International Business Machines Corporation
Training a cognitive system on partial correctness

Abstract:

A set of components is computed from performing NLP on a question in an input. An actual answer is computed corresponding to the question by a cognitive system. the actual answer corresponds to an actual subset of the set of components, and an expected answer corresponds to an expected subset of the subset of components. The actual answer is mapped to an actual category in a hierarchy of answer categories. A distance between the expected answer and the actual answer is computed where the distance is a function of a path in the hierarchy from the actual category to the expected category, and a degree of correctness of the actual answer is another function of the distance. A self-learning operation in the cognitive system causes a revised actual answer on the question being at a shorter distance from the expected answer.