Automated Taxonomy Expansion and Tag Recommendation in a Knowledge Management System

Abstract

We investigate two problems in order to improve a knowledge management system. The first problem is insert newly created tags, used to classify documents, to their semantically correct position in a taxonomy. To solve this previously unexplored problem we try several techniques including association rules, Bayesian network learning and a custom approach. The accuracy of tag insertion was 23.3% on realistic scenarios and 71.0% on adapted scenarios. This score leads us to conclude that this approach is not practically applicable. Data set analysis gives a good insight in why this score is low and gives motivation for further research. The second problem is tag recommendation. Our custom approach finds words that have most mutual information with the tag and selects these words to train a classifier for each tag in order to make a recommendation. The classifiers used are naive Bayes and support vector machines. The best setting has a micro average F1 score of 0.197. This score leads us to conclude that this approach is not practically applicable.