Uses a Google search engine AI

RankBrain: The Evolution of the Google Algorithm

As a machine learning system, RankBrain draws on its experience with previous search queries, creates links and, based on this, makes predictions about what the respective user is looking for and how his query can best be answered. It is important to Ambiguities dissolve and the Meaning of previously unknown terms (e.g. neologisms).

However, Google does not reveal how the AI ​​system masters this challenge. However, SEO experts suspect that RankBrain can use Word vectors in a form that enables computers to interpret contexts of meaning.

Google has the open source machine learning software ready in 2013 Word2Vec published, with which semantic relationships between words can be converted into a mathematical representation, measured and compared. This analysis is based on linguistic text corpora.

In order to “learn” the meaningful connections between words, Word2Vec creates an n-dimensional one in the first step Vector space, in which every word of the underlying text corpus (one speaks of “training data”) is represented as a vector. Thereby there n in how many vector dimensions a word should be mapped. The more dimensions are chosen for the word vectors, the more relations to other words the program records.

In the second step, the created vector space is converted into a artificial neural network (KNN), which makes it possible to adapt it with the help of a learning algorithm so that words that are used in the same context also form a similar word vector. The similarity between word vectors is calculated using the so-called cosine distance as a value between -1 and +1.

In a nutshell: If you give Word2Vec any text corpus as input, the program delivers corresponding word vectors as output. These enable an assessment of the semantic proximity or distance of the words contained in the corpus. Will Word2Vec with new input confronted, thanks to the learning algorithm, the program is able to adapt the vector space and thus create new contexts of meaning or discard old assumptions: the neural network is “trained”.