Google recently began using a large-scale artificial intelligence model it calls MUM -- short for multitask unified model -- which it announced in May 2021.
It’s a more powerful language than BERT, short for Bidirectional Encoder Representations from Transformers, because it’s based on the T5 architecture and is capable of doing much more. It provides a deeper understanding to many search queries. BERT also was Google’s first experiment with the technology.
With it, the company aims to improve its search relevance through the RankBrain tool. For starters, it is not limited to text. It’s multimodal, so it also can use video and images as inputs.
Pandu Nayak, Google Fellow and vice president of search, explained how it works during the Google I/O conference in May.
“It’s pushing the boundaries of natural-language understanding,” Nayak said.
Trained across 75 different languages and multiple modalities simultaneously, MUM has the potential to transform the way people search and understand content and context.
Google first developed the technique in 2018, but its “dramatic demonstration came with last year’s GPT-3, a system developed by OpenAI that shocked many in the AI world with its ability to generate large blocks of coherent-sounding text,” according to the Financial Times.
The first uses of MUM, explains FT, focused on search tasks like ranking, classifying information, or extracting answers from text. But it’s difficult to measure the search results objectively, and equality difficult to judge the impact of efforts like the one Google took with MUM. It’s not clear if similar technology has lived up to the expectations.
Greg Sterling, vice president of market insights at Uberall, told the FT that “many search users will have failed to notice much improvement, and that product searches in particular remain highly frustrating.”
Google denies it, but the FT brings up an interesting point. As Google further develops sophisticated search technology to distill multiple searches into one result will it reduce even more traffic Google sends to other websites, something it's been working on for years. It already is under scrutiny by regulators worldwide for favoring its own technology.
Reducing the number of searches required to answer a query could reduce the number of ads served per person per session. Improving the relevance of the ad also could reduce the number required to trigger the purchase.