In this paper, we explain the design philosophy of a postprocessor that supports a large vocabulary Japanese speech recognition system (Fujitsu Model 2361A). The postprocessor improves performance of the speech recognition system by selecting syntactically plausible sentences from the sentence candidate list. Selection of the sentences is achieved by 1) searching through a bunsetsu lattice (group of short Japanese phrases) which is grouped based on distance scores, 2) use of a bottom-up parser, and 3) case checking based on lexical functional grammar. The result of the performance evaluation was that the postprocessor recovered 21 sentences out of 26 resulting in an 80% recovery rate.