Generalized term frequency scores in information retrieval systems

[0002] This patent application is a continuation-in-part of the following U.S. patent applications: “Weighted Term Ranking for On-Line Query Tool”, Ser. No. 09/282,730, now pending to Jay Ponte; and “Hybrid Category Mapping for On-Line Query Tool”, Ser. No. 09/283,268, now pending to Jay Ponte, having a common application date of Mar. 31, 1999, having the same inventor and assignee as herein named.

[0003] This invention relates to techniques for organizing material on computer networks for retrieval, and more particularly to methods of indexing material of interest to a user.

[0004] Computer networks have become increasingly important for the storage and retrieval of documents and other material.

[0005] The Internet, of which the World Wide Web is a part, includes a series of interlinked computer networks and servers around the world. Users of one server or network connected to the Internet may send information to, or access information on, other networks or servers connected to the Internet by the use of various computer programs which allow such access, such as Web browsers. The information is sent to, or received from, a network or server in the form of packets of data.

[0006] The World Wide Web port ion of the Internet comprises a subset of interconnected Internet sites which may be characterized as including information in a format suitable for graphical display on a computer screen. Each site may include one or more separate pages. Pages, in turn, may include links to other pages within the site, or to pages in other Web sites, facilitating the user's rapid movement from one page or site to another.

[0007] In view of the quantity of information and material available on computer networks such as the Web, and for other reasons as well, automated or semi-automated techniques for retrieving information that is thought to be relevant to a user at a given time may be employed. These techniques may be utilized in response to a specific user request, as when a search query by a user seeks information. These techniques also may be utilized when a user is accessing certain material, in order to make available material that it is thought may be of interest to a user who has accessed the original material. These techniques may also be utilized when a user, given access to particular material, requests other similar material. Other situations when these information retrieval techniques may be employed will also be apparent to one of ordinary skill in the art.

[0008] Some information retrieval techniques such as are employed in these circumstances choose documents for retrieval from among documents in a collection based upon the occurrence of specified terms in the documents in the collection. (Hereinafter, for simplicity, “document” shall be used to refer to the items, such as Web pages or Web sites, in the collection being analyzed.) There are a variety of different techniques for specifying the terms to be used. (A “term” may be any word, number, acronym, abbreviation or other collection of letters, numbers and symbols which may be found in a fixed order in a document.) In some methods, a search may be made among the documents in the collection for some or all of the terms in a search query generated by the user. In other methods, a search may be made for some or all of the text of a given document. (In some methods, all terms except certain common words, referred to as stop words, such as “the” or “and”, may be included in the search.) In other methods, a search may be made for index terms which have been associated with that document by various means. Still other methods will use a combination of the above techniques, and further approaches to selecting terms for which a search is to be made will be familiar to one of ordinary skill in the art.

[0009] After a list of terms for which a search is to be made has been compiled, many information retrieval techniques then proceed by calculating scores for each document in the collection over which the search is being made, based upon the occurrence of the terms on the list in the documents. These scores which are calculated may be referred to as term frequency scores, insofar as the score assigned to a document depends on the frequency of occurrence of terms in the document.

[0010] There are a variety of different formulae which may be used to calculate these term frequency scores, including for example the Robertson's term frequency score (RTF). Term frequency score formulae may assign varying weights to terms found in a document, depending upon such factors as the relative rareness or commonness of the term. Other factors which may be used to vary the weight assigned to a term in calculating a term frequency score will also be apparent to one of ordinary skill in the art.

[0011] Documents in a collection which is being searched may be divided into different sections or segments, such as an introduction or summary, a main body, footnotes, captions, and the like. Other divisions of documents will be apparent to one of ordinary skill in the art.

[0012] A Web site may permit a user to obtain lists of relevant items of interest, such as Web sites, other documents or names of merchants carrying merchandise in particular categories. The site may be organized so that an item of interest may be considered to be in more than one category. The site may be organized so that the categories presented to the user may vary, depending on a term or terms specified by the user. If this approach is utilized, the user may input terms that relate to the merchandise in which he is interested, such as “automobiles”, and in return he may be presented with several categories, such as “automobiles, manufacturers” or “automobiles, sales” or “automobiles, service.” The categories presented may be chosen by any one of a number of techniques that will be familiar to one of ordinary skill in the art.

[0013] It may be desirable present additional material to a user who is searching for items of interest. For example, it may be desirable to present the user with banner advertisements which relate to the item of interest for which he is searching.

[0014] The above-mentioned and other features of the invention will now become apparent by reference to the following description taken in connection with the accompanying drawings in which:

[0015] FIG. 1 is a schematic diagram of a computer system that may be operated according to the present invention.

[0016] FIG. 2 illustrates a relationship between terms and documents.

[0017] FIG. 3 is a flow chart which illustrates a process, according to the present invention, of automatically assigning index terms to documents.

[0018] FIG. 4 illustrates a relation ship between terms, documents and index terms when some but not all documents in a collection have had index terms manually assigned to them.

[0019] FIG. 5 is a flow chart which illustrates an overall process, according to the present invention, of automatically assigning index terms to documents, where some documents have previously had index terms assigned to them.

[0020] FIG. 6 illustrates a relationship between terms, documents and index terms after documents in a collection have had index terms assigned to them automatically.

[0021] FIG. 7 illustrates a relationship between items of interest, categories and supercategories when some but not all categories in a collection have been manually assigned to supercategories.

[0022] FIG. 8 is a flow chart which illustrates an overall process, according to the present invention, of automatically assigning categories to supercategories, where some categories have previously been assigned to supercategories.

[0023] FIG. 9 illustrates a relationship between items of interest, categories and supercategories when categories in a collection have been assigned to supercategories.

[0024] FIG. 10 illustrates a relationship between categories and supercategories.

[0025] FIG. 11 is a flow chart which illustrates a process, according to the present invention, of assigning a supercategory to a query.

[0026] According to the present invention, a method and device for selecting documents which may be of interest from among documents in a collection comprises: choosing terms to be used in selecting documents which may be of interest; dividing each document D in the collection into S₀segments; for each document D in the collection, assigning a weight to each segment S_iof the document; for each document D in the collection, determining which of the terms chosen to be used in selecting documents are found in each segment S_iof the document D; for each document D in the collection, calculating a generalized term frequency score S_D; and selecting documents from among the documents in the collection based upon the scores S_Dachieved by the documents. The documents may be Web pages, Web sites or other collections of material. The terms chosen to be used in selecting documents which may be of interest may comprise the terms used in a search query. Preselected stop terms may be eliminated. Additional terms may be chosen to be used in selecting documents which may be of interest based upon the terms used in the search query. The terms chosen to be used in selecting documents which may be of interest may comprise the terms in a document identified as having characteristics that are desired in the documents to be selected. Preselected stop terms may be eliminated. Additional terms may be chosen to be used in selecting documents based upon the terms which occur in the document identified as having characteristics that are desired in the documents to be selected. The terms chosen to be used in selecting documents may comprise the terms in a document being accessed by a user. Preselected stop terms may be eliminated. Additional terms may be chosen to be used in selecting documents based upon the terms which occur in the document being accessed by the user. The documents in the collection may be ranked in order of their generalized term frequency scores S_D, and a predetermined number N of documents with the highest scores may be selected. All documents with generalized term frequency scores S_Dabove a predetermined cutoff score S may be selected. The generalized term frequency score S_Dfor a document D may be based upon Robertson's term frequency score the weights assigned to the i^thsegment of the documents in the collection may be equal. The weights assigned to the segments of a document D in the collection may sum to 1.0. The weights may be selected specifically for the collection of documents from which documents are to be chosen by carrying out test searches with different weights, and selecting for use the weights which yield the most useful results. Additional terms may be assigned to each document D in the collection, beyond the terms that occur in the document D, based upon the characteristics of the document D, without regard to the occurrence of the additional terms in the document D, and the additional terms may be placed in a segment S_iof the document D with no other terms. The additional terms may be assigned automatically by: creating a search query Q comprised of terms in document D; applying the search query Q to a collection of documents C₀; selecting the N₀documents from the collection of documents C₀which achieve the highest scores upon application of the search query Q; and selecting I_Tterms for automatic assignment from among terms T_nin the N₀documents based upon the co-occurrence of the terms T_nin the N₀documents with the terms T_iin the document D. The I_Tterms may be selected for automatic assignment by: calculating, for terms T_nwhich occur in the N₀documents selected, the co-occurrence of that term T_nwith each term T_iin document D; calculating, for terms T_nwhich occur in the N₀documents selected, the co-occurrence of that term T_nwith document D; and selecting I_Tterms for automatic assignment from among the terms T_nin the N₀documents based upon the scores achieved by the terms T_n. The weight assigned to term T in segment S_lof document D may be the co-occurrence score for that term T with the document D for all terms T automatically assigned. The weight assigned to term T in segment S_iof document D, may be 1.0 for all terms T which occur in the document D. The search query Q which is applied may comprise all of the terms in document D. Preselected stop terms may be eliminated. The search query Q may be applied to select documents from among the documents in the collection C₀by calculating for each document D in the collection C₀a score S_Dbased upon the occurrence in the document D of terms in the search query Q. In applying the search query Q to the collection of documents C₀the total score S_Dfor a document D in the collection C₀may be determined by using Robertson's term frequency score. The number N₀of documents chosen by application of the search query Q may be predetermined. In one embodiment, the number N₀may be 50. All documents whose scores upon application of the search query Q exceed a given cutoff score may be selected. Co-occurrences may be calculated for all terms contained in the N₀documents selected. Preselected stop terms may be are eliminated. The number I_Tof terms automatically assigned may be predetermined. In one embodiment, the number I_Tmay be 30. All terms whose scores exceed a given cutoff score may be automatically assigned.

[0027] Further according to the present invention, a method and device for presenting banner advertisements to a user who is seeking information about products and/or services comprises: defining a collection C₀of categories of products and/or services; for each of the categories in the collection C₀of categories, associating with it a set of terms which describe the product(s) and/or service(s) associated with the category, and a unique category identifier term; for each provider of products and/or services participating, assigning the provider to one or more categories based upon the products and/or services offered by the provider; grouping the categories into supercategories; associating one or more banner advertisements with each supercategory; for each supercategory, associating with it the sets of terms which describe the product(s) or service(s) associated with the categories assigned to it, and the category identifier terms which are unique to the categories assigned to it; in response to a user query Q, consisting of terms describing products and/or services of interest to the user, selecting every category, in the collection of categories C₀, which h as associated with it a descriptive term for the products and/or services therein which matches a term in the user query; preparing a new query Q′ consisting of the terms in the user query Q, the descriptive terms for the products and/or services associated with the categories selected, and the unique category identifier terms associated with the categories selected; applying the query Q′ to the collection of supercategories; selecting the supercategory with the highest score; and presenting to the user a banner advertisement associated with the supercategory selected. The search query Q′ may be applied to select a supercategory from among the collection of supercategories by calculating for each supercategory a score S_Cbased upon the occurrence in the supercategory of terms in the search query Q′. For example, in one embodiment there may be about 20,000 categories of products and/or services in the collection of categories, and there may be about 50 supercategories. Every category, in the collection of categories C₀, which has associated with it a descriptive term for the products and/or services therein which matches a term in the user query, may be presented to the user, and the user may be permitted to select from among said categories a category of interest for which a list of the merchants, stores or other sources of product(s) and/or service(s) associated with the category will be presented. In response to the user selecting a category from those presented: a new query Q″ may be prepared, consisting of the descriptive terms for the product(s) and/or service(s) associated with the category selected, and the unique category identifier term associated with the category selected; the query Q″ may be applied to the collection of supercategories; the supercategory with the highest score may be selected; and the user may be presented with a banner advertisement associated with the supercategory selected. The query Q′ may be applied to the collection of supercategories by utilizing Robertson's term frequency score. The categories may be assigned to the supercategories by a plurality of methods, the supercategories may be considered to comprise multiple segments, each segment of a supercategory may comprise the terms and term identifiers associated with the categories assigned to the supercategory by a single method, each segment may be assigned a weight, each term in a segment may be assigned a weight, and in applying the query Q′_ito the collection of supercategories a generalized term frequency score may be used. A subset of the categories may be assigned to supercategories manually, while the remainder may be assigned utilizing an automatic or semi-automatic index(term augmentation technique based upon the co-occurrence of terms between the manually-assigned categories and the categories being automatically or semi-automatically assigned. The remainder of the terms not manually assigned to supercategories may be assigned by: selecting a category C_ifrom among the categories in the collection not yet assigned to supercategories which has not yet been processed; selecting a supercategory S_jfrom among the set of supercategories; calculating a likelihood function for the category C_iand a category C_kin the collection which has previously been assigned to the supercategory S_jby manual means, which likelihood function is based upon the likelihood that a term occurring in the category C_ialso occurs in the category C_k; repeating the above step for a plurality of other categories C_kin the collection which have previously been assigned to the supercategory S_jmanually; calculating a total score for the category C_ifor the supercategory S_j, which total score is based upon the likelihood functions for the category C_iand the categories C_kin the collection which have previously been assigned to the supercategory S_jmanually; repeating the above steps for a plurality of other supercategories S_j; assigning category C_ito the supercategory for which the total score calculated for the category C_iis the highest; and repeating the above steps for a plurality of other categories in the collection which have not yet been assigned to supercategories and which have not yet been processed. The likelihood function for the category C_iand a category C_kin the collection which has previously been assigned to the supercategory S_jmanually may be the log likelihood ratio for the category C_iand the category C_k. The total score for the category C_ifor the supercategory S_jmay be the sum of the log likelihood ratios. For example, in one embodiment there may be about 20,000 categories, there may be about 50 supercategories, and approximately two thousand of the approximately 20,000 categories may be manually assigned to supercategories. In one embodiment, the weight assigned to the segment of the supercategories which comprises the terms and term identifiers associated with the categories assigned to the supercategory manually may be 1.0, and the weight assigned to the segment of the supercategories which comprises the terms and term identifiers associated with the categories assigned to the supercategory semiautomatically may be 0.4.

[0028] Referring to FIG. 1, a computer system 2001 includes a workstation 2002 having local storage 2003. The workstation may also be connected to a local area network 2004 and may access to the Internet 2005. The Internet 2005 may include or be coupled to remote storage 2006. The workstation 2002 may be any one of a variety of commercially available computers capable of providing the functionality described in more detail below. The local storage 2003 may include ROM, RAM, a hard disk, a CD, and/or any other media capable of containing data and/or programs for the workstation 2002 or other data. The local area network 2004, which is coupled to and exchanges data with the workstation, may also contain data and/or program information for use by the workstation 2002. The Internet 2005 may be accessed in a conventional manner by the workstation 2002. Alternatively, the workstation 2002 may access the Internet 2005 through the local area network 2004, as shown by the dotted line of FIG. 1. The remote storage 2006 if may also contain data and/or program information for the workstation 2002 or may contain other information, as will become apparent from the description below.

[0029] The system described herein permits a user (utilizing the computer system 2001 which includes the workstation 2002) who has accessed the Internet 2005, either directly or through the local area network 2004, to be given access to material that may be of interest to him. It will be appreciated by one of ordinary skill in the art that the system may be implemented using a variety of computers and programming languages. The system may be accessed by the user through the Internet 2005 from his workstation 2002 using a Web browser of conventional design, as would be familiar to one of ordinary skill in the art.

[0030] In the prior art, it is well known that information retrieval techniques may be utilized to identify documents, such as Web pages or sites, or portions of documents which may be of interest to a user. (Hereinafter, for simplicity, “document” shall be used to refer to the items, such as [but not limited to] pages or sites, in the collection being analyzed.) These techniques may be called into play in response to a search query initiated by the user. Alternatively, they may be called into play when a user requests additional documents that are similar to a document to which he has been given access. Alternatively, they may be called into play when a user is accessing a particular document, an d it is desired to make available to him other documents that are related to the document being accessed. Other circumstances where it may be desirable to utilize information retrieval techniques to identify documents that may be of interest to a user will be apparent to one of ordinary skill in the art.

[0031] Information retrieval techniques may choose documents from among the documents in a collection based upon the occurrence in the documents of specified terms. The terms to be utilized in this process may be selected by a number of methods that will be apparent to one of ordinary skill in the art.

[0032] One technique that may be employed to select terms to be utilized in the process is to permit the user to specify terms by defining a search query. Another technique that may be employed is to select some or all of the terms in a document being accessed by the user. Another technique that may be employed is to select some or all of the terms in a document identified by the user as being of interest to him, or as having characteristics he wishes to have found in documents made available to him. (In these techniques, all of the terms may be used, or certain common words, referred to as stop words, such as “the” or “and”, may be omitted.) Another technique that may be employed is to select index terms which have previously been associated with the document being accessed or selected by the user. Still other techniques may use a combination of the above approaches. Other techniques for selecting terms to be utilized will be apparent to one of ordinary skill in the art.

[0033] Once a list of terms has bee n generated, by the above methods or any other, information retrieval techniques may proceed by calculating, for each document in the collection from which documents of potential interest are to be chosen, a score which reflects the occurrence in the document of the terms on the list. Based upon the scores achieved by the documents in the collection, the documents may be ranked, and a predetermined number of documents may be presented to the user, or all documents which achieve scores above a predetermined cutoff may be presented.

[0034] These scores which are calculated for documents are sometimes referred to as term frequency scores, in that the scores depend in part upon the frequency of occurrence of terms in the document.

[0035] The formula for calculating a total score S_Dfor a document D may be written generally as: $S_D=\sum _{T=1}^{T_0}{\mathrm{TF}}_{\mathrm{TD}}$

[0036] where: T₀is the number of terms T which occur in the collection of terms included in the search, and

[0037] TF_TDis the term frequency score for document D based on the frequency of occurrence in document D of term T.

[0038] One particular formula in the prior art which may be used to assign a total score S_Dto a document D utilizes Robertson's term frequency score: $S_D=\sum _{T=1}^{T_0}{\mathrm{TF}}_{\mathrm{TD}}*{\mathrm{IDF}}_T,$

[0039] where: T₀is the number of terms which occur in the collection of terms included in the search,

[0040] TF_TDis Robertson's term frequency for term T in document D,

[0041] =N_TD/(N_TD+K₁+K₂*(L_D/L₀)),

[0042] where: N_TDis the number of times the term T occurs in document D,

[0043] L_Dis the length of document D,

[0044] L₀is the average length of a document in the collection being searched, and

[0045] K₁and K₂are constants

[0046] and IDF_T=log((N+K₃)/N_T)/log(N+K₄)

[0047] where:

[0048] N is the number of documents in the collection

[0049] N_Tis the number of documents containing the term T in the collection, and

[0050] K₃and K₄are constants.

[0051] Whatever particular formula is used, documents are ranked in order of their total scores S_D, and those which achieve the highest score are presented, typically in order of their scores, to the user.

[0052] In order to improve the effectiveness of information retrieval methods, additional terms may be associated with documents before term frequency scores are calculated. For example, index terms or key words may be associated with each document in a collection, and the calculation of term frequency scores may take into account the index terms or key words as well as terms that occur in the documents themselves, or may be based solely on the index terms or key words.

[0053] These additional terms may be assigned to a document by means of manual review of the document or by automatic means, Or by a combination of manual review and automatic means. Methods for doing so by manual means will be apparent to one of ordinary skill in the art.

[0054] The manual assignment of index terms to a document may be time consuming, and this may make it impractical to assign index terms to large collections of documents by this method. In addition, manual assignment of index terms may fail to reveal underlying relationships between documents. It may therefore be useful to utilize automatic techniques to generate appropriate index terms for documents, based upon analysis of the characteristics of the terms which occur in the documents.

[0055] In one embodiment of the system described herein, additional index terms are added to a set of documents D in a document collection automatically. In this embodiment, terms are chosen to be added as index terms to a given document Di automatically according to their co-occurrence to a high degree with terms already found in the document D_i, according to the method of local context analysis. This method has been described by Xu and Croft, in Improving the Effectiveness of Informational Retrieval with Local Context Analysis, which is incorporated herein by reference.

[0056] FIG. 2 illustrates a collection of Documents D 2020 which contain Terms T 2010. As illustrated in FIG. 2, each Term T 2010 may occur in one or more Documents D 2020, and each Document D 2020 contains one or more Terms T 2010.

[0057] According to FIG. 3, this method 2050 proceeds first at a step 2060 to select a Document D_iwhich has not yet had index terms assigned to it. At step 2070, a search Query Q_iis created, consisting of Terms T_jfound in Document D_i. In one embodiment of the system, the set of Terms T_jin the Document D_iused to create the Query Q_icomprises all of the Terms in the Document D_i. In another embodiment, the set of Terms T_jcomprises all of the Terms in the Document D_iexcept certain common words, referred to as stop words, such as “the” or “and.”

[0058] In this embodiment, after the query Q_iis prepared at step 2070 it is applied at step 2080 to a chosen collection C₀consisting of N documents. This collection of documents C₀may be the set of documents for which index terms are being generated by automatic means, it may be a larger set of documents including those documents for which index terms are being generated by automatic means as a subset, or it may be another set of documents, such as the set of documents over which searches will be done utilizing the index terms. It is helpful if the collection C₀has the property that the usage of terms in documents in it is characteristic of the usage of terms that will be found in documents over which searches will be carried out using the additional index terms added to the documents.

[0059] In applying the query at the step 2080, a total score S_Dfor a document D in the collection of documents C₀searched may be written generally as: $S_D=\sum _{T=1}^{T_0}{\mathrm{TF}}_{\mathrm{TD}}$

[0060] where: T₀is the number of terms T which occur in the query Q_i, and

[0061] TF_TDis the term frequency score for document D based on the frequency of occurrence in document D of term T.

[0062] While any one of a number of formulas for term frequency and inverted document frequency which will be known to one of ordinary skill in the art may be used without departing from the spirit and scope of the invention, in one embodiment of the system, Robertson's term frequency score is used to assign a total score S_Dto a document D: $S_D=\sum _{T=1}^{T_0}{\mathrm{TF}}_{\mathrm{TD}}*{\mathrm{IDF}}_T,$

[0063] where: T₀is the number of terms which occur in the query Q_i,

[0064] TF_TDis Robertson's term frequency for term T in document D,

[0065] =N_TD/(N_TD+K₁+K₂*(L_D/L₀)),

[0066] where: N_TDis the number of times the term T occurs in document D,

[0067] L_Dis the length of document D,

[0068] L₀is the average length of a document in the collection C₀being searched, and

[0069] K₁and K₂are constants

[0070] and IDF_T=log((N+K₃)/N_T)/log(N+K₄)

[0071] where:

[0072] N is the number of documents in the collection C₀

[0073] N_Tis the number of documents containing the term T in the collection C₀, and

[0074] K₃and K₄are constants.

[0075] After the query is run at step 2080, at a step 2090 a number of documents N₀in C₀which achieve the highest scores under the search query Q_iare selected. For example, in various embodiments the number N₀may be between 10 and 300, but it may vary depending on operational considerations which will be apparent to one of ordinary skill in the art. In one embodiment, the number of documents N₀selected is 50. This set of N₀documents has the property that Documents in it contain Terms also found in Document D_i, the document which is having index terms assigned to it. The next steps 2100 to 2150 in the process 2050 then attempt to determine which other terms in the N₀documents occur most frequently with the Terms T_jin the Document D_i.

[0076] After the N₀documents are selected in the step 2090, the system continues at step 2100 by choosing a Term T_kfrom among the Terms found in the N₀documents. In one embodiment, all terms in the N₀documents are used. In another embodiment, all terms in the N₀documents except certain common words, referred to as stop words, such as “the” or “and,” are used.

[0077] At a step 2110, the system then chooses a Term T_jfrom among the Terms in the Document D_iwhich is having index terms assigned to it.

[0078] At a step 2120, the system then proceeds by calculating the co-occurrence C_n(T_j, T_k) of the Term T_kfrom the N₀documents with the Term T_jfrom the Document D_i. The co-occurrence C_n(T_j, T_k) of a given Term T_kwhich occurs in the N₀documents, with a Term T_jin Document D_i, is determined as follows:

[0079] C_n(T_j,T_k)=log₁₀(co_ki(T_j,T_k)+1)*idf(T_k)/log₁₀(N₀),

[0080] $\mathrm{where}\text{:}{\mathrm{co}}_{\mathrm{ki}}\left(T_j,T_k\right)=\sum _{n=1}^{N_0}\mathrm{tf}\left(T_k,n\right)*\mathrm{tf}\left(T_j,n\right),$

[0081] tf (T_k, n)=number of occurrences of term T_kin Document n in the N₀documents,

[0082] tf (T_j, n)=number of occurrences of term T_jin Document n in the N₀documents,

[0083] idf(T_k)=the inverted document frequency for the term T_k,

[0084] =min (1.0, log₁₀(N/N_T)/5.0)

[0085] N=the number of documents in the collection C₀over which the query Q_iwas run, and

[0086] N_T=the number of documents in the collection C₀over which the query Q_iwas run, which contain the term T_k.

[0087] The system then proceeds to a step 2130. If it is determined at the step 2130 that not all Terms T_jfrom among the Terms in the Document D_i. have had their co-occurrences calculated with the Term T_kfrom the N₀documents, control returns to step 2110, and the co-occurrence of another Term T_jfrom among the Terms in the Document D_i. is calculated with the Term T_kfrom the N₀documents.

[0088] If it is determined at the step 2130 that all Terms T_jfrom the Document D_i. have had their co-occurrences calculated with the Term T_kfrom the N₀documents, control passes to a step 2140, at which a score f_D(T_k) is calculated for the term T_kwith respect to the document D_i: $f_D\left(T_k\right)=\prod _{j=1}^{T_0}{\left(\delta +C_{\mathrm{ni}}\left(T_k,T_j\right)\right)}^{\mathrm{idf}\left(T_j\right)}$

[0089] where T_j=a term in document D_i,

[0090] T₀=the number of terms in document D_i,

[0091] idf (T_j)=the inverted document frequency for the term T_j,

[0092] =min (1.0, log₁₀(N/N_j)/5.0),

[0093] N=the number of documents in the collection C₀over which the query Q_iwas run,

[0094] N_J=the number of documents in the collection C₀over which the query Q_iwas run, which contain the term T_jand

[0095] δ=a constant. (In one embodiment, δ may be assigned a value of 0.01, but this value may be varied without departing from the spirit and scope of the invention.)

[0096] After f_D(T_k) is calculated at the step 2140, control passes to a step 2150. If it is determined at the step 2150 that not all Terms T_kfrom among the Terms in the N₀documents have had their Scores f_D(T_k) calculated for the Document D_i, control returns to step 2100, and the process of steps 2100-2140 is carried out for another Term T_kfrom among the Terms in the N₀documents.

[0097] If it is determined at the step 2150 that all Terms T_kfrom among the Terms in the N₀documents have had their Scores f_D(T_k) calculated for the Document D_i, control passes to a step 2160, at which index terms are chosen for the Document D_i. To do so, in this embodiment the values of f_D(T_k) for the Document D_iare compared for the terms T_kin the N₀documents, and the terms T_kwith the highest values of f_D(T_k) for the Document D_iare chosen as additional terms to be added as index terms to the Document D_i. While the number of terms added may vary without departing from the spirit and scope of the invention, in one embodiment 30 terms are chosen to be added as index terms.

[0098] After index terms are assigned to Document D_iat the step 2160, control passes to a step 2170. If it is determined at the step 2170 that not all Documents D_ihave had index terms assigned, control returns to step 2060, and the process of steps 2060-2160 is carried out for another Document D_i.

[0099] If it is determined at the step 2170 that all Documents D_i, have had index terms assigned, this portion of the system is completed.

[0100] The system described herein may be employed via a Web site which presents a user with, or permits a user to obtain, specific documents or lists of documents, such as Web sites, names of merchants or stores carrying merchandise in particular categories, or other documents, and which uses index terms assigned to documents to assist in the process of identifying documents for presentation to the user, or for inclusion in a list to be presented to the user.

[0101] A further aspect of the system described herein may be employed when some of the documents in the collection from which the selection(s) are to be made have had index terms assigned to them manually (or by other automatic methods), but index terms have not been assigned to all documents, and it is desired to assign index terms to the remaining documents automatically.

[0102] According to FIG. 4, in one embodiment of the system described herein, there may be a very large number of Documents D 2420 which contain Terms T 2410.

[0103] In this embodiment of the system, it is desired to assign an Index Term I 2440 or Index Terms to each Document D.

[0104] It may desirable in this embodiment of the system to associate each Document D 2420 with one and only one Index Term I 2440, or it may be desired to associate a plurality of Index Terms with a Document D.

[0105] Index Terms may be associated with Documents manually. However, manual association is time consuming and therefore costly, and this is particularly the case if the Documents and/or Index Terms may change frequently. The system described herein therefore permits Documents to be assigned Index Terms automatically, after an initial group of Documents have been assigned manually. FIG. 4 illustrates the relationship of Terms, Documents and Index Terms, when some Documents have been assigned Index Terms manually, and others have not had Index Terms assigned. (It will be understood by one of ordinary skill in the art that the system here described may also be applied where an initial group of documents have had Index Terms assigned by another automatic method, rather than manually.)

[0106] According to FIG. 5, the process 2450 of assigning Index Terms 2440 to Documents 2420 begins at a step 2460 in which an (as-yet-unprocessed) Document D_ito which no Index Terms have been assigned manually is selected. Control then passes to a step 2470 at which an (as-yet-unanalyzed for the selected unprocessed Document D_i) Index Term I_jis selected. (The Index Terms may consist of a set of terms chosen from among the Terms T which occur in the collection of documents, or they may be chosen independently of whether they occur among the Terms in the document collection.) Control then passes to a step 2480 at which a Document D_k, which has been manually assigned Index Term I_jis selected.

[0107] At a step 2490, the process 2450 then calculates the log likelihood ratio L (D_i, D_k): $L\left(D_i,D_k\right)=\log \left(\sum _{m=1}^{M_0}{\Pi }_m\left(D_i,D_k\right)/\sum _{m=1}^{M_0}{\Pi }_m\left(D_i\right)\right),$

[0108] where:

[0109] Π_m(D_i, D_k)=1, if Term m is in Document D_iand to Document D_k

[0110] =0 otherwise

[0111] Π_m(D_i)=1, if Term m is in Document D_i

[0112] =0 otherwise

[0113] M₀=the number of Terms which are in Document D_i.

[0114] Control then passes to a step 2500, at which it is determined if there remain any further Documents D_k, manually assigned the Index Term I_jbeing analyzed, for which the log likelihood ratio of that Document D_kto the Document D_ibeing processed has not yet been calculated. If any such Documents D_kremain at the step 2500, control returns to the step 2480 at which a further Document D_k, which has had Index Term I_jmanually assigned to it, is chosen for calculation. If no such Documents D_kremain at the step 2500, control instead passes to a step 2510 at which is calculated the total score T ( D_i, I_j) for the unprocessed Document D_ifor the Index Term I_j: $T\left(D_i,I_j\right)=\sum _{k=1}^{K_0}W\left(D_k,I_j\right)*L\left(D_i,D_k\right)/\sum _{k=1}^{K_0}W\left(D_k,I_j\right),$

[0115] where

[0116] K₀=the number of Documents manually assigned Index Term I_j,

[0117] W(D_k, I_j)=the weight assigned to Index Term I_jfor Document D_k

[0118] This system permits varying weights to be assigned to different Index Terms I_jassociated with a given Document D. The weights assigned to the index terms associated with a given Document D may be equal, or they may be varied to reflect the degree of importance associated with the Index Term, or they may be varied to reflect the degree of confidence with which the Index Term is believed to represent the characteristics of the document. Other reasons and methods of varying the weight assigned to an Index Term associated with a Document will be apparent to one of ordinary skill in the art.

[0119] In the case where each Document D has assigned to it only a single Index Term I_j, then W(D_k, I_j)=1 for the one and only one Index Term I_jassigned to Document D_k, and the formula for the total score T (D_i, I_j) is simplified: $T\left(D_i,I_j\right)=\sum _{k=1}^{K_0}L\left(D_i,D_k\right)/K_0,$

[0120] where

[0121] K₀=the number of Documents manually assigned Index Term I_j,

[0122] Control then passes to a step 2520, at which it is determined if there remain any further Index Terms I_jfor which total scores T (D_i, I_j) have not yet been calculated for the Document D_ibeing processed. If any such Index Terms I_jremain at the step 2520, control returns to the step 2470 at which a further Index Term I_jis chosen for analysis with the Document D_i. If no such Index Terms I_jremain at the step 2520, control instead passes to a step 2530 at which an Index Term or Index Terms are selected to be associated with the Document D_ibeing processed.

[0123] In one embodiment of the system, a single Index Term I_M, whose total score T (D_i, I_j) for the Document D_ibeing processed is the highest, is selected as the index term for Document D_i. In another embodiment of the system, a predetermined number R of Index Terms may be selected as index terms for Document D_i. In this embodiment, the R Index Terms with the highest total scores T (D_i, I_j) are selected. In another embodiment, all Index Terms whose total scores T (D_i, I_j) exceed a predetermined cutoff score T₀are selected as index terms for Document D_i. (In any of these embodiments, if no co-occurrences were found between the Document D_ibeing processed and any document which has been manually assigned index terms, then no index terms are assigned to the Document D_i.)

[0124] Control then passes to a step 2540 at which it is determined if there remain any further Documents D_i, which were not assigned index terms manually, which have not yet been processed. If any such unprocessed Documents D_iremain at the step 2540, control returns to the step 2460 at which a further as-yet-unprocessed Document D_iis chosen for processing. If no such unprocessed Documents D_iremain at the step 2540, the process 2450 is concluded, and each Document D_ito which no Index Terms had been assigned manually, either has been assigned Index Terms or has been found not to have co-occurrences with any Document which had index terms manually assigned to it. According to FIG. 6, when the process 2450 has been completed, Index Terms 2440 will have been assigned to Documents 2420 containing Terms 2410, except for Documents 2420 which could not be assigned Index Terms 2440 because they lack any co-occurrences with any Document 2420 which had Index Terms 2440 manually assigned to it.

[0125] The system described herein may be utilized in one embodiment in connection with the assignment of categories consisting of items of interest into categories of categories, or supercategories.

[0126] In this embodiment, an item of interest may be considered to be a merchant, store or other source for a product or service, or a number of (related or unrelated) products or services. Each variety of product or service may be considered to be a category (such as, for example, “Auto Dealers, Used Cars”). In this embodiment, items of interest (merchants or stores, such as, for example, “Lannan Chevrolet, Oldsmobile”) may be assigned to more than one category (variety of product or service).

[0127] In this embodiment, it is desired to present categories to a user in response to his request. The categories presented to the user may vary, depending on a term or terms (such as, for example, “automobiles, used”) specified by the user in the request. The categories presented may be chosen by any one of a number of techniques that will be familiar to one of ordinary skill in the art.

[0128] In this embodiment of the system described herein, it is desired to present additional material to a user who is searching For items of interest. For example, it may be desired to present the user with banner advertisements (such as for automobile financing sources) which relate to the item of interest (such as used cars) for which he is searching.

[0129] According to FIG. 7, in one embodiment of the system described herein, there may be a very large number of individual items of interest 2810 to be organized into categories 2820 for presentation. While the number may vary without departing from the spirit and scope of the invention, there may be about 20,000 categories.

[0130] In this embodiment of the system, it is desired to choose a banner advertisement to present to a user. The banner advertisements in turn may be divided into categories 2840. While the number may vary without departing from the spirit and scope of the invention, there may be about 50 categories 2840 into which the banner advertisements may be divided. (To avoid confusion with the categories into which the items of interest are divided, these banner advertisement categories 2840 are referred to herein as “supercategories.” 2840)

[0131] It is desirable in this embodiment of the system to associate each category 2820 of items of interest 2810 with one and only one supercategory 2840 of banner advertisements, such that when a user is accessing that category 2820 of item he is presented with banner advertisements from the corresponding supercategeory 2840. (For example, in one embodiment the category “Auto Dealers, Used Cars” may be assigned to a supercategory also comprising other categories related to automobiles, such as “Automobile Dealers” and/or “Auto Repair & Service.”)

[0132] Categories may be associated with supercategories manually. However, manual association is time consuming and therefore costly, and this is particularly the case if the categories and supercategories may change frequently. This embodiment of the system described herein therefore permits categories to be assigned to supercategories automatically, after an initial group of categories have been assigned manually. FIG. 7 illustrates the relationship of items of interest, categories and supercategories, when some categories have been assigned to supercategories, and others remain unassigned. While the number may vary without departing from the spirit and scope of the invention, in one embodiment there may be about 2,000 categories manually assigned to supercategories.

[0133] According to FIG. 8, the process 2850 of assigning categories 2820 to supercategories in this embodiment of the system 2840 begins at a step 2860 in which an (as-yet-unprocessed) unassigned category C_iis selected. Control then passes to a step 2870 at which an (as-yet-unanalyzed for the selected unassigned category) supercategory S_jis selected. Control then passes to a step 2880 at which a category C_k, which has been manually assigned to supercategory S_jis selected.

[0134] At a step 2890, the process 2850 then calculates the log likelihood ratio L (C_i, C_k): $L\left(C_i,C_k\right)=\log \left(\sum _{m=1}^{M_0}{\Pi }_m\left(C_i,C_k\right)/\sum _{m=1}^{M_0}{\Pi }_m\left(C_i\right)\right),$

[0135] where:

[0136] Π_m(C_i, C_k)=1, if item of interest m is assigned to category C_iand to category C_k

[0137] =0 otherwise

[0138] Π_m(C_i)=1, if item of interest m is assigned to category C_i

[0139] =0 otherwise

[0140] M₀=the number of items of interest which are assigned to category C_i

[0141] Control then passes to a step 2900, at which it is determined if there remain any further categories C_k, manually assigned to the supercategory S_jbeing analyzed, for which the log likelihood ratio of that manually assigned category C_kto the category being processed C_i, has not yet been calculated. If any such manually assigned categories C_kremain at the step 2900, control returns to the step 2880 at which a further manually assigned category C_kis chosen for calculation. If no such manually assigned categories C_kremain at the step 2900, control instead passes to a step 2910 at which is calculated the total score T (C_i, S_j) for the unprocessed category C_ifor the supercategory S_j: $T\left(C_i,S_j\right)=\sum _{k=1}^{K_0}L\left(C_i,C_k\right)/K_0,$

[0142] where

[0143] K₀=the number of categories manually assigned to supercategory S_j

[0144] Control then passes to a step 2920, at which it is determined if there remain any further supercategories S_jfor which total scores T (C_i, S_j) have not yet been calculated for the category C_ibeing processed. If any such supercategories S_jremain at the step 2920, control returns to the step 2870 at which a further supercategory S_iis chosen for analysis with the category C_i. If no such supercategories S_jremain at the step 2920, control instead passes to a step 2930 at which is selected a supercategory S_Mwhose total score T (C_i, S_j) for the category C_ibeing processed is the highest. The category being processed C_ithen is assigned to the supercategory S_M. (If no co-occurrences have been found between the category C_ibeing processed and any category manually assigned to a supercategory, the category C_ibeing processed is not assigned to any supercategory.)

[0145] Control then passes to a step 2940 at which it is determined if there remain any further unassigned categories C_inot yet processed. If any such unprocessed categories C_iremain at the step 2940, control returns to the step 2860 at which a further as-yet-unprocessed category C_iis chosen for processing. If no such unprocessed categories C_iremain at the step 2940, the process 2850 is concluded, and each previously-unassigned category C_ihas either been assigned to a supercategory S_j, or it has been determined that it has no co-occurrences with any manually-assigned category, and hence no supercategory S_jassignment has been made for it. According to FIG. 9, all categories 2820 containing items of interest 2810 will have been assigned to supercategories 2840, except for those categories 2820 as to which it has been determined that the category 2820 has no co-occurrences with any manually-assigned category 2820.

[0146] When additional terms such as index terms or key words are assigned to a document, such as by the system described herein, the additional terms may be considered as terms along with the terms that occur in the document itself for purposes of calculating term frequency scores. The original terms and the index terms may be used together in searches, or the index terms alone may be used.

[0147] It may be thought that the occurrence among the additional terms of a term for which a search is being made may be more or less important as a predictor of the utility of the document than the occurrence of a term found in the document itself. A technique for taking into account whether a term occurs in a document itself or among the additional terms associated with the document, in the calculation of a term frequency score for that document, therefore may be useful.

[0148] In addition, documents in a collection which is being searched may consist of various segments or sections. The segments or sections may include a title, an abstract or introduction or summary, captions, and footnotes. Other sections or segments into which a document may be divided will be apparent to one of ordinary skill in the art.

[0149] In some circumstances, it may be thought that the occurrence of a term in one segment of a document may be more predictive of the utility of that document than its occurrence in another segment. A technique for taking into account the segment of a document in which a given term occurs, in the course of calculating a term frequency score for that document, therefore may be useful.

[0150] According to the system being described herein, a weight W_SDmay be assigned to each segment S_iof a document D containing S₀segments. In one embodiment of the system: $\sum _{S_i=1}^{S_0}{W}_{\mathrm{SD}}=1.$

[0151] In one embodiment of the system, an entire document itself is considered a single segment, and the additional index terms associated with the document, such as by the system described herein, are considered a second segment. In that embodiment, there are a total of two segments in a document, including the new segment containing the index terms.

[0152] In a further embodiment of the system, the index terms associated with the document, such as by the system described herein, are considered a segment, and the text of the document itself may be divided into a number of separate segments which may include a title, an abstract or introduction or summary, captions, and footnotes. Other sections or segments into which a document may be divided will be apparent to one of ordinary skill in the art.

[0153] In a further embodiment of the system, where additional terms such as index terms have been associated with a document by more than one method, for each method used the additional terms associated with the document by that method may be considered a separate segment of the document.

[0154] In a further embodiment of the system, where no additional terms have been associated with the document, the text of the document itself may be divided into a number of separate segments which may include a title, an abstract or introduction or summary, captions, and footnotes.

[0155] The weights W_SDassigned to the segments of documents may be chosen arbitrarily.

[0156] In one embodiment of the system, the weights W_SDassigned to the segments S_iof a document D may be individually determined in advance, based upon a decision about the relative utility of various segments of the document D in determining the relevance of the document under various criteria.

[0157] In an embodiment of the system, a given segment S_imay be required to have equal weight W_SDin all documents.

[0158] In a further embodiment, the weight W_SDof a given segment S_iof different documents may be different, based upon the relative utility of that segment of each document in predicting whether that document will be of interest to a user.

[0159] The weights assigned to the segments S_Iof a document containing the additional terms assigned to the document may be varied based upon the method used to assign the additional terms, and the degree to which the additional terms are considered to be highly related to the content of the documents. In an embodiment of the system, a segment S_Imay be required to have equal weight W_SDin all documents. In a further embodiment, the weights W_SDof the segments S_Iof different documents may be different, based upon the method used to assign the additional terms, and the degree to which the additional terms are considered to be highly related to the content of each document.

[0160] In one embodiment of the system, the weights W_SDare varied depending on the results of experiments which vary the weights for test searches and evaluate the utility of the results returned, either in terms of precision (the ability of the search formula to avoid returning documents that are not useful), or of recall (the ability of the search formula to avoid omitting documents that are useful), or of a combination of the two.

[0161] When it is determined to calculate a term frequency score under a given search query Q for a document D with S₀segments in the collection of documents C₀being searched under the system, a generalized term frequency score may be calculated as follows: $S_D=\sum _{T=1}^{T_0}\sum _{S_i=1}^{S_0}{\mathrm{TF}}_{\mathrm{STD}}$

[0162] where: S_Dis the total score for the document D,

[0163] T₀is the number of terms which occur in the search query Q, and

[0164] TF_STDis the score for document D based on the occurrence of term T in segment S_iof document D.

[0165] In one embodiment of the system, scores are assigned to documents utilizing Robertson's term frequency score, and the generalized term frequency score S_Dfor a document D may be calculated as follows: $S_D=\sum _{T=1}^{T_0}\sum _{S_i=1}^{S_0}{\mathrm{TF}}_{\mathrm{STD}}*{\mathrm{IDF}}_{\mathrm{ST}}$

[0166] where: S_Dis the total score for the document D,

[0167] T₀is the number of terms which occur in the search query Q,

[0168] S₀is the number of segments in the document D,

[0169] TF_STD=Robertson's generalized term frequency score for Term T in Segment S_iof Document D

[0170] =G_STD/(G_STD+K₁+K₂*W_SD*(H_SD/H_SO)),

[0171] where:

[0172] G_STD=the generalized term count for Term T in Segment S_iof Document D,

[0173] =W_SD*W_STD*N_STD,

[0174] where:

[0175] W_SDis the weight assigned to segment S_iof document D,

[0176] W_STDis the weight assigned to term T in segment S_iof document D, and

[0177] N_STDis the number of times the term T occurs in segment S_iof document D,

[0178] H_SD=the generalized length of segment S_iof document D, $H_{\mathrm{SD}}=\sum _{T=1}^{L_{\mathrm{SD}}}{W}_{\mathrm{STD}}*N_{\mathrm{STD}}$

[0179] where:

[0180] L_SDis the number of different terms in segment S_iof document D,

[0181] H_SO=the generalized average length of segment S_iof documents in the collection C₀being searched, $H_{\mathrm{SO}}=\sum _{N=1}^{N_0}\left(W_{\mathrm{SD}}*H_{\mathrm{SD}}\right)/\sum _{N=1}^{N_0}{W}_{\mathrm{SD}},$

[0182] where: N₀is the number of documents in the collection C₀, and

[0183] K₁and K₂are constants (In one embodiment, K₁may be assigned a value of 0.5, and K₂1.5, but these values may be varied without departing from the spirit and scope of the invention.)

[0184] In this system, IDF_ST=the generalized inverted document frequency for term T,

[0185] IDF_ST=log((₀+K₃)/N_ST)/log(N₀+K₄)

[0186] where:

[0187] N₀is the number of documents in the collection C₀

[0188] N_STis the number of documents in the collection C₀containing the term T in the segment S_i,

[0189] K₃and K₄are constants. (In one embodiment, K₃may be assigned a value of 0.5, and K₄1.0, but these values may be varied without departing from the spirit and scope of the invention.)

[0190] In one embodiment of the system, each segment S_iof a document D consists of a portion of the text of the document D, and there are no segments containing index terms. In this embodiment, the weights W_STDassigned to terms T in the segments S_iof the document D are equal. In this embodiment, the factors W_STD, the weights assigned to terms T in segment S_iof document D, may all be considered to be equal to 1.0, and the formula simplifies to: $S_D=\sum _{T=1}^{T_0}\sum _{S_i=1}^{S_0}{\mathrm{TF}}_{\mathrm{STD}}*{\mathrm{IDF}}_{\mathrm{ST}}$

[0191] where:

[0192] TF_STD=Robertson's generalized term frequency score for Term T in Segment S_iof Document D

[0193] =G_STD/(G_STD+K₁+K₂*W_SD*(H_SD/H_SO)),

[0194] where: $\begin{array}{c}{G}_{\mathrm{STD}}=\mathrm{the}\mathrm{generalized}\mathrm{term}\mathrm{count}\mathrm{for}\mathrm{Term}T\mathrm{in}\mathrm{Segment}{S}_i\\ \mathrm{of}\mathrm{Document}D,\\ =W_{\mathrm{SD}}*N_{\mathrm{STD}},\\ H_{\mathrm{SD}}=\sum _{T=1}^{L_{\mathrm{SD}}}{N}_{\mathrm{STD}}\\ H_{\mathrm{SO}}=\sum _{N=1}^{N_0}\left(W_{\mathrm{SD}}*H_{\mathrm{SD}}\right)/\sum _{N=1}^{N_0}{W}_{\mathrm{SD}},\end{array}$  IDF_ST=log((N₀+K₃)/N_ST)/log(N₀+K₄)

[0195] In this embodiment, if the document has only a single segment, then W_SDmay be considered to be equal to 1.0 for that segment, and the formula further reduces to: $S_D=\sum _{T=1}^{T_0}{\mathrm{TF}}_{\mathrm{TD}}*{\mathrm{IDF}}_T$

[0196] where:

[0197] TF_STD=Robertson's generalized term frequency score for Term T in Segment S_iof Document D

[0198] =G_TD/(G_TD+K₁+K₂*(H_D/H_O)),

[0199] where: $\begin{array}{c}{G}_{\mathrm{TD}}=\mathrm{the}\mathrm{generalized}\mathrm{term}\mathrm{count}\mathrm{for}\mathrm{Term}T\mathrm{in}\mathrm{Segment}{S}_i\\ \mathrm{of}\mathrm{Document}D,\\ =N_{\mathrm{TD}},\\ H_D=\sum _{T=1}^{L_D}{N}_{\mathrm{TD}}\\ H_O=\sum _{N=1}^{N_0}\left(H_D\right)/N_0,\end{array}$  IDF_ST=log((N₀+K₃)/N_T)/log(N₀+K₄)

[0200] This is the conventional Robertson's term frequency score for an unsegmented text document.

[0201] In another embodiment of the system, in which a segment S_Iof a document D contains index terms automatically associated with the document D according to the system, the weight W_STDassigned to an index term T_nin segment S_Iof a document D is $W_{\mathrm{STD}}=f_D\left(T_n\right)/\sum _{T=1}^{L_{\mathrm{SD}}}{f}_D\left(T_t\right),$

[0202] where f_D(T_n) has the value set forth above, and L_SDis the number of index terms in segment S_Iof document D.

[0203] In this embodiment of the system, other segments of a document D may contain the text of the document D itself, or portions of the text, or other index terms associated with the document by other methods.

[0204] In the embodiment of the system in which only the index terms automatically associated with the document by the system are utilized to carry out a search query, the formula for the score assigned to a document according to the system reduces to the following: $S_D=\sum _{T=1}^{T_0}{\mathrm{TF}}_{\mathrm{TD}}*{\mathrm{IDF}}_T$

[0205] where: S_Dis the total score for the document D,

[0206] T₀is the number of terms which occur in the query Q,

[0207] TF_TD=Robertson's generalized term frequency score for Term T of Document D

[0208] =G_TD/(G_TD+K₁+K₂),

[0209] where:

[0210] G_TD=the generalized term count for Term T of Document D,

[0211] =f_D(T_n)

[0212] In this embodiment, because the weights assigned to the index terms in a document are normalized, the “length” of every document is 1.0, and the denominator of Robertson's term frequency score considerably simplifies.

[0213] In a further embodiment of the system described herein, it is desired to present further information to a user who has visited a Web site, when the Web site has permitted the user to enter terms describing an item of interest to the user. In this embodiment, an item of interest may be considered to be a product or service, or a number of (related or unrelated) products or services. In response to the user providing terms related to the product(s) or service(s) which he is seeking (such as, for example, “automobiles, used”), the Web site may display for the user a list of categories. Each category (such as, for example, “Auto Dealers, Used Cars”) may contain information about merchants, stores or other sources (such as, for example, “Lannan Chevrolet, Oldsmobile”) for a particular variety of products or services which may relate to the product(s) or service(s) which the user is seeking. In this embodiment, merchants or stores who carry products or services may be assigned to more than one category (variety of product or service). The user then may select a particular category from the list of categories displayed to him, and the items of interest (merchants or stores) in that category will be displayed for him.

[0214] In this embodiment of the system described herein, it is desired to present additional material to a user who is searching for particular products or services, in addition to the list of categories which contain merchants or stores who may carry the desired product or service. For example, it may be desired to present the user with banner advertisements, such as for automobile financing, which relate to the product or service, such as automobiles, for which he is searching.

[0215] In one embodiment of the system described herein, there may be a very large number of individual merchants or stores to be organized into categories of products or services for presentation.

[0216] In this embodiment of the system, there are fewer categories of products or services than individual merchants or stores. While the number may vary without departing from the spirit and scope of the invention, in one embodiment of the system there may be about 20,000 categories. Each category has associated with it a set of terms (such as, for example, “Auto Dealers, Used Cars”) which describe the product(s) or service(s) which the merchants, stores or other sources associated with the category may provide. Each category further has associated with it a category identifier term which is unique to it, and serves to identify the category.

[0217] In this embodiment of the system, it is desired to choose a banner advertisement to present to a user. The banner advertisements in turn may be divided into categories. While the number may vary without departing from the spirit and scope of the invention, there may be about 50 categories into which the banner advertisements may be divided. (To avoid confusion with the categories into which the items of interest are divided, these banner advertisement categories will be referred to hereafter as “supercategories.”)

[0218] As illustrated by FIG. 10, it is desirable in this embodiment of the system to assign each category 2210 of merchants or stores to one and only one supercategory 2220 of banner advertisements. In this embodiment of the system, each supercategory has associated with it the sets of terms (such as, for example, “Auto Dealers, Used Cars”) which describe the product(s) or service(s) which the merchants, stores or other sources associated with the categories assigned to it may provide. Each supercategory further has associated with it the category identifier terms which are unique to the categories assigned to it.

[0219] According to FIG. 11, this method 2230 proceeds first at a step 2240 to select every category C_iof merchants or stores 2210 which has associated with it a term or terms (such as, for example, “Auto Dealers, Used Cars”) describing the product(s) or service(s) which the merchants, stores or other sources associated with the category may provide, that matches any term or terms in the user query Q_i(such as “automobiles, used”).

[0220] After every such category C_iof merchants or stores 2210 is selected at the step 2240, control passes to a step 2340. At the step 2340, a new Query Q′_iis prepared, consisting of the original user Query Q_iwith the addition of all terms which describe the product(s) or service(s) which the merchants, stores or other sources associated with the said categories C_imay provide, and with the further addition of the unique category identifier terms T_iwhich identify the categories C_i.

[0221] After the new Query Q′_iis prepared at the step 2340, control passes to a step 2350, at which the new Query Q′_iis run, on the collection C′_oof supercategories 2220. There are a number of methods of running the query Q′_ion the collection C′_oof supercategories 2220, which will be known to one of ordinary skill in the art.

[0222] In one embodiment the query is run by utilizing Robertson's term frequency score, where the score for a supercategory S_Cis determined by: $S_C=\sum _{T=1}^{T_0}{\mathrm{TF}}_{\mathrm{TD}}*{\mathrm{IDF}}_T,$

[0223] where: T₀is the number of terms which occur in the query Q′_i,

[0224] TF_TDis Robertson's term frequency for term T in supercategory S_C,

[0225] =N_TC/(N_TC+K₁+K₂*(L_C/L₀)),

[0226] where: N_TCis the number of times the term T occurs in supercategory S_C,

[0227] L_Cis the length of supercategory S_C,

[0228] L₀is the average length of a supercategory, and

[0229] K₁and K₂are constants

[0230] and IDF_T=log((N+K₃)/N_T)/log(N+K₄)

[0231] where:

[0232] N is the number of supercategories in the collection

[0233] N_Tis the number of supercategories containing the term T, and

[0234] K₃and K₄are constants.

[0235] In another embodiment of the system, the categories have been assigned to supercategories by a combination of methods. In this embodiment, a certain number of the categories may have been assigned to supercategories manually, while the remainder may have been assigned to supercategories utilizing a variety of automatic or semi-automatic index term augmentation techniques. While the number of categories assigned manually may vary without departing from the spirit and scope of the invention, and the number and type of automatic and semi-automatic index term augmentation techniques utilized may vary without departing from the spirit and scope of the invention, in one embodiment about 2,000 out of about 20,000 categories are assigned manually and the remainder by the semi-automatic technique of this system, which utilizes the co-occurrence of terms between the categories assigned manually and an unassigned category to help assign the unassigned categories.

[0236] In one embodiment of the system, it is further desired in evaluating queries Q′_iand selecting a supercategory to assign differing weights to the terms and term identifiers associated with categories, depending on whether the category has been manually assigned to a supercategory, or assigned automatically or semi-automatically. While the weights thus assigned may vary without departing from the spirit and scope of the invention, in one embodiment the terms and term identifiers associated with categories manually assigned to supercategories are assigned a weight of 1.0, while the terms and term identifiers associated with categories assigned to supercategories by the semi-automatic method of the system described herein which utilizes the co-occurrence of terms between the manually-assigned categories and an unassigned category to help assign the unassigned categories are assigned a weight of 0.4.

[0237] In this embodiment, in order to evaluate the query Q′_ithe supercategories are considered to comprise multiple segments. In one segment are the terms and term identifiers associated with the categories assigned to the supercategory manually, while each of the other segments comprises the terms and term identifiers associated with the categories assigned to the supercategory by a particular automatic or semi-automatic method. In this embodiment the generalized term frequency score for a supercategory S_Cwith respect to the query Q′_imay be calculated as follows: $S_C=\sum _{T=1}^{T_0}\sum _{S_i=1}^{S_0}{\mathrm{TF}}_{\mathrm{STC}}*{\mathrm{IDF}}_{\mathrm{ST}}$

[0238] where: S_Cis the total score for the supercategory S_C,

[0239] T₀is the number of terms which occur in the query Q′_i,

[0240] S₀is the number of segments in the supercategory S_C,

[0241] TF_STC=Robertson's generalized term frequency score for Term T in Segment S_iof supercategory S_C

[0242] =G_STC/(G_STC+K₁+K₂*W_SC*(H_SC/H_SO)),

[0243] where:

[0244] G_STC=the generalized term count for Term T in Segment S_iof supercategory S_C,

[0245] =W_SC*W_STC*N_STC,

[0246] where:

[0247] W_SCis the weight assigned to segment S_iof the supercategories,

[0248] W_STCis the weight assigned to term T in segment S_iof supercategory S_C, and

[0249] N_STCis the number of times the term T occurs in segment S_iof supercategory S_C,

[0250] H_SC=the generalized length of segment S_iof supercategory S_C, $H_{\mathrm{SC}}=\sum _{T=1}^{L_{\mathrm{SC}}}{W}_{\mathrm{STC}}*N_{\mathrm{STC}}$

[0251] where:

[0252] L_SCis the number of different terms in segment S_iof supercategory S_C,

[0253] H_SO=the generalized average length of segment S_iof the supercategories, $H_{\mathrm{SO}}=\sum _{C=1}^{C_0}\left(W_{\mathrm{SC}}*H_{\mathrm{SC}}\right)/\sum _{C=1}^{C_0}{W}_{\mathrm{SC}},$

[0254] where: C₀is the number of supercategories and

[0255] K₁and K₂are constants (In one embodiment, K₁may be assigned a value of 0.5, and K₂1.5, but these values may be varied without departing from the spirit and scope of the invention.)

[0256] In this system, IDF_ST=the generalized inverted document frequency for term T,

[0257] IDF_ST=log((C₀+K₃)/C_ST)/log(C₀+K₄)

[0258] where:

[0259] C₀is the number of supercategories

[0260] C_STis the number of supercategories containing the term T in the segment S_i,

[0261] K₃and K₄are constants. (in one embodiment, K₃may be assigned a value of 0.5, and K₄1.0, but these values may be varied without departing from the spirit and scope of the invention.)

[0262] In the embodiment of the system in which the terms and term identifiers associated with categories manually assigned to a supercategory are assigned a weight of 1.0, and are assigned to one segment of the supercategory, while the terms and term identifiers associated with categories assigned to the supercategory by the semi-automatic method of the system described herein, which utilizes the co-occurrence of terms between the manually-assigned categories and an unassigned category to help assign the unassigned categories, are assigned to the other segment of the supercategory, and are assigned a weight of 0.4, the generalized term frequency score for a supercategory S_Cwith respect to the query Q′_imay be calculated as follows, where all terms in a segment are assigned equal weight W_STC: $S_C=\sum _{T=1}^{T_0}\sum _{S_i=1}^2{\mathrm{TF}}_{\mathrm{STC}}*{\mathrm{IDF}}_{\mathrm{ST}}$

[0263] where: S_Cis the total score for the supercategory S_C,

[0264] T₀is the number of terms which occur in the query Q′_i,

[0265] TF_STC=Robertson's generalized term frequency score for Term T in Segment S_iof supercategory S_C

[0266] =G_STC/(G_STC+K₁+K₂*W_SC*(H_SC/H_SO)),

[0267] where:

[0268] G_STC=the generalized term count for Term T in Segment S_iof supercategory S_C,

[0269] =W_SC*N_STC,

[0270] where:

[0271] W_SC, the weight assigned to segment S_iof the supercategories,

[0272] W_SC=1.0 for the segment which comprises the terms and term identifiers associated with the categories manually assigned to the supercategory S_i,

[0273] W_SC=0.4 for the segment which comprises the terms and term identifiers associated with the categories assigned to the supercategory S_iby the semi-automatic method of the system described herein, which utilizes the co-occurrence of terms between the manually-assigned categories and an unassigned category to help assign the unassigned categories, and

[0274] N_STCis the number of times the term T occurs in segment S_iof supercategory S_C,

[0275] H_SC=the generalized length of segment S_iof supercategory S_C, $H_{\mathrm{SC}}=\sum _{T=1}^{L_{\mathrm{SC}}}{N}_{\mathrm{STC}}$

[0276] where:

[0277] L_SCis the number of different terms in segment S_iof supercategory S_C,

[0278] H_SO=the generalized average length of segment S_iof the supercategories, $H_{\mathrm{SO}}=\sum _{C=1}^{C_0}\left(W_{\mathrm{SC}}*H_{\mathrm{SC}}\right)/\sum _{C=1}^{C_0}{W}_{\mathrm{SC}},$

[0279] where: C₀is the number of supercategories and

[0280] K₁and K₂are constants (In one embodiment, K₁may be assigned a value of 0.5, and K₂1.5, but these values may be varied without departing from the spirit and scope of the invention.)

[0281] In this system, IDF_ST=the generalized inverted document frequency for term T,

[0282] IDF_ST=log((C₀+K₃)/C_ST)/log(C₀+K₄)

[0283] where:

[0284] C₀is the number of supercategories

[0285] C_STis the number of supercategories containing the term T in the segment S_i,

[0286] K₃and K₄are constants. (In one embodiment, K₃may be assigned a value of 0.5, and K₄1.0, but these values may be varied without departing from the spirit and scope of the invention.)

[0287] After the new Query Q′_iis run on the collection C′_oof supercategories 2220 at the step 2350, control passes to a step 2360, at which the supercategory 2220 which achieves the highest score S_Con the Query Q′_iis selected. The process then continues, and a banner advertisement associated with the supercategory chosen at the step 2360 is displayed to the user who has presented the Query Q_i. In addition, the user is presented with the set of categories C_i, of merchants or stores 2210 which have associated with them a term or terms describing the product(s) or service(s) which the merchants, stores or other sources associated with the category may provide, that matches any term or terms in the user query. The user then has the opportunity to select any of the categories presented, and to have displayed to him the list of merchants, stores or other sources associated with the category.

[0288] In this system, when a user, who has been presented with the list of categories C_i, selects a particular category C_Sfor presentation of its list of merchants, stores or other sources, control returns to the step 2340, with the collection of categories C_ireplaced by the single category C_S.

[0289] At the step 2340, a new Query Q′_iis prepared, now consisting of the terms which describe the product(s) or service(s) which the merchants, stores or other sources associated with the single category C_Smay provide, and with the further addition of the unique category identifier term T_Swhich identifies the category C_S.

[0290] After the new Query Q′_iis prepared at the step 2340, control passes to a step 2350, at which the new Query Q′_iis run, on the collection C′_oof supercategories 2220.

[0291] After the new Query Q′_iis run on the collection C′_oof supercategories 2220 at the step 2350, control passes to a step 2360, at which the supercategory 2220 which achieves the highest score on the Query Q′_iis selected. The process then concludes, and a banner advertisement associated with the supercategory chosen at the step 2360 is displayed to the user. In addition, the list of merchants, stores or other sources of the product(s) or service(s) associated with the category C_Sis presented to the user.

[0292] While the invention has been disclosed in connection with the preferred embodiments shown and described in detail, various modifications and improvements thereon will become readily apparent to those skilled in the art. Accordingly, the spirit and scope of the present invention is to be limited only by the following claims.