A new method of searching large dictionaries given uncertain inputs is
described, based on the lazy evaluation of a syntactic neural network
(SNN). The new method is shown to significantly outperform a conventi
onal trie-based method for large dictionaries (e.g. in excess of 100,0
00 entries). Results are presented for the problem of recognising UK p
ostcodes using dictionary sizes of up to 1 million entries. Most signi
ficantly, it is demonstrated that the SNN actually gets faster as more
data is loaded into it.