An evaluation of keyword, string similarity and very shallow syntactic matching for a university admissions processing infobot

Peter Hancox, Nikolaos Polatidis

“Infobots” are small-scale natural language question answering systems drawing inspiration from ELIZA-type systems. Their key distinguishing feature is the extraction of meaning from users’ queries without the use of syntactic or semantic representations. Three approaches to identifying the users’ intended meanings were investigated: keywordbased systems, Jaro-based string similarity algorithms and matching based on very shallow syntactic analysis. These were measured against a corpus of queries contributed by users of a WWW-hosted infobot for responding to questions about applications to MSc courses. The most effective system was Jaro with stemmed input (78.57%). It also was able to process ungrammatical input and offer scalability.