Synonyms in long queries and SEO: how Google checks a synonym against words far apart in the query

Synonym validation by co-occurring terms is a method that decides whether a synonym can replace a query term by looking at the other terms of the query, including terms that are not next to it. Patent US 8,538,984 B1, "Synonym identification based on co-occurring terms," describes it. Google filed it on April 3, 2012, and the patent was granted on September 17, 2013.

What does patent US 8,538,984 describe?

Patent US 8,538,984 describes a query reviser that evaluates candidate synonyms in successive contexts and revises the query only when a context supports the synonym. Its 4 inventors are Abhijit A. Mahabal, Takahiro Nakajima, Zachary A. Garrett and Kenji Inoue. The patent holds 26 claims and 6 drawing sheets, and it cites the query synonym patent by John Lamping and Steven Baker.

The abstract defines the core: identify "a candidate synonym for the particular query term in context with an other non-adjacent query term," obtain "a respective confidence value for the other non-adjacent query term," check that the confidence value "satisfies a threshold," and "revise the original search query to include the candidate synonym."

What is a synonym rule?

A synonym rule is a statement that one term can stand for another. The patent's notation: "cat→feline" means that "the term 'feline' is considered to be a synonym for the term 'cat.'" Its counterexample: "banana" is not a synonym for "cat." "Synonym rules are typically unidirectional": feline can stand for cat without cat standing for feline.

Rules can apply "generally," without any other term in the query, or "only when particular conditions, or 'query contexts,' are satisfied." The synonym engine identifies synonyms by evaluating terms of previously received queries stored in query logs. These logs "can include previous queries that led to search results that users indicated were of high quality, or that were determined by the search system to satisfy users' information needs."

Before any context check, a general context evaluation engine can give a pair a high confidence value when a general rule identifies the candidate as a synonym, or when the query term and the candidate "occur frequently together" in related user queries and query results.

How does Google parse the query first?

Google parses the query to find the terms worth evaluating. The parser identifies 2 kinds of elements:

1. Skip words: words ignored when looking for synonyms and contexts. The patent's example: in "dry food for my cat," the parser identifies "my," "for" and "dry" as skip words and keeps "cat" and "food." Adjacent terms can then skip over skip words.
2. Compound phrases: in "I want to read a book by Shakespeare Romeo and Juliet," the parser can identify "Romeo and Juliet" as one n-gram, "instead of or in addition to" its words "Romeo," "and" and "Juliet." Treated as a single term, a compound's words may then not be evaluated separately, and the query is read as "read book Shakespeare Romeo and Juliet."

What are adjacent and floating contexts?

Adjacent and floating contexts are 2 ways to check a synonym against the rest of the query:

1. Adjacent context: the terms right before or after the query term, possibly skipping over skip words. The system looks up a confidence value for the candidate pair in the context of each adjacent term.
2. Floating context: terms that are "not adjacent to the query term." The system looks up a confidence value for the pair "when the query term occurs in the context of another term that is not adjacent."

The patent explains why floating contexts matter: in a long query, the context for a term at the beginning can be a term at the end. The floating context uses more of the query than the adjacent context, which happens when the query holds more than 3 terms (the query term and its 2 neighbors). A context can also rest on several query terms at once, close to each other, whether or not one of them touches the query term: this can increase "the decision strength." Links receive a comparable treatment: Google reads the rare words around a link to build its link context and spot link spam.

The position of an adjacent term can matter. For "AA" before "meeting," "American Airlines" is a definitive non-synonym; in "I would like a meeting with the AA pilots," where "meeting" comes before "AA," both "Alcoholics Anonymous" and "American Airlines" can be identified as synonyms in that adjacent context.

Different thresholds can apply "when evaluating a candidate term/synonym pair using confidence values from the adjacent context database" and "from the floating context database." Each engine thus sets "how narrow the evaluation and selection process will be."

How does the "Chicago style pizza" example work?

The patent illustrates both contexts with the query "What is the best place for Chicago Style Pizza," the term "place" and 2 candidates: "restaurant" and "finish." The threshold is 0.75.

1. Adjacent context: after "for" is set aside as a skip word, the neighbors of "place" are "best" and "Chicago." No value passes the threshold: "restaurant" gets 0.30 and 0.20, "finish" 0.70 and 0.10. Both candidates receive a low confidence.
2. Floating context: after "is," "the" and "for" are set aside, the non-adjacent terms are "pizza," "style" and "what." In the context of "pizza," "restaurant" passes the threshold and receives a high confidence. "Finish" does not pass it with any of the 3 terms.

To give a pair a single score, the engine can take the average of the values (0.25 for "restaurant" and 0.40 for "finish" in the adjacent context), the largest (0.30 and 0.70, or 0.90 for "restaurant" in the floating context) or the smallest.

The same logic applies to "book→play" in "read book Shakespeare Romeo and Juliet." In the patent's floating context example, the adjacent terms "read" and "Shakespeare" do not support "play," but the non-adjacent n-gram "Romeo and Juliet," at any position in the query, does. The patent's explanation: Shakespeare also wrote sonnets and narrative poems, while "Romeo and Juliet" is one of his plays.

How does Google decide to apply a synonym?

Google decides in up to 3 steps, according to the patent's flowchart (FIG. 6):

1. Check the adjacent context: if it supports the revision, apply the synonym.
2. Check for a definitive non-synonym: if the adjacent context shows that the candidate is "a definitive non-synonym for the particular query term," do not apply it.
3. Check the floating context: if a non-adjacent term supports the revision, apply the synonym; otherwise, do not.

The patent illustrates each branch with "AA":

• "AA meeting": next to "meeting," "American Airlines" is a definitive non-synonym, because of the strong association of "Alcoholics Anonymous" with "AA" in that context. "Alcoholics Anonymous" is applied: revised query "Alcoholics Anonymous meeting."
• "AA pet flight guidelines": the adjacent term "pet" supports neither candidate, and neither is a definitive non-synonym. The non-adjacent term "flight" gives "American Airlines" a high confidence value: revised query "American Airlines pet flight guidelines." "Alcoholics Anonymous" is not applied.

A definitive non-synonym can also come from the floating context. In "What is the number of pilots that are female at AA as of January 2011?", "Alcoholics Anonymous" "can be considered a definitive non-synonym" for "AA" "when the term 'pilots' is included anywhere in the original query."

The query revision engine can then apply its own threshold to cut the number of candidates, for example in a system "with limited search bandwidth," which reduces the number of revised queries.

The query can then be revised in 3 ways: "adding to the original query additional terms that are synonyms," "substituting terms that are synonyms," or "adding synonyms as a disjunction for the query terms," an OR between the term and its synonym. The search engine "can use the original query and the revised queries to identify and rank search results." On the results page, the original query terms or the synonyms that appear in the results can be formatted in bold or italics.

How does this patent relate to earlier synonym patents?

This patent extends context from neighbors to the whole query. The 2005 contextual synonyms patent validated a synonym with the words right around it, such as "download" for "free" before "photos." This patent goes further with non-adjacent terms and an explicit definitive non-synonym check. The semantic distance patent applies a related idea to documents: the structure of a page decides which terms belong together.

What does co-occurring term validation change for your SEO?

Co-occurring term validation means that Google interprets each word of a long query in light of the whole query. 5 consequences follow:

1. Write for full questions, not keyword fragments. Long, natural queries give Google more context to choose the right synonyms: pages that answer the full question are better placed to match.
2. Use the precise terms of your field. A synonym is applied only when the context supports it: the exact term remains the safest match.
3. Cover the compound phrases of your topic. Titles and named phrases such as "Romeo and Juliet" can be treated as units, and they can serve as context for other words: use them in full.
4. Do not expect all variants to merge. Some candidates are definitive non-synonyms in a given context, like "Alcoholics Anonymous" for "AA" when "pilots" appears in the query: a variant can need its own coverage.
5. Think in query context, not single words. The same word can take different synonyms depending on the other words of the query: map your content to whole query patterns.

The patent describes what Google's system can do. It does not confirm how Google revises queries with synonyms today.