15 comments

[ 4.2 ms ] story [ 50.5 ms ] thread
Is Knowledge Graph that thing that tries to directly answer questions you type into Google Search? I haven't found it to be very useful; what is the thinking behind cloning it? Where does it add value?
>Where does it add value?

It keeps people on the Google site, with Google ads, instead of on different websites that may not have Google ads.

I think in addition to the box that you see it's also used to improve search results.
Let's say a user is on your site, and they do thing. You might want to then show them either content or adds for related thing.
Knowledge Graph is used for lots of stuff, not just answering facts. It helps to know what an entity is for many reasons.

Let's say you're searching Google Photos for "Mammals". Photos might have neural tags tagged for dogs, for dog breeds, and for very specific animals, but they might not be tagged for higher level concepts, like that a Dog is a Mammal.

The KG tells you if a given term is a Point of Interest, a location, a landmark, a person, an actor or character, a film, food, and what not. This lets stuff like Now on Tap provide contextually relevant apps like IMDb for films, actors, etc or Open Table for Restaurants.

I've found KG to be very useful for some queries, especially queries that produce Carousels, like "top ten movies in 1996"

The article is full of examples.

From the article:

If you’re listening to a Skrillex song, you can say “What’s his real name?” and your phone will give you his real name. If you open an email asking if you want to see Tomorrowland, you can tap on it and get instant reviews, ratings, and trailers for the latest sub-par George Clooney flick. If your get a text suggesting dinner at some hip restaurant you’ve never heard of, you can tap again for reservations and directions.

Could someone explain what differentiates the Google Knowledge Graph from a Markov Logic Network implementation? Is that what it is?

For reference: https://sites.google.com/site/slgworkshop2013/accepted_paper...

MLNs are one possible way to implement the inference component that any knowledge graph needs.

Google's Knowledge Vault uses a fusion of a number of different extraction methods. Their exact methodology is laid out in their "Knowledge Vault" paper[1].

If you want to go deeper (ha!), then Deep Dive[2] is open source, and pretty much state-of-the art. It does inference using Gibbs sampling on a Factor Graph (Markov models/MLNs can be represented as Factor Graphs).

[1] http://www.cs.ubc.ca/~murphyk/Papers/kv-kdd14.pdf

[2] http://deepdive.stanford.edu/doc/general/kbc.html

The comparison to Google's Knowledge Graph was misleading. The article frames Diffbot as _a_ database until the last section, where the article introduces the capability for customers to build their own proprietary knowledge graphs. That's much more interesting.
Founder here. I wouldn't describe what we are doing as cloning Google's Knowledge Graph. Rather, we're trying to build a machine that autonomously learns from reading the web, and makes that structured data accessible and queryable to any app, smart device, or business that uses data. You can also tell Diffbot what part of the web to crawl.

Our APIs power consumer services including Instapaper and DuckDuckGo to enterprise applications like content management and market intelligence.

The article mentions that you power the Bing infoboxes (or whatever the equivalent Bing term is).

How does that relate to Bing's use of Probase (their concept-graph)?

Unfortunately, I can't speak about this customer integration without permission. You can try to ask questions about Diffbot, though!
You said that Diffbot reads the web autonomously. Is it able to learn a language on its own? I'd be very interested in this (business-wise) if it could provide results in my mother tongue.
Working on it. The reliance on visual features means it works decently well on extracting from international pages. Try putting in foreign-language product pages or article pages into our homepage testdrive or with a developer token to see an example.
It seems like it is both, since the article also says

  Diffbot says its database now spans about 600 million
  objects, and the hope is ...