A Semi-supervised Approach for Natural Language Call Routing
Tatiana Gasanova, E. M. Zhukov, Ruslan Sergienko, Eugene Stanislavovich Semenkin, Wolfgang Minker · 2013
Natural Language call routing remains a complex and challenging research area in machine intelligence and language understanding. This paper is in the area of classifying user utterances into different categories. The focus is on design of algorithm that combines supervised and unsupervised learning models in order to improve classification quality. We have shown that the proposed approach is able to outperform existing methods on a large dataset and do not require morphological and stop-word filtering. In this paper we present a new formula for term relevance estimation, which is a modification of fuzzy rules relevance estimation for fuzzy classifier. Using this formula and only 300 frequent words for each class, we achieve an accuracy rate of 85.55 % on the database excluding the “garbage ” class (it includes utterances that cannot be assigned to any useful class or that can be assigned to more than one class). Dividing the “garbage” class into the set of subclasses by agglomerative hierarchical clustering we achieve about 9 % improvement of accuracy rate on the whole database. 1