How Google BERT Changed Search Understanding

Bert was originally described as affecting around one in ten English searches at launch.

BERT aimed to improve the relevance of search results, which could affect the queries through which people discover a website.

What Is BERT?
BERT stands for Bidirectional Encoder Representations from Transformers.

What does that mean in practice?

Google adapted its systems to understand natural language more effectively.

For context, Google already offered features such as flight status for flight-number queries, a calculator for mathematical expressions and stock charts for ticker symbols.

Another familiar example is autocomplete, which suggests a search before you finish typing.

Those features existed before BERT. The following example illustrates the language understanding discussed in the original article.

How Was BERT Useful?
The article introduces examples of BERT’s effect.

Consider the query “2019 brazil traveler to usa need visa”.

Before the update, a prominent result concerned how US citizens could travel to Brazil without a visa. Looking carefully at the query reveals a small but important difference.

The search concerns someone travelling from Brazil to the United States, rather than a US citizen going to Brazil.

The result shown after BERT was more relevant to that intent.

The example highlights the importance of prepositions such as “to”, whose meaning changes with query context.

To be continued…