Learning Subsequential Transducers for Pattern Recognition Interpretation Tasks

The “interpretation” framework in pattern recognition (PR) arises in the many cases in which the more classical paradigm of “classification” is not properly applicable generally because the number of classes is rather large or simply because the concept of “class” does not hold. A very general way of representing the results of interpretations of given objects or data is in terms of sentences of a “semantic language” in which the actions to be performed for each different object or datum are described. Interpretation can therefore be conveniently formalized through the concept of formal transduction, giving rise to the central PR problem of how to automatically learn a transducer from a training set of examples of the desired input-output behavior. This paper presents a formalization of the stated transducer learning problem, as well as an effective and efficient method for the inductive learning of an important class of transducers, namely, the class of subsequential transducers. The capabilities of subsequential transductions are illustrated through a series of experiments that also show the high effectiveness of the proposed learning method in obtaining very accurate and compact transducers for the corresponding tasks.