Google has updated the algorithm of its search engine that would ensure that the search engine now “performs ten percent better” for English speakers. The new algorithm uses an ai language model that understands searches better than previous models.
The improved artificial intelligence of Google Search is the result of the implementation of BERT, or full Bidirectional Encoder Representations from Transformers, according to Google. This new neural network, announced in May of this year, has an architecture specially developed to process sequences of data. This would make BERT suitable for understanding language, reports The Register based on an interview with Pandu Nayak, vice president of Google Search.
To understand sentences, it is necessary for the search engine to view all words in their specific order. According to Google, previous transformer models were only able to analyze the words in one direction, from left to right. “BERT can also look at a sentence in the other direction and take the previous and next word into account. In this way, the search engine gets a better picture of the full context and also the purpose of the search,” says Google Search vice president Pandu Nayak.
According to the company, some language models that BERT makes are so complex that traditional hardware is no longer sufficient for the calculations. That is why Google uses generation 3.0 of its self-developed tensor processing units, or tpus, for machine learning for the updated Google Search service.
According to Nayak, Google Search understands one in ten searches better thanks to BERT. “This result currently applies to English-language searches in the United States, but we will add more languages and regions in the future.” At this stage, BERT also works in 24 other countries and in Dutch, in addition to languages such as Portuguese, Korean and Hindi, but only for featured snippets. These are answers to frequently asked questions that appear at the top of the Google search results page, with information being picked from other websites.