Google's Amit Singhal Interview: Developing The Knowledge Graph

[engine]

Google's Amit Singhal Interview: Developing The Knowledge Graph [mashable.com]

The future of Google Search, though, could be a very different story. In an extensive conversation, Singhal, who has been in the search field for 20 years, outlined a developing vision for search that takes it beyond mere words and into the world of entities, attributes and the relationship between those entities. In other words, Google’s future search engine will not only understand your lake question but know a lake is a body of water and tell you the depth, surface areas, temperatures and even salinities for each lake.

Google now wants to transform words that appear on a page into entities that mean something and have related attributes. It’s what the human brain does naturally, but for computers, it’s known as Artificial Intelligence.

It’s a challenging task, but the work has already begun. Google is “building a huge, in-house understanding of what an entity is and a repository of what entities are in the world and what should you know about those entities,” said Singhal.

[santapaws]

is the most common clicked result for a search the same as the best result? It makes me despair as this continual dumbing down of the internet. Imagine sitting in a classroom at your first maths lesson. The teacher asks what is 2+2 and 75% say 5. So the teacher says ok, from now on i will suggest the answer is 5.
Or perhaps the teacher says, well my best guess is.......

[rlange]

santapaws wrote:
is the most common clicked result for a search the same as the best result?

Not necessarily, but it could be a strong indicator when combined with other signals.

It makes me despair as this continual dumbing down of the internet. Imagine sitting in a classroom at your first maths lesson. The teacher asks what is 2+2 and 75% say 5. So the teacher says ok, from now on i will suggest the answer is 5.
Or perhaps the teacher says, well my best guess is.......

Basic math is a horrible analogy so I'll stick to the height of the Empire State Building. Now, how can any human know the height of the building? Either they've 1) asked a single, highly trusted source, 2) asked numerous less trusted sources and tentatively accepted the most popular value, or 3) they've measured the height directly.

Google's algorithm obviously can't make direct observations of the real world (#3) to determine the correct answer to "How tall is the Empire State Building," so it either needs to get that information from a single, highly trusted source (#1), or from multiple, less trusted sources (#2). Most people here seem to be rabidly against anything that would even come close to #1, leaving #2 as the only viable option for answering a question.

Using #2, the best thing Google's algorithm can do to give the user an answer they can likely trust is to base that "best guess" on sources with the highest trust value as determined by their ranking algorithm. Popularity almost certainly plays into that trust value.

Now, if several very popular websites incorrectly state that the height of the Empire State Building is 1000m, how is Google's algorithm to know that's false? It essentially can't.

Google Search isn't an educator and it's not the source of information; it's an aggregator. Everybody loves pointing that out. It can only give you its best guess. Hell, that's what the results in any search engine are, have been, and always will be... nothing more than a best guess.

--
Ryan

[santapaws]

except this is the direction Google wants to go. One hit answer based on user past actions and authority weighting. Where in this attempt at AI is the validating process that the answer is actually correct? My point is real AI does not work like this. If your in a clasroom then the teacher is taken as a validated source of authority, by a real world test (they have been interviewed and hired). If i ask a person on the street a question the human intelligence comes from narrowing down an answer and/or source by two way conversation. Even if i just ask for directions i will probably ask some question for my own validation such as, do you live here, have you been there, are you sure.
Search engines are basing everything on most likely scenarios and one way replies to a single question, yes this is clever stuff but it can never be AI.
Even if the stranger was the most geographically knowledgable human on the planet and i asked where is the garage. he/she is not going to say well ok the two times i was asked this before was for a garage 100 miles away so i will go with that answer. They will narrow the garage required by further question. So searchg engines will move on the day they stop this mind set that results must be a one time served affair. They need to start asking questions back to reduce the guess work.

[nickreynolds]

Are most people searching for things that have a simple answer (such as the height of a building)? Seems to me that often people are looking for sites to browse for information on, not for a speficic answer. People like to browse. Not sure how truly accurate it is but Yahoo Uk has a list of topics people are currently searching - most of them are news based or the latest celebrity gossip

[Panthro]

I don't know why interaction can't and won't be involved in this evolution. I saw some Simpson's episode where Google of the future talked to Lisa. No reason that the Gorg couldn't reply to the searcher with questions that help narrow and define the query that helps it calculate an answer. To be accurate, and to be AI, it obviously has to and I'm sure will.

Its like real AI or Fusion Power it is always 10 years away :-)

But I agree with this.

[rlange]

santapaws wrote:
except this is the direction Google wants to go. One hit answer based on user past actions and authority weighting.

I don't think that's entirely true. The linked article doesn't really suggest that Google is going for simple answers to, say, [interpretations of Shakespearean plays]. In fact, the one example the article gives is that of lakes and the answer returned consisting of simple data about specific lakes.

Where in this attempt at AI is the validating process that the answer is actually correct? My point is real AI does not work like this. [...] If i ask a person on the street a question the human intelligence comes from narrowing down an answer and/or source by two way conversation. Even if i just ask for directions i will probably ask some question for my own validation such as, do you live here, have you been there, are you sure.

None of those questions validate the information itself. All they do is increase or decrease (rightly or wrongly) your trust in the source of the information.

Your criticism, though, seems to be focused on the use of the term artificial intelligence to describe the goal mentioned in the article. You seem to be bothered by the fact that they call it AI but seem to be ignoring a conversational ability on the frontend. The problem is that they're talking about the backend—the organization and collation of data into entities and the identification of relationships between those entities.

The frontend and backend are two different things. Your typical soda machine doesn't require a conversational interface, but it does require a human in the restocking part of the process to organize things properly so that when you push a certain button, the correct can drops out. By requiring the "complex" task of organization on the backend, the frontend can be much simpler.

Even if the stranger was the most geographically knowledgable human on the planet and i asked where is the garage. he/she is not going to say well ok the two times i was asked this before was for a garage 100 miles away so i will go with that answer.

You're veering off into location-based searches now. Of course it would be bad if Google returned results for garages 100+ miles away from your current location. They know this. It's precisely why they take your current location into account for "Where is the garage"-type searches.

--
Ryan

[santapaws]

heres a quick pointer, if i ask someone randomly on the street, "wheres the garage?", would you expect the answer to be:
the nearest garage is on third street
or
you mean the gas station?

[Robert Charlton]

"wheres the garage?"

Suppose you spelled it gasage?

[santapaws]

its a verbal conversation. or do you mean on google? Spell checks are one of the few times google attempts at refinement by asking did you mean. But of course often even when you spell something correct you get results for the wrong word simply because often people search for something similar with a different spelling or google is using assocation. I might search for a famous actor and gets results for his agent for example. Not much help if i ask wheres the garage and geta result for their head office.

[lexipixel]

if i ask someone randomly on the street, "wheres the garage?", would you expect the answer to be:
the nearest garage is on third street or you mean the gas station?

Depends if I was expecting them to direct me to:

a). Park in a parking garage

b). Get [petrol, gas, fuel], directions, or use the [restroom, loo, lav, facilities], at a garage.

c). My [car, auto, vehicle] was in obvious need of repair at a garage.

d). The person I was asking was standing in front of a residential home surrounded by a tall hedge, and there were two or more driveways and I wanted to find the one that lead to the home's garage.

e). I'm in a steam-punk type urban area and there's a [bar, club, venue], called "The Garage".

f). I'm in Cambridge, MA, USA, (in Harvard Square), and I'm looking for the indoor-mall called "The Garage", (which used to actually be a multi-level paid parking garage -- [yelp.com...] -- true story).


Without much of a definitive query --- context is everything.

[santapaws]

you had me at

depends

. Exactly!

[lexipixel]

you had me at depends

...sounds like an incontinent senior-citizen's romantic interest proclamation.

[jmccormac]

Of course it could be some Google Public Relations fluffery to get the term "knowledge graph" into the same category as Facebook and the Social Graph/network.

Regards...jmcc

[lexipixel]

fluffery to get the term "knowledge graph" into the same category as Facebook

It's what I first thought when I saw the term, (e.g., Knowledge Graph v. Social Graph).

The two are different animals. One is the connection of words to knowledge, the other is the connections of people to information.

Knowledge Graph Theory is nothing new -- there are citations that date back three decades...

Abstract. The project on knowledge graph theory was begun in 1982. At the initial stage, the goal was to use graphs to represent knowledge in the form of an expert system. By the end of the 80's expert systems in medical and social science were developed successfully using knowledge graph theory. In the following stage, the goal of the project was broadened to represent natural language by knowledge graphs.

(PDF) [doc.utwente.nl...]

[jmccormac]

Yep but Facebook has really put Google on the defensive and people are more likely to have heard of the Social Graph and the "Social Network". This is edging Google out of the limelight as a very hackishly driven startup. With all the bad publicity about Google snarfing user data, Google is rapidly assuming the position of evil empire that Microsoft had throughout the 1990s and 2000s. As for Amit Singhal's promotion of Wikipedia and plundering ideas from Star Trek, it is all very contrived.

Regards...jmcc

[tedster]

Sure it's PR - it's only reported as Google's "developing vision", according to the story. And to accomplish something like this will take a long time, even if the basic approach proves to be sound in the long term.

What might be possible in developing and maintaining a "knowledge graph"? If we think about IBM's achievement with Watson in recent years, then we know that something pretty dramatic is possible. However, Watson was not developed in an intensively competitive like the world wide web.

I would say we can see some signs of this thinking already in the search results. we've got a recent thread where people are talking about essentially relevant results that do not include the same vocabulary as the original query. So I do think this general vision is accurately represented. They've got a lot of smart linguists and engineers working on this recipe long term - no doubt in my mind.

[lexipixel]

It's not just Facebook's "Social Graph" that Google is attempting to out-PR, I'm sure part of the AI talk has to do with building out an (Apple) Siri-like speech-based AI search enhancement blended into the Google algorithm for use in Android -- and it all has to perpetuate Adwords sales or.......

[tedster]

Good point - Siri is much more the target here than Facebook. The "social graph" has very little to do with a knowledge graph. The two phrases just share a vocabulary word.

[tedster]

Now the WSJ has tried to use this interview to stir up something or other a month later reference [webmasterworld.com] Sheesh, these media wars!

One thing becomes clear to me - change of any kind is unwelcome to many people, and yet change is inevitable. With this direction (which will be unfolding in a many year process) Google is taking on a major challenge in natural language processing. I think that's great, even as I brace for even more disruption in SEO coming own the pike.

After all, information retrieval (a better term for what Google does than "search engine" these days) is a tough nut to crack, as anyone who has played with large-scale IR can tell you. I'm happy that they are not just resting on their past - that they are not ONLY trying to maximize advertising profit or push their social offerings. They certainly are doing those things, but this is almost in the realm of pure academic progress for IR.