Every second more than 90,000 queries are made on Google, whose search engine accounts for 90% of the market share, according to the statistics portal Statista. These stratospheric figures show the success of the search engine that many confuse with the internet gateway. But the Mountain View company has decided to perfect it. The result is MUM, an acronym for Multitask Unified Model (something like Unified Multitask Model), an artificial intelligence tool “a thousand times more powerful” than the previous one, according to the company, trained in 75 languages at the same time with one purpose: understand more precisely what the user wants to know, and offer the most relevant results from the most reliable sources. The company is already testing the tool in searches that have to do with the pandemic and vaccines.
Pandu Nayak, vice president of the Google department in charge of the search tool, gave details about the new tool on Monday by video call to a group of international journalists. According to Nayak, MUM can understand and generate language and is a “much larger model with many more parameters and, therefore, capable of doing much more” than BERT, the algorithm that currently establishes the results that the search engine shows to each user.
The challenge of MUM has to do with the different ways in which users refer to the same term. For example, Google has found that searching for information on COVID-19 vaccines uses more than 800 different words or expressions in more than 50 languages: from “Coronavaccin Pfizer” to “mRNA-1273” or “CoVaccine”. Identifying all of these terms would take hundreds of hours for one person. But with MUM it can be done in seconds.
“The reality is that people have complex needs and formulating inquiries from those complex needs is not always so easy,” Nayak explained. In fact, the technology calculates that a user makes up to eight queries on average to respond to complex searches. The goal is for MUM to simplify this process. “When users come to us with a question about a vaccine, we must recognize what they are asking about it and generate the most appropriate experience.” Google has used MUM to correctly identify how users refer to different vaccines and then try to provide them with the “latest and most reliable” information.
Trained with “high quality content”
Another example is the different ways in which users have referred to SARS-CoV-2 during the pandemic. “What is the coronavirus” was the question most asked by Spanish Internet users in the first weeks of 2020. There are also many products or topics that have different names depending on the language or the speech of each country or even depending on cultural nuances. This is the case of soda or soda, jersey or sweater or football — American or European. To overcome this language barrier, MUM has been trained “with high quality content” in more than 75 different languages at the same time. Something that, according to Nayak, serves to create robust systems in all languages and “provides many benefits when it comes to transferring knowledge of data-rich languages to those that are poor.”
The result is that MUM does not have to learn a new ability or skill in each new language, but can transfer what it has learned from one language to another. This is useful for incorporating all kinds of improvements “quickly and on a large scale, even when there is not much data to train the tool with.” In the case of vaccines, with only a small sample of their official names, MUM was able to quickly identify the variations between different languages, according to the expert.
The real potential of this tool is still unknown, but a priori it is quite promising. Nayak believes that it will serve to create completely new ways of searching and analyzing information in the medium term. “If I had to summarize what Google does in one sentence, it would be to produce relevant results from reliable authorized sources whenever possible,” said the company’s vice president.
Being multimodal, this system is capable of associating text comprehension with images. Google intends that in the future it will also work with other modalities such as video or audio. “We could take a photo of some hiking boots and ask the seeker if those boots are good for climbing Mount Fuji,” explained Nayak. Since the system can understand both images and text, ideally it should immediately analyze whether these boots are a good fit. It could take into account, for example, whether or not the ankles are covered. Google has assured that in the coming months and years it will launch new functions and improvements of this tool for its products.
Eddie is an Australian news reporter with over 9 years in the industry and has published on Forbes and tech crunch.