Google’s Use of Neural Matching and the Search Results Algorithm
https://www.searchenginejournal.com/google-neural-matching/271125/
During our in-class discussion of Chapter 14, “Link Analysis and Web Search,” we covered a few of the methods used by search engines to rank search results. Among these methods were the hub and authority process and the PageRank algorithm. The hub and authority process ranks search results based on an authority score, which is assigned based on the strength and number of hubs which link to each result. The strength of a hub is determined by the number of authorities the hub links to. Therefore, a network with many links is more likely to have more accurate search result rankings than those which only have a few links to analyze. Further, the PageRank algorithm considers ranking scores as a set amount of “energy” which flows through the network. Pages which are linked to more frequently are sent more of this “energy,” and therefore have a higher ranking.
Both the hub and authority process and PageRank algorithm rely on the structure of the links formed in a network to rank search results. However, this reliance brings about its own set of limitations. For example, in networks which contain loops or small sets of nodes that can be reached from the rest of the graph but have no paths back up to the rest of the network, the energy may become “stuck” to a few nodes in a way does that accurately represent the true structure of the network. A similar limitation arises in the hub and authority process when a hub is newly formed and only points to just one authority.
Over the past few months, Google has refined its search result ranking algorithm with hopes of combating these limitations. To do so, Google has begun using neural matching to better understand synonyms. In fact, 30% of search queries have already been affected by the change.
Neural matching is a concept taken directly from artificial intelligence research within computer science. By understanding how the words in a search query relate to each other, search results are better directed towards the content a user is hoping to find, rather than being centered around the exact syntax used.
In an example given by Google employee, Danny Sullivan, imagine typing the question, “Why does my TV look strange?” into the Google search bar. An algorithm which relies on network structure might return pages which relate to keywords “why,” “TV,” and “strange.” These results might discuss different examples of a TV having a strange look and what causes this problem. However, upon asking “Why does my TV look strange?” humans understand that the user is truly asking for how he could fix the display of his TV. To make the computer understand the implications of spoken language, we turn to neural matching. Using Google’s new algorithm, search results might include troubleshooting manuals or support pages detailing the steps required to fix the display of the TV. Overall, Google’s implementation of neural matching generates search results which are more relevant to a user’s query than ever before.