EP2035972A2 - Procédés et appareil pour rechercher un contenu - Google Patents

Procédés et appareil pour rechercher un contenu

Info

Publication number
EP2035972A2
EP2035972A2 EP07798456A EP07798456A EP2035972A2 EP 2035972 A2 EP2035972 A2 EP 2035972A2 EP 07798456 A EP07798456 A EP 07798456A EP 07798456 A EP07798456 A EP 07798456A EP 2035972 A2 EP2035972 A2 EP 2035972A2
Authority
EP
European Patent Office
Prior art keywords
content
search
aggregate
relevance
expression
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
EP07798456A
Other languages
German (de)
English (en)
Other versions
EP2035972A4 (fr
Inventor
Samuel S. Epstein
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zalag Corp
Original Assignee
Zalag Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zalag Corp filed Critical Zalag Corp
Publication of EP2035972A2 publication Critical patent/EP2035972A2/fr
Publication of EP2035972A4 publication Critical patent/EP2035972A4/fr
Withdrawn legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/951—Indexing; Web crawling techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/903—Querying

Definitions

  • provisional application 60/813,246 filed June 12, 2006, and claims priority to said provisional application.
  • the specification of the 60/813,246 provisional application is hereby fully incorporated by reference, to the extent it is consistent and supports the present specification.
  • Embodiments of the present invention relate to the field of data processing, in particular, to methods and apparatuses for searching electronic documents.
  • a Web page that satisfies a given search expression typically includes constituents that do not satisfy the search expression. In many cases, a small proportion of the page's total content will be relevant to the search. If the user's goal is information that corresponds to the search expression, then delivering the entire Web page to the user entails a waste of download bandwidth and a waste of screen real estate. It also presents the user with the task of finding the relevant constituents within the Web page. Highlighting search terms on the page eases this task only slightly. The problem of presenting search results on mobile devices is especially acute.
  • Standard Web search engines return links to Web pages.
  • Various search engines handle search requests that specify categories or instances of sub-document constituents. These may be called "sub-document" search engines. Some sub-document search engines are limited to returning text constituents. Other sub-document search engines return constituents that belong to non-text categories, but are limited to non-text categories that can be characterized by very simple markup properties. Some sub-document search engines use string-based algorithms to determine which constituents to extract. Other sub-document search engines use tree-based algorithms that examine very simple properties of markup trees. Yet other sub- document search engines support highly expressive languages for specifying constituents.
  • search engines handle search requests that specify proximity relationships. Some search engines are fundamentally limited to string-based proximity relationships. Other search engines recognize constituent boundaries in order to ignore these boundaries. Other search engines recognize when search terms occur within the same constituent. None of these search engines effectively exploits structural proximity relationships that are based on properties of the tree structures (or other graph structures) and layout structures of documents. Co-occurrences of search terms within documents are evidence that the search terms are mutually relevant. Moreover, relevance is transitive. Current systems use learning algorithms that leverage these principles to enable responses to search requests where in some cases, the response doesn't include any of the words contained in the request. These systems require a learning process.
  • the very limited download bandwidth and screen real estate associated with mobile devices has motivated the creation of the WAP (Wireless Access Protocol) network. Because building a WAP site is labor intensive, the WAP network remains extremely small, in comparison to the World Wide Web, and has correspondingly less to offer users. For purposes of search, the World Wide Web is a vastly more powerful resource than the WAP network. Limited download bandwidth and limited screen real estate has also motivated the creation of browsers that reformat HTML files for presentation on mobile devices. These mobile browsers reformat content so that horizontal scrolling is reduced. They may introduce page breaks into tall pages. They may remove or replace references to large files. They may replace fonts. They may offer distinctive user interfaces. Similar functionality is also offered by server transcoders that intercept user requests for HTML files. Such a server transcoder may be applied to reformat Web pages that satisfy search criteria. Current mobile browsers and server transcoders offer at most very rudimentary content extraction facilities, based on limited ranges of simple criteria.
  • a given Web page contains a constituent Ni that contains a single occurrence of the term haydn but doesn't contain the term boccherini.
  • the page contains a constituent N 2 that contains a single occurrence of the term boccherini but doesn't contain the term haydn.
  • the page contains just this one occurrence of haydn and just this one occurrence of boccherini.
  • a user searches the Web with the intention of finding information that pertains to both haydn and boccherini. While the Web page contains occurrences of both haydn and boccherini, the page may or may not satisfy the user's search request.
  • FIG 1 illustrates an overview of the methods and apparatuses of the present invention, in accordance with various embodiments
  • Figures 2-4 illustrate selected operations of the structured content search engine of Figure 1, in accordance with various embodiments;
  • Figure 5 illustrates an example computer system, suitable for use to facilitate practice of the present invention, in accordance with various embodiments.
  • Illustrative embodiments of the present invention include but are not limited to content search methods and apparatuses, in particular, content search methods and apparatuses that examine content structures.
  • A, B and C means "(A), (B), (C), (A and B), (A and C), (B and C) or (A, B and C)".
  • (A) B means "(B) or (A B)", that is, A is optional.
  • the present invention permits the results of searches performed over sets of Web pages or other content to correspond more accurately to users' requests.
  • searches return relevant document constituents, rather than entire documents.
  • constituents from documents may be combined in a single display.
  • the present invention thus enables the display of search results on mobile and other devices without wasting download bandwidth and screen real estate on irrelevant Web page constituents.
  • the present invention also permits more accurate results for searches based on combinations of search terms, and provides mechanisms for increasing the accuracy of search results through analysis of search match distributions. Further, the present invention supports content requests that specify content categories, in addition to specifying search expressions to be matched.
  • the present invention's methods and apparatuses can be applied in advance of content-request time, annotating content that can then be cached with its annotations for faster processing at content-request time.
  • FIG 1 a block diagram illustrating an overview of the content search methods and apparatuses of the present invention, in accordance with various embodiments, is shown.
  • content request engine 111 controls the processing of user content requests.
  • Content request engine 111 directs one or more document search engines 108 to identify content 106 from a variety of sources 101, including but not limited to applications 102, Web 103, and content databases and caches 104.
  • Document search engines 108 represent a broad category of application that includes, but is not limited to, Web search engines, content management systems, and database management systems. With documents containing desired content identified, content request engine 111 directs one or more document retrieval engines 107 to retrieve documents. In some cases, document search and document retrieval functionality may be combined in a single engine. Content request engine 111 directs retrieved documents to one or more document parsers 112, which provide the tree or other graph structures associated with retrieved documents. Content request engine 111 then directs parsed documents to content search engine 114, incorporated with the teachings of the present invention. In alternate embodiments, web search engines etc. may return documents that have been pre-parsed or equivalently prepared.
  • content request engine 111 may direct the pre-parsed documents directly to content search engine 114.
  • Content search engine 114 may call string search engine 109 and measurement engine 113 directly, or may access the functionality of these engines through the mediation of content request engine 111, as shown in Figure 1.
  • Content search engine 114 may access category repository 110 directly, or may access category repository 110 through the mediation of content request engine 111, as shown in Figure 1.
  • the functionality of one or more of the following may be combined in a single engine: document retrieval engine 107, document search engine 108, string search engine 109, document parser 112, measurement engine 113.
  • Measurement engines 113 provide information related to the intended layout and rendering of retrieved documents and their constituents.
  • the output of the content search engine 114 is provided as a content constituent catalog 115.
  • the content constituent catalog is directed to a content selection engine 116, which works with a layout engine 117 to construct display presentations 118.
  • Content search engine 114 may communicate with content selection engine 116 directly, or through the mediation of content request engine 111, as shown in Figure 1. While for ease of understanding, the functions performed by content request engine 111 and content search engine 114 are illustrated as distinct components, in practice, their functions may be partitioned and assigned to different smaller modules and/or tasks. Alternatively, they may be combined in a single module. The various modules and/or tasks may be executed as a single thread, or as multiple threads where appropriate.
  • the execution of document retrieval engine 107, document search engine 108, string search engine 109, content request engine 111, document parser 112, measurement engine 113, content search engines 114, content selection engine 116, and layout engine 118, and the storage of category repository 110 may be on the same system, and in other embodiments, they may be on different systems, e.g. with 107, 109, 111, 112, 113, 114, 116, and 117 on one server, and document search engine 108 on a different server.
  • HTTP Hypertext Transmission Protocol
  • HTTPS Hypertext Transmission Protocol Secured
  • category repository 110, content request engine 111, content search engine 114, content selection engine 116, and layout engine 117, together with document retrieval engine 107, string search engine 109, document parser 112, and measurement engine 113 may be implemented as part of a "larger" product offering.
  • all nine components 107, 109, 110, 111, 112, 113, 114, 116, and 117 may be implemented as part of a Web search service.
  • 107, 108, 109, 110, 111, 112, 113, and 114 may be part of a Web search service, while content selection engine 116 and layout engine 117 may be part of an enhanced Web browser or publishing tool.
  • other implementation arrangements may also be possible.
  • Markup files and other content sources are viewed as structured content, in tree, graph or other like forms. Important categories of content sources are intended to be laid out by specific classes of layout engines. For example, HTML files are generally intended to be laid out by browsers whose layout engines conform to W3C standards. For ease of understanding, the invention will be primarily described with markup files and other content sources structured as trees, and content search engine 114 shall also be referred to as structured content search engine 114, however the description should not be read as limiting on the invention. Embodiments of the present invention use tree structures (or more generally, graph structures), layout structures, and content category information to capture within search results relevant content that would otherwise be missed, to reduce the incidence of false positives within search results, and to improve the accuracy of rankings within search results.
  • Embodiments of the present invention further use tree structures (or more generally, graph structures), layout structures, and content category information to extend search results to include sub-document constituents.
  • Embodiments of the present invention also support the use of distribution properties as criteria for ranking search results.
  • embodiments of the present invention support search based on structural proximity.
  • structured content search “structured search,” and “structure search” will be used interchangeably to refer to embodiments of the present invention.
  • an "atomic search term” is either a quoted string of characters, or a string of characters that doesn't contain a designated delimiter (such as space, period, and quotation mark).
  • atomic search terms and search expressions more generally will appear in italics.
  • “Fr ⁇ nz Joseph H ⁇ ydn” and H ⁇ ydn are examples of atomic search terms.
  • Atomic search terms may include wildcards.
  • a “search expression” may be formed from atomic search terms with various operators, such as the standard conjunction, disjunction, and negation operators. In what follows, AND denotes the standard conjunction operator, OR denotes the standard disjunction operator, and NOT denotes the standard negation operator. Using parentheses for grouping, (haydn AND NOT mozart) OR (boccherini AND pleyel) is an example of a search expression.
  • embodiments of the present invention employ a recursive procedure that calls another recursive procedure, as illustrated in Figures 2-3.
  • the outer procedure walks the parse trees associated with markup or other content, from bottom to top. In various embodiments, these parse trees may be enhanced with information derived from layout structures.
  • the inner procedure walks the simple parse trees associated with search expressions, also from bottom to top.
  • the subroutine illustrated in Figure 2 is part of the structured content request search engine 114 illustrated in Figure 1. In various embodiments, the subroutine "calculate densities and r- centers for N for all sub-expressions of E" 206 shown in Figure 2 corresponds to the subroutine illustrated in Figure 3.
  • the subroutine "calculate node deviation for E for N" 207 corresponds to the subroutine illustrated in Figure 4.
  • embodiments of the present invention define matching as a function that takes a content constituent and a search expression and returns a real number between 0 and 1 inclusive, with 1 corresponding to the best possible match, and 0 corresponding to no match. Taking values between 0 and 1 is a matter of computational convenience. It involves no loss of modeling power. Content constituents may be aggregates of lower content constituents.
  • the 9 occurrences of haydn are at word positions 99, 202, 301, 397, 499, 601, 706, 798, and 899. It's reasonable to proceed on an assumption that S 1 , i is highly relevant to haydn, and that S 2 as a whole is more relevant to haydn than is Si as a whole.
  • the even distribution of haydn within S 2 guarantees that no part of S 2 is more than 105 words away from an occurrence of haydn, while the entire first half of Si is more than 300 words away from any occurrence of haydn. It's reasonable to proceed on an assumption that most of Si has little relevance to haydn.
  • a search that misses this text is inadequate.
  • the text can be included in a response to a request for texts that contain an occurrence of haydn within 7 words of an occurrence of boccherini, but such a request will miss texts with similar distribution patterns where the occurrences o ⁇ haydn and boccherini are a little farther apart.
  • prior algorithms don't take adequate account of density and distribution within strings, and don't take advantage of tree structures (or more generally, graph structures), layout structures, and content category information.
  • boccherini is an example of a search expression with the structural proximity operator. Given that the structural proximity operator is available, it makes sense to also provide classical logical operators. Thus a content constituent matches haydn AND boccherini if and only if it matches both haydn and boccherini. Negation and disjunction may likewise be interpreted a strictly classical sense. Search expressions may be constructed recursively with structural proximity and other operators.
  • the structural proximity operator may be alternatively referred to as the "structural proximity conjunction” operator, to emphasize its distinctness from the structural proximity complement operator and the structural proximity disjunction operator.
  • search requests may also include various scalar-valued (fuzzy) logical operators.
  • search requests may include operators which will be denoted here as &&, 11, and /.
  • the score on N of Ei && E 2 is the minimum of Si and S 2
  • E 2 is the maximum of Si and S 2
  • the score on N of !Ei is (1 - Si).
  • a Boolean-valued expression E is embedded under a scalar- valued operator, then a value for E of true is converted to a scalar value of 1 , and a value for E of false is converted to a scalar value of 0.
  • embodiments of the present invention assign a "relevance value" ("r-value” or simply “value”) to each word in S according to the following conditions: (i) if S contains no occurrences of E, then each word in the string
  • r-value (corresponding to a position in the string) is assigned an r-value of 0; (ii) if S contains at least one occurrence of E, then for any word W in S, the r-value assigned to W is ⁇ i ⁇ i ⁇ k (1/(1+ di) x ), where k is the number of occurrences of E in S, where x (the "distance attenuation exponent") is a positive real number, and where U 1 is the distance in words between W and the i-th occurrence of E, the distance between a pair of adjacent words taken as 1, and so on.
  • the distance between a word W and an occurrence of an atomic search term E that comprises more than one word is the maximum of the distances between W and the words in E.
  • Embodiments of the present invention proceed to normalize r- values assigned to words so that these r- values lie between 0 and 1 inclusive — the r-value assigned to W according to paragraph 0038 above is divided by ⁇ i ⁇ i ⁇ n (1/(1+ di) x ), where n is the number of words in S, where x is the attenuation exponent, and where U 1 is the distance between W and the i-th word of S.
  • prior art includes methods for calculating or estimating these normalization factors with closed forms.
  • Various embodiments of the invention use various distance attenuation exponents. A distance attenuation exponent of 1 gives reasonable results.
  • Alternative embodiments of the invention measure distances within strings in characters, rather than in words.
  • the distance between a word W and an occurrence of an atomic search term E that comprises more than one word is the minimum of the distances between W and the words in E. According to other alternative embodiments of the invention, the distance between a word W and an occurrence of an atomic search term E that comprises more than one word is the arithmetic mean of the distances between W and the words in E.
  • a r-value is assigned to each word in S according to the following conditions: (i) if S contains no occurrences of E, then each word in the string is assigned a r- value of 0; (ii) if S contains at least one occurrence of E, then for any word W in S, the r- value assigned to W is the 1/(L X ), where L is the length in words of the shortest substring of S that contains both W and an occurrence of E, and where x is a positive real number.
  • distances computed in accordance with paragraph 0038, or lengths computed in accordance with paragraph 0042 are calculated in terms of characters, rather than in terms of words.
  • Embodiments of the present invention capture the density of matches for atomic search expression E in string S as the arithmetic mean of the r- values assigned to the words in S.
  • Alternative embodiments of the present invention capture the density of matches for atomic search expression E in string S as the median, or as the geometric mean, of the r- values assigned to the words in S.
  • ⁇ distributed score
  • ⁇ will take values between 0 and 1.
  • a high value for ⁇ indicates an even distribution of E in S.
  • a low value for ⁇ indicates an uneven distribution of E in S.
  • median absolute deviation, or standard deviation, or variance may be used in placed of average absolute deviation in evaluating evenness of distributions. The characterization of distributions may be refined in terms of higher moments. D. Relevance center
  • Embodiments of the present invention capture the "relevance center" ("r-center") of occurrences of E in S according to the following formula: ( ⁇ i ⁇ i ⁇ n (V 1 * i))/ ( ⁇ i ⁇ i ⁇ n V 1 ), where n is the number of words in S, and where V 1 is the r- value assigned to the i-th word, with the first word in the string counting as the 1 st word, rather than the 0-th word, and so on.
  • the r- values assigned to words for purposes of calculating the relevance center may use a distance attenuation exponent that differs from the distance attenuation exponent that's used to assign r-values to words for purposes of calculating density and distribution.
  • Alternative embodiments of the present invention capture the relevance center of occurrences of E in S according to the following formula: ( ⁇ i ⁇ i ⁇ k p ⁇ /k, where k is the number of occurrences of E in S, and where P 1 is the position of the i-th occurrence of E, with the position of the first word in S counting as 1, and so on.
  • Embodiments of the present invention assign an overall score for S as a match for E according the following formula: C 1 * D + C 2 * ⁇ , where D is density as defined in paragraph 0044, where ⁇ is as defined in paragraph 0046, and where cl and c2 are positive real numbers such that C 1 + C 2 ⁇ 1. Note that 0 ⁇ D ⁇ 1 and 0 ⁇ ⁇ ⁇ 1, so 0 ⁇ (C 1 * D + C 2 * ⁇ ) ⁇ 1.
  • the values of C 1 and C 2 can be tuned as desired to adjust the relative importance of density and distribution in judging the relevance of S for E. Note that in ranking search results, properties in addition to D and ⁇ , such as string size, may be taken into account.
  • each property P takes values between 0 and 1, and using a formula of the form ⁇ i ⁇ ⁇ m (C 1 * P 1 ), where there's a total of m properties and where ⁇ i ⁇ i ⁇ ⁇ A ⁇ 1, to calculate overall score.
  • the r-value assigned to W is the geometric mean of the r- values assigned to W for E 1 ,..., E m .
  • the r-value assigned to W for case (ii) in accordance with paragraph 0053 is the arithmetic mean of the r- values assigned to W for E 1 ,..., E m .
  • the r-value assigned to W is the minimum of the r-values assigned to W for E 1 ,..., E m .
  • the r-value assigned to W is the geometric mean of the r- values assigned to W for E 1 ,..., E m .
  • condition (i) in accordance with paragraph 0053 is omitted and condition (ii) uses some function other than the geometric mean. According to these embodiments, S can match haydn ## boccherini even if it doesn't match haydn.
  • embodiments of the present invention assign a r-value to each word W in S according to the following condition: the r-value assigned to W is the maximum of the r- values assigned to W for E 1 , ... , E m .
  • Ei and E 2 are atomic search terms. IfEi has ki occurrences in S, and E 2 has k 2 occurrences in S, consider the ki * k 2 distinct pairs formed by taking an occurrence of Ei as the first member of the pair and an occurrence of E 2 as the second member of the pair. In alternative embodiments of the present invention, each of these pairs is taken as a "virtual match" for Ei ## E 2 . Similarly, virtual matches for Ei ## E 2 ## ... ## E m , where E 1 , ... E m are atomic search terms, are taken as m-tuples of occurrences, where the i-th member of such an m-tuple is an occurrence OfE 1 .
  • Embodiments of the present invention identify the position of a virtual match with its relevance center. Embodiments of the present invention also assign "weights" to virtual matches. Weights assigned to virtual matches compare with weights of 1 that are assigned to occurrences of atomic search terms. Embodiments of the present invention assign a relevance center ("r-center") to a tuple that's a virtual match for Ei ## E 2 ## ... ## E m , where E 1 , ..., E m are atomic search terms, according to the formula ( ⁇ i ⁇ i ⁇ m Pi)/m, where P 1 is the position of the i-th member of the virtual match, with the position of the first word in S counting as 1, and so on.
  • Embodiments of the present invention assign a weight to a tuple that's a virtual match for Ei ## E 2 ## ... ## E m , where E 1 , ... , E m are atomic search terms, as ⁇ 1 ⁇ j ⁇ m (l/(l + dj) x ), where x (the "distance attenuation exponent") is a positive real number, and where U 1 is the distance from the i-th member of the virtual match to the r-center of the virtual match.
  • E 1 , ...
  • E m are search expressions built up from atomic search terms through applications of ##, embodiments of the present invention assign a relevance center to a tuple that's a virtual match for Ei ## E 2 ## ... ## E m as ( ⁇ 1 ⁇ i ⁇ m(wi * Pi))/ ( ⁇ i ⁇ i ⁇ m(wi)), where W 1 is the weight assigned to the i-th member of the virtual match, and where P 1 is the r-center of the i-th member of the virtual match, with the position of the first word in S counting as 1 , and so on.
  • Embodiments of the present invention assign a weight to a tuple that's a virtual match for Ei ## E 2 ## ...
  • Embodiments of the present invention assimilate virtual matches and occurrences of atomic search terms.
  • Virtual matches may be treated with methods of the present invention that apply to occurrences of atomic search terms. They may also be treated according to prior art methods that apply to occurrences of atomic search terms.
  • Embodiments of the present invention calculate densities, relevance centers, distributions, and overall scores based on virtual matches.
  • densities, relevance centers, distributions, and overall scores are calculated on the basis of assignments of r- values, as described in paragraphs 0052 - 0059 above and paragraphs 0085 - 0098 below.
  • Alternative embodiments of the present invention evaluate a string S as a match for atomic search term E by evaluating S as a match for E ## E.
  • Other alternative embodiments evaluate a string S as match for atomic search term E by evaluating S as a match for other search expressions composed from E and ##.
  • embodiments of the present invention calculate densities, distributions, relevance centers, and overall scores as in paragraphs 0044 - 0051.
  • embodiments of the present invention walk the parse trees associated with markup or other content from bottom to top.
  • Alternative embodiments walk trees according to other orderings of tree constituents.
  • Applicable content includes but is not limited to Web pages, XML documents, text documents, and database records and other database structures.
  • constituents of these parse trees will be called "content constituents.”
  • Measurement engine 113 in Figure 1 is a module that provides measurement information.
  • Web browser layout modules provide measurement engine functionality. Measurement information includes, but isn't limited to, width and height, horizontal and vertical position, length in characters (for texts), and size of referenced file (for images, videos, and so on). Some measurement information is readily available. Text lengths are apparent in the tree itself. Widths and heights may be specified as node attributes, in pixels or in other units. The size of a file can be obtained by downloading the file.
  • embodiments of the present invention insert "virtual constituents" into parse trees.
  • Some documents include constituents that are far apart in the parsed structure for the document, but near each other in the window (or printed page) when the document is rendered.
  • an HTML Web page may contain article text in a table cell, while a captioned image that's associated with the article is in a table cell belonging to a different table, with the two tables having no common ancestor below the body constituent, and with each of these two tables having multiple levels of ancestor intervening between it and the body constituent.
  • the captioned image is rendered to the immediate right of the article text.
  • Embodiments of the present invention posit virtual constituents that contain constituents of the parse tree as sub-constituents.
  • an article constituent is posited that includes both the article text and the captioned image.
  • Embodiments of the present invention insert virtual constituents into the parse tree under the root of the parse tree, or elsewhere. Sub-constituents of virtual constituents are then deleted from their original positions in the parse tree. In alternative embodiments, immediate sub- constituents of virtual constituents retain their original parents in the parse tree. In these embodiments, virtual constituents are inserted in what becomes (if it is not already) a directed acylic graph that is not a tree, with some nodes having multiple parents.
  • an "deduced semantic tree” is built as a data structure distinct from the markup parse tree. Nodes of the deduced semantic tree correspond to instances of content categories, where articles, captioned images, and captions are examples of content categories. In these embodiments, while markup parse trees and deduced semantic trees are distinct, nodes of deduced semantic trees may be annotated with pointers to nodes of markup parse trees, and vice-versa.
  • a markup file that is intended to be laid out by a specific class of layout engines typically has a string structure, a tree structure, and a layout structure.
  • Other tree-based content sources at least have string structures and tree structures.
  • embodiments of the present invention assign relative sizes to the children of N, assign relative distances between the children of N, and correlate distances within the children of N with distances between the children of N. These assignments of sizes and distances capture what may be called a "geometry of relevance.”
  • the relative sizes of children of N influence the relative degree to which children's matches for E affect N's match for E.
  • the relative distances between children of N affect matches for structural proximity, and affect the evenness of match distributions.
  • String structures, tree structures, and layout structures may suggest very different relative sizes of constituents in particular cases, and may suggest very different relative distances between constituents in particular cases. For example, for text laid out in columns, the last word in the leftmost column is adjacent to the first word in the next column in string structure, but these two words are far from adjacent in layout. For another example, an image that occupies a large area in layout may correspond to a very small substring of the terminal string of a markup file (whether or not the image node has an attribute similar to HTML ALT, with a value that's subject to string search).
  • two words that are separated by 20 words in the terminal string of an HTML file where 10 of these words correspond to end tags and the other 10 correspond to simple start tags, are much farther apart in tree structure than are two words with a common TEXT parent that are separated by 20 words.
  • Embodiments of the present invention assign r-sizes and r-distances in a bottom-up traversal of the parse tree, where the parse tree has been annotated with measurement information, and where virtual constituents have possibly been inserted, as discussed in paragraph 0074 above.
  • Alternative embodiments assign r-sizes and r-distances in a bottom-up traversal of the deduced semantic tree, where the deduced semantic tree has been constructed as discussed in paragraph 0074 above.
  • Embodiments of the present invention assign r-sizes and r-distances, and also calculate match densities, match r-centers, match distributions, and match scores, in a single bottom-up traversal of the enhanced parse tree, or in a single bottom-up traversal of the deduced semantic tree.
  • Alternative embodiments walk trees according to other orderings of tree constituents.
  • the terminal nodes of the enhanced parse tree are texts and graphics. All texts use the same font. The sizes of texts correspond to their word counts. The relative sizes of texts and graphics are determined by a fixed conversion factor c, so that a text with n words is counted as the same r-size as a graphic with a rendered area of c * n square pixels.
  • the r-size of a non-terminal node is the sum of the r-sizes of its children. Graphics are considered to have a single internal position. Every word of a string is considered to be a position within the string. Given a pre-terminal node N that dominates some text nodes and some graphic nodes, for purposes of calculating r-distances between positions in child nodes, the child nodes are considered to be concatenated together from left to right, with a graphic with area A counting for the same distance as A/c words. For example, suppose that N's children in left-to-right order are a text node with 500 words, a graphic with area of 20,000 square pixels, and a text node with 1000 words.
  • N has an r-size equivalent of 1600 words.
  • the r-distance between word position 700 in the last child, and the sole position in the graphic is 700 word- unit equivalents.
  • N corresponds to an encyclopedia article.
  • N has children N 1 , ..., N 6 in left-to-right order.
  • Ni is a header constituent that includes the title of the article.
  • N 2 , N 3 , and Ns are sections of the article.
  • N 4 is a photograph and N 6 is a video.
  • E matches within header Ni are more significant in determining the relevance of N to E than are matches within any of the other children.
  • N 2 for example
  • its r-size is much greater. This illustrates what will be called "boosting" in the discussion below.
  • the video is rendered in a smaller area than the photograph, but because it's a video, in this scenario its r-size is considered to be greater than the r-size of the photograph.
  • Sections N 2 and N3 concern somewhat different sub-topics of the article topic. Therefore, for purposes of calculating r-distance, there's a sort of gulf between N 2 and N 3 , with the last word of N 2 considered to be much more than one word away from the first word OfN 3 . This illustrates what will be called "warping" in the discussion below.
  • the r-distance between p 2jl and p 3jl equals the r-distance between p 2i2 and p 3j2 .
  • the triangle inequality (d(pi, p 3 ) ⁇ d(pi, p 2 ) + d(p 2 , p 3 )) may not apply either.
  • Selected further illustrations of distance branching in this scenario are as follows: the r-distance between N 2 and N 3 equals the r-distance between N 2 and N5 equals the r- distance between N 3 and N 5 .
  • the r-distance between Ni and N 2 (which is smaller than the r- distance between N 2 and N 3 ) equals the r-distance between Ni and N 3 equals the r-distance between Ni and N 4 equals the r-distance between Ni and N5 equals the r-distance between Ni and N 6 .
  • a full specification of relevance geometry for an enhanced parse tree or for a deduced semantic tree specifies (i) relevance centers ("r-centers") for nodes and (ii) how the spaces corresponding to child nodes embed within the spaces corresponding to parent nodes. (ii) implies specification of the r-distances between arbitrary positions in arbitrary sibling nodes.
  • the relevance geometries of nodes can be modeled as subsets of R n , where spaces corresponding to child nodes are possibly transformed and then embedded in the spaces corresponding to parent nodes. In some cases, more abstract geometries will be more natural.
  • Boosting, warping, distance branching and other like deformations may be inferred with more or less confidence from knowledge of markup or other content authoring languages (for example, text under an HTML Hl node is boosted to a greater extent than text under an H2 node), or may be inferred with more or less confidence from layout structures (for example, as intervening white space increases, or as intervening border treatment becomes more prominent, the warped distance between adjacent constituents increases), or may be inferred with more or less confidence from knowledge of content categories (for example, within a constituent that comprises multiple product descriptions, the r-distance between any two product descriptions is the same).
  • guidance on r- sizes and r-distances may be provided by supplemental authoring. This supplemental authoring may be reflected in the content sources themselves, or may be stored independently for use with specific content sources (for example, for use with all news articles from a particular Web site).
  • Embodiments of the present invention evaluate search expressions on sub-constituents of documents, and thereby support applications that return document sub-constituents in response to search requests.
  • embodiments of the present invention walk the tree associated with U, from bottom to top.
  • Alternative embodiments walk trees according to other orderings of tree constituents.
  • E At each node, it's first determined whether the node is a candidate for evaluation of E. If it is, the relevance geometry presented by the node's children is then calculated. Then E is evaluated for the node, starting with E's terminal sub-expressions, and working from bottom to top. Note that a node may be a candidate for evaluation of E even if it's not a candidate response to the search request — evaluation of the node may be a necessary step in the evaluation of a higher constituent that's a candidate response.
  • the result of evaluation is assignment of overall scores to constituents of U as matches for E.
  • various embodiments of the present invention make various data structures calculated for lower nodes available for use in calculating data structures for higher nodes. Some embodiments of the present invention store all the data structures that have been calculated for lower nodes on those lower nodes, so that when data structures for higher nodes are calculated, the lower nodes may be re-traversed to retrieve relevant data structures. Other embodiments of the present invention eliminate re-traversals by passing up encapsulations of the data structures that have been calculated for lower nodes. These encapsulations take various forms in various embodiments of the invention.
  • inventions of the present invention make available the following data calculated for descendants of N when evaluating E for N: for each child N' of N, the node deviation of E for N', and for each sub-expression E' of E, the density of E' for N' and the relevance center of E' for N'. Yet other embodiments of the present invention supplement the node deviation of E for N' with more detailed information about the distribution of matches for E within N', as will be discussed below. Yet other embodiments of the present invention use alternative encapsulations of matches for E and sub-expression of E in lower nodes.
  • Embodiments of the present invention proceed by assigning a r- value for each sub-expression E' of E to each child N' of N.
  • assignment of r- values begins with the atomic search terms that are the terminal sub-expressions of E, and works from bottom to top.
  • embodiments of the present invention assign a pre -normalized r- value for E' to N' according to the following conditions: (i) if the density of E' is zero for all children of N, then N' is assigned a prenormalized x- value of 0 for E'; (ii) if at least one child of N has non-zero density for E', then the pre-normalized r- value assigned to N' for E' is ⁇ i ⁇ i ⁇ k (A 1 * D 1 Z(I+ d;) * ), where k is the number of children of N, where x (the "distance attenuation exponent") is a positive real number, where A 1 is the r-size of the i-th child of N, where D 1 is the density of E' in the i-th child of N, and where U 1 is the r-distance between the r-center of N' and the r
  • Embodiments of the present invention calculate the r- value for atomic expression E' for child N' of N from the pre-normalized r-value v for E' for N' as v/( ⁇ i ⁇ 1 ⁇ k (A 1 Z(I+ di) x )), where k is the number of children of N, where x is the distance attenuation exponent, where A 1 is the r-size of the i-th child of N, and where U 1 is the minimum r-distance between constituents of N' and constituents of the i-th child of N. Given a sub-expression Ei ## E 2 ...
  • E m of E the na ⁇ ve "indirect" method calculates prenormalized r- values for Ei ## E 2 ... ## E m for the children of N from the previously calculated densities and r-centers of Ei ## E 2 ... ## E m for the children of N, as follows: (i) if the density of Ei ## E 2 ... ## E m is zero for all children of N, then child N' is assigned a prenormalized r-value of 0 for Ei ## E 2 ... ## E m ; (ii) if at least one child of N has non-zero density for Ei ## E 2 ...
  • E m then the pre-normalized r-value assigned to child N' for Ei ## E 2 ... ## E m is ⁇ i ⁇ ! ⁇ k (A 1 * D ⁇ (I+ d;) x ), where k is the number of children of N, where x (the "distance attenuation exponent") is a positive real number, where A 1 is the r-size of the i-th child of N, where D 1 is the density of Ei ## E 2 ... ## E m in the i-th child of N, and where U 1 is the r-distance between the r-center of N' and the r-center of Ei ## E 2 ...
  • the prenormalized r-value assigned to child N' is the geometric mean of the pre-normalized r-values assigned to N' for E 1 , ..., E m .
  • the pre-normalized r-value for Ei ## E 2 ... ## E m that is calculated according to this na ⁇ ve direct method will be referred to below as the "direct" prenormalized r-value.
  • N has two children, where the first child Ni has a density of 0.1 for haydn, a density of 0 for boccherini, and a density of 0 for haydn ## boccherini, and where the second child N 2 has a density of 0 for haydn, a density of 0.1 for boccherini, and a density of 0 for haydn ## boccherini.
  • the children of N both receive pre-normalized r-values of 0 for haydn ## boccherini.
  • N Given that normalization does not affect pre-normalized r-values of 0, and given that the density of haydn ## boccherini for N will be calculated from the r-values for haydn ## boccherini assigned to children of N, N will be assigned a density of 0 for haydn ## boccherini. But N features occurrences of haydn (within Ni) in proximity to occurrences of boccherini (within N 2 ), so the na ⁇ ve indirect method is clearly inadequate.
  • N has two children with identical r-sizes, where the first child Ni has a density of 0.1 for haydn, a density of 0.05 for boccherini, and a density of 0.06 for haydn ## boccherini, and where the second child N 2 has a density of 0.05 for haydn, a density of 0.1 for boccherini, and a density of 0.06 for haydn ## boccherini.
  • the occurrences of boccherini are relatively sparse, and these occurrences are not in proximity to the relatively abundant occurrences of haydn.
  • the occurrences of haydn are relatively sparse, and these occurrences are not in proximity to the relatively abundant occurrences of boccherini.
  • the densities of haydn and boccherini remain the same for Ni and N 2 , and let the density of haydn ## boccherini be 0.08 for both Ni and N 2 .
  • haydn and boccherini are in greater proximity within N than they are within the first case. The na ⁇ ve direct method is unable to distinguish between these contrasting cases, and so is inadequate.
  • Embodiments of the present invention calculate pre-normalized r- values for Ei ## E 2 ... ## E m via the direct method, with corrections applied according to the indirect method, thereby incorporating the advantages of both methods without incurring their respective disadvantages. More precisely, given a sub-expression Ei ## E 2 ... ## E m of E, embodiments of the present invention calculate the pre-normalized r- value for Ei ## E 2 ... ## E m for child N' of N from the previously calculated densities and r-centers of Ei ## E 2 ...
  • E m for the children of N and from the previously calculated pre-normalized r-values for E 1 , ..., E m for the children of N, as follows: (i) let W 1 be A 1 * D 1 Z(I+ dj) x , where x is the distance attenuation exponent, where A 1 is the r-size of the i-th child of N, where D 1 is the density of Ei ## E 2 ... ## E m in the i-th child of N, and where U 1 is the r-distance between the r-center of N' and the r-center of Ei ## E 2 ... ## E m in the i-th child of N.
  • W 1J be A 1 * D 1J Z(I+ d lo ) x , where x is the distance attenuation exponent, where A 1 is the r-size of the i-th child of N, where D 1J is the density of E j in the i-th child of N, and where d Ui is the r-distance between the r-center of N' and the r-center of E j in the i-th child of N; (iii) let P be the product of the pre-normalized r-values of Ei, ..., E m for N'; (iv) the pre-normalized r-value assigned to N' for Ei ## E 2 ...
  • E m is (P + ⁇ i ⁇ i ⁇ kWi - ⁇ i ⁇ i ⁇ kdli ⁇ j ⁇ mWij))" m , where k is the number of children of N. Note that when P is expanded, fji ⁇ j ⁇ mW 10 corresponds to a capture that is not the most accurate capture of the weight of Ei ## E 2 ... ## E m at the i-th child of N, as this weight is felt at N'. (iv) in paragraph 0092 replaces this capture with the more accurate W 1 .
  • Embodiments of the present invention calculate the r-value for Ei ## E 2 ...
  • E m for child N' of N from the pre-normalized r-value v for Ei ## E 2 ... ## E m for N' as vZ( ⁇ i ⁇ i ⁇ k (A 1 Z(I+ di) x )), where k is the number of children of N, where x is the distance attenuation exponent, where A 1 is the r-size of the i-th child of N, and where U 1 is the minimum r-distance between constituents of N' and constituents of the i-th child of N.
  • embodiments of the present invention set the r-value for N' for Ei %% E 2 %% ... %% E m as the maximum of the r-values assigned to N' for E 1 ,..., E m .
  • embodiments of the present invention set the r- value for N' for ⁇ E as the one minus the r- value assigned to N' for E.
  • N is a Web page that presents information on composers of the classical period.
  • a central box contains a description of the musical characteristics of the classical style, without mentioning any composers.
  • the periphery of the page contains capsule biographies of various classical composers, including Haydn and Boccherini. No composer's biography mentions any other composer.
  • the r-distance between the central box and any composer biography is much less than the r-distance between any two composer biographies.
  • the central box receives non-zero r-values for haydn and for boccherini, and therefore receives a non-zero r- value for haydn ## boccherini. This constituent is therefore a candidate response to the search request haydn ## boccherini.
  • the relevance of the musical description to the search request is deduced entirely from the Web page that contains the musical description. No learning process over text corpora is involved.
  • the user, or information worker is provided with a means to explicitly formulate search requests that leverage cooccurrences of search terms. (Suppose that the central box in this example is replaced by an advertisement that is not a valid response to the search request haydn ## boccherini.
  • Most advertisements in Web pages can be identified by well-known earmarks in HTML code.
  • the relevance geometry of the Web page can be calculated so that the advertisement is placed at a suitably large r-distance from the composer biographies.
  • an orthogonal mechanism can exclude the advertisement as a response to the search request.
  • the same alternatives apply for other content whose relevance can be judged independently of how the content is placed in tree and layout structures.
  • the ("r-center") of matches for search expression E for a content node N is ( ⁇ i ⁇ ⁇ ⁇ n (A 1 * V 1 * P 1 ))/ ( ⁇ 1 ⁇ ⁇ ⁇ n (A 1 * V 1 )), where n is the number of children of N, where A 1 is the r-size of the i-th child of N, where V 1 is the r- value for E assigned to the i-th child of N, and where P 1 is the position of the r-center for E for the i- th child of N in the space corresponding to N.
  • the relevance geometry for the parent node may supply for each child node a function (the "displacement" function) whose inputs are a match value and a position within the region corresponding to the child node, and whose output is a displacement vector that applies to positions within the region corresponding to the parent node.
  • a position within a child node may be specified as a displacement vector applied to the r-center of the child node itself.
  • the relevance geometry for the parent node may be such that applications of displacement vectors are associative and commutative.
  • Embodiments of the present invention then capture the "relevance center" ("r-center") of matches for search expression E for a content node N by calculating the displacement vectors ⁇ (V 1 , P 1 ), where fi is the displacement function corresponding to the i-th child of N, where V 1 is the r- value for E assigned to the i-th child of N, and where P 1 is the position of the r-center of E in the i-th child of N, and then successively applying these displacement vectors, applying the first vector to the r-center of the parent node itself.
  • Embodiments of the present invention capture the density of matches for search expression E for a content node N as ( ⁇ i ⁇ i ⁇ n (A 1 * V 1 ))/ ( ⁇ i ⁇ j ⁇ n A 1 ), where n is the number of children of N, where A 1 is the r-size of the i-th child of N, and where V 1 is the r-value for E assigned to the i- th child of N.
  • G Distribution for parent content node Given that content node N has density D for search expression E, given that text child N' of N has density D' for E, and given that B' is the average absolute deviation from D' for the r- values assigned to the words of N' as matches for E, the average absolute deviation from D for the r-values assigned to the children of N' as matches for E cannot be deduced from D, D', and B' without additional information on how r-values for matches for E are distributed among the words of N'. For example, consider the case where D ⁇ D'. Suppose that ni words in N have r-values less than or equal to D.
  • n 2 words have r-values greater than D and less than or equal to D' and that the r-values of these n 2 words are X 1 , ..., Xn 2 .
  • n 3 words have r-values greater than D'. Then the average absolute deviation from D for the r-values assigned to the children of N' is B' + n 3 /n(D' - D) - ni/n (D' - D)+ 2/n* ⁇ i ⁇ , ⁇ ⁇ x, - n 2 /n(D' + D).
  • Various embodiments of the present invention pass varying degrees of detail concerning the distributions of r-values for search matches up the tree (which may be a parse tree, an enhanced parse tree, or a deduced semantic tree), or store varying degrees of detail concerning distributions of r-values on lower nodes in the tree, where this information can be accessed by re-traversing lower nodes when distributions are calculated for higher nodes.
  • the tree which may be a parse tree, an enhanced parse tree, or a deduced semantic tree
  • store varying degrees of detail concerning distributions of r-values on lower nodes in the tree where this information can be accessed by re-traversing lower nodes when distributions are calculated for higher nodes.
  • the tree which may be a parse tree, an enhanced parse tree, or a deduced semantic tree
  • store varying degrees of detail concerning distributions of r-values on lower nodes in the tree where this information can be accessed by re-travers
  • the distribution bands may be more or less finely grained.
  • distribution bands might be as follows: r- value less than .5 * density of parent, r- value greater than or equal to .5 * density of parent and less than density of parent, r- value greater than or equal to density of parent and less than 2 * density of parent, r- value greater than or equal to 2 * density of parent.
  • the following may be passed up from text node N' for use in processing N, the parent node of N' : the number of words of N' with r- values in the band.
  • preterminal content node N has density D for search expression E
  • child N' of N has density D' for E
  • the absolute average deviation from D' for the r- values assigned to the words of N' as matches for E embodiments of the present invention provide an exact or estimated "corrected average absolute deviation" from D for the r-values assigned to the words of N', as described in paragraphs 00102 - 00103.
  • Embodiments of the present invention calculate the "weighted corrected absolute average deviation" of matches for search expression E for preterminal content node N as ( ⁇ i ⁇ i ⁇ n (A 1 * C 1 ))/ ( ⁇ i ⁇ ⁇ ⁇ n A 1 ), where n is the number of children of N, where A 1 is the r-size of the i-th child of N, and where C 1 is the corrected average absolute deviation from D for the i-th child of N.
  • the "node deviation" of a preterminal content node N for search expression E is the weighted corrected absolute average deviation of N for E. Recall that the node deviation for a text node N is the absolute average deviation of the r-values of the words of N from the arithmetic mean of the r-values of the words of N.
  • Embodiments of the present invention provide exact or estimated corrected node deviations for nodes whose children include higher nodes than text nodes as described in paragraphs 00102, 00103, and 00105, except that for a child N' that is higher than a text node, in place of the average absolute deviation, the node deviation of N' is used.
  • embodiments of the present invention calculate the "weighted corrected node deviation" of matches for E for N as ( ⁇ i ⁇ i ⁇ n (A 1 * C 1 ))/ ( ⁇ i ⁇ i ⁇ n A 1 ), where n is the number of children of N, where A 1 is the r- size of the i-th child of N, and where C 1 is the corrected node deviation from D for the i-th child of N.
  • embodiments of the present invention pass up the "node deviation" for N, defined as this weighted corrected node deviation, for use in calculating the node deviation of the parent of N.
  • a high value for ⁇ indicates an even distribution of E in N.
  • a low value for ⁇ indicates an uneven distribution of E in N.
  • Embodiments of the present invention assign an overall score for content node N as a match for search expression E according the following formula: C 1 * D + C 2 * ⁇ , where D is the density of E on N, where ⁇ is the distribution score for E on N, and where cl and c2 are positive real numbers such that C 1 + C 2 ⁇ 1. Note that 0 ⁇ D ⁇ 1 and 0 ⁇ ⁇ ⁇ 1 , so 0 ⁇ (ci * D + C 2 * ⁇ ) ⁇ 1.
  • the values of C 1 and C 2 can be tuned as desired to adjust the relative importance of distribution and density in judging the relevance of N for E. Note that in ranking search results, properties in addition to D and ⁇ , such as layout size, may be taken into account.
  • each property P takes values between 0 and 1, and using a formula of the form ⁇ i ⁇ i ⁇ m (Ci * P 1 ), where there's a total of m properties and where ⁇ i ⁇ i ⁇ m (Ci) ⁇ 1, to calculate overall score.
  • Alternative embodiments of the present invention work as described above, except that they forego calculations of r- values for atomic search terms for words in strings, and/or forego calculations of r- values for proximity search expressions for words in strings, and/or forego calculations of r- values for search expressions for children of content nodes.
  • these embodiments work like paragraphs 0038 - 0051 above, with the difference that words that match an atomic search term are treated as though they were assigned value 1, while all other words are treated as though they were assigned value 0.
  • These alternative embodiments are somewhat simpler conceptually. They skip computation steps, with slight savings in computation times. However, by themselves these alternative embodiments don't capture evenness of match distributions.
  • a "maximal hitless sub-string" of S for E is a sub-string of S that (i) contains no occurrences of E and (ii) is not properly contained in a sub-string of S that contains no occurrences of E.
  • Other alternative embodiments of the present invention capture evenness of match distributions within strings in terms of the lengths of maximal hitless sub-strings. These alternative embodiments conflate distributions that the embodiments described in paragraphs 0046 - 0047 and paragraphs 00102 - 00107 are able to distinguish.
  • Embodiments of the present invention extend the methods described above to sets of documents.
  • sets of documents include the following: (1) a set of documents within a file directory, (2) the set of Web pages within a Web site, or within a well-defined sub-site of a Web site, (3) the set of documents obtained by starting with a Web page, adding the Web pages that this Web page links to (perhaps following only those links that belong to a certain category of link), adding the Web pages that those pages link to, and so on, with a bound placed on the size of the set of documents, or a bound placed on the length of the link- chain connecting members of the set to the initial page.
  • tree organization of documents in a file directory can correspond to the tree organization of the directory.
  • Tree organization of a set of Web pages obtained by following links can place directly linked-to documents as children of directly linking documents.
  • relevance geometry it may be considered that there's a fixed distance d such that for any position p in any Web page, and for any position p' in any sibling of this first Web page, the distance between p and p' is d.
  • relevance geometry derives from a particular two-dimensional layout that incorporates sibling Web pages.
  • occurrences of structural proximity conjunction in E may be replaced by occurrences of standard conjunction and occurrences of structural proximity disjunction may be replaced by occurrences of standard disjunction.
  • Sub-expressions in the scope of the structural proximity complement operator are deleted.
  • the resulting query E' is then submitted to an external search engine, which may be coupled with the full Web, or with a database of annotated cached documents, or with some other content source.
  • Structured search based on the original search expression E is then applied to the results returned by the external search engine, respecting any ordering suggested by the external search engine.
  • sub-expressions of E that lie in the scope of the structural proximity complement operator are deleted before submission to the external search engine, so that the external search engine will not miss content that includes sub- constituents that match deleted sub-expressions and also includes sub-constituents that match E. VIII. Complementary content selection criteria
  • a response to a user's search request may be influenced by criteria other than quality of search match.
  • Content constituents that render in smaller areas may be preferred to content constituents that render in larger areas, especially if the target device is small.
  • content constituents may be disfavored because they're too small. For example, a text constituent that consists solely of the word h ⁇ ydn is a poor candidate response to the search request h ⁇ ydn.
  • Embodiments of the present invention support explicit user requests for content that derives from a specific source or set of sources, including sets of sources that correspond to the results of previous content requests.
  • Embodiments of the present invention also support explicit user requests for content that belongs to a specified category.
  • the user may request product descriptions that match boccherini.
  • Categories of content may be characterized in terms of tree and/or layout structures. Such characterizations may be generic, or may be specific to particular content sources.
  • Web product descriptions may be characterized generically in terms of parsed and/or rendered HTML. Product descriptions may be similarly but much more narrowly characterized for a particular Web site. Characterizations of content categories may be stored in a category repository, as illustrated in Figure 1. Cached content may be annotated to reflect which sub-constituents belong to which categories.
  • Embodiments of the present invention also store information on the relevance geometry of constituents in category repositories.
  • Search requests may include additional predicates.
  • search requests may include predicates that specify string-based relationships, such as fixed-distance proximity relationships.
  • Search requests may also include predicates that specify tree-based relationships, such as predicates that specify node properties, and properties that specify inter-node relationships.
  • Tree-based relationships may refer to markup parse trees, to trees derived from string, markup, layout, and category information, or to other trees.
  • Search requests may also includes predicates that specify graph-based relationships more generally. Embodiments of the present invention use the && operator to interpret search requests that specify content categories, and to interpret search requests that include various predicates.
  • article matching "counterpoint” is interpreted as a request for N such that (N is an article) && (N matches "counterpoint”).
  • membership in content categories is evaluated as true ox false (although these embodiments are compatible with evaluating membership in content categories as a matter of degree).
  • the score for an article as a response to the search request article matching "counterpoint” is the same as score for the article as a match for counterpoint.
  • Search requests may also include explicit quantifiers, as in article that contains at least one captioned image, and as in article that contains exactly three captioned images. Search requests may also include predicates that correspond to specified search algorithms, such as the algorithms described in paragraphs 0038 - 00111 above. For example, article that contains captioned image may be interpreted analogously to article that matches
  • counterpoint As the number of occurrences of counterpoint within an article increases, as the distribution of occurrences becomes more even, and so on, the score for the article as a match for counterpoint increases. Similarly, according to the stated interpretation, as the number of captioned images within an article increases, as the distribution of captioned images within the article becomes more even, and so on, the score for the article as containing captioned images increases.
  • Search requests with Boolean and scalar-valued logical operators, with quantifiers, with predicates that specify string-based, tree-based, and graph-based relationships, with predicates that specify category membership, and with predicates that correspond to specified algorithms, may be embedded recursively.
  • Embodiments of the present invention apply the algorithms of paragraphs 0052 - 00111 to the resulting complex search requests. For example, consider article matching "counterpoint” that contains captioned image matching "haydn, " where contains is interpreted as described in paragraph 00121 above, and where && is used to interpret relative clauses.
  • the constituents of this search request (which may be represented as match(N2,”haydn"), captionedImage(N2), match(N2,”haydn”) && captionedImage(N2), article(Nl), match(Nl, "counterpoint”), contain(Nl, (N2
  • Embodiments of the invention use category information to optimize the evaluation. For example, when evaluating the search request article matching "counterpoint" that contains captioned image matching "haydn, " if a constituent is known not to be a possible sub-constituent of an article, then the constituent need not be evaluated as a match for counterpoint. For another example, if a constituent is known not to be a possible super- constituent of captioned images, then the constituent need not be evaluated as to its containment of captioned images matching haydn. It should be noted that complex search requests may be embedded under the structural proximity operator ##, as in (product description matching "suit”) ## (product description matching "tie”). Complex search requests may be annotated with indications of which constituents are to be returned.
  • a constituent that's assembled from unrelated smaller constituents, or assembled from marginally related smaller constituents is less readily perceived as integral.
  • an HTML table that includes a news article, lists of links to other articles, and advertisements may be very apparent visually when the containing page is rendered, but not readily perceived as an integral constituent.
  • Some integral constituents retain all or most of their perceived integrity when certain of their sub-constituents are expurgated. (Thus the term "constituent integrity" is based on an imperfect metaphor.)
  • the text of a news article may include an embedded advertisement that can be expurgated and delivered separately from the news article without reducing the total information conveyed.
  • constituent integrity For some simply specified constituent categories, characteristics related to constituent integrity are fairly immediate. For example, paragraphs readily combine to form higher integral constituents. In the absence of full information about constituent integrity, some general principles can be used to infer with more or less confidence which constituents are integral. For example, the lowest ancestor constituent of a header constituent that meets one of (or better yet both) of the following two tunable conditions is likely to be integral: (1) the total amount of text contained in the higher constituent is much greater than the total amount of text contained in the header; (2) the rendered area of the higher constituent is much greater than the rendered area of the header. If the header constituent is rendered at the top of the higher constituent, that increases the likelihood that the higher constituent is integral.
  • a constituent is unlikely to be integral if it contains a large number of similar complex integral constituents, where these contained constituents are not known to be possible sub-constituents of higher integral constituents.
  • Concerning candidates for expurgation, embodiments of the present invention allow a category repository to include annotations as to members of which distinguished constituent categories may be expurgated from members of which other distinguished constituent categories.
  • Embodiments of the present invention organize the output of a structured search into an annotated catalog of content constituents, as illustrated in Figure 1 ("content constituent catalog" 115).
  • Annotations may include indications of quality of search match, rendered size, constituent integrity, and degree to which a constituent corresponds to a user-specified constituent category.
  • Annotations may also include indications of which sub-constituents of a constituent returned by structured search are candidates for expurgation.
  • Embodiments of the present invention direct content constituent catalogs to a content selection engine that trades off among various content selection criteria to determine which content constituents to present in response to a user search request, and to determine the order in which content constituents will be presented.
  • the content selection engine communicates with a layout engine, as illustrated in Figure 1.
  • the layout engine places content constituents according to layout criteria, it may generate very specific requirements that it can communicate to the content selection engine. For example, a layout in progress may have room for a constituent of a specific size.
  • the layout engine can request a constituent of this size from the content selection engine, which can then supply the constituent of the required size that best satisfies the selection criteria in effect.
  • FIG. 2 a block diagram is shown illustrating the evaluation of a search expression E for a content constituent N, in accordance with various embodiments of the present invention.
  • a content node may be identified with the constituent that it dominates.
  • Content node and “content constituent” are therefore used interchangeably.
  • N is a terminal node 202
  • densities and relevance centers are calculated for N for all the sub-expressions of E 206.
  • Figure 3 illustrates in more detail these calculations of densities and relevance centers.
  • the node deviation for the root search expression E for N is then calculated 207.
  • Figure 4 illustrates in more detail this calculation of the node deviation for N for E.
  • the score for E for N is then calculated 208. In various embodiments, this score calculation proceeds as described in paragraphs 0051 and 00108 above.
  • FIG. 3 wherein a block diagram is shown illustrating the calculation of r- values of search expression E for children of content node N, and illustrating the calculation of density and relevance center for E for N from these r- values, in accordance with various embodiments of the present invention.
  • the subroutine corresponding to Figure 3 has not already been called all immediate sub-expressions of E 302, then this subroutine must first be called for all immediate sub-expressions of E 303. Thus in these embodiments, calculation proceeds recursively over the sub-expressions of E, from bottom to top.
  • r- values are calculated for E for children of N 304, 305, 307, 308, 309, 310, 311. IfN is a text node 304 and if E is an atomic search term 305, then r- values for the words of E are calculated from the occurrences of E 307. In various embodiments, these r-value calculations proceed as described in paragraphs 0038 - 0043 above. IfN is a text node 304 and if E is not an atomic search term 305, then r- values for the words of E are calculated from the r-values of E's immediate sub-expressions 308.
  • these r-value calculations proceed as described in paragraphs 0052 - 0059 above. IfN is not a text node 304 and if E is an atomic search term 309, then r-values for E for the children of N are calculated from densities and relevance centers for E for the children of N 310. In various embodiments, these r-value calculations proceed as described in paragraphs 0085 - 0098 above. IfN is not a text node 304 and if E is not an atomic search term 309, then r-values for E for the children of N are calculated from densities and relevance centers for E for the children of N 311.
  • these r-value calculations proceed as described in paragraphs 0085 - 0098 above.
  • N is a text node
  • the children of N are the words of N.
  • the density and relevance center for E for N is calculated from these r-values 312.
  • calculation of density and relevance center proceeds as described in paragraphs 0044 - 0045, paragraphs 0048 - 0050, and paragraphs 0099 - 00101 above.
  • Figure 4 wherein a block diagram is shown illustrating the calculation of the node deviation for search expression E for content node N, in accordance with various embodiments of the present invention.
  • N is a text node 402
  • the node deviation for E for N is calculated from the r- values assigned to the words of N 403. In various embodiments, this node deviation calculation proceeds as described in paragraphs 0046 - 0047 and 0071 above.
  • N is not a text node 402
  • node deviations of the children of N are corrected to reflect deviations from the density of N 404.
  • the node deviation of N is then calculated as the weighted average of the corrected node deviations of the children of N 405.
  • the calculation of corrected node deviations for the children of N, and the subsequent calculation of the node deviation for E for N proceed as described in paragraphs 00102 - 00107 above.
  • An "atomic search scoring function” takes as input an atomic search expression and a structure, and outputs a score corresponding to the degree to which the structure matches the search expression.
  • structures include strings of words, markup strings, trees corresponding to parsed markup, enhanced markup trees as described in paragraph 0074 above, deduced semantic trees as described in paragraph 0074 above, database records, and other database objects. Structures may be built recursively from lower structures (markup constituents from lower markup constituents, Web sites from markup documents, corpora of texts from text documents, XML repositories from XML documents, databases from records, and so on).
  • a “distance function" for a structure takes as input two sub-structures of the structure, each of which can at least hold one atomic search expression, and outputs a distance.
  • inter-word distance corresponds to a distance function.
  • a sub-structure is "atomic" if it can be fully occupied by an atomic search expression.
  • Two sub-structures So and Si of S are "similarly located” in S if the set of distances between So and other sub-structures of S is identical to the set of distances between Si and other substructures of S. (For example, in a four- word string with inter- word distance as the distance function, the first and fourth word positions are similarly located, and the second and third word positions are similarly located.
  • S comprises n atomic sub-structures, that there are no sub-structures of S except these atomic sub-structures, and that the distance function for S can be captured by equally spacing the n atomic substructures around a circle, with the distance between two atomic sub-structures corresponding to the Euclidean distance between the corresponding points on the circle. Then all the atomic sub-structures of S are similarly located.
  • an "effective move" of E in S exchanges the contents of So and S 1 .
  • An "effective move" for a set of atomic expressions Ei ,..., E n in S is an effective move of some E 1 (where 1 ⁇ i ⁇ n) that does not exchange an occurrence OfE 1 with an occurrence of some E j (where j ⁇ i and 1 ⁇ j ⁇ n).
  • E n in S is a sequence of effective moves for Ei, ... , E n in S, where no consecutive or non-consecutive sub-sequence of moves exchanges an occurrence of some E 1 (where 1 ⁇ i ⁇ n) with an occurrence of some E j (where j ⁇ i and 1 ⁇ j ⁇ n).
  • An effective move for Ei, ... , E n in S is considered to be special case of an effective tandem move E 1 , ... , E n in S.
  • an atomic search scoring function on the structure has "positional sensitivity" just in case an effective move of an atomic search expression E within the structure is guaranteed to change the score for E on S.
  • a “basic proximity search scoring function” takes as input two or more atomic search expressions and a structure, and outputs a score corresponding to the degree to which the atomic search expressions are in proximity within the structure.
  • a basic proximity search scoring function has "positional sensitivity" just in case an effective tandem move of atomic search expressions Ei , ... , E n in S is guaranteed to change the score for Ei, ... , E n on S.
  • a scoring function for a search expression language that doesn't include a proximity operator has "positional sensitivity” if the scoring function has positional sensitivity for the atomic search expressions within the language.
  • a scoring function for a search expression language that includes a binary or n-ary proximity operator has "positional sensitivity" if (i) the scoring function has positional sensitivity for the atomic search expressions within the language, and (ii) the scoring function has positional sensitivity for expressions in the language where the proximity operator is applied to atomic search expressions.
  • Search expressions associated with advertisements may include structural proximity operators, or additional operators and predicates as described in paragraphs 00118 - 00122 above.
  • the content constituent Given a content constituent, and given a set advertisements, each with one or more associated search expressions, the content constituent can be scored for each of these search expressions according to methods described in paragraphs 0038 - 00113 above.
  • the advertisements can then be ranked according to which advertisements have the best single associated search expression scores, or according to according to which advertisements have the best average associated search expression scores.
  • the content constituent can then be delivered together with the highest ranking advertisements, subject to space, size, and other constraints.
  • Various embodiments of the present invention use proximity relationships to rank advertisements for a given user search request E u and a given content constituent N.
  • an advertisement is associated with search expressions E a j, ..., E a , n .
  • prox is a proximity operator.
  • various embodiments identify the score of the advertisement for E u and N as the maximum of the scores over 1 ⁇ i ⁇ n of (E ajl prox E u ) on N, or an average of the scores over 1 ⁇ i ⁇ n of (E ail prox E u ) on N.
  • various embodiments identify the score of the advertisement for E u and N as the maximum of the scores over 1 ⁇ i ⁇ n of (E a1! ## E u ) on N, or an average of the scores over 1 ⁇ i ⁇ n of (E ajl ## E u ) on N, where ## is the structural proximity operator as described in paragraphs 0052 - 00111 above.
  • E ajl are atomic
  • E u comprises a sequence or set of atomic search expressions E u j, ..., E u , m
  • alternative embodiments identify the score of the advertisement for E u and N as the score on N of other expressions composed from E a j, ..., E a , n and E u j, ..., E u , m by applying proximity operators.
  • FIG. 5 illustrates an architecture view of a computing device 700, such as a desktop computer or a PDA, suitable for practicing the present invention in accordance with one embodiment.
  • Computing device 700 may be a server or a client. Whether as a server or client, computing device 700 may be coupled to clients or server via a wireless or wireline based interconnection, over one or more private and/or public networks, including the famous public network "Internet”.
  • computing device 700 includes elements found in conventional computing device, such as micro-controller/processor 702, digital signal processor (DSP) 704, non- volatile memory 706, display 708, input keys 710 (such as keypad, select button, D-unit), and transmit/receive (TX/RX) 712, coupled to each other via bus 714, which may be a single bus or an hierarchy of bridged buses.
  • non- volatile memory 706 includes operating logic 720 adapted to implement selected or all aspects of the earlier described content request engine 111, structured content search engine 114, content selection engine 116, and/or layout engine 117, in and of itself/themselves or as part of one or more larger components.
  • the various engines may be implemented on one or more computing systems.
  • the computing systems may be directly coupled, through Local and/or Wide Area Networks.
  • the implementation(s) may be via any one of a number programming languages, assembly, C, and so forth.
  • all or portions of the operating logic 720 may be implemented in hardware, firmware, or combination thereof.
  • Hardware implementations may be in the form of application specific integrated circuit (ASIC), reconfigured reconfigurable circuits (such as Field Programming Field Array (FPGA)), and so forth.
  • ASIC application specific integrated circuit
  • FPGA Field Programming Field Array

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

La présente invention concerne des procédés et des appareils pour rechercher des contenus, y compris une recherche structurée. Les modes de réalisation de la présente invention utilisent des structures arborescentes (ou, plus généralement, des structures graphiques), des structures de formatage et/ou des informations de catégorie de contenu pour prendre dans les résultats de recherche le contenu approprié qui, autrement, ne serait pas trouvé, pour réduire l'incidence des faux positifs dans les résultats de recherche et pour améliorer la précision des classements dans les résultats de recherche. Les modes de réalisation de la présente invention utilisent, en outre, des structures arborescentes (ou, plus généralement, des structures graphiques), des structures de formatage et/ou des informations de catégorie de contenu pour étendre les résultats de la recherche pour inclure des constituants de sous-document. Les modes de réalisation de la présente invention supportent également l'utilisation de propriétés de distribution comme critère pour classer les résultats de recherche. Les modes de réalisation de la présente invention supportent enfin une recherche basée sur la proximité structurelle, des expressions de recherche avec des opérateurs intégrés de manière récursive, des prédicats et/ou des quantificateurs et des applications pour sélectionner des publicités.
EP07798456A 2006-06-12 2007-06-12 Procédés et appareil pour rechercher un contenu Withdrawn EP2035972A4 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US81324606P 2006-06-12 2006-06-12
PCT/US2007/071026 WO2007146951A2 (fr) 2006-06-12 2007-06-12 Procédés et appareil pour rechercher un contenu

Publications (2)

Publication Number Publication Date
EP2035972A2 true EP2035972A2 (fr) 2009-03-18
EP2035972A4 EP2035972A4 (fr) 2011-06-15

Family

ID=38832793

Family Applications (1)

Application Number Title Priority Date Filing Date
EP07798456A Withdrawn EP2035972A4 (fr) 2006-06-12 2007-06-12 Procédés et appareil pour rechercher un contenu

Country Status (3)

Country Link
EP (1) EP2035972A4 (fr)
CN (1) CN101501688B (fr)
WO (1) WO2007146951A2 (fr)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9705888B2 (en) 2009-03-31 2017-07-11 Amazon Technologies, Inc. Managing security groups for data instances
US8713060B2 (en) 2009-03-31 2014-04-29 Amazon Technologies, Inc. Control service for relational data management
US8332365B2 (en) 2009-03-31 2012-12-11 Amazon Technologies, Inc. Cloning and recovery of data volumes
US9207984B2 (en) 2009-03-31 2015-12-08 Amazon Technologies, Inc. Monitoring and automatic scaling of data volumes
US8676753B2 (en) 2009-10-26 2014-03-18 Amazon Technologies, Inc. Monitoring of replicated data instances
US8074107B2 (en) 2009-10-26 2011-12-06 Amazon Technologies, Inc. Failover and recovery for replicated data instances
US9471693B2 (en) * 2013-05-29 2016-10-18 Microsoft Technology Licensing, Llc Location awareness using local semantic scoring
CN109101503A (zh) * 2017-06-20 2018-12-28 北京微影时代科技有限公司 一种创建组织结构层级关系树的方法及装置
CN110209663B (zh) * 2018-02-14 2023-06-20 阿里巴巴集团控股有限公司 搜索范围确定的方法、装置和存储介质

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1462952A1 (fr) * 2003-03-27 2004-09-29 Exalead Méthode pour l'affichage de résultats dans un moteur de recherche

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6285999B1 (en) * 1997-01-10 2001-09-04 The Board Of Trustees Of The Leland Stanford Junior University Method for node ranking in a linked database
US7761448B2 (en) * 2004-09-30 2010-07-20 Microsoft Corporation System and method for ranking search results using click distance
US7827181B2 (en) * 2004-09-30 2010-11-02 Microsoft Corporation Click distance determination

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1462952A1 (fr) * 2003-03-27 2004-09-29 Exalead Méthode pour l'affichage de résultats dans un moteur de recherche

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of WO2007146951A2 *

Also Published As

Publication number Publication date
CN101501688B (zh) 2013-07-24
EP2035972A4 (fr) 2011-06-15
WO2007146951A3 (fr) 2008-10-30
WO2007146951A2 (fr) 2007-12-21
CN101501688A (zh) 2009-08-05

Similar Documents

Publication Publication Date Title
US7987169B2 (en) Methods and apparatuses for searching content
US8489574B2 (en) Methods and apparatuses for searching content
US8140511B2 (en) Methods and apparatuses for searching content
US9047379B2 (en) Methods and apparatuses for searching content
US11036814B2 (en) Search engine that applies feedback from users to improve search results
US11803582B2 (en) Methods and apparatuses for content preparation and/or selection
Cohen et al. XSEarch: A semantic search engine for XML
WO2007146951A2 (fr) Procédés et appareil pour rechercher un contenu
RU2386997C2 (ru) Способ и система согласования схем баз данных web
US9192684B1 (en) Customization of search results for search queries received from third party sites
US10755179B2 (en) Methods and apparatus for identifying concepts corresponding to input information
US20110119262A1 (en) Method and System for Grouping Chunks Extracted from A Document, Highlighting the Location of A Document Chunk Within A Document, and Ranking Hyperlinks Within A Document
US20040064447A1 (en) System and method for management of synonymic searching
US20180300410A1 (en) Methods and apparatuses for searching content
EP1618503A2 (fr) Reseau notionnel
CA2637239A1 (fr) Systeme permettant d'effectuer une recherche
US10157222B2 (en) Methods and apparatuses for content preparation and/or selection
Bouras et al. Noun retrieval effect on text summarization and delivery of personalized news articles to the user’s desktop
Alghamdi et al. Extended user preference based weighted page ranking algorithm
Bjelland et al. Web link analysis: estimating document’s importance from its context
Wang et al. Data Crawling and Research Based on Topic Web Crawler
Jaiganesh et al. Enhancing Link Evaluation through a Coordinated Structure: A Comparative Analysis of PageRank and HITS Algorithms
Abdulrahman Web Pages Ranking Algorithms: A Survey
Kamali Querying Large Collections of Semistructured Data
Lee et al. Query based optimal web site clustering using simulated annealing

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

17P Request for examination filed

Effective date: 20090109

AK Designated contracting states

Kind code of ref document: A2

Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IS IT LI LT LU LV MC MT NL PL PT RO SE SI SK TR

AX Request for extension of the european patent

Extension state: AL BA HR MK RS

DAX Request for extension of the european patent (deleted)
A4 Supplementary search report drawn up and despatched

Effective date: 20110512

17Q First examination report despatched

Effective date: 20130909

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN

18D Application deemed to be withdrawn

Effective date: 20160713