Well, there’s been a bit of a gap since my last post, but no gap in my thinking about what to post on! Back in mid-June I hosted a panel discussion at the SemTech conference with colleagues and friends Jeff Jonas and Marc Davis. Our (unstructured) panel was entitled “True Semantic Reasoning: Self-Constructing Ontologies.” The discussion was fun, but in the feedback from the panel there was confusion on why we focused our discussion on context-awareness and not on the self-constructing ontologies (as implied in the title of the panel).
I’d like to explain!
The premise of our discussion was that Semantic reasoning in the absence of preexisting ontologies is done all the time, in fact has to be done in most cases today as ontologies simply don’t exist for most domains. In the absence of formal ontologies, it is still possible to create semantic graphs, albeit with significantly less inherent structure than a formal ontology. In the case of our technology at Atigeo, the graph doesn’t express classes of relationship. Without formal ontologies, reasoning can be achieved by using contextual information to help collapse the path in the semantic graph down to one that is less ambiguous or even completely unambiguous for the reasoning computation.
Think about how you reason to the name of the person I am describing;
· An Irish man
· Somewhat of a political activist
· Interest in 3rd World Charities
· Often seen wearing yellow sunglasses
· Is in a rock band
· Is friends with The Edge
Well, the person in question is Bono. Of course in the presence of a music ontology we could easily perform the reasoning needed, although we may find some of the pieces of information in the description above are tertiary to a music ontology and as such not understood in the semantic reasoning process. "Yellow sunglasses," for example, tends to trigger recognition with many people I try this example with.
The type of self-constructing ontologies we have been working on leverage large volumes of unstructured data related to a domain, in order to discover association-based relationships between the concepts expressed in that domain. The approach generally induces a large and widely connected graph associating concepts which can be queried by defining a list of concepts that contextualize the query. In the case of the example above, the list of concepts would include: Irish, political activist, 3rd World Charities, yellow sunglasses, rock band, The Edge, and U2. This contextualization allows us to collapse the modeled semantic ambiguity in the problem (the complex graph on the left in the image below) into a much simpler graph on which inference can be performed (on the right, below);
Another compelling example can be built around Apple’s new 4G iPhone. Self-constructing ontologies could be leveraged to assist customer support specialists in determining when an incoming complaint is associated with the “death grip” issue – one that has a lot of ambiguity in how users will describe it. In this case a list of concepts that could contextualize the query would be: antenna, dropped call, hand position, and phone case. It would be labor intensive to keep a traditional ontology up to date with the latest breaking information regarding this issue.
On our SemTech panel, Jeff Jonas spoke similarly about how increasing contextualization allows technologies to resolve ambiguity in identity or to connect multiple, apparently distinct identities. This is akin to discovering new content expressing a new semantic relationship in the graph, allowing previously isolated sub-graphs to be connected. Jeff writes very eloquently about this in his blog post “Puzzling: How Observations Are Accumulated Into Context.” Associating “death grip” and “dropped calls” is a great example of this. Persisting the context becomes the persistence of the ontology expressing current knowledge of the problem domain.
Mark Davis illustrated this same concept in the space of location information, with the notion that location-based context implies a certain semantic conceptualization (as illustrated by Yahoo! TagMaps). Essentially, tagged and location stamped images uploaded to Flickr generate semantic associations with locations, which embed the concept of location into an ontology.
So the topic of our panel was relevant because self-constructing ontologies must leverage context awareness in order to discover association-based relationships between the concepts in a particular domain. Additionally, real-time maintenance of such ontologies relies heavily on contextual relevance because they do not have inherent structure.
As the demand for mobile access to information and services continues to increase, true contextual awareness will become a more pervasive requirement for ontology utility (scalability and real-time capacity). According to Gartner,
"By 2015 context will be as influential to mobile consumer services and relationships as search engines are to the Web. Whereas search provides the ‘key’ to organizing information and services for the Web, context will provide the ‘key’ to delivering hyperpersonalized experiences across smartphones and any session or experience an end user has with information technology. Search centered on creating content that drew attention and could be analyzed. Context will center on observing patterns, particularly location, presence and social interactions. Furthermore, whereas search was based on a ‘pull’ of information from the Web, context-enriched services will, in many cases, prepopulate or push information to users."