The Growth of Machine Learning & AI in Search 

19.09.18 08:35 AM By Sonia Kisbee

When we talk about Artificial Intelligence and Machine Learning you may naturally think back to all those films you’ve watched over the years.  The ones where computers or rogue cyborg robots tried to take over the world or generally make life rather unpleasant for their human masters until we finally destroy them (or simply ask them to play Tic Tac Toe.)


Maybe we should be worried as even Elon Musk, probably one of the most forward thinking and innovative people out there, is a great advocate of AI but is extremely wary of its ability to do harm if not carefully controlled.

“AI is the rare case where I think we need to be proactive in regulation instead of reactive, because I think by the time we are reactive in AI regulation, it’ll be too late. AI is a fundamental risk to the existence of human civilization.”


The Cambridge Analytica saga also showed us that the misuse of data takes AI to another level that can be exploited at the highest levels. However to most of us, Machine Learning all seems rather futuristic and not something we should be worried about, but in reality it’s already here and making a massive (and generally positive) difference. 


One area where this is especially true is in the fast moving world of Search.  In this article we take a look at some of the recent search developments that have only been possible through AI and Machine Learning and where it could take us in the future.


So if it’s so great (but potentially going to wipe out the human race) I suppose we had better define what AI actually is and then how it differs from Machine Learning.

 

AI and Machine Learning - what are they?

Artificial Intelligence is neither artificial nor intelligent - it is simply the action of machines being able to carry out tasks that we would consider as “Smart”.  Machine Learning is a sub-set of AI whereby we give a computer data and a desired result and then let it try to find the best possible solution through trial and error - basically learning from its mistakes until it cannot improve upon the results. 

This machine learning can either be supervised (i.e. human input is given as to whether the computer gave a good answer or not) or unsupervised where it works out the answer itself based on predetermined success indicators.

 

So how is this used in Search?

As with many things these days, technology has been moving so fast that the only thing holding it back is the limitations of individuals or industries to fully understand the potential of it - and AI / Machine Learning is no exception. 

However Search, being at the forefront of computing and innovation, is one area that has started to really utilise this technology - and naturally companies like Google are right up there pushing the boundaries. 

In 2010 Google’s CEO, Sundar Pichai, announced that Google was a “mobile first” company - reflecting the massive growth and potential it saw from Mobile search usage.  By 2017 this message had radically changed and he recently announced that “Google was AI first”.  He stated that:

Machine learning is a core, trans-formative way by which we’re rethinking how we’re doing everything. We are thoughtfully applying it across all our products, be it search, ads, YouTube, or Play. And we’re in early days, but you will see us — in a systematic way — apply machine learning in all these areas.”

 

 

 

So let’s look in a bit more detail at some of the areas we are seeing Machine Learning integrated into Search products and the impact it has had.

 

Standard Google Searches:

Those outside of the Search industry probably have no concept of the massive changes that have taken place over the past few years in the Search results.

Google’s Search Engineers have been constantly developing and refining the Google Algorithm over the years to give us the best possible results it can.  If you think that Google has around 60 trillion web pages indexed and there are around 2.3 million Google searches every second (of which an astonishingly 15% have never been searched for before) you can see the true scale these engineers are up against to deliver the best results every time.  So this is where machine learning is now helping them.

In October 2015 Google introduced RankBrain, an “algorithm learning AI system” designed to help Google better work out the context and the meaning behind complex, rare or never-seen-before queries.  RankBrain also assesses the relevance of pages it chooses to bring back for these results and as such, after the content on your site and your backlink profile, it is now the third most important ranking factor in Google’s algorithm. 


What these means for SEOs is that, as Google gets better at understanding the context and relationships between words, the power of optimising a page for a specific keyword or phrase becomes less important - though it is not something you should start giving up anytime soon.

Other areas where this machine learning is improving the search results we get back is through monitoring engagement and behavioural metrics.

We all use search engines in pretty much the same way.  We type in a search query and if we get the result we want then we click through to that site and consume the information we need.  If however we don’t get the results we want then we might redo or refine the search query; we might click through to page 2 or 3 of the results or we might click though to a website and quickly bounce back out again after realising it did not fulfil our needs.

These are all signals that Google’s Machine Learning computers can analyse for trends over time to see if they gave a good result or not for a given search. They can continually test and adapt what sites they display accordingly to try and generate a more positive response next time.

Interestingly, other massive Google updates such as Penguin (penalising dodgy backlink profiles) and Panda (penalising poor content) are apparently not using any machine learning technology to monitor and improve on their accuracy on catch rates at the moment.

 

Voice Search:

Coming off the back of the growth in mobile use, one of the biggest changes in the Search landscape has been explosion in Voice search - firstly on phones but increasing via home assistants such as Google Home or Amazon Alexa devices.  In order for these to work to an acceptable level the computers need to be able to understand us properly.

Maybe it was just me but using voice search on my Android phone used to be a complete nightmare.  A simple request such as “Call Mum” would result in multiple incorrect “Did you Mean?” responses until I realised I needed to say it in a strong Texan drawl (seriously) and dictating a text was equally frustrating.

One of the biggest problems was regional variations in terms of people’s accents but also how they asked for things.  Teams of humans were constantly tweaking the algorithm to make it more accurate - a very slow and laborious process.  That is until the Machine Learning team at Google knocked on the door of the Google Speech team and offered their services. 

Using a cluster of 800 machines over the course of just 5 days the speech recognition software was “trained” to better understand voice searches.  The results were a staggering 30% reduction in word error rates for English searches.

 

Google Translate: 

Another area where Google has utilised machine learning to great effect in 2016 was in the growth and refinement of its language-to-language translation service The computers stopped trying to translate the sentence word by word and instead looked at the sentence as a whole in order to understand the actual context.  This allowed it to better understand the sentence and bring back a more relevant and natural translation that uses the most appropriate terminology and grammar.



 

Image Search:

Google can’t see pictures - that is why SEO’s add alt text and keyword rich file names to images.  Well, actually it can now and this is another area where machine learning is turning things on its head.  

Computers are being taught to “see” pictures without having to be told what is being shown.  This is possible by using a combination of things such as an object’s typical colour, shape, gradients, perspective (and naturally any supporting information it can garner such as alt tags or surrounding text if available).

However this has been a fairly slow process where training data (such as pictures of cats) were fed in and then the computer was asked to identify when cats were being shown.  By working out what was correct or wrong it was gradually able to reduce the margin of error and confidently identify pictures with cats in it.

The benefits of this are huge in terms of being able to bring back image search results, better understanding images on a web page, helping you search through your photos hosted on Google Photos and much more.

 

Video Search:

A video is just a load of image frames so it is perhaps not surprising that in March 2017, Google announced they could even start understanding the content of videos.  It is already offering the API to businesses so that they can search their video catalogues any that contain nouns and verbs (the example they use being “Tiger”) 


Another immediate use of Google’s “Cloud Video Intelligence” is to help it identify and remove extremist content from YouTube - something that has seen a raft of big name advertisers pull their advertising off the platform in recent months.

Yet there are inefficiencies in the current image-recognition software used by Google - the need for large numbers of images for objects to be recognised reliably in different situations. So the Google Brain Team, led by Geoffrey Hinton, perhaps better known as the Godfather of AI, have developed cutting edge AI technology, known as capsule networks, to help machines better understand the world through images and video.

 

Let’s not forget Google Paid Search:

Another innovation from Google that derives from Machine Learning has been in the area of Google Adwords.  They are now using it to analyse staggeringly large amounts of data and search signals in order to bring back better ads at the exact time that benefits both the user and the advertiser.

Their in-market audiences option “uses the power of machine learning to better understand purchase intent. It analyzes trillions of search queries and activity across millions of websites to help figure out when people are close to buying and surface ads that will be more relevant and interesting to them”.

 

What else…?

Finally, another way Machine Learning can help search marketers is in the form of testing and optimising their landing pages for conversions.  One tool I recently saw demonstrated at a conference was Sentient Ascend which is like a multi-variant testing tool on steroids.  This system can “test all your ideas at once, actively finding winning concept…. it compresses years’ worth of testing into a month by combining high-level AI with your inherent creativity.”


So, that’s just a small taster of where machine learning is helping to drive the Search industry.  It’s just the start and it’s going to revolutionise all our lives (and probably our jobs) in the next few years.

 

Author: Matt Lester, Senior Search Manager, Fidelity International


Disclaimer: Please note that this blog only contains general information and insights about legal matters. The information is not advice, and should not be treated as such. Kompass.com