Fact-Based AI. Improving on a Knowledge Graph

photo 1495511623436 ba44aaee07cf

Viev, produces Boston, software for Fact-Based Modelling (FBM). We feel that Fact-Based Modelling will be influential in the next wave of conversational user interfaces (CUI), voice user interfaces (VUI), and artificial intelligence (AI) in general.

CUIs and VUIs have become popular within a relatively short period of time. The introduction of Apple’s Siri VUI heralded the start of a race among Apple, Microsoft, Google and Amazon to produce VUIs within their respective computer operating systems and smart speakers the likes of which I proffer has not been seen in the information industry since the introduction of object-oriented programming and relational databases [1]. Competition seems fierce and VUIs are rapidly becoming more intelligent in terms of their usefulness. The technologies are pegged as becoming, in short order, commonplace within household appliances and even cars [2, 3, 15].

In terms of research and applied application, CUIs in the form of chatbots are also in an area of rapid growth. Chatbots within Facebook’s Messenger app and within Microsoft’s Skype have seen a sizeable uptake of development with reports of over 100,000 chatbots on the Messenger platform alone [16]. Early consumer experiences were reported as disappointing [16]. However it seems that at least one business is reporting monthly revenues of over USD$1 million from its chatbot implementation [16] and there are continued reports of their successful application forthcoming [14]. My personal analysis of chatbots leads me to feel that the missing elements are a deep understanding of conversational context and natural conversation flow.

The underlying premise of all CUIs/VUIs is that it is considerably easier for people to talk and use natural language than it is to communicate in other forms. By way of example, I recently purchased an Android based phone and now rarely make phone calls or send text messages by pressing successions of virtual buttons, but rather I just speak to my phone and it understands what I want to do. To send a text message to someone I simply say, “OK Google, send Peter a text message”. Google Assistant asks me in natural language what it is that I would like to say. My spoken message is converted into text. Google Assistant asks me if I would like to send the message. Answering, “Send it”, Google Assistant dutifully sends the text message. All this without pressing a single button on my phone. Incredible, apart from the fact that because this type of technology is now so good, it is credible. I feel that the technology is here to stay. As time goes by, I feel that the technology will only get better and even easier to use, increasingly becoming a part of everyday life for those with access to the technology.

To those intimately familiar with the technology as a consumer, and especially to those who develop the technology, such an example will rapidly become unremarkable. Even today I feel that because such technology is so readily available and desirable that it has become not only a differentiating feature of a product, but rather an expected technology. I believe that consumer demand for CUIs/VUIs will drive rapid adoption of AI and, to a large extent, already has.

From the dawn of time of the computer age man has pondered the ability to converse with machines in natural language and even judge the intelligence of a machine by how well a machine can converse in natural language (viz The Turing Test ) [4]. A recent and interesting article reaffirms this viewpoint and aims to cement ‘ Conversational Cognition ’ as a benchmark for AI [5]. The Turing Test has been so pervasive in the world of AI that it has spawned the Loebner Prize , an annual competition which will end when judges are simply unable to distinguish as to whether they converse with a machine or another human being [6].

If we peg humans as providing the ultimate benchmark of intelligence, then it seems only natural and fair that artificial intelligence should be judged by that benchmark.

The relative, and only seemingly, rapid rise of CUIs and VUIs today does not, however, reflect the long-term research and development that has gone into making such technologies available. From a consumer point of view such technology is relatively new, but from a research point of view it is the cumulative result of decades of research.

But it is the future of research we are interested in here. Where is technology leading? What form factor will the best technology take? These are the questions I will broach here and proffer a way forward.

If we accept that all CUIs and VUIs represent some form of artificial intelligence and of varying degrees of actual intelligence, then we can limit our discussion to talking about AI in general. This, especially if we accept Conversational Cognition as a benchmark of AIs of worth within our discussion. That is, we are not talking about AI in regards to computer vision or suchlike, only those AIs that manage conversation with people.

Research into AI has broadly taken two main streams, one algorithmic and the other based on what I will call intuition. The first deterministic, where we can peer into the workings of an AI and follow its reasoning to come up with an answer to a problem; the second so complex in its construction that we may never know exactly how it derived solutions, only that it did (rightly or wrongly) [7, 8].

I include within deterministic AIs those that incorporate not only a knowledge base of some description, but a set of logic rules and algorithms both of which are completely transparent to an outside observer interested in the workings of the AI. By default I include all those AI systems that have some algorithmic perception of first-order logic (FOL). Such AIs garner so little press that if there is a general term that represents them, it escapes me.

I include within intuitional AIs those for which the inner workings of the AI are so complex that determining exactly how an answer is derived could be that time consuming that it is easier to accept that the AI either does or does not produce the desired result. That is not to say one could not break down the inner workings of a transaction within the AI, but rather that it would be so complex as to be unreasonable to expect someone to try. These types of AIs do have recognisable names, such as Neural Networks and those AIs that incorporate Deep Learning and Machine Learning. For simplicity, I will lump them all under the term, Neural Network, here.

A casual and causal observer of both types of AI might well argue that the inner workings of a Neural Network are, of course, well known, otherwise how else would they be built. In finer detail though, here we only consider the daunting complexity of analysing what steps were taken for a suitably complex Neural Network based AI to come up with a single answer. Of course, with infinite time and resources one can analyse anything, however it seems to be that we simply accept or reject a Neural Network based on whether its data model returns a favourable result or not. If the result is favourable we might be less inclined to analyse why, for each iteration of problem solving.

For example, when Google released a paper announcing a ‘breakthrough’ in the automated playing of the game, chess, by their AlphaZero Neural Network based AI, world class players labelled the result remarkable [9]. Others remarked how ‘intuition’ has rivalled over deterministic game play [10]. I know of nobody who has analysed exactly how the decision was made for each remarkable move made by AlphaZero.

In this respect the difference between deterministic and stochastic AIs is distinct.

A cursory search of online literature easily shows that there is currently much activity in resolving a combination of deterministic and stochastic systems within an AI to produce a more intelligent whole.

If we refocus our discussion on those AIs that converse with people, and how they would mimic human capacity for discourse, it isn’t unreasonable to look inside yourself and ponder just how much human capacity is required to generate speech and answer questions of the outside world predicated on applying logic as opposed to falling back on trained pathways of thought.

For example, responding to a greeting of “Hello” with a suitable “Hi” may be so ingrained in the neural pathways of the mind that it is easy to extrapolate those functions as being trivially trained within a Neural Network AI.

Less trivial is the construction of a Neural Network that ponders the question “Do I need an umbrella today?”, “It is raining outside”, “OK, yes, you will need an umbrella today” and extend that capability to providing answers to questions that require traversal of tens or hundreds of conditional predicates to reach a suitable result.

None of which is to say that Neural Networks cannot handle questions that require some form of predicate logic, but more to say that however an AI handles analysis of natural language, if that AI is suitably intelligent at conversation it will require data with rich semantic information associated with that data, including data which incorporates some expression of predicate logic.

It's all about the data

I recently highlighted a sentence in a medium article about AI , which reads:

“There are broadly three parts to an A.I. approach: generating data, interpreting that data and
making judgment about that data”

We arrive at the core of this article and discuss why it is that I believe that Fact-Based Modelling of some description will be central to future AIs that master Conversational Cognition, and how I would personally use the methodology to improve upon Google’s Knowledge Graph.

At their core, every AI relies on data. It is the rich embedding of semantic information within Fact-Based Models, including the expression of predicates of first-order logic, that makes Fact-Based Modelling an attractive offering to define the structure of that data in such a way that it may be stored, analysed and judged over.

From this point forward we no longer consider how (deterministic or stochastic) decisions are made over the data but focus on what is the data, including logical predicates stored as data.

Fact-Based Modelling Fact-Based Modelling is a form of conceptual modelling that captures the structure of information presented as ‘Facts’. A Fact may be written in the form “Person ‘Barack Obama’ is
‘1.85’ Meters tall’”, from which a Fact Type captures the structure, “Person is Height (in Meters) tall”. ‘Person’ may be analysed as being an Entity Type.

We express ‘Height (in Meters)’ as a special type of Entity Type that represents a value of unit measurement. Usually, however, values are analysed as Value Types, and “Person ‘Barack Obama’ has first-Name ‘Barack’” would have Name represented as a Value Type.

The Boston software uses a graphical type of Fact-Based Modelling called Object-Role Modelling (ORM), and the information above is conceptually captured in the diagram below:

What differentiates Fact-Based Modelling and ORM from other, and more simplistic conceptual modelling languages (e.g. Entity Relationship Diagrams and UML Class Diagrams) is that ORM
is semantically rich in the information that it captures about Facts within Fact Types. Complete information to reconstruct Facts in natural language is captured within the respective Fact Type.

The “... is … tall” Predicate Reading (above) allows its Fact Type to be extrapolated as “Person is Height (in Meters) tall” when considering the joined Object Types (Person, Height) and the sample population/data allows us to extrapolate the Fact, “Barack Obama is 1.85 Meters tall”.

ORM allows us to query over a model in natural language and ask questions like, “Is Barack Obama 1.85 meters tall?”, with relative ease because of the rich semantic information that is captured along with the model.

An Entity-Relationship Diagram for the same conceptual model is limited in the amount of information that it is able to capture, as below:


No extra semantic information about the underlying predicate of each Entity<->Attribute (e.g. Person<->Height) pair is captured in an ERD of the type shown. E.g We do not know that Height is in Meters using that type of ER Diagram, and we have not extra information to form the Fact, “Barack Obama is 1.85 meters tall” in natural language.

I feel that Entity-Relationship Diagrams and UML Diagrams have their place, however what we are seeking here is how to capture rich semantic information to create an AI with Conversational Cognition. I feel that in order to do that, we need to not only capture data , but also information about the underlying predicates that bind the data together.

An observant reader may ponder what value a diagram has in realisation over data. Here we realise that Fact-Based Models, represented as graphs (such as an ORM diagram) have realisations at the data storage level as effective graphs of knowledge. Let's examine further the utility of Fact-Based Modelling.

Our example becomes one where predicate logic is required if we add to the model that the set of people who have a height, is a subset of people who are alive (e.g. Cremated people have, of course, no ‘height’ as we normally interpret such things). In ORM that subset constraint is represented as follows:

Querying over the model with the question, “How tall is Barack Obama?”, should return “Barack Obama is 1.85 meters tall”. Querying “Is Barack Obama 1.85 meters tall?” should return, “Yes”. But querying a suitably intelligent AI over the same model with, “Is Napoleon Bonaparte 1.69 meters tall?” should return, “No, because Napoleon Bonaparte is not alive”.

NB If we accept that ORM only caters for Open World Semantics , and where not recording that someone is alive doesn’t necessarily mean they are not, we may at least cater for the same result by adopting Closed World Semantics within our practice. Or simply modify the model to reflect that a person may not be dead and have a height.

PersonHeigthAliveLargerHaving semantically rich data plus logical predicates allows an AI to explore the underlying model using Semantic Queries posed as questions in natural language [12]. Exactly how that is achieved is the subject of current research, however an environmental scan of the research and development indicates that there is little resistance to the concept that Semantic Knowledge is required within or over a data store if Semantic Queries are to be achieved [11].

Looking at The Knowledge Graph of the future

If we accept that Conversational Cognition includes the ability to answer questions posed of the AI, then we can examine real-world attempts to implement such AI.

Google Search has a feature called, “Knowledge Graph” (Microsoft’s Bing implements something similar), which refers to the underlying technology. Knowledge Graph is an ensemble of gathered knowledge about various topics of interest; for interest, ‘Famous People’. Knowledge Graph is presumably conceptualised and stored as a graph of data. We know, however that sources of that data are stored in a more relational fashion . For our purposes it matters not how Google stores data gathered into the Knowledge Graph, but more what it stores.

We know that Knowledge Graph captures information about famous people and their height, for example.

What makes this information somewhat useful is that you can do a Google search, “How tall is Barack Obama?” and a suitable result of 1.85 meters is returned. Because information about Barack Obama is stored in the Knowledge Graph, and extra box of information about Barack Obama is displayed on the search results page. Such a search result is shown below.

KG Obama

The problem that we analyse becomes apparent if we do a Google search, “Is Barack Obama 1.85 meters tall?”. No Knowledge Graph information is shown (as below). If we were expecting a confirmation, “Yes”, we are out of luck at the time of writings this article, accompanied by whatever level of disappointment we generalise over state of the art.

KG IsObama

The problem becomes more compound if we do a Google search, “How tall is Napoleon?”. A search result is given (on my iPad at least) with Knowledge Graph information that might indicate that Napoleon is still alive! (below) It takes digging a little further and clicking on the “Quotes and overview” button to reveal a death date. Apparently though, Napoleon is still 1.69 meters tall. May hap, but here we look for a result provided in natural language to what is a question posed in natural language.

KG Napoleon

So we know that, in the opening months of 2018, Knowledge Graph does not contain suitable rules of predicate logic that allow for a level of intelligent question and answering that we would find suitable when conversing with an AI implementing Conversational Cognition of worth. Some may argue that Knowledge Graph represents no attempt at Conversational Cognition, but if so, why respond to a question, “How tall is Barack Obama?”, at all? Some attempt at Conversational Cognition has been made. To be fair, Microsoft’s Bing search engine fairs no better. ‘State of the art today’ simply doesn’t reflect the potential.

What needs to change?

What I feel is that we are in the early stages of analysing and implementing Conversational Cognition. I feel that before long Google/Bing searches will return answers like, “Napoleon is not 1.69 meters tall because he is not alive”. What seems very clear to me is that, in the current environment of competition, if Microsoft Bing were to return such a result, there would be no question, Google would soon follow suit, and visa versa. I’ll be the first to fall off my chair when that happens; it seems inevitable.

What is needed is an architecture that supports Semantic Queries and Conversational Cognition.

An Architecture for Conversational Cognition

By now you perhaps see my thesis. Semantic Queries require a Semantic Data Store where both data and rich semantic information about that data is stored within one logical data store, including conditional predicates of first-order logic. In order to achieve Conversational Cognition of worth, semantic data stores need the ability to resolve natural language predicates along with logical constraints over the underlying data.

The architecture in its simplest form is depicted in the diagram below:


Current research to develop Semantic Data Stores (SDS) is well under way. For example, the “Knowledge Representation Database Research Centre” of the Free University of Bolzano (Italy) is dedicated to developing what they call a Knowledge Representation Database (KRDB).

Fact-Based Modelling, such as Object-Role Modelling, is a natural fit for an SDS. Fact Type Readings may include ‘Inverse Readings’; where “Part is in Bin in Warehouse”, say, has the complementary readings, “Warehouse stores Part in Bin” and “Bin houses Part in Warehouse”. More simplistic forms of Fact Modelling only store one reading, and even try to truncate predicates, such as, “Person isFriendOf Person”. I am not convinced of the value of such simplistic representations of predicates.

What I feel is important is that it does not matter how full predicates in natural language are stored (as a graph, a relation or a hierarchy of data), because if we adopt the metamodel of Fact-Based Modelling to store the semantics of our predicates, Fact-Based Modelling is implementation agnostic when it comes to databases.

What also seems important is that two forms of query can be made over an SDS, a Semantic Query over the data, and a Semantic Analysis Query (SAQ) over the rich semantic data that defines and governs the underlying data in order to facilitate the Semantic Query. An SAQ returns the logical predicates which help the AI refine the Semantic Query and result.

For example, a semantic query “Does Bin ‘1’ store Part ‘456’ in Warehouse ‘Main Warehouse’?” may entail, in its execution, a resolution by way of a Semantic Analysis Query over the Fact-Based Model that the initial query is referring to the Fact Type, “Bin stores Part in Warehouse’ and its underlying data. Similar SAQs would resolve the question, “Which Part is stored in which Bin in Warehouse ‘Main Warehouse’?” to return a list of data meeting the request by first identifying the appropriate predicate for which the data is stored.

Because of the algorithmic gymnastics required to resolve such queries using a combination of parsing and Natural Language Processing (NLP) implementing such technology is no small feat, but it seems clear to me that without storing Fact Types, Fact Type Readings and conditional rules expressed as predicate logic, the goal of intelligent Conversational Cognition will not be achievable. Our first goal is to store the appropriate predicates within the data model underlying the AI.

The goal of achieving Semantic Data Stores is not new ([13]) and it seems to me that it is rather inevitable that such databases will eventuate and prevail within the sphere of AI. An open area of research is how best to get an AI to update its own semantic layer of its SDS. It seems to me that if we are to develop Artificial General Intelligence, this requirement is indispensable.

Viev is making the tools to help build the semantic layer of Semantic Data Stores and we reach out to the AI community to help make Conversational Cognition a reality.

Thank you for reading. I hope you found his article rewarding of your time. As time permits I will offer more information about Fact-Based Modelling and Object-Role Modelling.

1. Wollerton, M., “Voice wars: Siri vs. Alexa vs. Google Assistant”, , , Accessed at 24/02/2018
2. Westlake, M., “Smart Home & Connected Cars - Voice Activation”, , Accessed at 24/02/2018
3. “Innovative Voice Controlled Devices On The Market”,,, Accessed at 24/02/2018
4. “Turing Test”, , , Accessed at 25/02/208
5. Perez, C., “Conversational Cognition: A New Measure for Artificial General Intelligence”, ,  , Accessed at 25/02/2018
6. “Loebner Prize”, ,, Accessed aat 25/02/2018
7. Rankin K., “The Dark Secret at the Heart of AI”, MIT Technology Review,  , Accessed at 23/02/2018
8. Perez, C., “Deep Learning is Non-Equilibrium Information Dynamics”,, , Accessed at 25/02/2018
9. “AlphaZero: Reactions From Top GMs, Stockfish Author”,,, Accessed at 25/02/2018
10. Perez, C., “AlphaZero: How Intuition Demolished Logic”,, , Accessed at 25/02/2018
11. “Semantic Queries by Example”, Google Research,
12. “Semantic Query”,, , Accessed at 25/02/2018
13. Date, C.J., “Constraints and Predicates: A Brief Tutorial (Part 1)”, ,
14. Karma, J., “Case study: How a real estate bot helped sell 3 apartments in 10 days.”,
15. "Voice assistants are making the smartphone redundant", Binary District Journal,
16. Mai, M., "From Gloom To Glam: The Evolution Of Facebook Messenger Chatbots", Forbes,
-facebook-messenger-chatbots/#3b3869263abd, Accessed at 25/02/2018