, concepts are abstracted for operations, e.g., displaying, moving, and deleting information systems; and constructing the operation knowledge base. This knowledge based comprises operational concepts and the relationships between them. These relationships are assigned the foregoing classifications for intentions. In addition, we construct the knowledge base of objects for the target of operations, e.g., files, figures, strings. This knowledge base contains the relationships: a) antonym; b) exclusive; c) part‐of, between the objects; and uses these relations to infer the transitions between the objects. An algorithm is the proposed to infer concepts for operations and concepts for target objects of operations that may represent the user's actual intentions.This proposal scheme was tested with requests on UNIX and a commercially available Japanese Word Processor. The system successfully inferred the intentions for approximately 80 percent of the user's indirect speec