The Higher Reasoning Unit is responsible for making those conceptual leaps that will be required for this system to be very useful. It takes as input an array of RCF-encoded files or documents, and outputs a single file of Highly Reduced Conceptual Form (HRCF) text. This is basically just RCF that has been created from a larger base of other RCF rather than from natural language.

The purpose of the HRU is to attempt to construct higher, perhaps abstract concepts from the lower-level concepts available to it. This would be done by comparing similar information and attempting to relate it (a relation is a higher concept). Comparing contradictory information and resolving two opposing views based on their reliability weights would help to strengthen your own reliability in yourself, even if it may redice the reliability of an individual idea or concept. An important thing to note about the HRU is that it will never throw any information away. It stores sets of beliefs and disbeliefs, each with associated reliability weights. Something you believe to be very true would have a very high weight, and something very false, a very negative weight.

As an example, consider a corporate page giving you information about the cost of Nike shoes, and another politically slanted page giving you information about how little it costs Nike to make the shoes because of low labor costs and poor working conditions. The HRU would be able to compare the cost of the shoes to buy and the cost to make them and come up with the concept of exploitation. This might be a bit of a logical jump, but with multiple examples, it could be made.

The HRU is by far the most undeveloped section of this design, in terms of what research has been done, and what work needs to be done. It would work in a similar manner as to a logical planner. Beliefs would be stored as true clauses, and disbeliefs as false clauses. Concepts are stored as relations between clauses, which become clauses in themselves, hopefully also being resolved with the vocabulary to make the clauses somewhat readable. To answer a question, one would only need to assert it, and try to resolve to nothing. The steps taken during the resolution would represent the answer. The natural language parser could then take that RCF and turn it back into a natural english answer.

I think this module of the design would be best developed with a learning algorithm. The dependency sets are not simple to come up with, and ones that we might never think of might be derived by the system as well.