Topic detection in read documents

Abstract

This paper addresses the problem of topic annotation in the speech retrieval domain. It describes an algorithm developed to perform automatic topic annotation of broadcast news (BN) speech corpora. The adopted approach is based in Hidden Markov Models (HMM) and topic language models, solving the topic segmentation and labelling tasks simultaneously. To overcome the lack of topic labelled material for training statistical models, a two-stage unsupervised clustering was developed. Both stages are based on the nearestneighbour search method, using the Kullback-Leibler distance. On-going experiments to evaluate the system performance are also described.