When a natural language specification is translated into a formal one, it is important for objects and operations appearing in the natural language specification to be appropriately classified according to the framework of data types in the formal specification. In this paper, we propose a semi-automatic method of constructing a context-free grammar (cfg) representing an assignment of data types to words in a given natural language specification. In our method, a cfg is mechanically constructed from sample sentences in a natural language specification, where the cfg represents type declarations of expressions and type hierarchy. Then, the cfg is appropriately modified by adding nonterminals/production rules that represent type inclusion relations. In this modification process, candidates for the productions to be added are presented to the user. Finally, the cfg is simplified based on structural equivalence. The result of applying this method to a part of the OSI session protocol specification (39 sentences) is also presented. There was an example in which ambiguity of anaphoric bindings was solved by type checking based on the resulting cfg.