Learning Languages with Rational Kernels

Abstract

We present a general study of learning and linear separability with rational kernels, the sequence kernels commonly used in computational biology and natural language processing. We give a characterization of the class of all languages linearly separable with rational kernels and prove several properties of the class of languages linearly separable with a fixed rational kernel. In particular, we show that for kernels with transducer values in a finite set, these languages are necessarily finite Boolean combinations of preimages by a transducer of a single sequence. We also analyze the margin properties of linear separation with rational kernels and show that kernels with transducer values in a finite set guarantee a positive margin and lead to better learning guarantees. Creating a rational kernel with values in a finite set is often non-trivial even for relatively simple cases. However, we present a novel and general algorithm, double-tape disambiguation, that takes as input a transducer mapping sequences to sequence features, and yields an associated transducer that defines a finite range rational kernel. We describe the algorithm in detail and show its application to several cases of interest.