Figure 1

The overview of the method. The training data (X) comes from the labelled corpus (L), while the unlabelled data (UL) is transformed using semantic models (SEM) to produce smoothing matrices (S). The training data is then projected into the semantic subspace (XS) and passed into one or more of the available kernel functions. We use cosine (κ _c ), Gaussian (κ _g ), and polynomial (κ _p ) kernels. We combine the resulting kernels with a weighting β _s into a single combined kernel (K).