I've been seeing this for plenty of multi-word phrases. Search 'hip hop music' in Google. It gives you these refinements at the bottom:

music
hip hop music history
hip hop music online
hip hop songs
hip hop artists
bet
hip hop radio

So how is Google a semantic engine?

"the purist view of ‘The Semantic Web’ where everything is tagged and filed neatly into nice little packages."

good:

"The Google API could conceivably be used to tag every single word on a page, creating a page that covers every single keyword possibility."

So, I take it you have problems with the nice little packages? Doesn't every categorization system including Google's ultimately have similar problems?

In any case, I have trouble seeing ways that having a coherent, consistent system for tagging would be a bad thing since that does not block other systems for doing things.

On a related matter, you're right to say bad things about yourself for using the term Web 4.0. I have yet to encounter a serious usage of Web 3.0 or any other such variation that spoke to a shift of the depth and complexity of Web 2.0.

Not to sound like a religious nut or anything!

I agree with your insight on Google and your generalization of semantics as sense information. I am not very impressed with Hakia as there is too much empty jargon in their explanations of what they are doing to make any kind of evaluation. I also like your definition of semantics as well.

As you have implied, word sense can be explicated, yet, if you are someone who asserts that meaning is personal and takes place in the mind. That essentially means that meaning is ignored for you when using today's technology.

Proclaim that the semantic web is about making sense of what words mean to individual browsers of resources on the Internet, and almost every computer professional involved in the semantic web would say that is not what the semantic web is about. This is because one of the tenets of the symbol grounding problem in computing suggests a fundamental decision be taken to neglect meaning and concepts of the mind altogether. It will take another fundamental decision to change the status quo.

You probably would agree, Phil, that mind does not register or record every word imagined and do statistics on them. It works with much smaller sets of cognitivistic concepts that such large numbers of words attempt to represent in some ways. For this reason, I do not think Google, or any other search engine using keyword based indexes and some NLP or AI, could achieve an opinion search as you have outlined. Readware software can.

Readware is a mature semantic framework that delivers and utilizes sense information from words. It can also 'discern' concepts and topics and categories from the discourse on web pages read by the analyst sever. While Readware was not listed among the alternate search engines listed by Read/Write articles, it may be the only search engine that realizes how to seek, identify and compute semantic relevance from the words used in discourse and context.

While many researchers and developers may cringe at the prospect that the mind is where meaning is found, we can show that it possible to use cognitive objects and a conceptual model to identify sense information and precisely weigh and measure relevance among variable words forms and grammars in various forms of text and discourse.

To use your "citizen" example: if Google would ask me if I am looking for the financial institution "Citizen bank" of for the Crystal manufacturer "Citizen crystels", than I would call this (the first step of) semantic search.

By the way, I don't think that manual tagging / annotation is necessary for the semantic web to become reality. A lot of work has already been done in order to perform automatic annotion on webpages. Performance is getting better and better and I think it will be a matter of years before the big search engines start annotating webpages using (semantic ) ontologies.

From my site:

I'm reading about the semantic web and ontologies in last week's Economist (Tim Berners Lee) and a question arose in my head. Aren't we already creating this? Aren't del.icio.us and technorati examples of tools that allow for the emergence of a organized structure of information the same way a mass brain would? Doesn't the tagging going on today by millions of users ultimately allow for and supply the underpinnings of the pattern recognition algorithms our brain uses every day? The principles described in Jeff Hawkin's book "On Intelligence" have direct implications to this and may very well redefine how we get information from the web both personally and on a machine to machine basis.

see: www.theatomicweb.com

I don't see a semantic search a reality until the basic infrastructure of the web is changed.

I think that another way of describing the semantic engine is a tool that allows you to search at the topic level. Once you have the topic of interest, then drill down to the specific document.

Just how you categorize documents by topic is the big unknown, but that's the basic concept as I see it.

Make sense?

Some of these engines are quite excellent, and compared to LORD GRAND POOPA of GOOGLE, possibly even more effective.
I shutter to think at how much nay saying we would have if we were talking about inventing something that could stream text and pictures through little wires over thousands of miles at the speed of light!

Is there anyone out there that was not made to feel superior or cool because they had the distinction of being the first in their neighborhood to type Google?

OMFG. I am so amazed that millions of people can talk, inspect, research and finger manipulate a new cell phone ring tone or wideget so insignificant it will be worthless in 3 months, yet people attempt to create something astounding and some geeks just want to their little cubicle to stay safe.

Take some time, email the people involved, get them to explain what they are doing and perhaps understand that anything is possible still. Go to hakia or Powerset and talk with the people, perhaps they can delineate themselves from delicio lmao.

BTW the basic infastructure of the web is still the page, which typically has words on it. So, if one can make a several data sets from those combinations of words and use semantics to return less than the "billions" of permutations that Google's algorithms produce, then it is possible to do a much more effective and even faster search.

Always, Phil

I don't see a semantic search a reality until the basic infrastructure of the web is changed."

Yes, this is exactly what the evangelists tell us. However, it's not necessary! The basic structure of the web is not going to change and that's a good thing.

"Aren't we already tagging the web?"

Yes we are, del.icio.us is a good example - but ultimately, everything delicious does could be done by a machine. I've successfully programmed applications that will automatically tag in the same way Google suggests alternatives.

"You probably would agree, Phil, that mind does not register or record every word imagined and do statistics on them. It works with much smaller sets of cognitivistic concepts that such large numbers of words attempt to represent in some ways. For this reason, I do not think Google, or any other search engine using keyword based indexes and some NLP or AI, could achieve an opinion search as you have outlined. Readware software can."

I very much like this comment, and a few of you have said similar things. In fact, this is a psychological argument and when I was programming my first semantic tools it was something I grappled with a lot. I came to the conclusion that the human brain, as say a young child just learning to talk begins to derives senses from words in 2 basic ways.

- Association
- Questioning

Questioning is of course... asking what a word means or what an object is called.

Association is the statistical part, when a child frequently hears the words cat and kitten together, combined with questioning they build their own semantic ontology that changes and is redefined with age. Which is a very important reason for programmers NOT to use restrictive defined ontologies needing them to 'restructure the web' to 'solve' the problem. Instead they need to rely more heavily on association, with the limited questioning (through search queries) they receive.

You could run the same semantic search on each keyword.

You could run the semantic search over the keywords as a phrase.

You could do both and correlate results from each.

I know how mine works, but I couldn't tell you if it was 'right' because I don't exactly know how we as human beings do it yet.

"...tagging documents, pages and images to make them acceptable for a computer to read. Well, I’m sorry but I’m not going to waste my time tagging when a computer is able to derive context and do it for me."

I'm sorry, but documents must be marked up appropriately so that they can be consumed by a spider in a manner in which the spider is able to garner some information about the content of the document. The Semantic Web has absolutely nothing to do with tagging in the sense of the word that you have used.

I program web spiders and search engines for a living, and at the risk of offending the purists - such as you. It's been done by many others even apart from myself without the need to tag everything in sight.

I think I made the point quite clearly that 'The Semantic Web' has nothing to do with tagging in the sense of the word that I have used. I don't think it should have anything to do with it because it's a useless idea.

Taking what we know from early Google white papers and other research, when Google spiders the web it breaks a page down into 'barrels'. Which is just a collection of all the words on the page essentially.

If you then take these barrels and see how frequently words occur together on a page over say a 1000 pages, you build up a stastistical picture of words that are likely to be related. Of course the more pages you do this across the high the statistical probability is that they actually are related.

If you also store the position in the page at which the word occurs, you can factor in the distance between the words.

You can use this to speed up your algorithm substantially because effectively your select from database statement only has to pull in the say four words before and after the one you're looking for semantic links to. ie. the words occuring within sentence range of your keyword on a page.

I think I'm the first to relate that this in fact is generating a mass average opinion of a keyword's context and that this is itself something that could be used to show different context over subgroups of society, or even through dates.

I don't think you were very clear in that last post Phil. Word statistics accross these barrels may uncover co-occurrences (i.e. this is where a semantic relation exists) but little else. They do not tell you how those co-occurrences fit into the rest of the semiotics of the indvidual and his or her world: culture, society or clique, i.e. their 'worldview'.

Say, for example, you want to use your search engine to explicate your query for 'a good car' well-enough that it can include documents about 'high performance automobiles' and those that say 'clean, one-owner vehicles', 'a perfect ford for...' , 'a hot chevy with low milage', and documents with relative idioms I don't care to apprehend-- any methods depending on co-occurrence, noun or verb phrases, etc., LSI-wise or otherwise will fail and need to be re-trained for every case. Am I right or wrong?

I don't think you were very clear in that last post Phil. Word statistics across these barrels may uncover co-occurrences (i.e. this is where a semantic relation exists) but little else. They do not tell you how those co-occurrences fit into the rest of the semiotics of the individual and his or her world: culture, society or clique, i.e. their 'worldview'.

Say, for example, you want to use your search engine to explicate your query for 'a good car' well-enough that it can retrieve pages/resources about 'high performance automobiles' and those that say 'clean, one-owner vehicles', 'a perfect ford for...' , 'a hot Chevy with low mileage', and pages with many of the relative idioms I don't care to apprehend-- any methods depending on co-occurrence, noun or verb phrases, etc., LSI-wise or otherwise, will fail and need to be re-trained or re-indexed for every case.

Developers should also remember that meaning is built up from smaller elements. There documents semantics and sentence semantics and there must also be semantics at all levels of composition. Consider how the prefix /re/ changes the meaning of words above.

Now there are those that will argue that linguists know all about that 00and of course they do. What they do not do is consider or model how the addition of the prefix links to human perception. Neither do search engines.

While search engines do a great amount of indexing, they do not know what it means to index let alone what the significance or consequence (cause or effect) may be of having to reindex. I'll bet Google engineers do.

It is not too much analysis we do it in milliseconds while indexing. The problem with non-semantic search engines is that they do not register or index or link the semantics in phrases such as "a reliable car" and "this Chevy is dependable transportation" for example. Not in one or many pages...

In this example, reliable and dependible refer to the same concept. A semantic web product using WordNet should be able to discern that much. In addition, there is a pragmatic relation between car and Chevy that any search engine claiming to be semantic should be capable of capturing.

The difficulty is that are are milions, zillions, of words and phrases. They are not distilled or filtered by finding their mean in actuality. They are distlilled into topics and filtered into categories using a keen sense of semantics to recognize the essence of the message from the cultural code of the language.

It is not too much analysis; we do it in milliseconds while indexing. The problem with non-semantic search engines is that they do not register or index or link the semantics in phrases such as "a reliable car" and "this Volvo is dependable transportation" for example. Not in one or many pages...

In this example, reliable and dependable refer to the same concept. A semantic web product using WordNet should be able to discern that much. In addition, there is a pragmatic relation between car and Chevy that any search engine claiming to be semantic should be capable of capturing. Tagging, btw, has nothing whatsoever to do with any of this. Tagging is purely a data-processing crutch (that helps a little).

Think of the name Chevy as a tag, a socio-cultural tag, for a car. Names are tags. Names and tags are arbitrary. Individually indexing each and every occurrence or averaging all of them for their mean does what-- is useful for something? Finding phrases like those above by discerning the nature and import of the relations between specific names and other specific tags (in context) is a semantic process useful for discovery. Discovery is on the weak side of search engines

The difficulty is that are are millions, zillions, of words and phrases. They are not distilled or filtered by finding their mean, most frequent or least frequent, in actuality. To my mind, all that is read is distilled into topics and filtered into categories using a keen sense of semantics to recognize the essence of the message from the cultural code of the language.

Without going too far here, let me just say that it must be true that individuals may have different sensibilities, different modes of expression, different views, affinities, biases and beliefs. Each individual can have their own special way of interpreting what is going on while thoughts and perceptions race through their minds. Yet every individual has the same perceptual apparatus.

It is also true that shared (interpersonal) perceptions do not present themselves differently to each culture or society, let alone each individual. I mean that every individual in the world is subject to the same existential affairs. Unlike the (imperfect, changing) language, the semantics of such affairs must be universal that each individual may interpret as they will.

This is statistics, the more times you perform an operation the higher the percentage probability that the answer is correct.

It doesn't matter that people have imperfect language, or individuals have different semantics.

http://labs.hakia.com/

I spoke with Dr. Berkan about some more graphic representations and he told me he would get me some video and other information to supplement my pea brain lol. This is exciting stuff :)

Phil

http://demos.bitext.com/MSNen
user: readwriteweb
pw: bitext

Bitext's technology can be easily integrated with any search engine like Google Search Appliance, Autonomy, dtSearch, Lucene, etc.

For more info, check out www.bitext.com

Mark Johnson -
"It would be difficult, to classify dogs and cats as animals using that method - and it doesn't seem like Google can do that."
No, Google may not be able to explicitly decide that they belong to "animals" as such, however Google can realise that they belong to a group that relates to what we would call "animals" (see google sets).

"Also, you reference document classification (e.g. auto-tagging) and it's not clear that Google can do that either."
Google wouldn't be great at auto-tagging documents with tags that were english words, as the information about the page does not measure the meaning of the page using english words as symantics. It does "auto-tag" the document with semantics, and it is these semantics that are used in search.

"That seems like a very weak form of semantics."
Actually, compared to the raw semantics of language, this does a remarkably good job. This is because the layout of the text is taken into consideration, which effectively creates new semantics.

Phill Midwinter:
I think that the main problem you have seen is a problem with linguistics, not with the semantic technology behind it.
The concept space is, (it can be analytically proved - assuming it is using a SVD) making as good an approximation as possible. Likewise, we are only searching for the overall concept of a document, so the concept vector must be a very good representation of the document. The problem is in the search query, and in the wide distribution of how people use a language.

This is what google's personalised search is clearly going to try to do, determine how you use linguistics to build a map into the concept space that is more characteristic of your use of the language than the average map. Eventually they will probably be able to do this for the map from a website into the concept space too.

It's been good to come across you all, Phill - phillmidwinter.wordpress.com has gone in my RSS feeds, and I am sure I will be commenting on your blog as I am working on my own semantic search algorithms.

If you want to read my blog, click the link on my name.