Google’s Betting BERT Can Understand You

For those of us that use search engines a lot, you know that there are certain ways you have to phrase things to get the right results. According to Pandu Nayak, Google Fellow and Vice President of Search, Google receives 5.6 billion searches every day and of those, 15% are queries that they’ve never encountered before. (I can’t imagine a question that Google hasn’t seen, but if my math is correct, 15% of 5.6 billion is 840,000,000. So: a lot.)

Google Home Mini voice assistantGoogle has created ways to find results for those never-before-seen queries. Sometimes the questions are unique because the user isn’t sure how to phrase it correctly or how to spell a word. Face it, if we’re googling something, we probably don’t know much about it. It makes sense that we may struggle to write our queries perfectly. We learn to speak what Google calls “keyword-ese”.

What is Keyword-ese?

When using keyword-ese, you type related words together, but they aren’t necessarily grammatically correct. It’s kind of a computer baby-talk. For example, we may type “weather Friday Geneseo” to find out what the weather is going to be like for the high school football game. We don’t talk like that; unless we’re chatting with Tarzan or Frankenstein’s monster.

Google has announced that they’ve improved search to better understand the way that people think and speak. They are utilizing an open-sourced neural network to “teach” their search engine how to better process natural language. Bidirectional Encoder Representations from Transformers, or BERT, allows search algorithms to consider the linguistic context. In the past, they solely relied on keywords. Stop words like “to”, “for”, and “in” can give the search engine trouble.

An example that Google provided used the search query “2019 brazil traveler to usa need a visa”. In the past, search results would have been about U.S. citizens traveling to Brazil. But after BERT was implemented, the correct links were provided. (See image below)

Google search result

Voice Assistants on the Rise

The three big names in voice assistants, Google Assistant, Siri, and Alexa, are all becoming more popular and easier to use. Of the three, Google Assistant is considered the best at understanding and answering conversational questions and commands. I have a Google Home Mini and an Alexa Echo at home and the Google Home usually gets the results right the first time, without having to correct them. I hear my kids saying, “Alexa, no! Alexa, stop!” a lot when trying to get her to do something.

Google states that by using BERT, they can improve one in ten English language searches in the United States. It is only a matter of time before BERT is used to assist other language speakers.

Leave a Reply

Your email address will not be published. Required fields are marked *