Comments on: Exploring Semantic Means

By: Financial Times + Endeca = Newssift | The Noisy Channel

Financial Times + Endeca = Newssift | The Noisy Channel — Thu, 19 Mar 2009 00:46:27 +0000

[...] Now that the cat is out of the bag, I’m proud to tell readers here about an effort I’ve been involved with over the past few months. As reported in TechCrunch and Search Engine Land, the Financial Times just launched Newssift, a semantic search engine, powered by Endeca, that sifts through business news. Regular readers may recognize the application from an example I used in my presentation on exploring semantic means. [...]

By: MarkH

MarkH — Thu, 12 Mar 2009 18:13:53 +0000

I went on an Autonomy course many years ago. The mantra was keywords are “legacy” and that we should use example documents/passages to express intent.
As you suggest, I felt this was like trying to have a conversation with someone where you couldn’t speak for yourself but had to select a book from a shelf that best represented your intent.
I think the technique can have its place but I see issues trying to combine that with faceted result UIs.

By: Daniel Tunkelang

Daniel Tunkelang — Thu, 12 Mar 2009 18:01:22 +0000

For the user. My information needs usually aren’t in document form.

By: MarkH

MarkH — Thu, 12 Mar 2009 17:56:01 +0000

>>it strikes me as making the problem harder than it has to be

Harder for who? The user trying to specify the information need, the system that processes it or the user interpreting results?

By: Daniel Tunkelang

Daniel Tunkelang — Thu, 12 Mar 2009 17:50:22 +0000

I’m not morally opposed to it, but it’s not my preferred entry point into the information seeking process–at least not with a browser-based search engine. In general, it strikes me as making the problem harder than it has to be–turning the query into a similarity query on a document. I see that approach being more useful when it is later in the process, e.g., I like this document, show me more like it. Better still when the basis for similarity is transparent.

By: MarkH

MarkH — Thu, 12 Mar 2009 17:43:54 +0000

>>I assumed you[ were].. concerned about the ability to satisfy information needs like the one you described

Yes, but specifically using the retrieval technique I suggested- “given an example document”. Do I detect that you do not subscribe to that approach to specifying queries?

By: Daniel Tunkelang

Daniel Tunkelang — Thu, 12 Mar 2009 17:28:10 +0000

Indeed, and you didn’t even have to wait a month! I assumed that you weren’t concerned about the specifics of bank bonuses, but rather about the ability to satisfy information needs like the one you described. And I hope the interface there does take a step in the right direction.

You’re right that they aren’t showing numbers–that’s a design choice, but the technology is based on actual results, not something fuzzy. I’m not entirely sure how they are ranking them–but there’s no rule that you have to rank facet values based on their frequency in the result set, particularly if the results involve a search that favors recall over precision. That’s a topic for broader discussion, and I promise to post about it.

By: MarkH

MarkH — Thu, 12 Mar 2009 13:20:51 +0000

Hi Daniel,
So is slide 25 the “ask me in a month..” answer to my original question? http://tinyurl.com/c6y2p4

It’s certainly nice, but my original question was less about the specifics of bank bonuses and more about how fuzzier styles of query (using example documents as queries) can sit with a faceted search interface when all results are far from being equally relevant.
I suppose your start query could have been a document and not a keyword but the key here (unlike in previous faceted interfaces of yours I’ve seen) is NOT to show any totals for the number of matches in suggested groups. The numbers would be meaningless because they imply that you have either considered all results (including the very low-similarity matches) or you have introduced an arbitrary relevance cut-off that makes a nonsense of any precision in group totals.
Is the technology you show here exhaustive or sample-based in its grouping of results? The former would preclude a fuzzy “like this doc/para” style of query.