A Google patent granted this week, and filed in 2000, describes a reranking process for semantically meaningful phrases appearing in search queries.
This isn’t the phrase-based indexing that some have blamed a -950 penalty for, nor the Infoseek phrase-based indexing technology that Google acquired in 2005, but it does seem like a sensible way to do some reranking of results when there are semantically meaningful relationships between words which appear in a query together.
3 Comments
This isn’t the phrase-based indexing that some have blamed a -950 penalty for, nor the Infoseek phrase-based indexing technology that Google acquired in 2005, but it does seem like a sensible way to do some reranking of results when there are semantically meaningful relationships between words which appear in a query together.
3 Comments
Comments
This seems awfully tough on the other search engines. However, the RPGer in me is pleased to see Baldur’s Gate being used as an example!
Wow, Bill, that seems like a really important patent. And now it has been granted. Your excellent article does seem to indicate that this type of semantic unit processing could lead to some serious frustration for searchers. Your example of the country western migration is a perfect illustration of how difficult user intent can be to uncover. This will be especially tough if Google does, in fact, base user intent upon frequency or popularity. Will searches for "an apple a day" bring up SERPs for computers only? We already bump into this type of problem with non-compound words (9 of Google’s top 10 results for the word Amazon return references to the bookstore, not the river). It’s often those searches where one is really trying to drill down to some specific, hard-to-find, information that cause one to write those long tail queries. Can you see this patent causing trouble if Google actually applies it? Thanks for the great information! Miriam
Thanks, Miriam. The "country western migration" example is in the patent filing, rather than in my post, but it’s a good example of what can go wrong when trying to treat compounds as meaningful phrases. The patent application method actually intends to avoid that kind of problem, and it seems to me like a thoughtful approach. They show how this method would be able to distinquish between using "country western" or "western migration" to rerank results in a longer search query. "As an example of Acts 208 and 209, assume that the query "leaving the old country western migration" has four substrings having FRAC[s] greater than f, namely: "the old country" (f=0.53), "old country" (f=0.55), "western migration" (f=0.5), and "country western" (f=0.4). In Act 208, semantic unit locator 121 discards "old country," as it is completely contained in the longer substring "the old country," which has a FRAC value greater than the threshold. In Act 209, "country western" is discarded as it overlaps with "the old country" and "western migration" and has the lowest FRAC value of these three terms. Thus, in this example, the processed query contains the compounds "the old country" and "western migration," along with the single search term "leaving." Although the example described above included consideration of the word "the," other implementations may exclude the consideration of commonly-occurring words." In other words, a query is entered, and the search engine tries to find documents on the Web that have all of those words, but not necessarily in that order. This method is then used, and can cause the results to be reranked based upon one of the combination of words, which may seem to be a phrase that has meaning of its own. A simple example, "ice" means one thing, and "cream" means something else - but "ice cream," means something else entirely. A search for "ice cream" (without the quotation marks) may return a page where "ice cream" appears as a phrase rather than returning a result which only has the sentence, "I fell on the ice, going to the store to buy cream." The "leaving the old country western migration" is more complex, but it still involves finding meaning in a combination of words, and the example shows which phrase may be chosen first to use to rerank results. Good point, Andrew. If something like this can be patented, what do the other search engines do? I know that Yahoo has filed a number of patent applications that describe a "concept network" that is constructed differently I suspect that Microsoft has something similar in their arenal of patent filints.