How Google May Perform Entity Recognition

by Posted @ Feb 04 2016


Google replaced its head of search, Amit Singhal this week, with someone whom came to Google when the company acquired Metaweb. The person who founded Metaweb, John Giannandrea, has been a vice-president at Google, in charge of Google’s AI efforts, and will be taking over as head of search at Google. Metaweb is the company that ended up bringing the knowledge graph to Google, and filling Google search results with knowledge panels related to entities that show up in those search results.

We were first made aware of entities by Google back in 2012, in their video “Introducing the Knowledge Graph”:

The home to entities on Google is in The Knowledge Graph, and Wikipedia tells us what a knowledge graph is:

The Knowledge Graph is a knowledge base used by Google to enhance its search engine’s search results with semantic-search information gathered from a wide variety of sources. Knowledge Graph display was added to Google’s search engine in 2012, starting in the United States, having been announced on May 16, 2012. It provides structured and detailed information about the topic in addition to a list of links to other sites. The goal is that users would be able to use this information to resolve their query without having to navigate to other sites and assemble the information themselves. The short summary provided in the knowledge graph is often used as a spoken answer in Google Now searches.

Google displays information about Entities from its knowledge graph in knowledge panels that it displays to the right side of search results. If Google can identify that a specific entity is associated in some manner with a query that we as a searcher type into a search box, it will tell us more about that entity. That knowledge panel might include information about different aspects of that entity, from sources such as Wikipedia, Freebase and other knowledge bases. Google might also tell us about other entities that people sometimes search for in addition to the entity we searched for. Here is a knowledge panel result in Google for Nikoli Tesla:

Nikola Tesla Results in Google

So Google works to provide information to us about the many different entities that may appear in queries and searches. Google may associate names of entities with those entities when showing us knowledge panels for them. When they are asked about it in queries, Google may try to recognize what entity is being asked about to give us that information. For example, when Google is asked, “who is the CEO of Alphabet?”, it tells us “Larry Page” and provides information about him (in this case, from Wikipedia).

CEO of Alphabet Larry Page

A Google patent was granted this Tuesday that focuses upon entity recognition, and predicting entities when portions of sentences might refer to those entities. The knowledge bases that make up Google’s knowledge graph are filled with sentences about entities, which are likely a source of many answers to questions about them.

The patent provides some examples of how it might predict the identity of entities that have been referenced in complete sentences, by gathering information about those entities and then using that information from those sentences to identify an entity that might be asked about. Here’s an example from the patent:

The entity identification system obtains complete sentences that each include entity text that references a first entity. An example complete sentence may be, “In 1890, the President of the United States was Benjamin Harrison.” In the context of this example sentence, the entity text, “President of the United States” references the person, “Benjamin Harrison.” For each complete sentence, the entity identification system emulates typing the sentence and providing portions of the sentence to an entity identification model. The entity identification model determines a predicted entity for each portion of a sentence that it receives as input.

The patent builds upon this example by telling us how that information from a complete sentence might be used to identify a specific entity:

The entity identification system may continue to input portions of the complete sentence, updating the entity identification model for each portion based on the accuracy of the prediction. For example, the entity identification model may receive, “In 1890, the President of the Un,” as input. While this input may still refer to multiple entities, e.g., Benjamin Harrison, the President of the United States in 1890, or William Ellison Boggs, the President of the University of Georgia in 1890, the entity identification model may correctly predict Benjamin Harrison for this portion of the complete sentence. In this situation, the entity identification model may be updated, for example, by increasing a confidence and/or likelihood that Benjamin Harrison will be identified for a sentence that begins “In 1890, the President of the Un.”

The patent is:

Entity identification model training
Invented by: Maxim Gubin, Sangsoo Sung, Krishna Bharat, and Kenneth W. Dauber
Assigned to: Google
US Patent 9,251,141
Granted February 2, 2016
Filed: May 12, 2014


Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training an entity identification model. In one aspect, a method includes obtaining a plurality of complete sentences that each include entity text that references a first entity; for each complete sentence in the plurality of complete sentences: providing a first portion of the complete sentence as input to an entity identification model that determines a predicted entity for the first portion of the complete sentence, the first portion being less than all of the complete sentence; comparing the predicted entity to the first entity; and updating the entity identification model based on the comparison of the predicted entity to the first entity.


Search Engines like Google and Bing are transforming from providing information to searchers from an index of pages on the Web to providing information from an index of data they find on the Web. This means answering questions to searchers directly and being information resources that are focusing upon indexing information. Information about entities can be found in sources like Google’s knowledge graph, which can be used as training source to build entity recognition models, as described in this patent.

The patent provides a number or examples of how it might use complete sentence information to build models that can help with the search engine recognizing specific entities. One of these includes the sentence “The tallest building in the world is the Burj Khalifa”, and how that information changes when you are talking about the world in 1931, when the tallest building in the world was the Empire State Building. The patent tells us that the models to recognize entities it builds might use partial sentences, returning information about the tallest person in the world and the tallest mountain in the world. Reading this whole patent is recommended for more details on how models helping to recognize entities might be built from sources like Google’s Knowledge graph.

I thought it was interesting that one of the inventors listed on this patent was Krishna Bharat, who has done a lot of work behind Google News -another place we see entities spring up regularly. If Google is learning about entities, and how to answer questions about them, it might be doing that in part by learning from the News about what they are up to. We heard this past summer, about Google’s Deep Mind and their efforts to read CNN and The Daily Mail as training sources to learn about the world. They should be able to find information about a lot of entities from those sources.

And, with John Giannandrea, former founder of Metaweb, taking over as head of search at Google, we may see more knowledge graph information entering search results in the future, and more information about entities filling those results. The future of Google may involve lots of entities.

subscribe to our newsletter


  1. Gerry

    February 04th, 2016 at 2:03 pm

    Hey Bill cheers for your insight on this especially your takeaways, for me what stands out most outside of SEO in general and search as a way of obtaining leads for businesses, is their use of traning/learning anything from CNN and The Daily Mail, to me that’s quiet shocking to say the least, sounds like private interests are going to tell us what the news is in the near future what do you think ?


    • Bill Slawski

      February 04th, 2016 at 2:19 pm

      Hi Gerry,

      Thank you. The paper that tells us about Deepmind using CNN and the DailyMail as training sources tells us that those were selected because of the way they were formatted, which supposedly made it easier to do the kind of training described in that paper. I think those may be the starting points, but i expect Google to use more sources as time passes, and with Google News actively looking at a lot of news, they have lots of options.

  2. Gerry

    February 04th, 2016 at 2:41 pm

    It seems like a strange direction for Google to be taking to me when they hold all the data on nearly every click and search on the net, why they need to learn anything from other sources just doesn’t make sense or am I missing something ? Google news is my new daily place of interest cheers.


    • Bill Slawski

      February 04th, 2016 at 3:01 pm

      Hi Gerry,

      It seems that searchers have changed the kinds of things that they are searching for. According to a study done by Microsoft (pdf) people seem to be searching for information about entities more frequently:

      “According to an internal study of Microsoft, at least 20-30% of queries submitted to Bing search are simply name entities, and it is reported 71% of queries contain name entities”

  3. Gerry

    February 04th, 2016 at 5:10 pm

    Cheers Bill I will give those resources a good look at to get a better grasp of what it’s about, there is a whole lot of changes going on it’s always good to stop by here for your insights cheers again.


    • Bill Slawski

      February 04th, 2016 at 6:02 pm

      Thanks, Gerry.

      There have been a lot of changes to the Web, and to search. Trying to keep up with those is something that we try to do; and is why I find myself digging through the USPTO website looking at patents from Google, Apple, Microsoft, Yahoo, Facebook, and a number of others. Sometimes the things I end up running across are surprising, but this stuff with entities and the semantic web has been going on for a while now. We’ll be keeping a close eye upon it.

  4. Marianne Sweeny

    May 09th, 2016 at 10:48 pm

    Greetings Bill,
    I believe that Krishna Bharat was also involved in Topic Sensitive PR, another seminal algorithm of a semantic nature.


  5. Pingback: 2016 Important SEO Patents from Google - SEO by the Sea

  6. Pingback: How Google's Knowledge Graph Updates Itself by Answering Questions - SEO by the Sea ⚓

Leave a Comment