Can Google Read Your Mind? Processing Predictive Queries

A new patent application published this morning doesn’t involve ESP, but it does attempt to anticipate what searchers are looking for. The document has the names of some prominent Google employees on it.

Anticipated query generation and processing in a search engine
United States Patent Application 20050283468
Published: December 22, 2005
Filed: June 22, 2004

The focus of this document is on returning search results quicker, and enabling personalization to make those results more relevant for the person searching.

In my recent Google Acquisition post, one of the companies I mentioned, Kaltix, specialized in personalization and speeding up search results. I also linked to a patent application there, assigned to Kaltix, which covered those types of issues.

The inventors named on that patent application, Sepandar D. Kamvar, Taher H. Haveliwala, and Glen M. Jeh, are also the inventors listed on this application. So, it shouldn’t come as a surprise that it covers personalization and searching speed.

The abstract:

A search system monitors the input of a search query by a user. Before the user finishes entering the search query, the search system identifies and sends a portion of the query as a partial query to the search engine. Based on the partial query, the search engine creates a set of predicted queries. This process may take into account prior queries submitted by a community of users, and may take into account a user profile. The predicted queries are be sent back to the user for possible selection. The search system may also cache search results corresponding to one or more of the predicted queries in anticipation of the user selecting one of the predicted queries. The search engine may also return at least a portion of the search results corresponding to one or more of the predicted queries.

The problem this patent application addresses?

Searches don’t start until a searcher types in the full query. Can those begin before the query is fully entered?

The answer is yes. If the search engine captures keyboard strokes as they happen, and starts sending partial queries to the search engine based upon a prediction of what the searcher is looking for, it may speed up the process.

How does it work?

It’s easier to show than to tell. The patent application doesn’t mention it, but take a look at Google Suggest, which seems to be a good example of anticipating queries.

What we don’t see with Google Suggest is some of the technology used to create that query list it displays in a dropdown under the query window. We also don’t see the ability to personalize those predictive searches. We don’t know if Google Suggest uses the processes described in this patent application.

The anticipated queries shown could be based upon previous queries, from other users (as if their entries were in dictionaries), and may be able to incorporate a user profile to choose which dictionary to use, and which queries to fetch.

The predictions could be cached, to speed up retrieval. If a similar set of results is cached from a previous searcher, the search engine could grab results without sending the query to the full index of the search engine.

The patent application notes that this search could be done through a browser, or toolbar or other input devices.

The document does illustrate a number of mechanisms that would be used to make this search seem very quick to a user of the search engine, including a few alternative methods for making the predictions described.

Triggering and producing queries

The point at which predictive queries start could be triggered a number of ways:

  • After a certain number of characters are entered.
  • After a certain amount of time has passed since the searcher starts typing.
  • A space bar entry or something similar (like a hyphen).

The final result would be triggered by the pressing of a search button, or the keyboard entry button, or a similar action.

Predicted queries could be presented in a dropdown, like in Google Suggest.

Predicted results may even be presented before a searcher finishes typing and before they possibly select one of the predicted queries. As the patent application notes, that would make the search engine seem very responsive.

Producing search results on these predicted queries from a cache, which is possible, increases the overall performance of the search engine by making it work less hard. If those results aren’t in the cache, then the search engine will retrieve them from the search engine’s inverse term index and document database.

One method that could also be used to help create these predictive queries is the use of an “auto-complete” server, which might try to match dictionary entries to what a searcher is typing. When those are presented in a dropdown, again like in Google Suggest, they might be presented in alphabetical order, or maybe in an order “based on a metric or score representative of how likely each entry is to match the user’s search query.”

Using Dictionaries to predict queries

There may be more than one dictionary against which to match partial queries. That can allow for personalization of the entries in those dictionaries matching the queries.

Example:

The dictionary used could match one or more of the interests of the user in their profile, such as “sports, music, news, finance, food, popular culture, etc.”

These dictionaries could be created from queries made in previous searches by the user, or by a “community” of users. The benefit of that is these predictive searches could be made up of either or both commonly submitted searches, and recent searches that could still be cached by the search engine. That potentially speeds up the return of results.

Entries in a dictionary

Each entry would have a term portion (the single or multiple word terms which could be a query), and a popularity value based upon how popular that term might be at that time. We know that Google can track how popular search terms are for the global index when we look at the Google Zeitgeist. Keeping track of those values in a popularity score should be possible.

Example:

Typing in “Bri” might return the following as predicted entries (queries):

  • Britain
  • British
  • Britney
  • Britney Murphy
  • Britney Spears

Note that in this example, capitalization is taken into account. In an alternative version implementing this patent application, it might not be.

A couple of other examples are presented that describe other ways this prediction could work.

What determines popularity?

Popularity scores for results could be informed by such things as breaking news stories, or popularity fads, or even when the query was last selected. That last bit of popularity information could be stored in the cache, and could possibly include a “reuse count” showing how many times the query was made.

User profile information is stored, and might contain information such as a preference in entertainers. For the example above, the entries “Britney Spears” and “Britney Murphy” might be then given more weight in (1) matching, (2) selecting, and (3) ordering than other terms.

Which terms are cached?

Some predicted queries are selected to have their search results cached, anticipating those being chosen as the final query.

Many factors could be used to select which are cached:

  • Popularity
  • Recency
  • A predefined metric, possibly considering a user profile.

When those queries are selected, they would be checked against results currently cached by the search engine. Single or multiple caches could be in place, and may need to be checked.

While many of these predicted queries may first be looked for in a cache, if there aren’t enough terms within the cache to reach a certain threshold, The full query (the final one), would be executed by the search engine, looking to the inverse document index and document database.

Those results could be cached for later use.

Combining multiple terms in a search

When a single term is being entered, and predictive results are generated, the searcher may include a second search term. This may trigger a new set of predictive queries, with their own set of results which may be assigned scores.

When the second query term is entered, scores for results of predicted queries based upon those could be combined with the predicted queries from the first term to form a better fit for the searcher.

This could continue if even more terms are added, enabling a set of search results to be built incrementally while the searcher is typing in a query.

When a searcher indicates that a search is complete, by hitting enter, or pressing the search button, or so on, those combined results may be more quickly available then if a full search was performed from the document database directly, without this prior processing. Some other methods of this combining of terms is described by the patent application.

Other issues

Some of the different hardware and software configurations that may be used to achieve this process are described in the remainder of the patent application.

These include such things as the makeup of a query processor, an auto-complete server, some details of the whole search engine, and how the search assistant mechanism used, such as a browser window or toolbar, would function.

There are other possible ways to incorporate this process into how a search engine works, and this patent application defines only some of the possible methods.

Share

23 thoughts on “Can Google Read Your Mind? Processing Predictive Queries”

  1. Thanks, Garrick.

    It is pretty hypnotic, seeing what the Google Suggest comes up with when you are typing away. The caching of results is interesting, too. Have to play with some searches there that involve popular news stories, and see if the queries suggested match those popular stories.

    I think I’ll play with that a little tonight, and then post it to the forums. :)

  2. Fascinating article, Bill. It’s pretty impressive how, with each character typed into the search query, information is being sent back and forth to create it’s suggestion. Thanks for posting that. You might like to share it at our favorite forum. =)

  3. Take a look at the options in the Google toolbar. You’ll see some of the concepts described in the patent at work there (I am not talking about the on-page word search feature).

  4. Those are in there, Michael.

    Which leads me to wonder if some of the caching, and incremental building of queries is also happening on the main search engine. I don’t know.

  5. Hiya Bill. I am trawling your archives a wee bit, and this one came up. The long tail of search has been a hot topic for quite some time now, and this post was written quite a while before the advent of the popularity of long tailed keyword phrases. In a way, whilst their predictive text system is definitely geared towards speeding up the information retrieval process, the long tail phenomenom was definitely fueled by Google Suggest and its ilk. Viewing things in retrospect can indeed be quite illuminating, I must say.

  6. Hi Jacques,

    Thanks. I like tying that concept to predictive queries.

    I know I was aware of the use of long tail terms at the time that I wrote this post. Chris Anderson’s article in Wired on the Long Tail http://www.wired.com/wired/archive/12.10/tail.html came out in October, 2004, and it did a nice job of popularizing the concept. I had seen the power of focusing upon long tail terms for a number of years even before that. I do very much agree with you that things like Google Suggest made even more people aware of them.

  7. This is obviously in effect right now with Google but I had always wondered how those “suggestions” were brought up. Was it based on my search history? Everyone’s search history? A bunch of information put together?

    Apparently it’s quite a few factors. :)

    Nevertheless, I do like that they do it. Helps from a searcher’s point of view.
    Thanks again Bill!
    Jason

  8. Hi Jason,

    It is potentially based on a number of different factors. Imagine a search engine using information from the social networks that you are a member of, and the words used by people whom you are connected to in those networks. Not sure how helpful that would be or not, but it would be another interesting approach that the search engines could take.

  9. Got to the old blog post from the September 2010 on instant search results or real-time ones and have to say this blog post has aged very very well. The more things change the more they stay the same. :)

  10. Hi Kimberly

    Thanks. I was wondering if Google would ever introduce immediate search results the way that they describe in this patent. I can mark this one off the list of “things that Google might do that they described in their patents.” :)

  11. Hi Matthew,

    I was thinking that Google might start showing immediate search results sometime soon after they introduced the drop-down query suggestions to their main web search in 2008.

  12. The long tail of search has been a hot topic for quite some time now, and this post was written quite a while before the advent of the popularity of long tailed keyword phrases.~

  13. Hi Amy,

    With these predictive query suggestions, long tail queries may be amongst the suggestions offered, but there’s more going on under the hood than that. The mystery here is often why Google picks some of the suggestions that it does for search, and with Google Instant, they are much more visible than they were previously.

Comments are closed.