What is Google’s Neural Matching Algorithm?
Google issued a blog posting detailing their new algorithm which implements their adaptation on Neural Matching – an artificial intelligence method used to stretch beyond word matching to instead match words to concepts. According to sources, Google is using document relevance ranking, or ad-hoc retrieval, which involves ranking documents for relevance to a search query using only the query and the text of the individual documents. This contrasts with the more standard process of using network structure for information retrieval, which relies on PageRank and links. The new algorithm combines both ranking methods, first ranking based on relevance and network structure and then ranking based on relevance and ad-hoc retrieval.
Using relevance to rank web pages in response to a search query means that the top search results are not a product of keywords or links but instead of “context-sensitive encodings” in the algorithm. The algorithm relies heavily on what Google calls “super synonyms” between queries and documents, attempting to evaluate web pages based on their content and subsequent relevance to a given query. “Super synonyms” therefore avoid giving high ranks to pages that are “spammed” with keywords and direct synonyms but rather to ones with the proper content and meaning for a given search. Google describes this as making the “major jump forward from understanding words to understanding concepts.” Their algorithm relies on neural embeddings, which extract the underlying concepts of words then compare the concepts found in queries with the concepts found in web page documents during the re-ranking stage of the search process.
This marks a development in the way that searching occurs, relying on more realistic and perhaps accurate ways of ranking pages based on their implicit relevance to a particular search. It also demonstrates a deeper level of artificial intelligence and utilizing computers and algorithms to interpret true meanings and intentions, moving away from more rudimentary ways of matching words and synonyms.
https://www.searchenginejournal.com/google-neural-matching/271125/