Going on 5 years ago, I wrote part 1. Now, finally, it’s time for the rest of the story.
The bottom line to this story:
- It’s time to trade rule technology dating back to the 80’s for state of the art AI.
The dominant rule/policy/decisioning systems are “production rule systems”. Wikipedia defines a “production rule system” as follows:
- A production system (or production rule system) is a computer program typically used to provide some form of artificial intelligence, which consists primarily of a set of rules about behavior. These rules, termed productions, are a basic representation found useful in automated planning, expert systems and action selection. A production system provides the mechanism necessary to execute productions in order to achieve some goal for the system.
Note the reference to out-of-favor “expert systems” and the lack of anything about reasoning…
The timeline below, which stops at the point of part 1, is intended to emphasize the following:
- 80’s technology dominates business rules, policy automation, and decision management
- artificial intelligence (AI) and natural language processing (NLP) have improved dramatically
- logical reasoning technology has advanced while production rule technology remains stagnant
It has more specific themes related to deep NLP, knowledge acquisition, and reasoning technology that the most competent AI systems all involve. Although not critical to the point of this confession, the advances in Prolog technology since XSB, Flora-2, and SILK are notable. More recently, deep learning is being applied to logical reasoning technology (for example see this Google Research).
Let’s pick up the story after this 30+ year timeline:
- 1981 Carnegie Mellon OPS5 – non-logical, forward-chaining, production rule system
- 1982 Japan’s 5th generation computer challenge –aims to build Prolog super-computers
- 1983 AI startups: Carnegie Group, Intellicorp, Teknowledge, Inference Corp
- 1983 Prolog WAM – Warren Abstract Machine compiles logic programming language
- 1984 MCC Cyc – acquisition of common sense knowledge project begins
- 1984 Inference Corp ART –forward/backward Automated Reasoning Tool (LISP)
- 1985 NASA CLIPS –forward-chaining clone of Inference’s ART for IBM PCs (C)
- 1985 Neuron Data Nexpert – expert system shell for Apple PCs
- 1988 SRI CLE –deep NLP: unification-based grammar & under-specified semantics
- 1989 SRI PTTP – Prolog Technology Theorem Prover
- 1990 Haley Eclipse – forward & backward chaining (C/C++)
- 1994 Cycorp Cyc – encyclopedic ontology & logical inference engine spins out of MCC
- 1995 Sandia JESS – Java clone of CLIPS
- 1995 XSB Prolog SLG – improved Prolog with tabling
- 1996 Ilog JRules –forward chaining (Java)
- 1997 XSB Prolog WFS– incorporates “well-founded semantics”
- 1998 Haley Authority – deep NLP translates English into production rules
- 1998 Zurich ACE –Attempto Controlled English for logic programming
- 2000 SUNY FLora-2 – higher-order, defeasible logic based on XSB Prolog
- 2002 Fair Isaac Blaze Advisor– via acquisition of Neuron Data via Brokat via HNC
- 2004 Vulcan Project Halo – Cyc, SRI, et al take advanced placement tests using logic
- 2005 OMG SBVR – standard based on the logic of English beyond BRMS vendors
- 2006 Red Hat Decision Manager – via acquisition of JBoss Drools
- 2007 C&C and Boxer – deep NLP translates English into first-order logic
- 2008 Microsoft Bing – acquires deep NLP of Xerox XLE by way of Powerset
- 2008 Oracle Policy Automation (OPA) – via acquisition of Haley
- 2009 IBM Operational Decision Manager – via acquisition of Ilog
- 2010 Vulcan SILK– enhanced Flora-2 and XSB Prolog for deep question answering
- 2011 IBM Watson – deep NLP for simple QA versus Jeopardy champions
- 2012 Automata Linguist – deep NLP translates textbook English into SILK knowledge
- 2013 Allen Institute for AI – zigs to deep learning and zags back to “common sense”
- 2013 Part 1 of this article
Part 1 addresses the logical inadequacies of production rule systems. It also addresses their inability to implement what the community specified in the SBVR standard. This is further discussed and graphically presented in a subsequent post on requirements for logical reasoning.
At the time of Part 1, I hoped Vulcan would release SILK to the public. SILK was experimental but promising. Vulcan’s funding over 5 years led to significant improvements in Flora and XSB, some of which were also incorporated into Cyc. Among them were a transformation of extended first-order logic into Flora-2 code (which is an extended form of Prolog). There was also a nice explanation facility, as shown below for a question answered with knowledge captured using the Linguist.
If you take even a cursory look at other articles here, it should be clear that the Linguist can generate sophisticated logic. So, I was most concerned with the functionality of SILK’s transform. Unfortunately, Vulcan never released SILK, so this part 2 was delayed until a replacement developed.
By the end of 2013, Coherent Knowledge had formed to develop Ergo, a successor to SILK. But starting from Flora and XSB meant that much of what had been developed in SILK had to be recreated by a smaller team. It took Coherent a few years to get to the point that we could target their product, Ergo.
We could not rely solely on our Coherent friends, however. We had no practical alternative but to develop an alternative. Now we can translate English into various theorem proving systems and into Flora or XSB (and, of course, Ergo).
- We can do this for any rule- or axiom-based technology (e.g., W3C OWL, CMU SCONE, and TPTP provers, ISO Common Logic, CycL, etc.).
- If you are interested, we could also generate production rules, such as for the ART/CLIPS syntax or other systems listed above.
For details, see Simply Logical English.
We have already done so for most of those listed here and welcome the opportunity to work with others to make the Linguist work for them. For example, Intel funded the work that produces OWL and SCONE.
Here is what you need to know: The Linguist makes it easy to eliminate all ambiguity from completely natural English.
- all sources of ambiguity become readily apparent and are easily resolved
- no linguistic or logical skills beyond mere literacy are required
- very good for business analysis and knowledge engineers
- great for linguistic and logic education and experts
- it’s generally available under a public-use license
- Machine translation of English into logic is fast, easy, and reliable. Doing so by hand is painstaking and error prone.
- highly skilled experts take time to translate even simple English into logic or rules
- the resulting “code” is likely to suffer from errors of omission or expression
As for automating the knowledge in a business process, intelligent agent, or QA system, we target any of the following:
- Theorem proving yields the most powerful reasoning.
- but it typically fails miserably given inconsistent axioms (i.e., rules)
- XSB Prolog provides more powerful reasoning than production systems.
- it can handle logical inconsistences unlike theorem provers
- it is more practical and easier to use than other Prologs
- Flora offers more robustness and usability than XSB alone.
Overall, we recommend Flora and XSB.
- XSB Prolog is commercial, open-source. It has been used in IBM Watson and by U.S. Customs.
- Flora-2 is academic, open-source. It has been used by Vulcan SILK, in SRI’s Sunflower, and others.
 All the while, Cyc advances…
 The author was Chief Scientist and Doug Lenat of Cyc was scientific advisor
 via under-specified semantic representation (USR) resulting from a unification-based grammar (CCG)
 Explanations provided by SILK for questions answered using knowledge captured by the Linguist: