Powered by the latest AI innovations, Google Search is set to become more efficient and intuitive than ever thought possible
With Google being the most popular search engine in the world for what seems like an eternity, it is easy to ask whether it is even possible to make it better? The simple answer is yes. Google Search is nowhere near perfect, and the problems it poses are far from solved.
During Google’s Search On event at the end of September, the company shared how the latest AI innovations are coming to play in Google Search and attempting to make the world’s information more helpful to users. They released several features that probably try the hardest to make Search more useful and contextual than now.
Google is still very actively working on Search. During the event, it flexed its ability to recognise an assortment of information and related topics with the help of advanced AI and machine learning and present to the users in a sorted way. To do this, Google is using the Multitask Unified Model (MUM), an algorithmic model far superior to BERT (Bidirectional Encoder Representations from Transformers). The new and improved model involves the creation of a continuous circle of information, where more contextual information is provided to the users, prompting them to search the internet with more complex queries and providing more context. It would result in more relevant search results and lead to a bidirectional and mutually beneficial flow of information between the system and the user.
The new and improved model involves the creation of a continuous circle of information, where more contextual information is provided to the users, prompting them to search the internet with more complex queries and providing more context.
To understand what all this means, it is first essential to understand what is problematic with Search in the first place. For an average user today, Search is still not the most straightforward solution to their questions or problems. Beyond the most precise questions (“What’s the weather in Dhaka today, Google?” or “How many calories are there in a large Snickers bar?”), one still has to go through several rounds of queries to find a complete answer to a specific question. For most people, even figuring out the right way to ask a question to a machine can as a formidable challenge.
Let’s look at a practical example of such a scenario. If someone had already travelled to Nepal last season and now wants to know how to plan differently for an upcoming trip to Bhutan next month, he/she would have to categorise the areas of divergence and individually search for each category. While the question “What do I need to plan differently for this particular trip?” may sound simple to a human listener, but for a machine, its solution is not as straightforward. Ideal search results in such situations could potentially include temperature and weather conditions (highlighting differences to pack clothes effectively), terrain topography (to understand if gear choices are appropriate or not), and many other types of relevant data. However, the same queries can be made to an expert travel agent or veteran traveller in the form of a simple question – “I’ve travelled to Nepal already, so if I want to take a Bhutan trip in a month, what would be the differences?” And sure enough, the answer would be complete, with all the nuances and appropriate contexts fully taken into account. Sounds relatable? To date, despite the vast improvements made to Google’s speech recognition and synthesis capabilities, this is where the difference lies between human cognition and Google search – personalisation and straightforwardness. It takes around eight search attempts on average for queries as complex as this.
MUM is Google’s solution to making Search richer and deeper, making complex tasks more straightforward for the user. The model will use the T5 text-to-text framework, which is almost a thousand times more powerful than BERT. The amazing part of this model is it has multi-language support as well as multimodality.
Multi-language Support
MUM supports 75 languages, not only understanding these languages but also responding using them. It is compelling because languages often act as huge barriers when accessing information, and multi-language support enables Google to have a more comprehensive understanding of information and facts. When a search is made in one language, Google can use MUM to retrieve information created in that particular language and relevant information created in other languages and present it to the user in an easily understandable form. For example, a search on sushi in English can bring in information from Japanese-language websites and present it to the user in English, making results more inclusive and, hopefully, relevant. Going back to the Bhutan trip planning example, the user can also see the best local stores, food and sights – information that would be more commonly found if search queries are made in Bhutanese.
Multimodality
MUM is also multimodal, i.e., it does understand not only text but also images, webpages and more. Audio and video learning to aid Search are also in the pipeline. Google is pushing its Lens feature much more aggressively now, increasing its identification capabilities to answer more diverse questions to solve problems for the user, going far beyond simple object recognition. Reverting to the travel plan example, a user can take a photo of a trekking shoe with Google Lens and ask Search whether it would be appropriate for the upcoming Bhutan trip or not, and receive more direct answers in return. Moreover, Google Search would also suggest other gear to better prepare the traveller for the trip, along with nearby stores that may be selling them, or recommend relevant travel blogs containing helpful information. Now that Google Lens is a built-in feature of the Google app in iOS, it will become more mainstream now, and it’s only a matter of time before it shows up on the Chrome browser on desktop computers.
Speaking of blogs suggestions, one of the first redesigns coming to Google will be a new ‘things to know’ section with different subtopics when a search is made, even taking the user to videos and specific parts of videos, when necessary, if the entire video is not related to the matter searched for. While on the topic of more shop results coming up in Search, Google makes shopping easier. The search engine sends users almost 100 million results on any given day. It connects customers to businesses that don’t even have websites with the help of phone calls, directions and location-centric data. A nifty addition is the display of inventory in nearby shops, which can play instrumental roles behind customer decisions when querying about product availability.
Google informs that features powered by MUM and improvements to their products are already in the pipeline, to be released over the coming months and years.
All of this involves using machine learning and AI to make Search closer to being more intuitive and solution-oriented, having made great strides in linking words and topics. But just like any other time, it is a critical test for objectivity. Google says that MUM will go through the same rigorous evaluation process as BERT did, keeping the human raters and its search quality rater guidance in place. But one of the problems that have arisen previously with Google’s AI was the prevalence of racial and gender biases, and it is now likely to become even more pronounced with Google providing more direct answers, so keeping things unbiased and objective is something that needs to be kept a careful eye on. With machine learning crawling the web and gathering heaps of information, it is not news that the brand has been plagued with not-so-pleasant problems when it comes to the issues mentioned earlier. Google says it is applying filters in place so that the data learned is not directly presented to the user, but that creates another fresh dilemma – if Google provides its version of answers, the users would then be at risk of being subjected to the point of view which is not objective.
This complex intertwining of algorithms, AI, language models and inputs is almost beyond our understanding at this point. Google’s claim of its Search system understanding the content of videos is a massive revolution in itself, but how clear and objective results would finally be once the model matures is something that remains to be explored in the future. It can be hoped that someday one would simply need to ask Google ‘how to plan for my Bhutan trip differently than my Nepal trip’ to get a relevant answer in line with what an actual human expert would provide.
Google informs that features powered by MUM and improvements to their products are already in the pipeline to be released over the coming months and years. But the domination of context is something to be very excited about, as Google gets better in the different ways of natural communication and information interpretation.
