Orion Semantic Technology
Here at Interactive Engineering, we spend our days building machines that can read text and build semantic structures, so that they “understand” the text and can make use of the knowledge in it. This isn’t Machine Learning as it is currently understood – a few words together can blow away millions of lines of code or learning based on thousands of trials – by comparison, it is “instant learning”, aka human learning
We are heavily focused on the development of Orion. If providing services in complex areas will help us to come to a better understanding of what we are trying to do, we would be glad to help.
Development of Orion
Greenfields development can take many directions. We can prioritise areas, as long as they don’t impact our longterm goals.
If you are into machine reading of text, you had better have a good idea what the word “knowledge” means, and how it should be managed.
Large, complex specifications for large, complex things is a good illustration of where humans can become overwhelmed by the sheer volume of detail.
Sometimes you don’t want to activate the knowledge in a document – just get it right, free of errors and inconsistencies. Orion can help with that too.
Here are some ongoing projects. They range from what you might expect, to applications where humans fear to tread.
We are working to improve the delivery of automated knowledge in the Health Insurance business, by reading policies and extracting their information.
We are working on providing automated knowledge expansion using an acclaimed dictionary. The system encounters a new word, looks it up and builds the definition into its structure, or finds a new nuance on a word it already knows. The possibilities are either exciting or frightening.
This is a long range project we are working on – to break the absurdly low human limit of four pieces of information in play at once, and thus eliminate its effects – one of which is tens of thousands of people killed in hospital each year due to boredom and inattention to detail.
You may have a project which seems impossible with the technology you are familiar with. Our technology can turn the impossible into the admittedly hard.
We started out converting mathematical formulae into structure – Active Structure, we called it. Values and logical states flowed through the structure in any direction – the structure was undirected as to purpose. This was a powerful solver of complex numerical problems, and was used on a few projects – scheduling of flights at a busy airport, asset analysis at a bank, but still it languished – people weren’t very good at describing complex situations using mathematical symbols. Still, an active, undirected structure was a powerful method of representing and activating a complex problem. Why not use it on natural language, where the problem is implicit in high-level text?
So we did.
Proof of Concept
Artificial General Intelligence is a slippery concept – if we show we can do something on a machine, then ipso facto it doesn’t require a human level of intelligence. So, for Proof of Concept, we had to find a reasonably bounded hard problem. Anti-Money Laundering probably meets the bill. A money launderer is a criminal who turns dirty money into clean by a variety of techniques – they can turn a large sum of money into a large number of small amounts that fly below the radar, then reassemble the money, possibly offshore to pay for importation of illegal drugs, or have a reputable company issue fake invoices, or move the money through a casino, so it becomes “winnings”. The money launderer is willing to lose 15% of the money in the process, so they can afford the best lawyers, or a large workforce of “smurfs” (seemingly innocent people, like pensioners).
A bank can implement systems to combat money-laundering, but the bank and its systems are slow and cumbersome, while the money launderer can “turn on a dime”. What is needed is a way of increasing the speed of response of a bank to a threat. One way to do that is to change from a programmed system to the use of English in determining and implementing the strategy used to fight money laundering – to change from a response time of six months at best to a response time measured in days. The application illustrates how a machine can be used to counter the activities of a well-resourced adversary with human-level intelligence, a determination to succeed, and the ability to see weaknesses in the bank’s defences, substantiating a claim for AGI.
The first step is for the machine to read and “understand” the Act. This already has substantial benefits, as a human reader without much legal experience can see exactly what is meant by every word and group of words, without getting lost in a forest of bullets and other punctuation. An important aspect is that the Act uses the word “reckless”, which is not a word a programmer would be happy with, as it introduces human frailties to what is otherwise simple database transactions.