Excellent clarification on semantic search

Today I received an update on one of te discussions on Semantic search on LinkedIn. Charlie Hull put up an excellent example on how semantic search works. This has to do with the capabilities of the search technology that is used in a specific situation, but also with the fact that a search application has to engage in a dialog with the user to assess his meaning or context. This has to do with the fact that most users just use 1 to 3 words to formulate a query. There's not much you can do with such a query in the first try. But... the search application has to pick up on those keywords and try to make something out of it.
The next step is to try to ask the user what he means.

Semantic search technology - does it actually exist? 33 comments »

Started by Charlie Hull

At Expert System we have been building semantic search systems for 20 years. Here is what we learned in serving 100's of corporate customers. A semantic search system must establish and store the CONTEXT of content. Then you need an interface to choose the CONTEXT you would like so a match can be made.

Establishing CONTEXT means the following processes must be followed. 1). word morphology (e.g. stems), 2). word roles (e.g. nouns, verbs, etc.) 3). word logic (e.g. subject - verb - object reduction) and 4). sense disambiguation (e.g. assignment of a definition for each word based on the best fit from available alternatives and in the context of the rest of the sentence(s). All 4 of these methods require the use of a semantic network that is both broad - covers the majority of the language to be used and deep - has many ways in which words relate to one another.

With the above approach will you reach a precision ("accuracy") and recall ("completeness") in search beyond the 80% mark. With further customization a 90% mark is easily achieved. Systems that rely on statistical / heuristic methods typically fall far short of these benchmarks. This is true since statistical / heuristic methods cannot fully establish logic and disambiguation.

Finally the interface must be constructed in a way that allows the user to tell the system what CONTEXT the query is in. Full natural language questions using the above methods can do this automatically. But the reality is we live in a 1-3 query word world. So allowing the user to select the word sense of one or more of the query words gives the system much more "to chew on" and is not generally an intrusion for the user. Similar user interface interactions include showing categorical, domain, people, places, organization outcomes from a search which are clickable, showing lists of semantic triples (subject-predicate-object) from which to choose, etc. All of these are at most 1-2 more clicks than a normal keyword based search but improve the experience immensely.

Such interfaces also allow what we call a 3-step walk through search where step 1 is about precision - less of a list, step 2 is an expansion of concepts - to include things related but that you did not know about, and step 3 another step of precision. This "ratcheting" effect therefore begins to bring into the Enterprise Search function other important aspect of corporate work like discovery, exploration and analysis. http://www.expertsystem.net By Brooke Aker

Semantic Search Engine: Inbeta... as in "not Alpha"?

Through a discussion on LinkedIn about real life examples of semantic search, I was pointed to the existence of Inbeta. I company with a curious name because it says that the company has a "Bèta"-status. I can't image what that says about there offerings.
But now for the offerings of the company. On their product page they have many products listed.
Of course the first one caught my eye because a "semantic search engine" is something that everybody dreams of. Imagine a search engine that gives you insight and context regarding the query of the user in relation to the information at hand and maybe also on external resources by using the sematic relations between information....

But wait... Before you think I found the holy grail of search, The sentence

"Natural Language: user will not need to search for keywords anymore, our Semantic Search understands the aim of every search query and suggests results that are relevant, thus increasing cross-selling and saving customer care costs"

had my feet put on the ground again.

This proposition on using natural language as query input and giving back relevant results based on the combinaton of words that most likely exist in the available search index, is something that has been here for years. Autonomy has marketed that concept with the name Meaning Based Computing. It all revolves around the concept of terms and weights withing documents and in relation to the words in the entire index (corpus) and matching the queried words to these calculated figures.

For a serious search engine a regard this technique almost as a must have.

But, back to the semantic side of this... Where is it?

When you want a good example of what semantics can do within a search application, take look at http://www.freebase.com/view/en/barack_obama.

It has everything to do with the context of the concepts that can be derived from a query. People have roles and jobs, names can be linked to artists, historical data etc.