Combining deductive and statistical explanations in the FRANK query answering system

Both symbolic and sub-symbolic AI have their limitations, but their combination can be more than the sum of their parts. For instance, statistical machine learning has been hugely successful at classification and decision-making tasks, but not so good at deliberative systematic reasoning nor at explanation. We argue that by combining symbolic and sub-symbolic reasoning into hybrid systems, the whole will be more than the sum of its parts.

To illustrate the potential of hybrid AI system, we describe the FRANK query answering system. FRANK infers new knowledge from the diverse and immense knowledge sources on the Web, using a combination of both deductive and statistical reasoning. This enables it to make predictions. For instance, to answer the question “Which country in Europe will have the highest GDP growth rate by 2032?”, it (i) decomposes Europe into its constituent countries, (ii) then for each country uses regression over their previous GDP growth rates to extrapolate each to 2032 and (iii) then returns the country which is predicted to then have the maximum value. The decompositions are explained deductively and the regressions by a prediction model that can be rendered graphically. This explanation of FRANK’s reasoning merges deduction and statistics.

In this paper, we highlight recent work on FRANK that focus on leveraging hybrid AI to tackle question answering with emphasis on explainability of the inference process and its inferred answers. We aim for whole system reasoning; that is, we
are automating the choices of knowledge sources and the planning that constructs the inference process from the facts found in these knowledge sources. We intend that these ‘engineering’ choices are also explained to the user.