Dictionary organizations for efficient similarity retrieval

Abstract

Searching dictionaries for an exact match to a given input string can be done very efficiently. Performance degrades, however, when inexact matches have to be located. Approximate string matching methods enable inexact matches to be located by comparing the search string with every dictionary entry. To avoid such exhaustive searches, dictionaries are generally partitioned according to some criteria. This paper examines some methods of dictionary organization that could lead to improved performances. The use of a variant of the n-gram method as a preselection technique is also discussed, along with a method using two levels of indexing. Experimental results show that of all the methods examined, indexing the string on each half provides the best result. A two-level indexing method combining two partitioning criteria also performs reasonably well. The n-gram technique appears to be close to these two in performance.