The Future of�

�Search

The Web

Improving Search�

The same, only more of it

Faster spiders

Bigger indexes (and directories)

More Boolean operators (like Lexis-Nexis)

The same, but improved functionality

Better query processors

Better translation engines

Faster delivery of Web pages, e.g., Inktomi

Improving Search:� Inktomi

Started at UC-Berkeley in 1996 by an engineering professor & one of his graduate students as a faster spider+query processor

Just another Web site, until they realized they could speed up delivery of pages by caching

Caching is storing of frequently-used information�

Inktomi�s secret:� Caching

Caching is storing frequently-used information:

on your PC

on a Web server

at a search engine

in a whole corporation

in the �server farm� of an ISP (e.g., AOL)

Inktomi�s real secret: Turning cache into cash

now owned by Yahoo!

Improving Search for Text

Natural-language queries

This is the AskJeeves model on steroids

Applied more narrowly in certain fields, e.g. finance and investing: http://www.iphrase.com

Understand the user � keywords, phrases, sentences, and extended prose including e-mails and IM sessions. But these can be incomplete, obscure, or �dirty� (e.g., misspellings)

Improving Search for Text

Natural-language queries

This is the AskJeeves model on steroids

Applied more narrowly in certain fields, e.g. finance and investing: http://www.iphrase.com

Understand the user

Search on context, not �silo�

Guide the user � dynamically build a page on-the-fly around search terms, not just a list of hits

Improving Search for Text

Natural-language queries

This is the AskJeeves model on steroids

Applied more narrowly in certain fields, e.g. finance and investing:� http://www.iphrase.com

Blue Cross/Blue Shield

TD Waterhouse

Lexis-Nexis (!!!)

Motorola

Purchased by IBM announced 11/1/05

Improving Search for Text

Text mining � also called �context analysis�

But to understand Text mining, we need to look at Data Mining

Improving Search for Text

Digression on Data mining

Example:� American Express special offers

As computing power and databases improved, AE could compare buying patterns and choose cardholders who might buy, say, a leather calendar

What began as a way to identify a few hundred customers, soon found a few tens, then a few�

And finally, circa 1995, a target population of one!

Improving Search for Text

Text mining � analogous to Data mining

Compare �word patterns� instead of buying patterns

Examine unstructured data (outside DBs) such as email, corporate portal (Web) pages, help files, Word documents, Excel spreadsheets, etc.

Identify patterns, hence information and knowledge, that the company didn�t know it knew!

One of the tools of �knowledge management�

Improving Search for Text

Text mining the Web ????

The Web is a giant collection of unstructured pages

Could text mining find knowledge that the Web (and directories and indexes) doesn�t know it has?

This would be the equivalent of asking, �What can you tell me about this topic that I don�t know to ask?�

Would it be done within an index?� Or by an outside user with a text mining application program?

Improving Search for Text

Text mining the Web ????

There are no answers at the moment, but there are some prominent companies with software that might be players:

ClearForest ClearTags (www.clearforest.com)

Entrieva�s SemioMap (www.entrieva.com)

Inxight�s Categorizer (www.inxight.com), which can also work with data�

Improving Search for Data

Structured information � some of the �stuff� that neither a typical search engine nor you, as a typical searcher, is equipped to handle�

Financial reports (public-domain)

Sports statistics

Web-based tools could give access to these kinds of data that is currently only possible for users who know highly specialized query languages.

Improving Search for Images

Remember:� current search capability, such as is offered by AltaVista and Google, is still largely based on text associated with the images:

Filenames, e.g., mydog.jpg

�alt� attributes within an image tag

The text on a Web page near where the image is displayed.

Improving Search for Images

Wait a minute!� Why not just turn loose all this computing power we have?� Let some server somewhere decide what the various images are.

Easier said than done, because unfortunately�

Improving Search for Images

A computer cannot distinguish between this:

And this:

although it is obvious to us.

Improving Search for Images

Technologies to tackle this problem have been around for some time and at least one, from Virage, was available on AltaVista in 1996.� It allowed the user to vary and try to match:

Color � general color impression

Composition � spatial arrangement of color

Texture � �patterns,� e.g., wood, granite, clouds

Structure � shape of objects in the image

The Virage Image Engine

Improving Search for Images

The problem to be solved is that image recognition is visual, not semantic.� Yet the thing we can do best on the Web is word-based searching.

Said differently: We still cannot distinguish an orange on a green tablecloth from a basketball on green grass.

But the problem is under attack�

Improving Search for Images

Three of the current players in the field are:

Pixlogic (www.pixlogic.com) - object-level info

BioImagene (www.bioimagene.com) - content

Visfinity (www.visfinity.com) - image management

Connections and Context

Three alternatives to conventional search:

Alexa (www.alexa.com)

Intelligent agents

Mapuccino (from IBM)


	The same, only more of it
		Faster spiders
		Bigger indexes (and directories)
		More Boolean operators (like Lexis-Nexis)
	The same, but improved functionality
		Better query processors
		Better translation engines
		Faster delivery of Web pages, e.g., Inktomi


	Started at UC-Berkeley in 1996 by an engineering professor & one of his graduate students as a faster spider+query processor
	Just another Web site, until they realized they could speed up delivery of pages by caching
	Caching is storing of frequently-used information�


	Caching is storing frequently-used information:
		on your PC
		on a Web server
		at a search engine
		in a whole corporation
		in the �server farm� of an ISP (e.g., AOL)
	Inktomi�s real secret: Turning cache into cash
		now owned by Yahoo!


Natural-language queries
	This is the AskJeeves model on steroids
	Applied more narrowly in certain fields, e.g. finance and investing: http://www.iphrase.com
		Understand the user � keywords, phrases, sentences, and extended prose including e-mails and IM sessions. But these can be incomplete, obscure, or �dirty� (e.g., misspellings)


	Text mining � also called �context analysis�

	But to understand Text mining, we need to look at Data Mining


	Digression on Data mining
		Example:� American Express special offers
		As computing power and databases improved, AE could compare buying patterns and choose cardholders who might buy, say, a leather calendar
		What began as a way to identify a few hundred customers, soon found a few tens, then a few�
		And finally, circa 1995, a target population of one!


	Text mining � analogous to Data mining
		Compare �word patterns� instead of buying patterns
		Examine unstructured data (outside DBs) such as email, corporate portal (Web) pages, help files, Word documents, Excel spreadsheets, etc.
		Identify patterns, hence information and knowledge, that the company didn�t know it knew!
		One of the tools of �knowledge management�


	Text mining the Web ????
		The Web is a giant collection of unstructured pages
		Could text mining find knowledge that the Web (and directories and indexes) doesn�t know it has?
		This would be the equivalent of asking, �What can you tell me about this topic that I don�t know to ask?�
		Would it be done within an index?� Or by an outside user with a text mining application program?


	Text mining the Web ????
	There are no answers at the moment, but there are some prominent companies with software that might be players:
		ClearForest ClearTags (www.clearforest.com)
		Entrieva�s SemioMap (www.entrieva.com)
		Inxight�s Categorizer (www.inxight.com), which can also work with data�


	Structured information � some of the �stuff� that neither a typical search engine nor you, as a typical searcher, is equipped to handle�
		Financial reports (public-domain)
		Sports statistics

	Web-based tools could give access to these kinds of data that is currently only possible for users who know highly specialized query languages.


	Remember:� current search capability, such as is offered by AltaVista and Google, is still largely based on text associated with the images:
		Filenames, e.g., mydog.jpg
		�alt� attributes within an image tag
		The text on a Web page near where the image is displayed.


	Wait a minute!� Why not just turn loose all this computing power we have?� Let some server somewhere decide what the various images are.

	Easier said than done, because unfortunately�


	A computer cannot distinguish between this:

	And this:






	although it is obvious to us.


	Technologies to tackle this problem have been around for some time and at least one, from Virage, was available on AltaVista in 1996.� It allowed the user to vary and try to match:
	Color � general color impression
	Composition � spatial arrangement of color
	Texture � �patterns,� e.g., wood, granite, clouds
	Structure � shape of objects in the image


	The problem to be solved is that image recognition is visual, not semantic.� Yet the thing we can do best on the Web is word-based searching.
	Said differently: We still cannot distinguish an orange on a green tablecloth from a basketball on green grass.
	But the problem is under attack�


	Three of the current players in the field are:
	Pixlogic (www.pixlogic.com) - object-level info
	BioImagene (www.bioimagene.com) - content
	Visfinity (www.visfinity.com) - image management


	Three alternatives to conventional search:
	Alexa (www.alexa.com)
	Intelligent agents
	Mapuccino (from IBM)