MUM: how new Google’s algorithm will change NLP

    April 27, 2022

    At the annual developer festival I/O in 2021, Google presented diverse updates for the search, including Lens, Augmented Reality in Search, and About the Results as a fast way to learn more about your search result.

    The search engine giant explained how to make the information even more helpful and announced the MUM algorithm – the future of search. MUM stands for Multitask Unified Model and is a new unique technology for Google Search that will replace the BERT algorithm (Bidirectional Encoder Representations from Transformers) and ensure better results in search queries.

    In this article, we will find out what to expect from the new Google algorithm, what will be the future of search, and how it will affect the NLP. But first, let’s discover what exactly the MUM is?

    What is MUM itself?

    Before the MUM, Google search was able to answer only simple questions, for example, “what is the weather in Cabo Verde in august?” This limitation gave the idea to review the search engine’s architecture to integrate new technologies such as machine learning and natural language processing with an intention to improve complex search query results in Google. And that’s how the MUM was born.

    Let’s dig deeper into the primary technology that is hiding in the MUM algorithm.

    NLP or Natural Language Processing refers to a branch of Artificial Intelligence that gives the machine the ability to read, understand, process, and derive sense from human language through training machine learning models to comprehend, divide into segments, and separate important details from different forms of information.

    Here we have a brief guide on how NLP works:

    1. The phrase you insert in the search field is going through semantic separation into units and classes;
    2. every word in each unit is tagged with their parts of speech like noun, verb, adjective;
    3. words are converting into their root form;
    4. later, the terms that provide less critical information than others are removing
    5. in the final step, the algorithm, like Multitask Unified Model, can be applied to the gathered data.

    The newest Google Artificial Intelligence used in MUM is revolutionizing the way how the search engine handles users’ requests.

    How Google has implemented Multitask Unified Model in its algorithms

    The new MUM technology is built on a transformer architecture that gives the ability to multitask. The algorithm is the most advanced and innovative in answering search queries 1000 times more efficiently than its predecessor and can handle various tasks in parallel.

    The Multitask Unified Model also is trained for 75 languages, unlike most AI models, which can train on one language at a time. According to Google, Multitask Unified Model allow source search results from across different languages. That makes sense because answers in another language from the asked may be more accurate if the little content is produced in your language. For instance, in the culinary sphere, a recipe for cooking a traditional cuisine of Ukraine in Ukrainian might be considered more authentic than the same recipe written by a Canadian food blogger. So, MUM overcomes language barriers, and if the appropriate information is found in another language – it will be collected and translated to the language of a search query.

    Another characteristic that makes MUM innovative – is the ability to understand different forms of information like text, video, or images. This feature allows combining various forms of information in one search query. Consequently, Google Multitask Unified Model develops a better understanding of information and provides complex search requests with precisely determined results. It sounds like a sci-fi movie, but this is an actual massive development of a global search engine performance.

    How did the new MUM algorithm beat BERT?

    BERT is a language model designed to understand language, our intent, and the way how we ask questions. MUM can understand and generate language to precisely comprehend what you want to ask to generate an organic response. Let’s overview the key advantages that helped MUM overcome BERT in detail.


    The first one is quite simple to guess. Since the MUM algorithm is working on T5 or Text To Text Transfer Transformer – a framework proposes reframing all NLP tasks into a unified text-to-text format where the input and output are always text strings, it allows to perform several actions concurrently.

    Google's MUM
    T5 enables translation (green), linguistic acceptance (red), sentence similarity (yellow), and document summary (blue). Source:

    Text Generation

    The new algorithm not only can look at the text and understand the language and what this text means, but it can also generate texts based on the given inputs.

    Breaking language barriers

    No matter in what language you put the request in a search engine. MUM is trained across 75 languages and can search for accurate answers through different sources from users’ language. As a result, Google MUM develops a more comprehensive understanding of information than the previous algorithm, and the user will see a more specific answer in the search result.

    MUM is multimodal

    It understands the information across its different forms – text or image, and in the near future, it will expand its capabilities to audio and video. It is another significant advantage of a search engine, and now it’s going to allow it to understand context further. The user will be able to use diverse content to ask a question like (for the simplest example): “how many calories are in this fruit? *photo of an apple*” and the algorithm will understand the photo and the text together and will respond with a relevant answer.

    How does the MUM implementation impact problem-solving in NLP?

    Since the MUM can simultaneously operate with different languages and media formats, it is a game-changer for natural language recognition.

    The way that new Google’s MUM algorithm presents rich results for complex queries for multimedia could drastically improve text and media processing into meaningful operable data while keeping the process fast and efficient because of its multitasking nature. Moreover, it solves an unprecedented challenge for multilingual systems that usually require multiple processing attempts for each language by merging it into one that resolves by a single query in the preferred language.

    Although the MUM algorithm is a tool that’s hidden under the hood, it changes the way we can communicate with computer systems by granting them robust decision-making regardless of our language and information format.

    No surprise that in Quantum, we all are passionate about technology. Subscribe to our newsletter, and let’s discover the new tech world together! We promise to keep you updated with the latest announcements and releases.

    • #Data science
    • #Google
    • #Google algorithm
    • #NLP

    Share Article

    Success stories

    LLM-based financial investment advisory chatbot
    #Large Language Model
    #Text analysis

    LLM-powered investment advisory chatbot for efficient investment decision making

    Digital financial market infrastructure platform
    #Distributed ledger technology
    #Transaction monitoring

    Building a scalable, secured system allows users to instantly create transactions in any asset, from anywhere on Earth, at any time.

    Transaction monitoring and suspicious data detection solution
    #Data analytics
    #Sensitive data
    #Transaction monitoring

    Transaction monitoring system development with complying data security standards


    Certification thumbnail