Spoken document classification with SVMs using linguistic unit weighting and probabilistic couplers

摘要

The task addressed by this paper is spoken document classification (SDC) of German TV news with support vector machines (SVMs). It shows the benefits of weighting different linguistic units when combined into one feature vector. Further experiments show that probabilistic SVMs (pSVMs) with couplers perform well on a SDC task. New couplers for multi-category classification, both for pSVMs and non-pSVMs, are discussed. They are easy to implement and show good and promising results. It turns out that using the distance instead of the decision value can be favorable. Theoretical justification is given for our approaches, and some results are explained theoretically.