Knowledge Graphs - Computerphile

we're going to talk about knowledge graphs knowledge graphs are a ai technology that is used pretty much anywhere online in any major platform or system or application that we're using on a daily basis anytime you ask google or bing or your favorite search engine a question it will most likely use a knowledge graph alongside many other types of techniques to retrieve information and answer questions so if for instance you are asking um about bush house which is the building where we are now any search engine will provide a list of documents maybe some ads of some form but somewhere on that result page it will give you some very specific information about bush house it will provide probably a picture of the entrance where you came in and then it will say well um this is an address of this thing bush house that me as a computer without any sort of um human understanding can nevertheless relate to so so that string of characters bush house all of a sudden has an address and it has perhaps opening hours and it has particular departments at king's college that are located in this building right now if i go and query national theater i will still get information about the address because it's still a building but perhaps it will not take tell me anything about any university departments because national theatre is not a university or it's not a location for a university right so how does google know all those things about the world around us about london well because it uses the knowledge ground and i say d because actually google is the company that has coined this term but actually the principles and the technologies have been around for for much much longer in fact decades and decades where um ai researchers were building so-called knowledge bases which are very similar types of of i.t systems that are encoding knowledge information data i don't think we need to go into a big discussion about what these three mean and how how they are different from each other they are in this case in the example i had it's about knowledge about london what connects all these different pieces of data together so this is the knowledge graph this is the knowledge graph and the knowledge graph um as it is today not just in the case of google but pretty much any large organization that that provides this sort of internet information services has one they are huge databases with billions and billions of facts like this bush house is located in london and at this address right um this information is extracted from documents um say on the king's website somewhere there is information about bush house and and it's written probably in text and google has the machinery to process all this text and recognize that this string bush house is probably a building and and and learn that information that structured information and store that in their knowledge graph and so then autograph ultimately um you can you can imagine it as a um as a database and that stores facts um about things in the world and relations between them um and we in in ai and in knowledge graph engineering we distinguish between the actual data so bush house national theatre and then abstract data which is theaters buildings universities so any university most universities in the world will have a set of common properties right they will teach students they will be located on a certain campus they will have a certain number of employees kings college london has three or four actually different campuses lse also university shares many types of properties also teaches students also has employees but has different campuses different address so there is the data about lse in king's college london and there's all these joint properties that all types of things like lse like king's college have and that helps in web search but in many other types of ai applications as well so in question answering in recommendations of movies of products and so on and so forth because it basically allows that recommender or search engine to provide very specific very accurate to the point answers rather than relying on extracting those answers from natural text and you see the difference actually in search when sometimes you type something and google says well people also asked for this and this right and then you have that question and you have that little drop down list and a snippet answer now some of those answers sometimes match what you're looking for but sometimes they're actually only remotely or not all relevant and that's the difference a computer has struggles to process and extract very specific types of information from what we call unstructured data which is text images videos but it has no issue whatsoever or much less issue to um work with data that is already structured in a graph for instance my simplistic view of graph being a thing you plot on a piece of paper you know two axes beyond paper you could have many more axes is this what we're looking at here so we're not talking about charts okay right so we're not talking about um visually displaying data in a 2 3 or multi-dimensional visual space we're talking about a graph in a mathematical sense so a graph consists of nodes and edges we know graphs from computer networks right so you have um computers which are the nodes and then the communication links between them we know graphs from um geospatial information systems maps right so you have points on a map and then you have roots you need all sorts of algorithms to compute edges for instance in the case of maps if you want to go from a to b you can do that by passing multiple intermediary nodes so when you say an edge i mean we kind of had the example of a map and that being a root in the map yes what how do you define an edge it's any type of relation right so for bush house it could be has address so let's just say that we represent the nodes in these ovals and then we represent the edges as lines between these nodes so we're talking about bush house and we're talking about king's college london right and then so what's the relationship between bush house and kcl campus off or something yeah campus of or we could call it building of kcl we can have another relationship that says bush house has picture and here's my picture of bush house and it could be an actual image all these lines can have different meanings so i can have the has picture edge from bush house from kcl you could have something like has logo and it points to another image that is the logo um you could also have some sort of has picture with some picture of kcl who knows and there's an image file here so you can have different types of edges this type of has picture edge and this one here are the same so if you want to have all pictures of things you can query just like you would query with sql in a database you can have any type of edge you want you label it so you give it a name mine have arrows that's mostly just for my understanding because i want to say that the image is something that belongs to which house and not the other way around sometimes you have edges that could have arrows on both ends so if you think about elena has sibling alex my brother so the edge has sibling is what we call symmetrical so if i'm his sister then he's my brother say the moment when i define the edge has sibling from the node elena to the node alex the computer where i can actually tell the computer by the way there is also the opposite edge if i have an edge or relationship like is located in so bush house is located in london you see how the graph actually starts to take shape in london and elena works at busch house and london is located in england you can tell the computer well is located in is a type of edge that is what we call in mass transitive so if i know that bush house is located in london i know london is located in england then i know that bushow is located in england so you don't actually have to look for that sort of question if someone ever would ask is bush house located in england you don't actually need to look for the data specifically you can what you call infer it so you can from all this information that i have written down you can now infer additional knowledge this is a bit like an ontology though yes it is an ontology well it is based on an ontology to be more specific the ontology will be something about cities and about buildings in cities and buildings belonging to institutions and people working in certain buildings is there an easy a sort of free way people could play with this or try this is there something out there people could download for developers the google knowledge graph api is something that then people can look at there's also various open source knowledge graphs available say wikipedia has one it's called wikidata and it's what they use to manage all the structured information they have so you know when you go on a wikipedia article and on the right hand side you see what they call an info box bush house go on wikipedia and then you'll see that comes from a knowledge graph that is publicly available and people can download it it's quite big um but at the same time you have lots of information that you can play with and i don't know train a neural network on and do all sorts of wonderful things so yeah it's available and and and there are smaller versions as well for for people to to play with it's super hard to know i mean there are different things going on here there's intelligence which i think you've got to say maybe some computer systems are intelligent at least at certain things there's sentience and there's consciousness and they're slightly different than intelligent things so that big pointy spike in that sunglasses sunglasses and if we

Channel: Computerphile

Views: 101,640

Keywords: computers, computerphile, computer, science

Id: PZBm7M0HGzw

Length: 12min 5sec (725 seconds)

Published: Wed Sep 07 2022