Google announces new search algorithm 'MUM' that can understand complicated sentences and images beyond the language barrier

Google, which operates the world's largest search engine, is holding a new search algorithm 'MUM (Multitask Unified Model, Multi) ' at the developer conference 'Google I / O 2021' held from May 19, 2021 Japan time. Task integration model) ”was announced.

MUM: A new AI milestone for understanding information
https://blog.google/products/search/introducing-mum/

Search, explore and shop the world's information, powered by AI
https://blog.google/products/search/ai-making-information-helpful-io/

When you have something you want to know, searching with a search engine such as Google or Bing and gathering the necessary information has become a common practice in modern times. But with traditional search engines, the more complex you want to look up, the more queries you will need. Also, since search engines are not experts, it is difficult to accurately select and display the information that users are looking for.

That's why Google developed MUM. MUM is built on a neural network architecture 'Transformer ' that excels in language understanding tasks, similar to the traditional natural language processing model ' BERT ', and is trained in 75 different languages and many tasks. Therefore, it is possible not only to understand the language but also to generate it. In addition, it understands the information of the entire text and images, and can process video and audio information in the future.

For example, if you want to know about Mt. Fuji vaguely, enter 'Mt. Fuji' in the input field of the search engine, and a large number of pages related to Mt. Fuji will be hit. However, if you are thinking, 'I have climbed Mt. Adams before, but now I want to climb Mt. Fuji, what should I do?', The conventional search algorithm gives an accurate answer. In order to obtain it, it was necessary to combine various words and search repeatedly.

However, after understanding the above question, MUM understood that 'comparing Mt. Adams and Mt. Fuji' and 'preparation for mountain climbing requires training and purchase of appropriate equipment', and 'Mt. Adams'. You can present information such as 'Mt. Fuji has almost the same altitude', 'Mt. Fuji tends to rain in autumn', and 'A waterproof jacket is required'.

In addition, information on Mt. Fuji is often summarized on Japanese sites, and conventional search algorithms required searching in Japanese. On the other hand, MUM goes beyond the language barrier and collects and presents knowledge from sources, so even Americans will be able to quickly find out about the scenic points of Mt. Fuji, nearby hot springs, and popular souvenir shops.

Furthermore, one of the features of MUM is that it can understand image information. For example, if you take a picture of your usual hiking boots and ask the search engine, 'Can you climb Mt. Fuji with this?', MUM understands the hiking boots and the question, and says, 'You can climb without problems with these hiking shoes. You can also display a blog that lists other recommended mountaineering shoes.

Google revealed that 'MUM is in the experimental stage at the time of article creation and will be put into practical use within the next few months to years.' 'MUM is a search engine of Google that interprets various communications and information that people naturally perform. It's an important milestone for the future to be able to understand. '

Google also plans to make educational content accessible in more than 100 languages from Google Lens, which can be searched by simply shooting with a smartphone camera, and to display AR (augmented reality) content in search results. I am announcing. In addition, Google I / O 2021 also revealed improvements to the route search function of Google Maps.

Five new features of Google Maps announced, high-definition map and visualization of congestion situation, etc. --GIGAZINE

The presentation of MUM in the keynote speech of Google I / O 2021 can be seen from about 43 minutes of the following movie.

Google Keynote (Google I / O '21) --YouTube