How Search Engines Can Learn From Looking at Sequences of Search Queries

Whenever someone searches at a search engine, they not only get information in response to their search, but they also provide information to the search engine about the things they are searching for – information which the search engine might find useful in helping other searchers.

If that searcher performs another search related to their first search, then the search engine might create an association between the two search phrases that the searcher used, if the two phrases appear to be related. If they perform a series, or sequence, of searches on a concept, then the search engine might take advantage of that information.

If a lot of people perform that first search, and then the same second search, or that same search within a search session, then the search engine might decide that the phrases are semantically related to each other. Knowing that relationship exists between search queries might help the search engine help people find things on the web, and it might help provide better advertisements from the search engine.

A patent application from Yahoo explores how the search engine might find semantically related terms by looking at queries searched for by people in search sessions, and describes some of the processes behind how the search engine might determine that phrases may be related to each other. It also describes how a search engine might identify whether a query comes from a person, or from a program.

How Can Looking at Queries Help Search Engines?

Search engines may be able to learn a lot from searchers, by looking at the words that people search with and related searches performed by those searchers in the same search session. Search engines can also learn from the words chosen by advertisers who may choose to advertise on a search engine.

If a search engine can determine that two words or phrases are related in meaning in a language or in logic, knowing that “semantic” relationship could allow advertisers “to broaden or focus their online advertisements to relevant potential customers and allows searchers to broaden or focus their Internet searches in order to obtain more relevant search results.”

Imagine a search engine setting up a system that can determine semantically meaningful relationships between certain terms when those terms are searched with, and people click on particular pages when receiving those pages in response to those related terms.

The same system might also determine that terms that advertisers choose to relate to each other when setting up an online advertisement campaign are semantically related.

This system could be used to:

1) Discover semantically related words for purposes of bidding on online advertisements

Advertisers could optimize their advertising campaigns by finding terms to advertise with that are semantically related to terms they planned on using in their advertisments. Those semantically related terms may also appear on webpages that their advertisements could appear upon in an content advertising program.

Providing an advertiser the ability to serve an advertisement based on semantically related terms can increase the relevance and efficiency of those ads to an advertiser, so that the advertiser does not have to determine every possible word combination where the advertiser would like their advertisement served to a potential customer.

2) Assist a searcher performing research at an Internet search engine

When someone performs research at a search engine, the queries that they use could be treated as seed terms, with a search engine suggesting semantically related words related to the terms either within the search results generated in response to the original query, or as search suggestions.

This can help a searcher to broaden or focus their future searches, and receive more relevant search results.

3) Automatically discover terms related to evolving products of a given product brand or related to new events happening at a given place of interest

If a search engine is looking at semantically related terms, and someone searches for a term such as “Sony DVD player,” the search engine might return a term such as “Sony <new model> DVD player” which the search engine found from sequences of search queries containing both the term “Sony DVD player” and “Sony <new model> DVD player.”

Someone searching for an event might search for the word “Burbank” and might get a suggestion to look at “Yahoo! Burbank” as an event that they might want to find out more about. This way of automatically discovering terms related to new events happening at a given place of interest could be based upon sequences of search queries containing both the term Burbank and Yahoo! Burbank.

People recently searching for “Burbank” may have discovered an event happening at Yahoo known as “Yahoo! Burbank” and decided to search for that term. The sequence of searches let the search engine know that the terms are semantically related.

The Yahoo patent application is:

System and method for determining semantically related terms based on sequences of search queries
Invented by Kevin Bartz, Vijay Murthi, Benjamin Rey, and Shaji Sebastian
Assigned to Yahoo
US Patent Application 20080120072
Published May 22, 2008
Filed: November 16, 2006

Abstract

The present disclosure is directed to systems and methods for determining semantically related terms based on sequences of search queries.

Generally, a semantically related term tool examines search logs to associate search queries with a user submitting the search query. The semantically related term tool establishes a plurality of sequences of search queries, each sequence of search queries comprising one or more search queries associated with a common user and relating to a common concept.

The semantically related term tool receives one or more seed terms and determines one or more terms related to the received seed terms based on the established plurality of sequences of search queries.

Information from Queries

When a searcher submits a query to a search engine, the search engine usually records information about the search in a search engine log file. At the very least, the information collected includes the terms used to search with and the search results that were shown to the searcher in response to those search terms. But the search engine log could include additional information, such as:

  1. The positioning of URLs in the search results,
  2. A URL the searcher clicked on after viewing the search results,
  3. The number of times the searcher clicked on the URL,
  4. The time the search query was received,
  5. The search query received before the present query, or;
  6. Any other type of information relating to the search query or search results desired by the search engine.

It’s possible that the search engine logs could contain a lot more information about how a searcher interacts with the search engine and search results. This information might be used to draw conclusions about whether or not phrases used in searches are related to each other in some manner.

Related Searches and Patterns

A program that attempts to see how semantically related terms might be to each other examines search engine logs to determine sequences of search queries by an individual user that are related to a common concept.

Someone searching for a particular concept may try many many related search queries, hoping that one of the variations of the search query is better at returning pages that are more relevant to the particular concept being searched.

This pattern of sequences of queries around the same concept could help the search engine try to examine how related those terms might be.

Switching Topics and Multi-tasking

The search engine might not look at every query during a session as being related to each other. Sometimes, when a searcher is done researching one concept, they might move on to a new concept. Or they may be multi-tasking, and looking for information on multiple concepts at the same time.

So, the search engine might examine all of the search queries from a searcher for a particular period of time to determine if one or more search queries should be grouped into a sequence.

The search engine starts sorting queries found in the search engine log (or during real time searches) from a user into sequences, and may decide that a query is similar to one or more previously entered queries, and may place that query into an “established sequence.” If the query term appears unrelated, the search engine may start a new sequence.

A search session from an individual searcher may contain multiple sequences. How does a search engine identify whether or not queries are similar to each other?

Determining Similarity of Queries

The search engine might decide that search queries from a session are semantically related by looking at one or more properties of changes in search queries used in that session.

These properties of queries in a session can include things such as edit distances, prefix overlaps, and suffix overlaps.

Edit distance — also known as Levenshtein distance, is the smallest number of inserts, deletions, and substitutions of characters needed to change one search query into another search query.

Word edit distance — the smallest number of insertions, deletions, and substitutions of words needed to change one search query into another search query.

Prefix overlap — occurs between two search queries when one or more words occur at the beginning of both search queries. For example, the search queries “Chicago Bears” and “Chicago Cubs” have a prefix overlap due to the fact the word “Chicago” occurs at the beginning of the search queries.

Suffix overlap — occurs between two search queries when one or more words occur at the end of both search queries. For example, the search queries “San Francisco Giants” and “New York Giants” have a suffix overlap due to the fact the word “Giants” occurs at the end of the search queries.

The search engine might look at how related terms are by using those features and looking at:

  • A minimum, average, or maximum edit distance between a search query and the search queries of a sequence;
  • A minimum, average, or maximum difference between a time a search query was received at the search engine and a time the search queries in the sequence were received at the search engine;
  • The minimum, average, or maximum prefix overlap between a search query and the search queries of a sequence;
  • The minimum, average, or maximum suffix overlap between a search query and the search queries of a sequence;
  • Whether a search query and the search queries of a sequence relate to the same geographic location; or,
  • Whether a search query and the search queries of a sequence relate to a common product brand.

Assigning Weights to Search Sequences

The search engine might look at a number of search sequences, and then apply different weights to the search sequences based on the number of users who have submitted similar search sequences.

For example, if many searchers submit the terms “basketball and “Lakers basketball” in a search sequence, the search engine may have more confidence in that search sequence than when only one searcher has submitted the terms “basketball” and “Lakers basketball” in a search sequence.

Did a Human or a Program Submit the Queries?

After looking at a number of query sequences, the search engine might examine them to determine whether a program instead of a person submitted any portion of the search queries to the search engine. Search queries submitted by programs, such as rank checkers and others, are often not helpful in establishing relationships between search queries, so the search engine might remove any sequences of search queries created by a program.

How might a search engine determine if a program submitted a query?

It could look at the number of search queries submitted for a given period of time from a certain IP address or from a cookie associated with the query.

For example, if 100 search queries were submitted to a search engine in less than five minutes with the same cookie, or from an IP address that is not known for providing a high volume of traffic, it could be assumed that a program is submitting the search queries and all of the sequences associated with the program should be removed.

The search engine might also determine whether a program has submitted the search queries based on lexical patterns of the search queries. For example, if the search engine sees sequences of search queries with the words “link,” “suggest,” and “add” in conjunction with a domain name, those could be considered indicative of lexical patterns that suggest a program submitted the search queries to the search engine.

Seed Terms and Semantically Related Terms

After looking at a number of sequences of search queries, and removing any potential sequences of search queries created by a program, the search engine might receive search queries from searchers or advertising terms to be used in an ad campaign (together referred to as seed terms) and determine a set of terms semantically related to those seed terms based on the sequences of search queries.

Each of the received seed terms may be a single word or a phrase.

Further, each of the seed terms may be a positive seed term or a negative seed term.

Positive Seed Terms — A term that an advertiser would like to bid on to have the advertisement ad provider serve an advertisement or the type of terms a searcher interacting with a search engine would like to receive search results relating to.

Negative Seed Terms — A term that an advertiser would not like to bid on to have the ad provider serve an advertisement or the type of term a search interacting with a search engine would not like to receive search results relating to.

In other words, an advertiser, searcher, or system within the ad provider such as the ad campaign management system submits seed terms to receive more terms like a positive seed terms, while avoiding keywords like a negative seed term.

The search engine might send semantically related terms to an advertising campaign management system to be used in a keyword suggestion tool. Semantically related terms might also be used by the search engine to broaden or focus searches.

Conclusion

One of the most important takeaways from this patent application is that search engines don’t just look at information on the Web to help them with searches, but are also paying attention to how people search, and what they search for, as well as the terms that advertisers might relate together when they advertise on a search engine.

The patent filing provides some useful terms about queries, and how searchers may change their queries when they are searching, such as edit distances, prefix overlap, and suffix overlap.

We also learn a little about how a search engine might identify whether a query comes from an actual human searcher, or from a program that is performing searches on the Web.

Share

7 thoughts on “How Search Engines Can Learn From Looking at Sequences of Search Queries”

  1. Hi David,

    A number of patent filings from Google describe their use of information from query logs to provide search suggestions, query expansions during searches, and keyword suggestions to advertisers.

    Yahoo also provides query suggestions based upon user data from query logs, and will expand queries that people search upon based upon that data.

    What I found interesting in this patent filing was that they provide some details of things that they might be doing, like looking at the edit distances between queries within a query session to determine if terms are semantically related, or making distinctions between queries conducted by humans and those conducted by programs. :)

  2. One of a few interesting tidbits from last week. Isn’t query analysis just all the rage these days? Interesting the attempts at identifying query source types though… look out query spammers… ;0)

    I also now find mentions of AdServing/Analysis points interesting given recent developments in using Google in that area.

    Nice write up – thanks for the reading materials… back to blogging again this week… the break was nice I must say…

  3. Hi Dave,

    Thank you. Looking forward to some more blog posts from you. Hopefully your hand is recovered enough so that it isn’t too painful.

    There were a few interesting patent filings from last week. This one captured my attention for a couple of reasons, one of which was the section on identifying queries made by programs. Hopefully I’ll have some more on query analysis to come…

    The information about ad serving was interesting, too.

  4. Hey Bill,

    Great post as usual. It just cracks me to see the kind of effort going into the analysis of related queries. How about a really simple way – Ask the user!

    In its primitive form, bookmarks and folders are just that – a way for users to save and categorize the weblinks that are related. Users get utility and the product company gets semantical information.

    Tags are an interesting concept, but I think the folders will make a comeback, because they allow a much better foundation for collaboration.

  5. Thanks, Sumit.

    This isn’t looking at tags or annotations of pages as much as it is how people change the queries that they look at if they may not seem to be quite satisfied with the search results that they see, or want more information, and continue searching on the same subject.

    The assumption behind the approach seems to be that if enough people conduct similar queries around a certain topic, that sharing those similar queries with other searchers may provide them with a better search experience.

    Bookmarks/tags/folders do have some limitations. Often, tags describe the relationship between the tagger and the content instead of just the content – things such as “toread,” or “homework,” or the name of someone whom they might want to share the page with. That doesn’t help others understand what a page is about.

Comments are closed.