— TONY SHAW, OLIVER (OLLY) DOWNS
Tony Shaw: Hi Olly, so to get us started, could you give me a high-level overview of what Atigeo does?
Olly Downs: Absolutely. Atigeo’s platform, xPatterns, enables enterprises to derive insights from large, disparate sources of unstructured data. In doing so, we’ve taken two approaches toward how our platform is productized. The first approach is around aspects of our core technology, which allows us to build simple ontologies for domains of unstructured data and then act upon the understanding of the data – in response to queries or profiles of entities, for example. The second aspect of our product enables enterprises to provide customers the ability to access and manage their profile or persona.
Tony Shaw: Looking at the first slide here, Challenge with the Growth of Unstructured Data, certainly I’m well aware of that broad problem as I do a lot of education in the large scale enterprise data management space. Talk to me please about what your solution is.
Olly Downs: In doing our discovery with enterprises, what they told us was that they absolutely believe in the potential of their unstructured data. They’re continually getting pressure to understand their own repositories of unstructured data, and now talk of a huge push, particularly when it comes to market intelligence in understanding real time, large volumes of unstructured data in social media as an example. The frustration we’ve found our potential customers having with natural text analytics and natural language processing systems is, for companies to extract information from large bodies of unstructured data, they already had to know what it was they were looking for and be able to codify that into semantic rules or a rich ontology for the specific domain of interest.
In the case of one of our customers, current approaches required them to create up to sixty different ontologies. And of course, having purchased those ontologies and semantic rules and dictionaries, you continually have to keep them up to date, which require human domain experts. To give the example of medicine, as you know, there are 19 different publically funded and curated ontologies in medicine. There’s value in performing the effort when it comes to medicine or when the product is ontologies and semantic rule sets for law. However, when you get to more informal knowledge bases, say music, movies, and social media, there aren’t formal curated ontologies and the effort put into creating them does not scale well.
Of course, the whole push of the semantic web is that there are open, editable, and manageable ontologies and open linked data. But as things are today, most ontologies that are available are incomplete or don’t have the volume or amount of traction that could really make them fully effective for users to rely upon. So to help us get solutions to the market faster we have taken a hierarchy-free approach to the creation and maintenance of ontologies.
Tony Shaw: As we look at slide 2 could you explain how you created this new approach to ontologies?
Olly Downs: The technology came about here at Atigeo as we attempted to understand expressed preferences of a given user in the simple domain of sports.
In our collective brainstorming, we threw it on the whiteboard and thought, ok, we’ve got different sports, sports leagues, sports teams that play at different sports venues, with different owners, managers, and players. One could easily create a taxonomy or an ontology on professional sports and curate that. Then of course, we expanded the conversation and thought, what if I wanted to look at a player through their career. Ok, well we’d have to create a new ontology specific to this player or we have to actually expand the ontology significantly to understand that player through time, across teams that they’ve played on throughout their career and so on.
And so we quickly realized, of course, that for other companies that want to understand many different aspects of an end user, we would need to come up with a technology that would allow us to scale in terms of generating ontologies. We couldn’t hire a domain expert for every single element of human concern or interest customers might want to express. So we came up with this approach for automating the discovery of ontologies from unstructured data.
The difference about the ontologies created by our technology is that they are not rich in terms of semantic types and they are generally hierarchy free. The kind of typing relationships generated in our model are “is associated with” relationships rather than “is a maker of” or “is a member of” types of relationships. However, we built the technology such that, we can imbed existing ontologies and semantic rules right into it. We also leveraged our partner IBM’s technology LanguageWare to help us do so.
Now what’s unique about our technology in addition to creating these ontologies automatically is that we can dynamically maintain them, and we maintain them by subscribing the technology to real time feeds of unstructured text based on a given domain. Persisting the example of sports, we subscribe the technology to a number of different sports authorities out there on the web including some community forums, blogs, as well as the major media publishers around sports and accumulate that data in real time. What that means, continuing with our sports domain example, is when a player changes teams, or is on the injured list, or when scandals emerge we can continually add those concepts and understand the relationship to current contexts in the ontology. On top of that dynamic updating, our technology has the capability to learn and refine in real time; finally, in an end user setting we can actually use crowd sourcing or learning feedback from individual end users to de-noise our representation of the ontology of a given domain.
Tony Shaw: Could you elaborate on what it means to de-noise?
Olly Downs: We discover an ontology by determining numerous statistics of the underlying unstructured data, which can result in a level of noise that you might not have if you had a domain expert created ontology. So what we do is, allow our ontology to make use of many possible domain experts (particularly the domain expert of mass action) and allow our domain ontology to adapt and refine in real time. And we’ve solved the issue of being able to do this in a learning model at scale. It’s very common today to implement already built machine learning methods at run time at large scale. What they do not do at large scale at run time is actually adapt the index. Usually the index is built offline and then snapshotted onto a search engine web farm. What our technology does is actually adapt the model that’s running in real time based on user-interaction and feedback. That allows us to refine, for example, ambiguous concepts like the association between Giants and the baseball team, Giants and the football team, and Giants in fiction.
Tony Shaw: So the competitive advantage or the unique selling point, whatever the appropriate terminology here is, seems to be that you can automate the creation of the ontology with a relatively modest expertise in ontology building. Is that a reasonable statement?
Olly Downs: That’s a very reasonable statement.
Tony Shaw: As we both know, given the dearth of expertise in this field, it is difficult and time consuming to get anywhere. But one of the things that I run into a lot when particularly talking to classic enterprise data folks is they say, well, a lot of semantic capabilities like this can be done with traditional technologies. Why make the switch? What is to be gained here?
Olly Downs: I’ll answer that with another good example that comes up. We used it with one of our telecom partners in a CRM situation, which is, they have a number of attributes of structured data that they understand for predicting churn such as dropped calls, current device and plan, tenure relative to their term commitment lifecycle, and then they usually have some unstructured data that they just don’t know how to act on.
Say for example I bought the White Room ringtone, a song by Cream, on T-Mobile, the carrier that has the Fender My Touch. Normally, how would a service representative target customers with an offer for the Fender My Touch edition Android phone? Well, they’d probably pick some demographics of people who might be interested in classic rock music. However, they can’t query directly for classic rock, unless they have all of their ringtones annotated with a genre and type of music and so on. However, understanding that I bought the White Room ringtone (it’s a song by Cream, Eric Clapton is a member of Cream, and plays a Fender guitar rather prominently) allows the customer service rep to make a direct, or at least computationally meaningful association between me, Cream, and My Touch. There’s no way they could have done that by querying a database for people who like classic rock (as is the case with using traditional technologies).
Tony Shaw: So you said the ontologies that you discover are generally relatively simple and non hierarchical. That sounds like it would have some limitations compared to something that’s more, shall we say robust? Maybe you could compare and contrast those two styles.
Olly Downs: Good question. Our approach to ontologies separates establishing semantic relevance from the set of structured operations that come with filtering by type. So if you’re thinking about answering a semantically posed question, you’d not only be looking for results that have a high assessed associated relevance with the query, but also the sub-concepts in the response to that query that are related to the type implied by the question. So with our technology, it postpones the action of filtering of a semantic type to after our system has responded with the associated concept set. It pushes that structured operation outside of the query.
Tony Shaw: Let’s move further into the technical explanation here.
Olly Downs: So the way we build our model is that we essentially allow a new loopy graph to be created that relates concepts to content, and then in the case of an ontology, creates a mapping from the content back to concepts again.
The way we actually create relationships in the loopy graph is by unfolding a single possible loop to a retrospectively simple neural-network model. This allows us to represent concepts and their relationship to other concepts, as intermediated by documents spanning the domain we’re talking about, whether movies or sports, etc. By creating this simple neural network structure, we can actually learn to find the relationships in a straight forward manner, in real time. Additionally, we can initialize this representation just by using a set of documents that we believe span the semantics of a particular domain that we’re trying to reflect.
In the case of a baseball ontology, one loop in a graph might be from Barry Bonds to steroids and steroids back to Barry Bonds, this would be one very simple closed loop. We resolve this general, cascaded set of looping relationships—Barry Bonds to steroids to San Francisco Giants to Jose Conseco and so on—by representing any particular loop or any particular edge between two concepts that actually might recursively describe a loop. I can have a term or a concept related to another term, and I’m going to establish the value or strength of that relationship (by a non-linear transformation of the co-occurrence of those two concepts) in the set of documents I’m using, as a proxy for describing the knowledge of that domain. So that single loop describes one part of the domain, tying your concepts on the left (input) of the neural network to another term or concept on the right (output) of the neural-network.
What we’ve done is try and flatten out this cascading network of looping relationships into a very flat structure, input-output type of structure, which is amenable to machine learning and then to recreate the semantic graph or re-derive the ontology, we actually fold this neural network back up, because the output terms express the same dictionary as the terms that are on the input side, and we ignore the self-associations. We ignore the association of Barry Bonds to Barry Bonds; we can leap through this neural network many times, from left to right. This path through the graph describe a Markov random field, which means we can essentially interpret, through the weights in the neural–network, the probabilistic association of one concept or term to another.
Tony Shaw: Is this the underlying science to all of your product categories or specific to any in particular?
Olly Downs: This is at the very heart of all of our products and deriving key knowledge from this takes us out into a number of different related technologies such as topic discovery, such as inference and cold-start recommendation (recommendation when you know very little about the entity you’re trying to recommend for), and so on. It also allows us to interpret and understand user profiles when the profile attributes for that user are expressed in a completely unstructured way.
Tony Shaw: So maybe I could ask you to elaborate a little bit on how you approach the opportunities and challenges of customer data?
Olly Downs: What we’ve found is that there are several bodies of customers that have consumer data, but they have been unable to unlock it until recently because of concerns about privacy and privacy management. What we observed is that there are companies in various industries (a good example today is Telcos) who are being dis-intermediated from their customer base by enterprises observing their end users, completely bypassing their technology. At the same time, there are relatively strict regulatory frameworks that prevent them from leveraging their own customer data.
Tony Shaw: Could you give me an example of what you mean by dis-intermediation?
Olly Downs: A good example would be Google, of course, on the Android phone, you can’t actually use the android phone without signing into Google. And as a result, any and all of the activity you have on that device becomes part of your Google profile whether that’s explicitly visible or not.
Now we were thinking of the consumer space, and we had this notion of trying to create very intuitive experiences for end users that reflect a genuine understanding of who you are and what you do, and of course to be able to do that, you need to be able to have very rich user profiles.
There’s never anything better than having end users explicitly tell you who they are across a multitude of dimensions, but of course even users who aren’t that concerned about privacy generally don’t put in that much effort to do that. This is of course the world that behavioral advertisers are in now, where they end up “spying” on users and maintaining third party cookies to track their behavior and so on.
What we realized was that to accumulate such rich user profiles, to create intuitive experiences, we needed to be able to facilitate the accumulation of that information based on implicitly observing behavior and actually expose the findings of that behavior to the user to whom they related.
The interesting differentiation here is, in the last six to twelve months, I’ve been able to look at my Google search history and actually delete records. However, when you go out and play with that solution, what you find is, you can’t see what knowledge was derived from an element of your search history. While you can remove a transactional record, you have no knowledge of the insight that has already been derived about you that that particular transaction generated from Google’s perspective.
What our technology enables enterprises to put in the hands of end users, is the ability to see the knowledge which has been derived about them, control the accuracy of that information, and also manipulate whether they want it in their profile and who should be able to act on that information.
Tony Shaw: Atigeo seems to be tackling many different markets simultaneously, and that seems to be a lot for a relatively new company to be taking on. As I was looking over your website in advance, you seem to have solutions in the privacy space, in the advertising space, in the enterprise integration space, this particular example you showed today seems to have business intelligence type applications and search based applications. What will you be emphasizing when you come to the SemTech conference next week?
Olly Downs: That’s a great question, Tony. The challenge and opportunity of building a technology platform is you always get a lot of pressure to demonstrate vertical applications on that platform. What we’ve done is focus our direct sales opportunities on developing some of those with customers currently. At the same time, we’ve been able to partner with a number of significant OEMs, some announced, some not announced, and one of our partners is IBM. IBM has helped us traverse a number of different verticals. We have a direct sales focus today in the telco space and in the advertising space. We have some CRM and BI style initiatives that have come as a corollary to the traction that our consumer product has in the telco sector, and a lot of our BI and CRM work has actually come in through our OEM relationships.
Our participation at SemTech will illustrate our solutions for leveraging unstructured enterprise data and enabling enterprises to provide customers the ability to access and manage their profile or persona.
Tony Shaw: You talk about platforms, then, would you like people to think of you principally as a platform provider for others to build on?
Olly Downs: Absolutely. We have built a web services-based platform that allows straightforward integration into a loosely coupled architecture that can enable a multitude of applications and solutions.
Tony Shaw: Great, thank you, Olly. We’re looking forward to seeing you present at SemTech next week.