Some call centres have an easy time the customer information can be extracted from a database and shown on a screen, or there is a single defining document that applies to all callers the tax law, say so it is worth going into great detail to spell out every nuance.
Some call centres have a much more difficult time knowledge about the relation between the organisation and the customer is held in text, and it can be anywhere from subtly to dramatically different for different callers. We will use Health Insurance as an example. The call centre operator has to answer questions and make decisions based on a lengthy and rather impenetrable document running to 150 pages together with other supporting documents guessing or using a previous answer to another caller could be an expensive mistake. We are proposing a semantic factory approach to this problem the machine can then answer questions based on the contents of the document.
It would take too long (at least several hours) for a machine to read one of these documents and extract its semantic structure. An alternative approach is to read the document beforehand to have a fleet of machines reading documents, and creating the semantic structure for each document in the form of a memory image. The images are stored in parallel with the documents. These images are not like pictures - the memory image is a mass of computing elements, which can hold and transmit states and values - it is much more like a working cognitive machine held in the memory of the machine. Part of the process of reading the documents would be to rid them of error - there is no point helping the operator to find the context of an answer if what they find is confusing or inconsistent or wrong.
For a new call, the caller is identified and the specific memory image is loaded. A database access is used to update the semantic structure the caller is up to date on payments, the term for a pre-existing condition has expired, the current out-of-pocket amount is loaded, etc. The caller speaks to the operator, and the operator enters questions in free text. The machine analyses each question, taking into account terms defined in the document and the contents of other documents, such as pre-authorisation lists, and returns the appropriate answer. Questions can take the form of discovering context, choosing between objects, logical states (Yes/No) and providing values (all the necessary pre-requisite steps would be checked before providing a value). Some examples:
Is this the secondary plan?
Is cruciate ligature repair covered?
My daughter needs a heart operation costing $42,000. How much will the out-of-pocket be?
There is a large semantic technology component to make this work that we have not touched on here - see NLP as an introduction.