Google's Query Refinements

Search engines aim to provide the most relevant results in response to queries but limitations can be seen on what is actually returned based on the queries used. Search queries can either be too specific or too general for search engines to recognize good results. Google has filed patent applications regarding alternative query terms or query refinements to offer a solution.

Search queries that are not too effective in providing good results include homonyms which are words that have the same sound or spelling but different meanings. Improper contexts in the choice of words can also be very confusing especially to search engines. Very general terms provide results that are too broad while very narrow terms can be very restrictive and may provide non-responsive search results.

Google presents a system and method that attempts to address this particular problem. In this system, a stored query and a stored document are associated as a logical pairing. The pairing is assigned a weight thus when a search query is issued, a set of search documents is produced. There is at least one search document that matches at least one document. Retrieval is done when the stored query and the assigned weight associated with it matches at least one stored document. A cluster is formed through this and scoring is done on at least one cluster relative to at least one other cluster. At least one such scored query is suggested as a set of query refinements.

The process starts when Google finds results by choosing the top 100 documents for clustering. During this phase, term vectors are computed for each of the said documents which were ranked by relevance score. The documents are matched to a stored document listed in an association database. Alternative query terms are found by looking at associations with queries that had been computed beforehand for the matched stored documents.

Term vectors are also created for alternative query terms. Clusters are created from both sets of term vectors to form groupings. Each cluster has a calculated cluster centroid. Search queries associated with a search document in the cluster are scored according to the distance from this centroid and the percent of stored documents occurring in the cluster. The best suggested query refinement contains the highest number search query terms and the most frequently seen in the documents in the cluster.

Other clusters and query names may be created to come up with additional suggested query refinements. Refinements are sorted by relevance scores. Alternative queries can include negated forms of terms appearing in the set of refinements but does not appear on the original search query. A number of predetermined search queries selected from past user queries can be used to arrive at a precomputed possible set of refinements. The predetermined queries would be issued while search results are maintained in a database for future user search requests. The refined queries would be provided to the user together with the results of the original search.

The precomputation stage happens before any query is entered into the search engine. It is best described with the use of at least four parts - associator, selector, regenerator and inverter.

The associator creates relevance-weighted relationships between stored queries and stored documents. The selector decides which stored documents and stored queries should be retrieved. The regenerator looks at query logs and selects stored documents based on previous searches. The inverter looks at the cached data and selects documents and associated queries based on the cached data.

The query refinements system itself has four parts. A matcher matches one or more stored documents to the actual search documents which have been generated by the search engine to answer a search query. It also identifies the stored queries and assigned weights using the associations corresponding to the matched stored documents. A clusterer forms one or more clusters using term vectors formed from the terms occurring in the matched stored queries and corresponding weights. The scorer computes centroids which represent the weighted center of each cluster's term vector. A presenter identifies the highest scoring search queries as one or more query refinements to the user. The interesting aspect about this approach is how user data is incorporated into results through the use of log files and cached information.

The patent application shows one way of achieving query refinements but no one really knows for sure exactly how Google comes up with alternative results. However, it offers some hints on how to create contents on websites and how to show up in these alternative results. By taking into careful consideration the words that people will probably search for and what appears in Google's results for search phrases, a clue can be provided on how the search refinements approach will treat a website.

Multi-Stage Query Processing

The determination of page relevancy in responding to queries from searchers considers how a term or phrase is used in the context of a page. A patent application that looks into the possible ways of considering the context of these words was likewise submitted by Google. It describes a multi-stage process that determines relevancy and finds results to a search.

The possible actions to be taken as described in this document can be divided into stages. The first stage deals with deletion of stop words, term stemming and expansion of queries to use things like synonyms and related terms that commonly co-occur with them. During this stage, the relevancy scores are created between query and each document computed with one or more scoring algorithms. The second stage uses adjacency and proximity of terms to rank documents. The third stage reviews the term attributes such as determining whether terms are titles, headings, metadata or whether these terms possess certain font characteristics. The fourth and last stage is the generation of snippets to return with results.

Interactive query refinements have shown that it can promote effective retrieval. Major search engines use the history of a user's actions such as queries or clicks to personalize search results. The query-specific web recommendations (QSRs) retroactively answer queries from the user's history as new results arise. Its main goal is to recommend new web pages for user's old queries. However, this will not be of any use unless the user has a standing interest in a particular query. Focus can also be shifted from individual queries to query sessions which includes all actions associated with a given initial query. A query is considered a query refinement of the previous one if both queries contain at least one common term.