GOING TO a search engine like Google and typing in a query may seem reasonably straightforward to the person doing the search, but presenting back a useful range of answers is increasingly complicated as Google endeavours to anticipate and then provide the answers people want.
“It turns out to be really, really challenging to do a lot of this stuff,” says Jack Menzel, director of product management at Google Search.
Speaking from Google headquarters in California, he says: “We really want search to be able to answer any question that you could possibly have. You have something that you want to get done in the real world, and we want to facilitate that. So we want to answer any question, and to make it really easy to interact with the world’s information.”
Search historically has been about getting searchers to the right information source, he says. “We want to go a step further than that. We don’t want to just get you to the best page; we want to point you right into that best page. The kind of answer that you get from someone who really understands the topic and can get you right to the exact information that you want.”
Understanding the question – already a difficult bit of programming – is only part of the challenge. “You then need to understand what information is out there. And then, giving answers is easy – but giving correct answers is really, really hard.”
Sometimes a question can be ambiguous, with a range of different answers, depending on the context of the question. So, Google has to have precise modelling of words to understand that a simple query – such as typing the word “kings” into the Google search box – returns the most likely items. If you are in California, that means several sports teams, as well as links to a TV show of that name, come up at the top of the list.
Tackling this kind of contextual challenge is exactly what a newly introduced search feature called Knowledge Graph does. Originally available only in the US, the service recently launched here, and has over 500 million different entities modelled, using 3.5 billion defining attributes and connections.
Alongside it, Google has also just launched a related feature called Knowledge Carousel.
Knowledge Carousel presents a range of images across the top of the search results page, not for every search, but for certain kinds of searches that might have several categories of answer. For example, searchers for “Disneyland rides” tend to then click in to individual rides in the results. Now, with Knowledge Carousel, the searcher is presented with a row of images across the top, each of a different ride.
If a searcher clicks into any of those rides, new information is presented below in the search results, but the carousel of images across the top stays the same, enabling the searcher to explore further without having to enter further information into the search box itself, or leave the page of carousel choices.
How does Google obtain such a range of information and analyse its appropriateness? Menzel says the company starts with what is actually already available on the web, and tries to understand what is there using an automated approach insofar as possible.
“We have intelligent algorithms that go out and understand this is associated with this, and that there are these kinds of connections in the world. We also take that general kind of understanding from the web, and we combine it” with other data Google holds, such as its maps and books data.
“Then the challenge is how do you combine these data sources in a way such that you ensure there is still quality in the data? How do you reconcile them all into a representation of the world that you know is correct?”
Personalisation also plays an increasing role in search. At the moment, this primarily involves layering knowledge about where a search is coming from onto the possible responses. So a search for “kings” brings up different responses in Ireland than in the US.
But, Menzel says, search could evolve to bring in information from a person’s own computer and their email account, to be able to answer a personal question such as, “Is my Amazon package going to arrive today?” or “What time do I need to leave for the airport to collect my wife?”
“Since we want to be able to answer all these kinds of questions, we want to be able to start to merge information and make search so universal” that you don’t have to go in and search your email for flight information, and then go search for the actual flight number, and then go search for real-time flight information, and traffic information, he says.
“We could make that search much simpler. We could help you combine all those points so that all you have to do is naturally ask the question that you would ask a human, and ask it in that format – ‘what time do I need to leave to pick up my wife?’”
Menzel says there is an initial field trial on Gmail that is “an initial step towards this kind of vision”, but notes that it is extremely challenging, not least because of privacy concerns (an area where the company has stumbled in some service offerings in the past, and which remains a focus for regulators in the US and Europe).
“How do we make sure that we really respect people’s privacy and make sure that data is only being used in these really useful ways and can’t be abused, and make sure that everything we do is very transparent to the user and that it is truly useful,” asks Menzel. The answer will in part come from the field trial, he says, where the company can gauge what level of personal data use people are comfortable with.
New services like Knowledge Carousel add more information, text and image, to search results for users. How do they weigh up the pros and cons of increasing clutter to the screen? “We try to balance how much information we’re taking up on the page with how useful it is to people. We’re always taking that kind of feedback . It’s no good to give an answer that no one cares about. And we want the interface to be fast and easy for people to use.”
One gradual development that will greatly improve the accuracy of search results, he notes, is the slow advent of the “semantic web” – an internet in which the items in it (text files, webpages, videos, images, anything) – are tagged with informative terms that define them in various ways.
Menzel says Google itself has a role in that evolution, even if it can’t yet precisely define exactly what it is.
“One of the challenges we have is how do we actually help the web itself evolve, because no one is going to create the semantic web just for kicks, right? It’s a lot of work to do this stuff, so people need to have some reason to do it. So we have this kind of bootstrapping problem where we can do cool things if people were marking up the web in a standardised, semantic fashion. But until we can do something with it, nobody is going to markup the web” – at least, not just to see if Google can do interesting things with it.
