We show that for each context-free grammar a new grammar can be
constructed that generates a regular language. This language
contains the original context-free language. Our construction
differs from existing methods of approximation (in particular,
the one by Pereira and Wright, 1991) in that the use of a pushdown
automaton is avoided. This allows better insight into how the
generated language is affected. The new method is also more
attractive from a computational viewpoint.