Elevator Pitch Friday: Truevert, The Green Search Engine

Getting someone to try a new search engine is not easy. In this Friday’s Elevator Pitch, Herbert Roitblat tries to entice you to try his new green search engine Truevert by asking, “Are you average?” Google, Yahoo, and Microsoft Live, he contends, give you average results. Truevert gives you special results—everything comes back through green-colored glasses. So a search for “SUV” brings back HybridSUV.com as the top result. A search for “building materials” brings back results for green building materials.

But what is really special about Truevert is not that it is a green search engine. It is a Yahoo BOSS search mashup. Truevert is actually just a demonstration of some powerful underlying semantic technology developed at OrcaTec, a company co-founded by Roitblat and Brian Golbère. Truevert gets its search results from Yahoo BOSS and applies its own text-analyzing software to generate the most relevant green results. OrcaTec’s software could just as easily be used to create a fashion search engine, a startup search engine, or any of a thousand other vertical search engines.

In an email, Roitblat explain how his technology works:

Our approach is to mimic the way that people learn language. When people learn a new word they learn its meaning by it relation to the other words in the sentence or paragraph. . . . Even if you learn it from a dictionary, its meaning is still from the context of other words. Similarly, when people understand a sentence, each word in the sentence helps to disambiguate the other words in the sentence. For example, consider the sentence, “the tree surgeon examined the young man’s palm.” By the time you get to the word “palm,” you have a pretty good idea what that word means.

Our system also learns meaning from the context. We provided the system with a set of green documents. These documents are broken into paragraphs and the word relationships within each paragraph are computed using our patent-pending modeling software. Each word in the paragraph becomes the context for the other words in the paragraph.

Then, when a user submits a query, that query is transmitted to the model and a set of additional terms, that are most closely related to that query are generated. All of these terms are then submitted to BOSS. The snippets that come back from BOSS are then re-ranked by according to the match between the snippet and the expanded and weighted list of query terms. The result is a set of pages that match the query in the context.

The big difference between the semantic technology underlying Truevert and those used by other semantic search engines such as Hakia or Powerset is that Truevert’s does not require massive and unwieldy ontologies, taxonomies, or even a thesaurus. It tries to learn the meaning of words from the text itself and the surrounding words. Roitblat distances Truevert’s approach from other semantic search technologies and explains their limitations:

Other semantic search engines may be based on 20 or more years of ontology building. Ontologies capture only the words and relationships that their designers think are important. They are usually limited to a single language and require substantial effort to extend to a new language. In contrast, the Truevert system learned about the green vertical in well under an hour. The technology works in any language because it learns the meaning of words directly from the pages that it reads.

Verticals can be as broad as you like, for example, consumer goods, travel, or as narrow as a single person’s interest. Because construction of the verticals can be automated, there is no intrinsic limit to the number that could be created very quickly. Users of general search engines can be offered a choice of verticals in the same way that they are offered a choice of related searches.

Other companies have offered what they call semantic search based on the semantic web approach. The difficulty with this approach is that it requires some person to actively determine what category or categories a document belongs in. (Or it requires machine classification.) But this is similar to the situation search engines used to have with meta tags. It was too easy for people to cheat and assign pages to categories that were inappropriate, but profitable. We don’t know what will stop a similar descent into chaos for an RDF framework. The Truevert approach is more resistant to such cheating because it depends on the actual content of the pages, rather than on someone’s description. It does not rely on an author’s honesty or reliability.

Human categories do not correspond to the kind of things that are likely to be represented by semantic web tags. People make up new words (the Jabberwocky effect, e.g., “podcast”) and use old words in new ways (the Humpty Dumpty syndrome, e.g., “twitter”). In short, vocabularies are constantly growing and human categories of meaning are always changing, so ontologies are always behind.

The semantic web approach also relies on fitting words into categories that are stable from time to time and from person to person, but people change how they categorize things based on their current context and needs. A given word or object can belong to an infinite number of categories. For example, what categories does a basketball belong in? Round things, bouncy things, brown things, etc. How long is this list? Do you ever reach a point where no one could add another category to it? Things with tiny dimples? Things that my brother hates? Things that Barack Obama likes? Things that float? For this last one, imagine that you are on a sinking ship, in this context that category is important and obvious.

The question is: Does Truevert do a better job? Check it out and tell Roitblat what you think in comments.