| Abstract |
The ALIRA system is an experimental prototype that borrows techniques from information retrieval and
artificial intelligence to accomplish retrieval of relevant documents and classification of user queries. The
retrieval of relevant documents is based on the Okapi BM25 ranking mechanism. Given the retrieved result
set that contains the k best-ranked documents for a specific query, the ALIRA system builds upon the
paradigm of a lazy learning scheme named nearest neighbours to classify this query. The primary goal of
this research project is to optimise the ranking and classification accuracy of the system with respect to its
base performance for two specific datasets. This is achieved by combining several components, which
optimise the original feature space and parameter settings of the Okapi BM25 algorithm and consequently
improve its mapping of features. The secondary goal consists of designing a virtual helpdesk framework that
uses the ALIRA system as a backbone and a FAQ model as its document reference collection. This
research is part of the M. Sc. Agents & Computational Intelligence course at Utrecht University. Most of the
practical work was done at Clockwork B.V., a Dutch company focused on interactive multi-channel ebusiness
solutions. Performance was measured using two real-life datasets, donated by Q-Go, a company
specialized in online marketing and customer interaction services. |
