In order to let computers to work for us, they must understand data: not just the grammar and the syntax, but the real meaning of things.
KEES is an architectural design pattern that establishes specific requirements for Semantic Web Applications. Its purpose is to formally describe domain knowledge with the goal of making it tradeable and shareable.
Domain knowledge refers to information about a specific subject (e.g., a range of products, commercial offerings, a social network, etc.). Knowledge domains are cumulative, with no defined limits on their scope or size.
With KEES, you can formalize and license all the components necessary to construct a knowledge domain. This includes:
- Methods for collecting accurate data.
- Establishing the context or significance of the data.
- Determining the reliability of the information.
- Deriving new insights from gathered data.
- Guiding the process of using information to address specific queries.
Both machines and humans can leverage this know-how to enhance their knowledge.
For more details, refer to the KEES presentation slides."
KEES is mainly a matter of definitions. A lot of concepts used by KEES refer to the well-known Semantic Web Standards published by the World Wide Web Consortium (W3C).
Data, as per KEES, encompasses any literal like words, numbers, or generally, any sequence of symbols. For instance, strings such as xyz, 123, 33.22, or http://LinkedData.Center exemplify data. Typically, data is linked with a data type that defines constraints on the sequence of symbols composing the data. For instance, the number 123, the float 33.22, or the URI http://LinkedData.Center each possess their respective data types, which specify rules regarding the permissible symbols sequence. Importantly, the data type does not inherently convey the meaning of the data itself.
Information, as defined in KEES (also known as facts), is characterized as data with meaning. This meaning can either be inferred from the context where the data is situated or explicitly defined. KEES adopts the RDF standard to delineate information through a triplet of three elements, forming an RDF statement, commonly referred to as a triple. This triple comprises a subject, a predicate, and an object. Indeed, a triple can be visualized and represented as an undirected labeled graph, and it can be stored within a Graph Store. This store serves as a repository for organizing and maintaining collections of these graph-based representations, allowing efficient storage and retrieval of triples.
KEES defines knowledge as a graph of linked information (i.e. linked data). This graph is possible because, in RDF, any URI can be both the object of a triple and the subject of another one or even the predicate for another.
In KEES, a Knowledge Graph is characterized as a Graph Store with a purpose. It represents a collection of interconnected information (i.e. Linked Data) organized in a graph-like structure that's designed specifically to be queried, providing answers to specific questions within its domain of related information.
From a theoretical perspective, a knowledge graph comprises information (facts), a formal logic system for knowledge representation, incorporates the Open-world assumption, and employs an inference engine to demonstrate theorems.
In a knowledge graph, information is partitioned into two datasets: TBOX and ABOX. ABOX statements delineate facts, while TBOX statements define the terms used to qualify the meaning of those facts. If you're familiar with the object-oriented paradigm, TBOX statements can be likened to associations with classes, whereas ABOX statements are linked to individual class instances.
TBOX statements are typically more enduring within a knowledge graph and are often organized into ontologies, describing specific knowledge domains such as business entities, people, goods, friendships, offerings, geocoding, and more. These statements are expressed using an ontology language that offers a formal semantic, like the W3C OWL.
ABOX statements, on the other hand, are related to instances of classes defined by TBOX statements. They possess a much more dynamic nature and are populated from datasets available on the web or through reasoning processes.
Theorems within a knowledge graph can manifest as rules or axioms. A rule signifies a generalized inference that establishes a logical correlation between propositions. Conversely, an axiom embodies a rule accomplished through entailment that is inferred by the semantics of existing factual information. Axioms and rules can implement deductive and abductive reasoning.
KEES assumes that knowledge graph is implemented by a SPARQL service supporting the SPARQL protocol
For instance, an axiom could be articulated using OWL (Web Ontology Language) and demonstrated through an OWL reasoner. Alternatively, a rule could be formulated using SPARQL QUERY CONSTRUCT and materialized with SPARQL UPDATE operations.
The Language Profile (also referred to as semantic Application Profile) forms the section of the TBOX comprising all terms recognized by rules and employed to comprehend and respond to inquiries. This profile defines the scope and vocabulary utilized within a Semantic Web Application, outlining the terms essential for understanding and generating responses within its semantic framework.
The internet is full of open data resources. Unfortunately, these resources very often have no SLA. So, when a business requires high data availability, it is best practice to mirror such resources in a data lake. During this download, some ETL processes may occur: for example, transforming raw data into Linked Data according to a defined language profile, adding provenance metadata, ACL, etc.
KEES does not concern the implementation of a Linked Data Lake, but this concept is very useful in any real-world application that uses Semantic Web Technologies.
In KEES , there are two types of reasoning:
- Deductive reasoning: A logical process in which conclusions are drawn from premises or general principles. Deductive reasoning involves computing axioms to generate new facts from existing ones through theorem demonstration based on the semantic properties of the Language Profile.
- Abductive reasoning: A form of reasoning where the best explanation or hypothesis for a set of observations or phenomena is inferred. Abductive reasoning is achieved through rule computation.
Note that in semiotics there is a third reasoning type: the Inductive reasoning that involves making generalizations based on specific observations or patterns. KEES does not jet able to support Inductive reasoning.
It is common practice to classify data into categories that aid in distinguishing the sources and ownership of the data, as well as the context in which it is collected and used:
- Axioms: information that is considered self-evident or universally accepted as true without requiring proof or validation
- First-party data: Information collected directly from individuals or users by the entity that owns the data, typically through interactions with their own website, app, or services.
- Second-party data: Data that is obtained directly from another organization, often through a mutually beneficial partnership or collaboration, where the data is shared directly between the two parties involved. In KEES second-party data ar use to classify deductions derived by semantic rules derived only by first-party data.
- Third-party data: Data collected or obtained from external sources or entities that are not directly affiliated with the organization using the data. This data is often purchased or obtained from data brokers, aggregators, or other external sources.
- Inferences: information resulting from a materialization of some reasoning about existing data
Trust holds significant importance in KEES due to the Open-world assumption and the inherent nature of RDF, which permits the amalgamation of diverse information, even when it might be incoherent or falsified.
In the context of KEES, trust serves as metadata linked to each RDF statement, ranging in value from 0 (indicating no trust) to 1 (indicating complete trust). Notably, this concept of trust extends to both ABOX and TBOX statements within the knowledge graph.
Moreover, the trust assigned to an inferred fact is intricately tied to the trust levels associated with all the involved facts and rules contributing to its derivation. Therefore, the trustworthiness of an inferred fact is contingent upon the collective trust levels of the underlying information and the processes involved in its generation within the knowledge graph.
The trust value of 0 or 1 signifies certainty in a fact. This principle extends to TBOX statements as well.
Note that the trust of an inferred fact MUST be less or equal to the minimum trust of the involved statements.
Inferred facts can be derived by the following reasoning type:
- deductions: when ABOX and TBOX facts are equally trusted
- abdutions: when TBOX facts are more trusted than ABOX facts
- inductions: when ABOX facts are more trusted than TBOX facts
An example
For instance, suppose that an axiom in your knowledge graph TBOX states that a property ":hasMom" has a cardinality of 1 (i.e. every person has just one "mom"), your knowledge graph could also contain two different facts (:jack :hasMom :Mary) and (:jack :hasMom :Giulia), perhaps extracted from different data sources. In order to decide who is Jack's mom, you need trust in your information.
If you are sure about the veracity of all data in the knowledge graph, you can deduct that: Mary and Giulia are two names for the same person. If you are not so sure, you have two possibilities: deduct that the one data source is wrong, so you have to choose the most trusted statement concerning some criteria (even casually if both statements have the same trust rank) or change the axiom in TBOX, allowing a person to have more than one mom. In any case, you need to get an idea about your trust in each statement, both in ABOX and in TBOX, in the knowledge graph. At least you want to know the provenance and all metadata of all information in your knowledge graph because the trust in a single data often derives from the trust of its source or in the creator of the data source.
So if:
- fact 1:
:jack :hasMom :Mary- fact 2:
:jack :hasMom :Giulia- axiom :
?restriction owl:qualifiedCardinality=1; owl:onProperty :hasMomThe trust helps to solve potential conflict according to this table:
trust in fact 1 trust in fact 2 trust in axiom derived facts derived fact trust reasoning type k <= 1 k <= 1 k <= 1 new fact (:Mary owl:sameAs :Giulia) k deduction x < 1 y > x y<=k<=1 fact1 wasInvalidatedBy (axiom + fact 2) <=y abduction x < 1 y < x x<=k<=1 fact2 wasInvalidatedBy (axiom + fact 1) <=x abduction k k z<=k<=1 axiom wasInvalidatedBy ( fact 1 + fact 2) <=z induction j < z j < z z<=1 fact 1 and fact 2 wasInvalidatedBy axiom z free will Note that the last row in the table is a just subjective illogic paranoic response, you could also decide to:
- invalidate the axiom (i.e. take the risk of fake data)
- randomly invalidate one of the conflicting facts (i.e, take the risk of wrong rule )
- derive that :jack is not an human and add a restriction on :hasMom property domain ( creative thinking )
- ... Resolving the trust conflict it depends from your risk appetite and from your free will.
The creation of knowledge requires the execution of certain activities. KEES distinguishes between:
- Booting Activities: which initialize an empty knowledge graph with the language profile.
- Ingestion Activities: which ingest facts and compute axioms. Ingestion activity are further classified in:
- 1st-party data learning
- 2st-party data learning
- 3st-party data learning
- Reasoning Activities: which infer facts. Reasoning activities are further classified in:
- Deductions
- Abductions
- Enriching Activities: which discover and inject new facts by leveraging existing facts, deductions, and abductions.
- Publishing Activities: which map the language profile into other ontologies (typically ontologies recognized by applications).
The life cycle of a KEES activity is based on three states:
- pending: when an activity is defined but not yet started
- started: when an activity is started but not finished
- ended: when an activity finished (with or without errors)
Taking into account the KEES activities it is possible to define the following states for a knowledge graph:
- created this is the initial state of a any existing booted knowledge graph;
- stable this is a state where there are no pending or running activities.
More specific states may be defined, for example:
- ingested: when all ingestion activities are ended
- reasoned: when all reasoning activities are ended
- enriched: when all enriching activities are ended
- published: when all publishing activities are ended
- etc,etc.
The generation of a knowledge graph involves a sequence of KEES activities organized into phases known as KEES windows, forming a workflow called the KEES cycle.
The KEES Cycle begins with an empty knowledge graph and provided facts, culminating in a fully populated knowledge base ready for application use. The KEES cycle windows are:
- Boot Window (Startup Phase): triggers KEES data processing upon specific events indicating new information availability.
- Learning Window (Population): Populates the knowledge graph with known facts, often extracted from the linked data lake. It computes Axioms
- Reasoning Window(Inference): Produces inferences about facts within the knowledge graph, utilizing specific reasoning rules. Reasoning could need to be recomputed on data changes.
- Enriching Window (Fact Enhancement): discover new third-party facts, enriching the existing data.
- Publishing Window (Optimization and Publication): Involves data cleansing, semantic conflicts resolution, versioning, and mapping to application ontologies.
Note that the step 4 may produce new facts and trigger a new boot of the knowledge graph, this is an interactive process that concludes only when no new data or inferences are discovered.
The KEES Cycle Convergence (or just convergence) is defined as the interactive process that, starting from initial facts, executes the KEES cycle until the knowledge graph reaches a stable state. The duration between the detection of new data and the reaching of a stable state is termed convergence time. It's important to note that the convergence time is always greater than zero; this means that the entire KEES cycle can be conceptualized as a big ETL (Extract, Transform, Load) process that produces an eventually consistent mirror of all the ingested data.
The term KEES Agent refers to a processor capable of understanding the KEES language profile . One or more KEES agents may be involved in the execution of a KEES cycle
There are seven types of KEES agents:
- Orchestrators: processors that directly support all the activities required to complete a KEES cycle, Could be orchestrator of other KEES agent
- Autonomous agents: processors that directly provide Linked Data in a language profile transforming raw data or mapping different ontologies. They do not interface the Knowledge graph, but often are used to feed a Linked Data Lake.
- loaders: processors that loads linked data into a Graph Store. They access the Knowledge graph in write only.
- Publisher: processors that access the Knowledge graph in read only organizing queries and producing data according different ontologies outside the knowledge graph
- Learning Agents: ETL processors that acts both as fact providers and loaders.
- Reasoners: processors that operate within the same knowledge graphs, focusing on applying inference rules while reading and writing data. They access the Knowledge graph in read/write.
- Enrichers: special fact providers that enhance existing facts by discovering and integrating third-party data as linked data. They access the knowlege graph in read only.
Here is a working draft for a KEES implementation proposal. LinkedData.Center SDaaS product is a commercial implementation of this specifications (also available with Open Source license)
A great way to contribute to the site is to create an issue on GitHub when you encounter a problem or something. We always appreciate it. You can also edit the code by yourself and create a pull request.
All stuff here in the Creative Common (unless otherwise noted).