Patent Grant
8548995
A. Field of the Invention
The present invention relates generally to the ranking of documents and, more particularly, to techniques for refining the ranking of an initial set of documents.
B. Description of Related Art
The World Wide Web (“web”) contains a vast amount of information. Locating a desired portion of the information, however, can be challenging. This problem is compounded because the amount of information on the web and the number of new users inexperienced at web searching are growing rapidly.
Search engines attempt to return hyperlinks to web pages in which a user is interested. Generally, search engines base their determination of the user's interest on search terms (called a search query) entered by the user. The goal of the search engine is to provide links to high quality, relevant results to the user based on the search query. Typically, the search engine accomplishes this by matching the terms in the search query to a corpus of pre-stored web pages. Web pages that contain the user's search terms are “hits” and are returned to the user.
In an attempt to increase the relevancy and quality of the web pages returned to the user, a search engine may attempt to sort the list of hits so that the most relevant and/or highest quality pages are at the top of the list of hits returned to the user. For example, the search engine may assign a rank or score to each hit, where the score is designed to correspond to the relevance or importance of the web page. Determining appropriate scores can be a difficult task. For one thing, the importance of a web page to the user is inherently subjective and depends on the user's interests, knowledge, and attitudes. There is, however, much that can be determined objectively about the relative importance of a web page. Conventional methods of determining relevance are based on the contents of the web page. More advanced techniques determine the importance of a web page based on more than the content of the web page. For example, one known method, described in the article entitled “The Anatomy of a Large-Scale Hypertextual Search Engine,” by Sergey Brin and Lawrence Page, assigns a degree of importance to a web page based on the link structure of the web page. In other words, the Brin and Page algorithm attempts to quantify the importance of a web page based on more than just the content of the web page.
In addition to conventional query-based engines, other systems exist that return documents in which it may be desirable to rank the documents. For example, a returned set of news articles about a particular news topic may be ranked. Postings gathered from message groups, such as Usenet groups, may also be ranked when returned to the user.
In any system that ranks documents, the general goal of the system is to rank the documents so that the more desirable documents are ranked higher. Thus, in such systems, improvements to the ranking technique are desirable.
The present invention is directed to a document ranking technique in which, for a given document, a set of related documents is determined. A related score is calculated based on the related document. The score may then be used to modify an original ranking of the given document.
One aspect of the invention is directed to a method for scoring documents. The method comprises obtaining an initial set of documents and generating a set of related documents for at least one document in the initial set of documents using a similarity criterion. The method further includes generating a related set score by applying a related set criterion to the set of related documents corresponding to the at least one document and scoring the at least one document using the related set score.
Another aspect of the invention is directed to a method for refining a ranking associated with a document. The method includes obtaining a set of documents related to the document, calculating a relevance ranking for the set of documents using a criterion, and modifying the ranking associated with the document based on the relevance ranking for the set of documents.
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate an embodiment of the invention and, together with the description, explain the invention. In the drawings,
The following detailed description of the invention refers to the accompanying drawings. The detailed description does not limit the invention. Instead, the scope of the invention is defined by the appended claims and equivalents.
As described herein, a ranking component refines the initial rankings of documents that are initially ranked via a scoring criterion. The ranking component may generate or receive a set of documents that are related to a given document. The ranking component may then apply a scoring criterion to the set of related documents to generate a related set score. The related set score may be used to refine the initial rankings.
Clients 110 may include client entities. An entity may be defined as a device, such as a wireless telephone, a personal computer, a personal digital assistant (PDA), a lap top, or another type of computation or communication device, a thread or process running on one of these devices, and/or an object executable by one of these device. Server 120 may include server entities that process, search, and/or maintain documents in a manner consistent with the principles of the invention. Clients 110 and server 120 may connect to network 140 via wired, wireless, or optical connections.
In an implementation consistent with the principles of the invention, server 120 may include a search engine 125 usable by clients 110. Search engine 125 may be a search engine such as a query-based web page search engine, a news server, or a Usenet message server or archiving source. One of ordinary skill in the art will recognize that other implementations of a search engine are possible. In general, in response to a client request, search engine 125 returns sets of documents to the client. These documents may be ranked and displayed in a ranking order determined consistent with aspects of the invention.
A document, as the term is used herein, is to be broadly interpreted to include any machine-readable and machine-storable work product. A document may be an email, a file, a combination of files, one or more files with embedded links to other files, a news group posting, etc. In the context of the Internet, a common document is a Web page. Web pages often include content and may include embedded information (such as meta information, hyperlinks, etc.) and/or embedded instructions (such as Javascript, etc.).
Processor 220 may include any type of conventional processor or microprocessor that interprets and executes instructions. Main memory 230 may include a random access memory (RAM) or another type of dynamic storage device that stores information and instructions for execution by processor 220. ROM 240 may include a conventional ROM device or another type of static storage device that stores static information and instructions for use by processor 220. Storage device 250 may include a magnetic and/or optical recording medium and its corresponding drive.
Input device(s) 260 may include one or more conventional mechanisms that permit a user to input information to client/server 110/120, such as a keyboard, a mouse, a pen, voice recognition and/or biometric mechanisms, etc. Output device(s) 270 may include one or more conventional mechanisms that output information to the user, including a display, a printer, a speaker, etc. Communication interface 280 may include any transceiver-like mechanism that enables client 110 to communicate with other devices and/or systems. For example, communication interface 280 may include mechanisms for communicating with another device or system via a network, such as network 140.
As will be described in detail below, server 120, consistent with the principles of the invention, performs certain searching or document retrieval related operations through search engine 125. Search engine 125 may be stored in a computer-readable medium such as memory 230. A computer-readable medium may be defined as one or more physical or logical memory devices. and/or carrier waves.
The software instructions defining search engine 125 may be read into memory 230 from another computer-readable medium, such as data storage device 250, or from another device via communication interface 280. The software instructions contained in memory 230 causes processor 220 to perform processes that will be described later. Alternatively, hardwired circuitry may be used in place of or in combination with software instructions to implement processes consistent with the present invention. Thus, implementations consistent with the principles of the invention are not limited to any specific combination of hardware circuitry and software.
Search engine 125 may include a document locator 330 and a ranking component 340. In general, document locator 330 finds a set of documents whose contents match a user search query. Ranking component 330 further ranks the located set of documents based on relevance. A more detailed description of the functionality implemented by search engine 125, document locator 330, and ranking component 340 will be described below.
Document locator 330 may initially locate documents from a document corpus by comparing the terms in the user's search query to the documents in the corpus. In general, processes for indexing documents and searching the indexed corpus of documents to return a set of documents containing the searched terms are well known in the art. Accordingly, this functionality of relevant document component 330 will not be described further herein.
Ranking component 340 assists search engine 125 in returning relevant documents to the user by ranking the set of documents identified by document locator 330. This ranking may take the form of assigning a numerical value corresponding to the calculated relevance of each document identified by document locator 330. Ranking component 340 includes main ranking component 345 and re-ranking component 347. Main ranking component 345 assigns an initial rank (a score) to each document received from document locator 330. The initial rank value corresponds to a calculated relevance of the document. There are a number of suitable ranking algorithms known in the art. One of which is described in the article by Brin and Page, as mentioned in the Background of the Invention section of this disclosure. Alternatively, the functions of main ranking component 345 and document locator 330 may be combined so that document locator 330 produces a set of relevant documents each having rank values. In this situation, the rank values may be generated based on the relative position of the user's search terms in the returned documents. For example, documents may have their rank value based on the proximity of the search terms in the document (documents with the search terms close together are given higher rank values) or on the number of occurrences of the search term (e.g., a document that repeatedly uses a search term is given a higher rank value).
The initial ranking scores assigned by main ranking component 345 may be refined by re-ranking component 347 to improve the relevance scores.
As previously mentioned, document locator 330 may return an initial set of documents, such as a set of documents generated in response to a user search query (act 401). Each of the documents in the initial set of documents may be initially ranked based on some scoring criterion that may generate a rank score or value for each document in the initial set of documents. The criterion may be based on, for example, a user search query, a topic (e.g., sports), a list of keywords, a geographical area, or similarity to another document or set of documents.
Based on the initial set of documents, re-ranking component 347 may then generate one or more sets of related documents (act 402). In one implementation, a set of related documents may be generated for each document, d, in the initial set of documents. The related sets of documents will be referred to herein as a related set D, which includes documents d1, d2, . . . , dn, where n is a positive integer greater than or equal to one. The related set D may be drawn from all documents known to search engine 125, or may be drawn from any desired subset of documents.
Based on the related set D and a scoring criterion (which may be any criterion, whether the same as, related or similar to, independent of, etc., the original scoring criterion), re-ranking component 347 may compute a related set score for D (act 403). The related set score may be based on a matching procedure between the document d and the related set D, or a subset of the related set D. The initial ranking score for document d may then be modified based on the related set score (act 404).
A number of techniques may be used to compute the set of related documents D for a particular document d in the initial set. In general, the documents in D are determined as documents that re-ranking component 347 determines to be somehow similar to document d.
In one implementation, similarity between document d and another document may be based on authorship or publication information. This type of similarity criteria can be particularly useful in the context of news articles. For example, if news articles are being ranked, one similarity criterion may be defined from the news source that published the news article d. The related set D may then be defined as a set of documents published by the same source as news article d. Thus, if d is an article from the New York Times then D may be the set of previous articles published by the New York Times.
Another news article based similarity criteria may be defined as news articles having the same author as the news article d. Under this criteria, other articles written by the same journalist may form related set D. For example, if document d is an article by Thomas Friedman, the related set D may be the set of previous articles by Thomas Friedman.
Yet another possible news article based similarity criterion may be defined as news articles from the same or similar publication sections. In the context of a newspaper, for example, the similarity criterion may be the same newspaper section. Other articles from the same section (e.g., “sports” or “business”) could be used to form related set D. If document d is an article in the “business” section then related set D could be defined as the set of other articles in the “business” sections of any newspaper.
Similarity criteria may be defined that are more specific to message groups, such as Usenet postings. In one implementation, a similarity criteria may be based on the author, where the author is defined by the email address of the poster. Thus, other postings by the same email address may form the related set D.
Search engine 125 may be used to search postings from multiple message groups. In this situation, a similarity criteria may be based on the group in which the posting appears. Other articles in the same message group would be considered related by this criterion. For example, if document d is a posting in the news group “soc.culture.Ethiopia,” the related set D may be defined as all the other postings from the news group “soc.culture.Ethiopia.” Another similarity criteria that can be used for message groups may be based on the thread in which the posting occurred. Thus, the related set D may be defined as all the other postings in the same thread as posting d. This similarity criteria will tend to return smaller sets than one based on the news group.
As mentioned previously, search engine 125 may implement a general web page search engine. In this situation, in which web pages are being ranked, the similarity criteria may define the related set D as other web pages from the same web site as document d, as web pages that link to document d, or as web pages to which document d contains a link. The same web site, as used herein, can broadly refer to documents on the same host or in the same domain. In some implementations, it may be desirable to define a same web site to cover multiple hosts or domains that are somehow related, such as, for example, a news publisher that publishes articles on multiple hosts or under multiple domain names.
As can be appreciated from the above discussion, the similarity criteria may be defined in many different ways, depending on the particular situation. In general, one of ordinary skill in the art will appreciate that numerous other classification or clustering techniques can be used to identify a set of related documents to a given document. The related documents in related set D can be pre-computed or generated when necessary. Further, one of ordinary skill in the art will recognize that combinations of the above-discussed similarity criteria or other similarity criteria may be used to define related set D.
Given the related set D and the original scoring criterion, in act 403, re-ranking component 347 computes the related set score for document d. In another implementation, the search query may be classified and the classification may be compared to the pseudo-document. For example, assume the original search query is “New York Yankees.” This query may be classified as a “Sports” query, and the pseudo-document may be compared based on how well it matches the topic of sports.
When the initial scoring criteria is based on geography, another possible technique for determining the related set score can include determining a geographical relevance vector of each document in related set D. For example, based on terms in the documents in D, a vector may be generated over all the documents in D, in which the vector defines a set of geographic scores. The geographic scores may represent a confidence level that the documents in related set D are relevant to a particular geographic region. For example, an exemplary vector for a particular related set D may include three non-zero confidence scores, such as (USA, 0.5), (Europe, 0.4), (Asia, 0.05). This vector may be matched with a geographic ranking criterion, such as the geographic search query “USA,” to produce the related set score. As a possible modification to this technique, the vector can be generated based on the source of each of the documents in D instead of on the documents themselves.
When the initial scoring criterion is based on topics, another possible technique for determining the related set score can include determining a topic of each document in related set D. The topics may be generated using automated classification techniques or drawn from either manually or automated pre-generated classification information (e.g., a hierarchical web directory tree). The topics for the documents in related set D may be combined to produce a vector of topic scores that defines confidence levels in the topics. The vector of topic scores can then be matched with the ranking criterion, in a manner similar to that described above, to produce the related set score.
In another possible technique for determining the related set score, a set of terms that are “strong” in each document in the related set D may be determined and combined to produce a vector of final strong terms for the documents in D. The determination of whether a term is “strong” can be based on, for example, a pre-determined list of terms or terms that are determined to have an inverse document frequency (idf) above a threshold level. The idf of each term in a document may be defined based on a ratio of the number of occurrences of the term in the document to the relative frequency of the term in the language or over the entire document corpus. Thus, terms that are generally less common in the corpus but that occur frequently in the document will have a high idf and may be classified as being strong. The final vector of strong terms can then be matched with the given ranking criterion (e.g., a search query) to produce the related set score.
As can be appreciated from the above discussion, the related set score can be defined in many different ways, depending on the particular situation.
The related set score and the initial ranking score for each document d may be combined to produce a modified (final) ranking score for document d. The initial set of documents returned in act 401 may then be re-ranked by re-ranking component 347 based on the modified ranking scores.
In one implementation, the modified ranking score may be calculated as a weighted sum as follows:
α(Initial_Score)+β(Related_Set_Score), (1)
where α and β are predetermined constants. The values to use for α and β may be determined by one of ordinary skill in the art through empirical trial-and-error techniques. Exemplary values for α and β may be 0.8 and 0.2, respectively. As a special case of formula (I), α could be set to zero and β could be set to one. In this case, the modified ranking score is equal to the related set score. Thus, in this case, the initial ranking scores may not even need to be calculated.
In another implementation, the modified ranking score may be calculated by using the related set score to boost the initial score as follows:
Initial_Score×(1+β×Related_Set_Score). (2)
As in equation (1), β may be a suitable predetermined constant.
A number of variations on the above-discussed techniques are possible. In the related set D, for instance, some documents may be weighted more heavily than others. For example, newer documents may be given more weight than older documents. Also, duplicate documents can be discarded or given less weight.
As previously mentioned, combinations of the discussed similarity criteria may be used to define related set D. Thus, for example, for news articles, the source, the journalist, and the section can all be used to define a related set D. Correspondingly, more than one related set score can be computed for each document. The multiple related set scores can then be combined by, for example, a summing or averaging function, and then used in formula (1) or (2).
In other alternate implementations, the related set score can be computed independently of the ranking criteria. For example, the related document set D can be scored based on the length of the included documents, on the timeliness of the documents, on the quality of the documents (determined by, for example, human evaluation or automated techniques based on grammar, spelling, or writing style), or based on popularity or usage characteristics of the documents.
In yet other implementations, there may be no initial criterion and all documents may be given the same initial score/ranking. Related-set ranking can then be applied to differentiate between the documents. Hence, related set ranking can be used (by itself) to create a ranking, or it can be used to modify other initial rankings.
The ranking component described above improves the ranking of documents initially ranked through a number of possible existing scoring criterion. The ranking component can be applied, for example, to the ranking of news articles, Usenet postings, or general web searches.
It will be apparent to one of ordinary skill in the art that aspects of the invention, as described above, may be implemented in many different forms of software, firmware, and hardware in the implementations illustrated in the figures. The actual software code or specialized control hardware used to implement aspects consistent with the present invention is not limiting of the present invention. Thus, the operation and behavior of the aspects were described without reference to the specific software code—it being understood that a person of ordinary skill in the art would be able to design software and control hardware to implement the aspects based on the description herein.
The foregoing description of preferred embodiments of the present invention provides illustration and description, but is not intended to be exhaustive or to limit the invention to the precise form disclosed. Modifications and variations are possible in light of the above teachings or may be acquired from practice of the invention.
No element, act, or instruction used in the description of the present application should be construed as critical or essential to the invention unless explicitly described as such. Also, as used herein, the article “a” is intended to include one or more items. Where only one item is intended, the term “one” or similar language is used.
The scope of the invention is defined by the claims and their equivalents.
| Number | Name | Date | Kind |
|---|---|---|---|
| 4823306 | Barbic et al. | Apr 1989 | A |
| 5855015 | Shoham | Dec 1998 | A |
| 5864846 | Voorhees et al. | Jan 1999 | A |
| 5918236 | Wical | Jun 1999 | A |
| 5920854 | Kirsch et al. | Jul 1999 | A |
| 5963940 | Liddy et al. | Oct 1999 | A |
| 5983221 | Christy | Nov 1999 | A |
| 6038560 | Wical | Mar 2000 | A |
| 6067552 | Yu | May 2000 | A |
| 6070158 | Kirsch et al. | May 2000 | A |
| 6070176 | Downs et al. | May 2000 | A |
| 6112203 | Bharat et al. | Aug 2000 | A |
| 6119114 | Smadja | Sep 2000 | A |
| 6138113 | Dean et al. | Oct 2000 | A |
| 6185550 | Snow et al. | Feb 2001 | B1 |
| 6189002 | Roitblat | Feb 2001 | B1 |
| 6285999 | Page | Sep 2001 | B1 |
| 6292830 | Taylor et al. | Sep 2001 | B1 |
| 6321228 | Crandall et al. | Nov 2001 | B1 |
| 6502081 | Wiltshire et al. | Dec 2002 | B1 |
| 6526440 | Bharat | Feb 2003 | B1 |
| 6581057 | Witbrock et al. | Jun 2003 | B1 |
| 6591261 | Arthurs | Jul 2003 | B1 |
| 6766316 | Caudill et al. | Jul 2004 | B2 |
| 6778997 | Sundaresan et al. | Aug 2004 | B2 |
| 6829599 | Chidlovskii | Dec 2004 | B2 |
| 6842876 | Schilit et al. | Jan 2005 | B2 |
| 6848077 | McBrearty et al. | Jan 2005 | B1 |
| 6901399 | Corston et al. | May 2005 | B1 |
| 6947920 | Alpha | Sep 2005 | B2 |
| 6965900 | Srinivasa et al. | Nov 2005 | B2 |
| 7003513 | Geiselhart | Feb 2006 | B2 |
| 7146409 | Bushee et al. | Dec 2006 | B1 |
| 7197497 | Cossock | Mar 2007 | B2 |
| 7254587 | Lee et al. | Aug 2007 | B2 |
| 7260571 | Amitay et al. | Aug 2007 | B2 |
| 7505964 | Tong et al. | Mar 2009 | B2 |
| 7610281 | Gandhi et al. | Oct 2009 | B2 |
| 7739264 | Jones et al. | Jun 2010 | B2 |
| 7836391 | Tong | Nov 2010 | B2 |
| 7962479 | Jones et al. | Jun 2011 | B2 |
| 7996396 | Leblang et al. | Aug 2011 | B2 |
| 8024326 | Tong et al. | Sep 2011 | B2 |
| 8024327 | Tunkelang et al. | Sep 2011 | B2 |
| 8086594 | Cao et al. | Dec 2011 | B1 |
| 8086619 | Haahr et al. | Dec 2011 | B2 |
| 8086631 | Gollapudi et al. | Dec 2011 | B2 |
| 8224827 | Dean et al. | Jul 2012 | B2 |
| 8250092 | Gollapudi et al. | Aug 2012 | B2 |
| 8266143 | Dean et al. | Sep 2012 | B2 |
| 8380705 | Tong et al. | Feb 2013 | B2 |
| 20030115188 | Srinivasa et al. | Jun 2003 | A1 |
| 20030217052 | Rubenczyk et al. | Nov 2003 | A1 |
| 20030220913 | Doganata et al. | Nov 2003 | A1 |
| 20040059726 | Hunter et al. | Mar 2004 | A1 |
| 20040128273 | Amitay et al. | Jul 2004 | A1 |
| 20040236725 | Amitay et al. | Nov 2004 | A1 |
| 20050010605 | Conrad et al. | Jan 2005 | A1 |
| 20050060311 | Tong et al. | Mar 2005 | A1 |
| 20050154761 | Lee et al. | Jul 2005 | A1 |
| 20060271524 | Tanne et al. | Nov 2006 | A1 |
| 20070174340 | Gross | Jul 2007 | A1 |
| 20070192369 | Gross | Aug 2007 | A1 |
| 20070234140 | Lee et al. | Oct 2007 | A1 |
| 20080010270 | Gross | Jan 2008 | A1 |
| 20080016050 | Stensmo | Jan 2008 | A1 |
| 20080114721 | Jones et al. | May 2008 | A1 |
| 20080126335 | Gandhi et al. | May 2008 | A1 |
| 20080140699 | Jones et al. | Jun 2008 | A1 |
| 20090216696 | Downs et al. | Aug 2009 | A1 |
| 20100131523 | Yu et al. | May 2010 | A1 |
| 20110029517 | Ji et al. | Feb 2011 | A1 |
| 20110040787 | Cierniak et al. | Feb 2011 | A1 |
| 20110191310 | Liao et al. | Aug 2011 | A1 |
| 20120016888 | Dean et al. | Jan 2012 | A1 |
| 20120089588 | Gollapudi et al. | Apr 2012 | A1 |
| 20120209838 | Dean et al. | Aug 2012 | A1 |
| 20120290926 | Kapadia et al. | Nov 2012 | A1 |
| Entry |
|---|
| Brian Amento, Loren Terveen, and Will Hill; “Does ‘Authority’ Mean Quality? Predicting Expert Quality Rating of Web Documents”; ACM-SIGIR; Jul. 2000, pp. 296-303. |
| Bharat, K. and Henzinger, M., “Improved Algorithms for Topic Distillation in a Hyperlinked Environment,” Aug. 1998, ACM SIGIR '08, Melbourne, Australia, pp. 104-111. |
| Osinski, Stanislaw, “An Algorithm for Clustering of Web Search Results,” Jun. 2003, Master Thesis, Poznan University of Technology, Poland, pp. 1-101. |
