BERT is the new search algorithm behind Google searches. The new tool uses artificial intelligence to understand the user’s requests, based on a construction of meaning that takes into account the context of each word and phrase. Thus, the search engine can offer more accurate results focused on what matters to the person. The technology is already applied in the 70 countries versions of Google’s search service.
What is an algorithm?
The term algorithm has its origin in mathematics and, from computer science, it serves to classify a set of steps that a given software needs to perform to reach a result.
In the case of a search algorithm, and in a simplified way, these steps could be to capture the text that the user inserts in the search box and search the Internet for results that are then displayed on the screen. Generally speaking, when someone quotes the Google search algorithm (or any algorithm, in fact), they are referring to a software engine that performs a specific task.
Before applying the new algorithm, Google searches gave more relevance to keywords than to the context of the questions. Moreover, they organized the results giving more evidence to the most accessed content: that’s why, sometimes, when searching for questions about a product, you can be directed to online stores before finding the information that matters.
These results, created by the relevance and importance of keywords, lead to inconsistencies, especially in a scenario where 15% of all searches performed per day are completely unprecedented. Because they were never done, the previous algorithm didn’t have good parameters to organize the answers, deepening the problem of imprecision of the results.
Regarding the search itself, the algorithm processed word by word independently and this created a habit in users. According to Google, people tend to search only using the keywords they consider important, without building a question with a beginning, middle and end – something Google believes is not only possible but desirable, for better results with BERT.
What does BERT do?
The image above compares the results of the same search with or without BERT. The question is, “2019 brazil traveler to USA need a visa”. Without BERT, Google gives as its first answer a news article in the Washington Post whose title is “American citizens can travel to Brazil without documentation or a visa”, which is exactly the opposite of what was searched for.
With BERT, the algorithm takes the formulation of the phrase into context and realizes that the search is relating Brazil as the starting point and the United States as the destination, which gives a more accurate result: the site of the American Embassy in Brazil.
The idea behind BERT is to offer a search tool that does not give so much weight to keywords, but rather to the context of searches. Thus, it tries to infer the real meaning of the search and offer results more in line with what the person is searching for. The tool is a complement to RankBrain, the first artificial intelligence technique associated with search results and implemented by Google five years ago.
To this end, Google has trained the tool with text extracted from Wikipedia articles. With this database in hand, Google developers began training the AI with tests to complete, similar to those children perform in the literacy phase: text passages had words removed and the algorithm needed to figure out the ideal term to complete the content, without damaging its meaning.
BERT should, therefore, enable more accurate and contextual results to Google’s search engine. This ability tends to deepen the quality of the tool since it is common that the user simply does not know exactly how to formulate a search correctly. As Google observes in its official blog, “we often use search tools to learn” something new that we do not know how to search efficiently.
In the image, one more comparison. The search is “beauticians can handle a lot at work”. Without BERT, the result is a comparison between two types of professionals. With BERT the answer is more relevant: the result points to a site that details what are the demands that are part of the daily professional life of a beautician.
Another development of the technology has to do with voice assistants and speech searches. According to Google, the application of BERT’s more contextual processing should allow users to interact more naturally with voice assistants, since the algorithm should be able to better understand the directions of instructions.