Google's Changes In Search Algorithms Will Affect 10% Of Result Ranking

Google is rolling out key changes to its search algorithms, which may drive the alteration of results ranking as many as one every ten results. The changes are based on NLP or natural language processing.

According to Google, it has improved the searches by having a better understanding of how words relate to each other in a sentence. What it usually does is to treat sentences like word bags. It picked out keywords and return results based on them. However, in reality, the same words with different combinations will mean different things. However, the new changes will allow Google to understand the context of the query, thus increasing the chance of the algorithms returning exactly what users ask for.

These improvements are driven by BERT (short for Bidirectional Encoder Representations from Transformers), which leaves no word out and looks at the sentences as a whole.

What Google did to make BERT understands what it should do is to train it with Mad Libs games. A corpus of sentences in English with 15% of the words removed was used for this task. What BERT had to do is to try to figure out the missing words. This way proves to be highly effective in helping the model of NLP gasps the context of sentences.

As stated by the company, the changes had been rolled out for searches in English in the U.S in the last few days. Other nations and languages would be added shortly.

The improvements have been tested thoroughly to make sure that they are actually useful in enhancing users’ search results. One of these tests involved human testers who graded the results the algorithms return. Live A/B tests were also conducted before the changes were applied in real life.

However, the enhancements will affect only certain searches. BERT’s abilities are the latest addition of a series of tools Google uses to return results and also rank them. So far, we have no idea how these tools work together to display results.

The process is a secret because Google does not want anyone to have the chance to game with its system. Another, no less important, reason is that once machine learning is used to decide something, there is no way to tell why it does what it does.

That is why if machine learning does something wrong, it is really challenging for us to know the causes. According to Google, the company is working to make sure that BERT will not make bias worse. This is a widespread problem of machine learning with the data used to train them are also biased. Since the English corpus Google starts with has its own inherent biases, Google needs to keep a close eye on this issue.

There are many companies whose livelihood depends on Google, and whenever it announces a change, they are the first to listen and worry. For established publishers, Google said that it did not predict critical differences in the volume of traffic or where the algorithm directs them.

This is the change that all of the publishers who rely on the search tool to make an income should notice, not only for results in English, as Google said that other languages would follow suit.

This is arguably the largest modification Google rolled out since the beginning of the company.