Michel Bauwens of the P2P Foundation just posted The future of the web: semantic, or just structured? that is comprised almost in its entirety of what I had written in a post about 10 months ago. The point of my earlier post was that structured Web content, short of the full-blown aspects and aspirations of the semantic Web, is where the Web was currently at and it represented the natural transition point between documents and meaning interpretable by machines.
I’m honored with Michel’s treatment of these comments and I thank him. It is always nice to see one’s name (and words) in lights, as it were.
But, I also had forgotten about that original post and its sentiments.
This re-surfacing caused me to go back and read and think about it again. (Not too bad, if I may say!) But, with the passage of now 10 months, I also don’t think I got it exactly right. As I commented back to Michel:
But being directed to these words by my feed reader (from a basis that is now, what, 10 months ago?) also got me to thinking. I stand by the words here (and in the full article that I just re-read after a long hiatus), but I would likely say them differently today.
Linked data and the semantic Web are moving at warp speed. Linked data, in particular, is showing the way to pragmatic, meaningful connections. And, for me and my company, Zitgist, that is also leading to a quicker exploration of semantics, class relationships, ontology mappings, and much that is closer to the "semantic" end of the spectrum rather than the "structured" end of the spectrum implied by my comments above.
I frankly did not see this rapidity of uptake and it is very, very exciting.
The lesson, I think, is that structure, yes, makes sense, but, once we taste it, we want to take it further. The role of structure driving the demand for semantics is pretty compelling.
In that regard, I would not equate "premature" nearly so strongly to the semantic Web.
The freight train is now moving through the yard but without any slowdown. The train left the station quite a bit ago. Like most trends, it is hard to see the start and initial pace.
Fortunately, we are also finding that items such as context, named entities, and quality are now becoming themes within our community’s posts. I say Hear, hear!, and Rah, rah! Good show, all. The great thing about opining is that we never get it right!
I remember one of my formative jobs at the American Public Power Association when we were running all of APPA’s technical activities. While mostly a lobbying outfit, we were after all in Washington, DC, and were perceived by many to be near some nexus of power and influence. And, because my group did technical stuff, we were a natural magnet for some inventors seeking an edge. And, believe me, some of those folks were real crackpots.
We’d hear claims how this person or that invented radar, or perpetual energy machines, or bendable concrete, and, once, even, cold fusion. Hehe.
In monitoring various ontology and semantic Web mailing lists, claims sometimes arise about the “ontology of everything” or the single, universal ontology that cures cancer or walks on one leg. It makes me smile and think about those past wild claims about radar or perpetual energy.
Some, I believe, when we mention the UMBEL lightweight subject concept reference structure, conjure up similar visions of a universal “ontology of everything”. That is wrong and not our intent. But, UMBEL is trying to straddle two different worlds and world views, and that can often lead to misunderstandings and misperceptions.
This posting is not the first and surely will not be the last on the subject, but it is worthwhile again to try to explain the role of UMBEL from these different angles and with slightly different analogies.
In prior posts, we have described UMBEL as a backbone, as a roadmap to related content, as a lightweight ontology or a lightweight reference structure, and as middleware. In this post, I am going to concentrate on its role as middleware, in its role as residing at the infocline between two different worlds and world views.
The Greek base -cline is often applied to gradual transition layers or changes in gradients or slope. A thermocline, for example, represents the layer between the deep and surface ocean. While there is mixing across this layer, it is slower than within the two parts that it separates. Both parts and the thermocline layer itself have quite different properties and temperatures, even though all are ocean and salty water.
The UMBEL infocline acts in a similar manner. On one side of the UMBEL layer is the Cyc knowledge base, with its self-contained, more-or-less closed world of higher order logics, microtheories regarding thousands of knowledge domains, rich predicates, and coherence. It is venerable, solid and proven, but with its own language and world view. Its purpose is also directed to reasoning and inference, driven from a foundation of (generally not codified outside of Cyc) common sense. It was designed well in advance of the creation of RDF or OWL, indeed in advance of the Internet and Web itself.
On the other side of the UMBEL infocline is the entire Web. This is a chaotic, decentralized, distributed knowledge environment representing untold numbers of world views. The specifications of the semantic Web and its languages and vocabularies have been designed expressly with these differences in mind and the means and structures to link and interrelate them. The Web environment — though not exactly incoherent — is also not in its ground state coherent. Indeed, it is the very purpose of existing semantic Web standards and UMBEL to help provide that coherence.
A key aspect of the Web is its “open world” assumption, defined in SKOS  as:
What this means for the Web is that we must assume that the system’s knowledge is incomplete, and that if a statement cannot be inferred from what is explicitly expressed, we still cannot infer it to be false. Adding new information never falsifies a previous conclusion, and most of what we can know about the world will remain unknown. Cyc, on the other hand, can make closed-world assumptions under appropriate conditions.
UMBEL thus must act as a mediator, or middleware, in its role as the interface between these world views. It can lead to tension and turbulence when contemplating or transiting this infocline layer.
The central purpose of UMBEL is to provide a context for relating information. Once such a purpose for context is embraced, the natural next question is: And what shall be the basis for this context?
A previous post discussed why Cyc was chosen over alternatives as this contextual basis. Ultimately, the reasons for choosing Cyc come down to real practical tools and capabilities such as helping to disambiguate the identities of named entities, mapping ontologies and schema, doing natural language processing, and the sheer provenness of the concept relationships that are at the core of UMBEL.
(Also, as noted many times, others could just as reasonably chose other bases for providing context. The important point, again, is to provide some context over no context.)
We can view the Cyc knowledge base as a complete, albeit large, world unto itself. Like the Earth, it is complex and varied and self-contained. It has its own atmosphere and perspective on the broader universe:
But, like other planets or celestial bodies, Cyc is a world, not the world. There are many different possible worlds with different atmospheres and gravities and temperatures and compositions. And, of course, Cyc is not a physical world at all, but a conceptual “world” representing knowledge and its relationships. We will, however, represent it with the Earth image below.
There is 25 years and perhaps close to 300 person-years of development behind Cyc. It has thousands of able practitioners around the world and has been used in hundreds of meaningful projects and engagements. Since its release in 2002, there have been well in excess of 100,000 downloads of its open-source OpenCyc version.
This legacy and history leads to distinct functional and terminology differences from current semantic Web perspectives. For example, the richness of Cyc predicates does not lead to simple mappings to existing OWL and RDFS properties. The notion of class is different than the closest analog in Cyc, the ‘collection‘. The concept and treatment of individuals and types is different. The 1000 or so microtheory domains in Cyc are not easily transferred or mapped to OWL constructs. Cyc uses reification aggressively to functionally combine concepts from constituent elements, such as “apple tree”. Higher-order logic is not transferable in all cases to the first-order logic (FOL) of the semantic Web. And so forth . . . .
Perhaps most importantly, however, is that Cyc has been designed, built and extended by professional ontologists and related researchers. This brings a degree of consistency and quality control that Web-broad initiatives can not hope to approach.
In our working with Cyc there has been nothing but good will and professionalism from the staff at Cycorp and the Cyc Foundation. But, there are clearly times when world view and terminology can differ, sometimes leading to translation problems and issues. Moreover, attempts to bridge from the Cyc world to the open world assumptions of the general Web means the translation is “lossy”, much like what happens in moving from a 16 million palette of 24-bit colors to something less.
|Reifiied Terms (Constants and NARTs)||263,332||303,340|
|Unique Predicates (Properties) ||17 (OWL)||~16,000|
|Disk Storage (KBs) ||495,104||566,400|
The sheer size and sophistication of either version is too great for easy comprehension and linkage by standard Web resources. Thus, the UMBEL project set out to determine and derive the most fundamental concepts from within OpenCyc. What was desired was a tractable set of subject concept “hub” nodes from within OpenCyc. A further design criterion was to maintain a 100% consistency with OpenCyc for this subset of subject concepts in order for UMBEL to preserve linkage into the Cyc knowledge base.
A subsequent post will relate in detail the nine-month (and continuing!) vetting and extraction process applied to Cyc, the result of which is currently the identification of about 21,000 subject concepts. These are schematically illustrated by the yellow dots on our Cyc Earth representation:
Of course, these yellow dots are not really physical locations on a globe. Rather, they represent important “hub” locations within the virtual Cyc knowledge “space”.
Once removed from the broader knowledge base, we now have a simple skein of these 21,000 subject concepts and their interrelations. We can show this lightweight structure as a ball of subject concept nodes (in red) connected to one another via their graph edges. We can represent this lightweight skein as follows, which has similarities to a hairnet with the nodes represented by the knots (in red) in the net:
This simplistic wireframe representation has been presented before for all 21,000 UMBEL nodes via the Cytoscape graph visualization software (see figure right; click for larger size).
The following table shows that the overall size and complexity of Cyc has been reduced by 1-2 orders of magnitude through this cleaning exercise, resulting in a lightweight UMBEL structure about 5-10% of the original size:
|Terms or Concepts||263,332||303,340||21,057|
|Unique Predicates (Properties) [2,5]||17 (OWL)||~16,000||18|
|Disk Storage (KBs) ||495,104||566,400||14,445|
Very striking is the predicate reduction, which is both a key source of “lossiness” and a challenge in maintaining a meaningful OWL and RDFS correspondence with the original Cyc. However, since the purpose of UMBEL is context and not reasoning or inference, this reduction is appropriate and understandable.
Metaphorically, we can now re-apply this UMBEL skein over the Cyc knowledge base. We have described this visual metaphor as the “hairnet over the basketball,” with UMBEL being the hairnet, and Cyc (Earth) the basketball:
Note that the UMBEL skein can act and be used fully independently from the underlying Cyc structure or not.
This UMBEL lightweight skein or wireframe structure now is ready to act as middleware, to play its role as an infocline. Each of UMBEL’s 21,000 subject concepts is, in effect, a “docking port” to which external Web data can “attach”. Once attached, this data can then be related to other Web data via the subject concept relationships in the UMBEL skein. This docking and attachment mechanism can be visualized as follows (click to enlarge):
If you mentally remove the Earth figure (Cyc) above, the UMBEL skein acts solely as a context reference structure for other Web data through its lightweight SKOS taxonomy structure (narrowerTransitive and broaderTransitive). These are the internal edge relationships of the wireframe structure with the red nodes above.
Though lightweight, this structure is surprisingly powerful in that it also enables tie-ins with external ontology classes — what Fred Giasson has called ‘exploding the domain‘ — and provides a reference context for Web data. Without these docking ports via UMBEL’s subject concepts, there is no contextual frame of reference and these Web data bits essentially tumble aimlessly in a dark knowledge space.
But one need not stop at the infocline wireframe layer of UMBEL. Because each subject concept (“docking port”) has a direct correspondence to Cyc, we can dive more deeply into the Cyc knowledge environment. First through OpenCyc and then (via licensing or other arrangements) into ResearchCyc or the full Cyc, another dimension of tools and capabilities can become available. We now have backup and support to assess mappings and assignments and inferences and reasoning.
Will everyone want such capabilities? Most will not.
But it also surely does not hurt to have these value-added pathways so readily available for use and exploitation.
Some perhaps in the Cyc community may look at this picture and say, Whoa!: We’re giving Web denizens loaded Cyc guns via the UMBEL infocline to harm themselves and others.
Perhaps so. But this is also why we have courses on firearms safety and practice ranges for gaining the experience. Ontology mapping of any nature in an open world requires attention and skill to maintain quality.
The open world circumstances have already shown challenges with sameAs assignments and will certainly be exacerbated as we extend to class mappings in ontologies and inferencing. Quality and provenance will assert their prominence. Who do you trust and who is capable? But haven’t these always been operative questions?
Some perhaps in the broader Web community may go, Whoa! We are free and independent actors who hate any sniff of possible centralized Big Brother crap. Why UMBEL? Why Cyc? I want to free-form tag and twitter to my heart’s content.
OK, well, sure. But how can the Web of data meaningfully expand without reference points, structure and context? Though we may have foundational semantic Web standards in place, if we are going to meaningfully inter-relate data, we also need context and semantics.
UMBEL and Cyc offer one set of contexts, semantics and tools. Whether they are the best or not is a matter for the market to decide. But I think it will rapidly become clear that future Linked Data that is published without context will remain largely unused data. The question now going forward is not the rejection of context but deciding what contextual frameworks work better, are easy to implement, and are readily understood.
So, I think the game has changed and I’d like to believe for the better. UMBEL has placed a marker down — and it’s smack dab in the middle.
Yes I’m stuck in the middle with you,
And I’m wondering what it is I should do,
It’s so hard to keep this smile from my face,
And knowledge, yeah, is all over the place,
Cyc to the left of me, Open Web to the right,
Here I am, stuck in the middle with you. 
In a recent blog post, Kingsley Idehen picked up on the UMBEL project’s mantra of “context, Context, CONTEXT!” as contained in our recent slideshow. He likened context to the real estate phrase of “location, location, location”. Just so. I like Kingsley’s association because it reinforces the idea that context places concepts and things into some form of referential road map with respect to other things and concepts.
To me, context describes the relationships and environmental proximities of what UMBEL calls subject concepts and their instance sub-concepts and named entity members, the whole of which might be visualized as a graph of reference nodes in the firmament of a global knowledge space.
Indeed, it is this very ‘cloud’ of subject concept nodes that we tried to convey in an earlier piece on what UMBEL’s backbone structure of 21,000 subject concepts might look like, shown at right. (Of course, this visualization results from the combination of UMBEL’s OpenCyc contextual framework and specific modeling algorithms; the graph would vary considerably if based on other frameworks or models.)
Yet in a comment to Kingsley’s post, Giovanni Tummarello said, “If you believe in context so much then the linking open data idea goes bananas. Why? Because ‘sameAs’ is fundamentally wrong.. an entity on DBpedia IS NOT sameAs one on GeoNames because the context is different and bla bla… so it all crumbles.” 
Well, hmmm. I must beg to differ.
I suspect as we now are seeing Linked Data actually enter into practice, new implications and understandings are coming to the fore. And, as we try new approaches, we also sometimes suffer from the sheer difficulty of explicating those new understandings in the context of the shaky semantics of the semantic Web.
Giovanni’s comment raises two issues:
Therefore, since UMBEL is putting forth the argument for the importance of context in Linked Data, it is appropriate to be precise about the semantics of what is meant.
What is context? The tenth edition of Merriam-Websters Collegiate Dictionary (and the online version) defines it as:
context \ˈkän-ˌtekst\ n.;ME, weaving together of words, Latin contextus connection of words, coherence, from contexere to weave together, from com- + texere to weave ( ca.1586)
1: the parts of a discourse that surround a word or passage and can throw light on its meanings
2: the interrelated conditions in which something exists or occurs: environment, setting <the historical context of the war>.
Both of these references, of course, base their perspective on language and language relationships. But, both also provide the useful perspective that context also conveys the senses of environment, surroundings, interrelationships, connections and coherence.
Context has itself been a focus of much research from linguistics to philosophy and computer science. Each field has its specific take on the concept, but I believe it fair to say that context is consensually used as a holistic reference structure that tries to put all worlds and views, including that of the observer and observed, into a consistent framework. Indeed, when that framework and its assertions fit and make sense, we give that a word, too: coherent.
Hu Yijun , for example, intersects the interplay of language, semantics and the behavior and circumstances of human actors to frame context. Yijun observes that an invariably-applied research principle is that meaning is determined by context. Context refers to environmental conditions surrounding a discourse and its parts which are related with it, and provides the framework to interpret that discourse. There are world views, relationships and interrelationships, and assertions by human actors that combine to establish the context of those assertions and the means to interpret them.
In the concept of context, therefore, we see all of the components and building blocks of RDF itself. We have things or concepts (subjects or objects) that are related to one another (via properties or predicates) to form the basic assertions (triples). These are combined together and related in still more complex structures attempting to capture a world view or domain (ontology). These assertions have trust and credibility based on the actors (provenance) that make them.
In short, context is the essence of the semantic Web and Linked Data, not somehow in variance or conflict with it.
Without context, there is no meaning.
While one interpretation might be that the characteristics of one individual (say, Quebec City) might be oriented to latitude and longitude in a GeoNames source, while the characteristics of that individual may have a different context (say, population or municipal government) in the different DBpedia (Wikipedia) source, we need to be very careful of what is meant by context here. The identity of the individual (Quebec City) remains the same in both sources. The context does not change the individual nor its identity, only the nature of the characteristics used to provide different coherent information about it.
With the growth in Linked Data, we are starting to hear the rumblings around possible misuse and misapplication of the sameAs predicate . Frankly, this is good, because I share the view there has been some confusion regarding the predicate and misapplications given its semantics.
The built-in OWL property owl:sameAs links an individual to an individual . Such an
owl:sameAs statement indicates that two URI references actually refer to the same thing: the individuals have the same “identity”.
A link is a predicate is an assertion. It by nature ties (“glues”) two resources to one another. Such an assertion can either: (1) be helpful and “correct”; (2) be made incorrectly; (3) assert the wrong or perhaps semantically poor relationship; or (4) be used maliciously or to deceive.
(Unlike email spam, #4 above has not occurred anywhere to my knowledge for Linked Data. Unfortunately, and most sadly, deceitful links will occur at some point, however. This inevitability is a contingency the community must be cognizant of as it moves forward.)
To date, almost all inter-source Linked Data links have occurred via owl:sameAs. If we liken this situation to early child language acquisition, it is like we only have one verb to describe the world. And because our vocabulary is relatively spare, we have tended to apply sameAs to situations and relations that, comparatively, have a bit of semblance to baby-talk.
So long as we have high confidence two disparate sources are referring to the same individual with the same identity, sameAs is the semantically correct RDF link. In all other cases, the use of this predicate should be suspect.
Simple string or label matches are insufficient to make a sameAs assertion. If sameAs can not be confidently asserted, as might be the case where the relation of individual referents is perhaps likely but uncertain, we need to invoke new predicates or make no assertion at all. And, if the resources at hand are not individuals at all but classes, the need for new semantics increases still further.
As we increase the size of the Linked Data ‘cloud’ or show rapid growth in Linked Data, we should be aware that quality, not size, may be the most important metric powering acceptance. The community has made unbelievable progress in finally putting real data behind the semantic Web promise. The challenge now is to add to our vocabulary and ensure quality assertions for the linkages we publish.
One of UMBEL’s purposes, for example, is to broaden our relations to the class level of subject concepts. As we move beyond the early days of FOAF and other early vocabularies, we will see further richening of our predicates. We also need predicates and predicate language that reflects the open-world nature  of public Linked Data and the semantic Web.
So, while sameAs helps us aggregate related information about the same identifiable individual, the predicates of class relations in context to other classes helps to put all information into context. And, if done right — that is, if the semantics and assertions are relatively correct — these desired contextual relations and interlinkages can blossom.
The new predicates forthcoming from the UMBEL project, to be published with technical documentation this month, and related to these purposes will include:
Assertions such as these that are open to ambiguity or uncertainty, while appropriate for much of the open-world nature of the semantic Web, may also be difficult predicates for the community to achieve consensus. Like our early experience with sameAs, these predicates — or others that can just as easily arise in their stead — will certainly prove subject to some growing pains.
Most people active in the semantic Web and Linked Data communities believe a decentralized Web environment leads to innovation and initiative. Open software, standards activities, and vigorous community participation affirm these beliefs daily.
The idea of context and global frames of reference, such as represented by UMBEL or perhaps any contextual ontology, could appear to be at odds with those ideals of decentralization. But one paradox is that without context, the basis for RDF linkages is made much poorer and therefore the potential for the benefits (and thus adoption) of Linked Data lessen.
The object lesson should therefore not be a rejection of context. Indeed, any context is better than no context at all.
Of course, whether that context gets provided by UMBEL or by some other framework(s) remains to be seen. This is for the market to decide. But the ability of contextual frameworks to richen our semantics should be clear.
The past year with the growth and acceptance of Linked Data have affirmed that the mechanisms for linking and relating data are now largely in place. We have a simple, yet powerful and extensible data model in RDF. We have beginning vocabularies and constructs for conducting the data discourse. We have means for moving legacy data and information into this promising new environment.
Context and Linked Data are not in any way at odds, nor are context and sameAs. Indeed, context itself is an essential framework for how we can orient and grow our semantics. Human language required its referents in the real world in order to grow and blossom. Context is just as essential to derive and grow the semantics and meaning of the semantic Web.
The early innovators of the Linked Open Data community are the very individuals best placed to continue this innovation. Let’s accept sameAs for what it is — one kind of link in a growing menagerie of RDF link predicates — and get on with the mission of putting our enterprise in context. I think we’ll find our data has a lot more meaningfully to say — and with more coherence.