“It’s not expected to become this question-answering system,” Nayak told Search Engine Land, adding that such a system is “simply not useful” for complex demands.
For the past two decades, search engines have mostly operated in the same way. They’ve gotten better at detecting intent, offering relevant results, and combining different verticals (such as image, video, or local search). But the principle is still the same type in a text query, and the search engine will return a mix of organic links, rich results, and advertisements.
Recent advances, such as BERT, have improved search engines’ language processing capabilities. Allowing them to better interpret searches and offer more relevant results. Google recently introduced its Multitask Unified Model (MUM), a 1,000-times more powerful technology than BERT. According to Google, it combines multitasking and multimodal input capabilities with language understanding.
In an interview with Search Engine Land, Pandu Nayak, Google’s vice president of search, discussed how MUM might radically alter how consumers interact with the company’s search engine. As well as the technology’s future and what Google is doing to ensure that it is used responsibly.
MUM, Google’s latest milestone in language understanding
It’s easy to think of MUM as a more advanced version of BERT. Especially since Google views it as a similar watershed moment in language understanding and touts it as significantly more powerful than BERT. While both are based on transformer technology, and MUM has BERT language understanding capabilities built in. There are some differences between the two.MUM is built on a new design (the T5 architecture) and has far more capabilities.
Training across more languages scales learning.
“This is helpful because it allows us to generalize from data-rich languages to languages with a scarcity of data,” Nayak said. “[MUM is] trained simultaneously across 75 languages.” This could mean that MUM’s applications will be easier to translate into other languages. If that’s the case, it might help Google Search gain traction in those markets.
MUM isn’t limited to text.
Another contrast is that MUM is multimodal, which means that it can accept video and image inputs in addition to text. “Imagine taking a picture of your hiking boots and asking, ‘Can I hike Mt. Fuji with these?’” During the MUM unveiling at Google I/O, Prabhakar Raghavan, SVP at Google, using the following hypothetical example: “MUM would be able to decipher the image’s content as well as the intent behind your query.”
Multitasking also facilitates scaled learning.
“MUM is inherently multitasking,” Nayak added. Ranking websites for a specific query, document evaluation, and information extraction are just a few of the natural language jobs it can do. MUM is capable of handling many jobs in two ways: training and use.
“By training it on various tasks, those notions are taught to be more robust and general,” Nayak continued, “that is, they apply across multiple activities rather than being brittle when applied to a different task.”
MUM will not be rolled out as a single feature or launch in search, according to Google.“The goal is that over the next few months, we’re going to see many, many teams within search adopting MUM to better whatever activities they were doing to help search, and the COVID vaccine example is a pretty nice example of that,” Nayak added.
Google’s roadmap for MUM
Where we are now, the short-term.
MUM’s short-term goals are primarily focused on knowledge transmission between languages, according to Google. MUM’s first public application, in which it discovered 800 different vaccine names in 50 languages in a couple of seconds, is a fair example of where it’s at now. It’s worth noting that Google already had a subset of COVID vaccine names that would activate the COVID vaccination experience in the search results, but MUM allowed it to receive a much broader range of vaccine names, allowing the search results to trigger in more instances when necessary. And, as part of this short-term stage, Google teams have started to incorporate MUM into their projects.
“We have tens of teams experimenting with MUM right now, and many of them are finding great utility in what they’re seeing here,” Nayak said, declining to provide more specific details at this time.
Multimodal features are planned for the medium-term future.
“We think multimodality is where the action is in the medium term — it’s going to be like a new capability for search that we haven’t had before,” Nayak added, building on Prabhakar Raghavan’s picture search example from Google I/O.
Nayak envisions a user interface for MUM in search in which users may upload images and ask text queries about them. Nayak sees Google giving appropriate results that bridge the gap between the uploaded image and the user’s question. Rather than a quick solution that may result in a zero-click search.
Despite the fact that Google’s MUM tests have given them confidence. Nayak was quick to point out that the particular implementation of these “medium-term” goals, as well as any specific timetables, is unknown.
Connecting the dots for users over the long term.
“We believe that the promise of MUM, in the long run, comes from its ability to understand language at a much deeper level. ” Nayak said, adding, “I believe it will support much deeper information understanding. we hope to be able to convert that deeper information understanding into more robust experiences for our users.”
Search engines are still struggling to surface relevant results for some particular and complex searches, such as “I’ve ascended Mount Adams and wish to hike Mount Fuji next fall.” “How should I prepare differently?” “Today, if [a user] simply typed that query into Google, there’s a strong chance it wouldn’t return any results. So what you’d have to do is split it down into separate inquiries, which you can then dig around for results and assemble together for yourself. We think MUM can help here,” Nayak explained.
“We think MUM can take a piece of text [the search query] like that, which is this complex information need and break it up into these sort of individual information needs,” he said, implying that MUM’s language understanding capabilities could help Google provide results related to fitness training, Mt. Fuji’s terrain, climate, and so on.
“Remember, we don’t have this working because this is a long-term project,” he explained,”.But this is exactly the kind of thing you’re doing in your head when you come up with individual inquiries, and we hope MUM can help us develop queries like this.” “As you may guess, we might send you results for many queries like this. Maybe some text that connects all of this to your original, more complex question essentially organizes this information… that shows what the connection is. So that you can now go in and read the article on the best gear for Mt. Fuji or the tips for altitude hiking or something along those lines in a richer way.”
According to Nayak, one of the reasons this is a long-term goal is because it necessitates a reconsideration of why people come to Google with complicated demands rather than individual searches. Google would also have to break down the complicated demand stated by a user’s search phrase into a subset of queries and then organize the results for those queries.
Who is driving development?
When asked who would be in charge of MUM’s development and execution. Nayak replied that Google wants to create new search experiences. While also allowing teams to use it for their own projects.
“We fully expect many search teams to use MUM in ways we had not even imagined,” he added. “But we also have attempted to have innovative, new search experiences. And we have people studying the potential for generating these new experiences.”This is immediately evident to everyone. Both existing teams and those looking for new experiences are that the fundamental technology appears to be highly powerful and promising. Now it’s up to us to turn that promise into fantastic search experiences for our users. Which is the current challenge,” he said.
MUM won’t be just a “question-answering system.”
“This concept that MUM would become a question-answering system. That is you come to Google with a query, and we just give you the answer. I’m here to tell you that is absolutely not the vision for MUM,” Nayak added.
Nayak contrasted MUM’s ability to help consumers traverse complex intent queries with easier, more objective searches. Those are frequently handled straight on the search results page. I completely understand that if you ask a basic question like, “What is the speed of light?” you will get a simple answer. That it deserves a simple, straightforward answer. But most people need this hiking example or you’re looking for a school for your child. Or you’re trying to figure out what neighborhood. You want to live in any sort of even moderately complex intent is just not well satisfied by a short, crisp answer,” he said.
“You’ve surely heard the statistic that every year since Google’s founding. We’ve sent more traffic to the open web than the year. Before we fully expect MUM to continue this trend,” he said, adding, “There is no expectation that it will become this question-answering system.”
Mitigating the costs and risks of developing MUM
The development of search models can have an ecological impact and necessitates huge datasets. Google says it is aware of these concerns and is taking steps to ensure that MUM is used appropriately.
Limiting potential bias in the training data.
“If there are undesired biases of any kind in the training data. These models can learn and perpetuate biases in the training data in ways that are not great,”. Nayak said, adding that Google is tackling the issue by monitoring the data that MUM is trained on.
“We don’t train MUM on the entire web corpus; we train it on a high-quality subset of the web corpus. So that all the undesirable biases in low-quality content. In adult and explicit content, don’t even have a chance to learn. Because we’re not even presenting that content to MUM. He said, acknowledging that even high-quality content can contain biases.
“When we launched BERT a year and a half ago. We undertook an unusual level of evaluation in the months preceding up to the launch just to make sure there were no troubling patterns,” Nayak explained. Before we have a large launch of MUM in search. I fully anticipate us to perform a significant amount of evaluation in the same way to avoid any worrying patterns.”
Addressing the ecological costs.
Large models can be costly and energy-intensive to construct, which can have a negative influence on the environment.
“Recently, our research team published a comprehensive and interesting paper on the climatic impact of different major models developed by our research team. As well as those models built outside of it, such as GPT-3, and the article. According to the study, depending on the model chosen. The processors and data centers used, the carbon footprint can be decreased by a thousandfold. “So, whatever energy is being consumed. The carbon impact has been offset only by Google. Nayak said, adding that Google has been carbon-neutral since 2007.
MUM has potential, now we wait and see how Google uses it
Nayak’s comments on MUM’s future and how he doesn’t see it as a “question-answering system” are significant. Because they acknowledge a concern that many search marketers have but it’s also a concern for regulators looking to ensure that Google doesn’t favor its own products over those of competitors.
Other search engines may be working on comparable technology. As we saw with Bing and its adoption of BERT nearly six months before Google. Right now, Google appears to be the first out of the gate. Which given MUM’s effectiveness in its maiden outing, could be an advantage that helps the business maintain market share.
Google’s MUM roadmap gives marketers context and a lot of options to think about. But nothing is clear enough to start planning for just yet. What we can assume, though, is that if the technology is implemented and follows Google’s lead, users’ search habits will change to make use of those capabilities. Marketers will have to seek new opportunities in search and change. Their strategies as a result of a shift in search behavior. Which is par for the course in this business.