US20050021545A1 - Very-large-scale automatic categorizer for Web content - Google Patents

Abstract

A method and apparatus for efficiently classifying and categorizing data objects such as electronic text, graphics, and audio based documents within very-large-scale hierarchical classification trees is provided. In accordance with one embodiment of the invention, a first node of a plurality of nodes of a subject hierarchy is selected. Previously classified data objects corresponding to a selected first node of a subject hierarchy as well as any associated sub-nodes of the selected node are aggregated to form a content class of data objects. Similarly, data objects corresponding to sibling nodes of the selected node and any associated sub-nodes of the sibling nodes are then aggregated to form an anti-content class of data objects. Features are then extracted from each of the content class of data objects and the anti-content class of data objects to facilitate characterization of said previously classified data objects.

Description

RELATED APPLICATIONS
• This application is a non-provisional application of the earlier filed provisional application No. 60/289,418, filed on May 7, 2001, and claims priority to the earlier filed '418s provisional application, whose specification is hereby fully incorporated by reference.
BACKGROUND OF THE INVENTION
• 1. Field of the Invention
• The invention relates to the field of data processing. More specifically, the invention relates to the automatic analysis of the content of electronic data objects and the categorization of the electronic data objects into one or more discrete categories.
• 2. Background Information
• The Internet consists of billions of discrete pages, which can be accessed from any browser-equipped computer or appliance connected to the World Wide Web (hereinafter “Web”). The availability of so many pages simultaneously represents both a boon and a bane to the user. Information, opinion, and news are available about a vast array of topics, but the challenge is to find those pages of the Web which are most relevant to the particular needs or desires of the user at any given moment.
• A number of search engines are available on the Web for free use. These search engines typically index some fraction of the pages available on the Web, and provide users with the ability to search for information on the Web using keywords or may not know how to correctly formulate a search query to find the most appropriate page(s).
• Another method of organizing the Web is the use of categorical hierarchies. Certain companies have analyzed the contents of tens or hundreds of thousands of web pages, placing each page into one or more of the categories in their particular subject hierarchy. Users can then browse such subject hierarchies, or search through them based upon keywords. Such searches provide results annotated with the subject area of the target page, which can assist the user in determining whether the page might be relevant to the actual topic being searched.
• FIG. 10 illustrates an exemplary priorart subject hierarchy1002 in which multiple decision nodes (hereinafter “nodes”)1030-1036 are hierarchically arranged into multiple parent and/or child nodes, each of which are associated with a unique subject category. For example,node1030 is a parent node tonodes1031 and1032, whilenodes1031 and1032 are child nodes tonode1030. Becausenodes1031 and1032 are both child nodes of the same node (e.g. node1030),nodes1031 and1032 are said to be siblings of one another. Additional sibling pairs insubject hierarchy1002 includenodes1033 and1034, as well asnodes1035 and1036. It can be seen fromFIG. 10 thatnode1030 forms afirst level1037 ofsubject hierarchy1002, while nodes1031-1032 form asecond level1038 ofsubject hierarchy1002, and nodes1033-1036 form athird level1039 ofsubject hierarchy1002. Additionally,node1030 is referred to as a root node ofsubject hierarchy1002 in that it is not a child of any other node.
• In general, search hierarchies are filled with pages by manual classification of individual web pages using the talents of experts in particular subject fields. This method has several problems, including the cost of finding experts to perform the classification, and the necessary backlog between the time a site is placed on the Web and the time (if ever) it enters the classification hierarchy, moreover a grader expert in one subject area may misclassify a page of another subject, which can make the page more difficult to find for the casual browser.
• Although this is an active area of research, existing systems typically work with only a limited number of subject fields and often display poor performance. Therefore, what is desired is an automatic system for classifying a large number of documents quickly and effectively into a large subject hierarchy.
BRIEF DESCRIPTION OF DRAWINGS
• The present invention will be described by way of exemplary embodiments, but not limitations, illustrated in the accompanying drawings in which like references denote similar elements, and in which:
• FIG. 1 illustrates an overview of the present invention, in accordance with one embodiment;
• FIG. 2 illustrates a functional block diagram of the classifier training services ofFIG. 1, in accordance with one embodiment of the invention;
• FIG. 3 illustrates the operational flow of relevant aspects of the classifier training services ofFIG. 1, in accordance with one embodiment of the invention;
• FIG. 4 illustrates the operational flow of relevant aspects of the feature extraction function ofFIG. 1, in accordance with the teachings of the present invention;
• FIG. 5 illustrates an exemplary N-gram feature extraction list obtained in accordance with one embodiment of the invention;
• FIG. 6 illustrates an operational flow of certain aspects of salient feature selection function ofFIG. 1, in accordance with one embodiment of the invention;
• FIG. 7 illustrates a graphical representation of one embodiment of the salient feature selection function described with respect toFIG. 6;
• FIGS. 8A and 8B together illustrate an operational flow of certain aspects of the classifier categorization services ofFIG. 1, in accordance with one embodiment of the invention;
• FIG. 9 illustrates an example computer system suitable for use to practice the present invention, in accordance with one embodiment; and
• FIG. 10 illustrates an exemplary prior art subject hierarchy including multiple decision nodes.
DETAILED DESCRIPTION OF THE INVENTION
• In the following description, various aspects of the present invention will be described. However, it will be apparent to those skilled in the art that the present invention may be practiced with only some or all aspects of the present invention. For purposes of explanation, specific numbers, materials and configurations are set forth in order to provide a thorough understanding of the present invention. However, it will also be apparent to one skilled in the art that the present invention may be practiced without the specific details. In other instances, well known features are omitted or simplified in order not to obscure the present invention.
• Parts of the description will be presented in terms of operations performed by a processor based device, using terms such as data, storing, selecting, determining, calculating, and the like, consistent with the manner commonly employed by those skilled in the art to convey the substance of their work to others skilled in the art. As well understood by those skilled in the art, the quantities take the form of electrical, magnetic, or optical signals capable of being stored, transferred, combined, and otherwise manipulated through mechanical and electrical components of the processor based device; and the term processor includes microprocessors, micro-controllers, digital signal processors, and the like, that are standalone, adjunct or embedded.
• Various operations will be described as multiple discrete steps in turn, in a manner that is most helpful in understanding the present invention, however, the order of description should not be construed as to imply that these operations are necessarily order dependent. In particular, these operations need not be performed in the order of presentation. Further, the description repeatedly uses the phrase “in one embodiment”, which ordinarily does not refer to the same embodiment, although it may.
Overview
• Referring first toFIG. 1, wherein a diagram illustrating an overview of the present invention, in accordance with one embodiment, is shown. As illustrated,classifier100 is provided to efficiently classify and categorize data objects such as electronic documents including but not limited to text files, image files, audio sequences and video sequences comprising both proprietary and non-proprietary formats, within very-large-scale hierarchical classification trees.
• Classifier100 includesclassifier training services105 for extracting classification rules from a previously categorized data hierarchy and fortraining classifier100 to categorize new data objects based upon those rules.Classifier training services105 includesaggregation function106,feature extraction function107, salientfeature selection function108, andnode characterization function109. In accordance with the illustrated embodiment, content from the previously categorized data hierarchy is aggregated at each node in the hierarchy, throughaggregation function106 for example, to form both content and anti-content groups of data. Features from each of these groups of data are extracted byfeature extraction function107 for example, and a subset of those features are determined to be salient and are selected by way of salientfeature selection function108 for example.Node characterization function109 is utilized to characterize each node of the previously categorized data hierarchy based upon the selected salient features, and to store such hierarchical characterizations indata store110 for example, for further use byclassifier categorization services115.
• Onceclassifier100 is trained,classifier categorization services115 facilitate categorization of new data objects input intoclassifier100.Classifier categorization services115 includenode selection function116 for successively selecting the hierarchical characterizations stored e.g. indata store110, for comparison against new data objects;classification function117 for comparing new data objects input intoclassifier100 against one or more of the selected node characterizations, and for generating a confidence rating (i.e. score) based upon each comparison; anddiscrimination function118 for categorizing the new data objects based upon the determined confidence ratings.
• Althoughclassifier training services105 includingaggregation function106,feature extraction function107, salientfeature selection function108,node characterization function109, as well asclassifier categorization services115 includingnode selection function116,classification function117, anddiscrimination function118 are each illustrated as separate elements, it will be readily apparent from the description to follow, that the present invention may be practiced with some or all of these elements combined or further subdivided.
Classifier Training
• FIG. 2 illustrates a functional block diagram ofclassifier training services105 ofFIG. 1, in accordance with one embodiment of the invention. As illustrated inFIG. 2, previously categorizeddata hierarchy202 is provided for input intoclassifier training services105 ofclassifier100. Previously categorizeddata hierarchy202 represents a set of electronic documents which have been previously classified and categorized into a subject hierarchy (typically through manual entry by individuals). For example, previously categorizeddata hierarchy202 may represent one or more sets of electronic documents previously categorized by a Web portal or search engine.
• Aggregation function206 aggregates content so as to increase differentiation between sibling nodes at each level of the hierarchy. In one embodiment,aggregation function206 aggregates content from previously categorizeddata hierarchy202 into content and anti-content data groups. In one embodiment, content aggregated from a selected node and all sub-nodes of the selected node is used to form the content data group, whereas content aggregated from the siblings of the selected node and all sub-nodes of such sibling nodes is used to form the anti-content data group. In other embodiments, content and anti-content data groups may each be based upon any combination of nodes within the hierarchy.
• Feature extraction function207 operates to extract features from the content and anti-content groups of data, while salientfeature selection function208 operates in the form of a filter to determine which of the extracted features is to be considered salient and to distinguish such salient features (209′) from the larger group of extracted features (209).
• Additionally,node characterization function109 ofFIG. 1 operates to characterize the content and anti-content groups of data. In one embodiment, the content and anti-content groups of data are characterized based upon the determined salient features. In one embodiment, the characterizations are stored indata store110, which can be implemented in the form of any number of data structures such as a database, a directory structure, or a simple lookup table. In one embodiment of the invention, the parameters for the classifiers for each node are stored in a hierarchical categorization tree having a file structure that mimics the previously categorized data hierarchy. However, the parameters for the classifiers need not be stored in a hierarchical structure.
Classifier Training Method
• In accordance with one embodiment of the present invention, the construction of a hierarchical categorization tree, through the training ofclassifier100 for example, is accomplished by making decisions at each node of the data hierarchy on a one-to-one basis. In another embodiment, decisions at each node of the data hierarchy are made in a one-to-many basis. That is, at each node of the hierarchy, the content of one or more documents corresponding to one sub node is compared to document content corresponding to all the other sub nodes of that node, rather than comparing document content of one sub node to document content of a single other sub node, or comparing document content of one sub node to all the documents in the entire tree. Accordingly, this method advantageously avoids adding undesirable “noise” to the decision process.
• FIG. 3 illustrates the operational flow of certain aspects ofclassifier training services105 ofFIG. 1, in accordance with one embodiment of the invention. Beginning atblock302, a first node of a previously categorized data hierarchy is selected. In one embodiment, the selected node is a root node, whereas in other embodiments, the selected node may be a child node. Once a node is selected, previously categorized content corresponding to the selected node and any sub-nodes (i.e. child nodes) is aggregated to form what is referred to as a content class of data, block304. Similarly, previously categorized content corresponding to non-content nodes are aggregated to form an anti-content class of data, block306. In one embodiment, the sibling nodes of the selected node (i.e., all other child nodes of the selected node's parent), as well as any sub-nodes of these sibling nodes are aggregated to form an anti-content class of data. Once the content and anti-content classes of data have been formed, feature sets are created from each respective class of data (e.g., content and anti-content) to facilitate characterization of the previously categorized content, block308, and features determined to be most salient are then identified from the content and anti-content feature sets, block310. In one embodiment, the feature sets are N-gram based feature sets.
• A document typically consists of a sequence of words, where the term “word” is commonly understood to represent a group of letters within a language, having some semantic meaning. In the present invention, the term “feature” is used to describe any group of alphanumeric characters, which may or may not form a whole or part of a word. More generally, a “word” or “feature” could be an N-token gram, where a token is one atomic element of a language. This would include N-letter grams and N-word grams in English, as well as N-ideogram grams in Asian languages such as Japanese.
• In various embodiments of the invention, feature extraction includes preprocessing of the content and anti-content data classes of the data object. Such preprocessing may include removal of formatting, markup and/or punctuation characters for example.FIG. 4 illustrates one embodiment of an operational flow of the feature extraction function ofFIG. 1 including preprocessing. In accordance with the illustrated embodiment, as part of the preprocessing of the data object, markup language specific tags, such as hypertext markup language (HTML) tags, standard markup language (SGML) tags, extensible markup language (XML) tags, and so forth are removed from the data object to be categorized, block402. Additionally, uppercase alphabetic characters are converted to lower case characters, certain selected punctuation such as the dollar sign (“$”) is eliminated, and redundant white space is removed from the data object, blocks404,406, and408 respectively. It should be noted that variations to the above described preprocessing are entirely possible and contemplated. For example, all white spaces could be removed from the data object instead of only redundant white spaces being removed, punctuation in addition to the dollar sign may be eliminated, or no punctuation need be eliminated, and so forth.
• Once the data object has been prepared through preprocessing (according to whether preprocessing has been deemed desirable), a list of unique character strings existing in the data object is constructed including a count of the number of occurrences of each data string, block410. In one embodiment of the invention, the character strings are N-gram based, although other methods of determining character strings may be utilized. N-grams are sequences of N consecutive characters in a document, for example, that can be generated by “sliding a window” N characters wide across the document's text one character at a time. In one embodiment of the invention, a list of unique 3-character, 4-character, and 5-character N-gram based strings is constructed. The respective strings may be combined into a single list or they each may be maintained in a separate list.
• FIG. 5 illustrates an exemplary N-gram feature extraction list obtained in accordance with one embodiment of the invention in which N=3. For explanatory purposes, sentence510 (e.g., ‘Submarine magazine “marches” to maritime market.’) is intended to represent a character string from which a feature extraction list of N-grams is to be constructed. Before the features are extracted fromsentence510, however, the sentence is preprocessed by converting uppercase characters (502) to lowercase, and removing punctuation (504 and508) and redundant white space (506) resulting incharacter string512. Fromcharacter string512, a list of all unique 3-character strings514 is created, along with the number of occurrences516 (i.e., frequency) of each unique 3-character string514, and the resultingrank518 of each unique 3-character string514 based upon each respective string'srank518. Sliding a 3-character wide window one character at a time acrosscharacter string512, results in the construction of a list of thirty-four unique 3-chracter strings each having a frequency of occurrence ranging from one to four.
• It should be noted thatsentence510 is presented for illustrative purposes only, as it constitutes only a small portion of content that a typical data object might contain. Accordingly, the number and frequency of 3-grams obtained fromsentence510 may not necessarily be typical of most data objects. More likely, the data objects will contain a far greater number of N-grams having greater frequencies than that shown. Notwithstanding this,sentence510 is nonetheless useful in illustrating certain principles of the present invention.
Salient Feature Selection
• As mentioned above in accordance with one embodiment of the invention, salient features are determined and selected from the extracted features, which have been ranked based upon each feature's number of occurrences (i.e. frequency) in the data object.FIG. 6 illustrates an operational flow of relevant aspects of salientfeature selection function108 ofFIG. 1, in accordance with one embodiment of the invention. To begin the salient feature selection process, a feature is randomly selected from the list of extracted features, block602, and a determination is then made as to whether the frequency of the selected feature is equal to the associated rank of the selected feature, block604. If the frequency of the selected feature is equal to the associated rank of the selected feature, the selected feature is designated as a “corner” feature, block606. Once the corner feature is established, a first set of features having a higher frequency of occurrence than that of the corner feature are identified and a second set of features having a lower frequency of occurrence than that of the corner feature are identified. In one embodiment, features having a higher frequency than that of the corner feature are accepted into the first set of features until the cumulative frequency of the first set of features equals twenty percent of the cumulative frequency of the features having a lower frequency than that of the corner feature, block608. Likewise, in one embodiment, features having a lower frequency than that of the corner feature are accepted into the second set of features until the cumulative frequency of the second set of features equals eighty percent of the cumulative frequency of the features having a higher frequency than that of the corner feature, block610. In one embodiment, the features included within the first and second sets of features are determined to be salient features.
• If, however atblock604, it is determined that the frequency of the selected feature is not equal to the associated rank of the selected feature, a further determination is made as to whether the frequency of the selected feature is greater than the associated rank of the selected feature, block614. If it is determined that the frequency of the selected feature is not greater than the associated rank of the selected feature (i.e., the frequency is less than the associated rank), then a new feature is chosen having either a higher frequency or higher rank than the originally selected feature, block612. Once a new feature is chosen, a new determination is made as to whether the frequency of the newly selected feature is equal to the associated rank of the newly selected feature, block604.
• If it is determined (i.e. at block614) that the frequency of the selected feature is greater than the associated rank of the selected feature, then a new feature is selected from the list of features having the next lower frequency than the frequency of the previously selected feature, block616. In many cases this new feature having the next lower frequency may be obtained by subtracting one from the previous frequency. Once a new feature having a lower frequency is obtained, a further determination is made as to whether the frequency of the new feature remains greater than the associated rank of the new feature, block618. If the frequency of the new feature is still greater than the associated rank of the new feature, another new feature is selected having an even lower frequency, block616. This procedure is repeated until the frequency of the feature becomes equal to or less than the associated rank of the feature. If the frequency of the feature becomes equal to the associated rank of the feature, block622, then that feature is then designated as a corner feature, and a first and second set of features are selected as described above with respect toblocks606,608, and610. If, however, the frequency of the feature becomes greater than the associated rank of the feature, the previously selected feature is once again selected, block622. The newly selected feature is then designated as a corner feature, and a first and second set of features are selected as described above with respect toblocks606,608, and610.
• In one embodiment, the number of features included within the first and/or second sets of features is limited by an absolute number rather than by a percentage of total features. For example, the first set of features may be limited to include features with a cumulative frequency equaling either twenty percent of the cumulative frequency of the features having a lower frequency than that of the corner feature or up to a total of 5000 features, whichever comes first. Similarly, both first and second sets of features taken together may be limited to include a total of 5000 features for instance, regardless of the respective percentages.
• FIG. 7 illustrates a graphical representation of one embodiment of the salient feature selection process described above with respect toFIG. 6. InFIG. 7, a plot of feature frequency values versus feature rank values is shown. In the plot, the selected “corner” node is indicated, as is the first set of features located to the left of the corner node, representing twenty percent of the cumulative frequencies to the right of the corner node, and the second set of features located to the right of the “corner” node, representing eighty percent of the cumulative frequencies to the left of the corner node as described above. This 80-20 operating rule (i.e. 80% of the cumulative frequencies above corner node and 20% of the cumulative frequencies below the corner node) is derived from the notion that features in the middle of the frequency distribution (e.g., those features corresponding to the 80-20 rule) are most representative of the content of a data object such as an electronic document. Often the features at higher frequencies are representative of the language of the document, whereas features at lower frequencies are representative of unique, rare, or misspelled words in the document.
• In an alternative embodiment, salient features may also be determined and selected from the content and anti-content classes of data by eliminating a certain number of the most frequent features. In one embodiment, the number of features eliminated is dependent upon the level of the corresponding node. For example, the 400 most frequent features may be eliminated for top-level nodes in the search hierarchy, whereas 150 most frequent features may be eliminated for the second level nodes, and 10 most frequent features may be eliminated for the third and lower level nodes of the hierarchy. Such an adjustment is useful because data objects found at the top level nodes of the hierarchy are characterized by more irrelevant N-grams for example, with a broader selection of keywords and themes encountered in the agglomerated subordinate nodes, whereas the data objects at lower level nodes are more similar to one another and thus exhibit more frequent use of relevant terms. In other embodiments, salient features may be selected by way of information theoretics including mutual information.
• Further methods of determining salient features are described e.g. in copending, contemporaneously filed U.S. Patent Application S/N<<51026.P007>>, titled “Method and Apparatus for Automatically Determining Salient Features for Object Classification”, and having a common assignee as the present application.
Classification/Categorization
• Onceclassifier100 has been trained by the methods described above, for example,classifier100 can be utilized to classify and categorize a newly presented data object. In addition to the training methods described above, several other classifier training techniques known in the art may also be utilized to trainclassifier100. Such training techniques known in the art include Trenkle's ad hoc rank analysis, Spearman's rank correlation coefficient, centroid analysis, vector dot product, Euclidean distance, and a support vector machine.
• FIGS. 8aand8btogether illustrate an operational flow of relevant aspects ofclassifier categorization services115 includingnode selection function116,classification function117, anddiscrimination function118 ofFIG. 1, in accordance with one embodiment of the invention. Each data object to be categorized is examined by each of the classifiers (i.e. stored characterizations) for the child nodes at given level and depending upon the outcome, is either categorized as belonging to the subject associated with the selected node or is further examined at the next level of the hierarchy.
• The categorization process begins by selecting the root node of the hierarchy to be the starting point, block802. In other embodiments, nodes other than the root node may be identified as the starting point. Once the starting point is selected, a determination is made as to whether the selected node is the parent of at least one child node, block804. If it is determined that the selected node is not the parent of at least one child node, the data object is categorized in association with the subject matter of the selected node, block806. If, however, it is determined that the selected root node is the parent of at least one child node, a first of the child nodes is then selected, block808. The data object is then classified using the classifier associated with the selected node, block810.
• A determination is then made as to whether the selected node has at least one sibling node, block812. As mentioned earlier, a first node is considered to be a sibling of a second node if the first and second nodes share a common parent node. If it is determined that the selected node has at least one sibling node, one of such sibling nodes is then selected and the data object is classified using the classifier associated with the newly selected sibling node, blocks814 and816 respectively. Each classification performed in an attempt to categorize the data object results in the generation of a confidence rating which is stored in association with each respective node, in a data store such asdata store110 for example. Once the data object is classified by a first (and possibly only) sibling node, a determination is again made as to whether any additional sibling nodes remain, block812. Each additional sibling node is selected, for classifying the data object resulting in a confidence rating for each sibling node. This process continues until no further sibling nodes exist, block812.
• Once it is determined that no further sibling nodes exist, a determination is then made as to whether the confidence rating of at least one of the classified child/sibling nodes exceeds a minimum threshold, block818. The minimum threshold may be set so as to depend upon various aspects of the categorization hierarchy. For example, if the categorization hierarchy contains many categorization levels, the minimum threshold may be set relatively high so as to encourage a more precise categorization, whereas if the categorization hierarchy contains relatively few categorization levels, the minimum threshold may be set relatively low so as to encourage any categorization.
• If it is determined that no child or sibling node exceeds the minimum threshold, a further determination is made as to whether the parent node of the selected child/sibling node is a root node, block820. If the parent node of the selected child/sibling node is a root node and the child/sibling node exceeds the established minimum threshold, the data object is unable to be categorized and is assigned to a default “unknown” category, block824. If, however, no child/sibling node exceeds the established minimum threshold and the parent node of the selected child/sibling node is not a root node, the data object is assigned to the subject category associated with the parent node, block822.
• If atblock818 it is determined that the confidence rating of at least one child/sibling node exceeds the established minimum threshold, then a secondary metric is calculated and applied to the child/sibling node(s) exceeding the minimum threshold. In the illustrated embodiment, the mean and standard deviation of the previously stored confidence ratings are calculated, block826.
• Next, comparisons between each of the previously obtained confidence ratings and the calculated mean and standard deviation are made. For each child/sibling confidence rating that exceeds the established minimum threshold, but does not exceed the value represented by mean minus the standard deviation, then the data object is categorized with respect to the selected node category, block828. On the other hand, for each child/sibling confidence rating that exceeds both the established minimum threshold and the value represented by the mean minus the standard deviation, then the process described above is repeated, block830. That is, each child/sibling node meeting the criteria established inblock830, is selected in parallel and a further determination is made as to whether that child/sibling node is a parent of at least one child node itself, block804. This process is repeated in parallel for each node meeting the established criteria until a leaf node is reached or the data object is able to be categorized, whichever comes first.
Example Computer System
• FIG. 9 illustrates an example computer system suitable for use in association with the categorization system of the present invention, in accordance with one embodiment. As shown,computer system900 includes one ormore processors902 andsystem memory904. Additionally,computer system900 includes mass storage devices906 (such as diskette, hard drive, CDROM and so forth), input/output devices908 (such as keyboard, cursor control and so forth) and communication interfaces910 (such as network interface cards, modems and so forth). The elements are coupled to each other viasystem bus912, which represents one or more buses. In the case wheresystem bus912 represents multiple buses, they are bridged by one or more bus bridges (not shown).
• Each of these elements performs its conventional functions known in the art. In particular,system memory904 andmass storage906 are employed to store a working copy and a permanent copy of the programming instructions implementing the categorization system of the present invention. The permanent copy of the programming instructions may be loaded intomass storage906 in the factory or in the field, as described earlier, through a distribution medium (not shown), or throughcommunication interface910 from a distribution server (not shown). The constitution of these elements902-912 are known, and accordingly will not be further described.
Conclusion and Epilogue
• Thus, it can be seen from the above descriptions, a novel method and apparatus for very-large-scale automatic categorization of web content has been described. The novel method/apparatus is advantageously scalable to support an ever-increasing amount of data objects requiring categorization.
• While the present invention has been described in terms of the above-illustrated embodiments, those skilled in the art will recognize that the invention is not limited to the embodiments described. The present invention can be practiced with modification and alteration within the spirit and scope of the appended claims. The description is thus to be regarded as illustrative instead of restrictive on the present invention.

Claims (24)

1. A method of training a classifier system by utilizing previously classified data objects comprising one or more electronic documents and organized into a subject hierarchy of a plurality of nodes, the method comprising:

selecting one node of the plurality of nodes;

aggregating those of the previously classified data objects corresponding to the selected node and any associated sub-nodes of the selected node, to form a content class of data objects, said content class of data objects comprising a content class of the one or more electronic documents;

aggregating those of the previously classified data objects corresponding to any associated sibling nodes of the selected node and any associated sub-nodes of the sibling nodes to form an anti-content class of data objects, said anti-content class of data objects comprising an anti-content class of the one or more electronic documents; and

extracting features from at least one of the content class of data objects and the anti-content class of data objects to facilitate characterization of said previously classified data objects.

2.-11. (canceled)

12. The method ofclaim 1, wherein said one or more electronic documents comprise at least one of a text document, an image file, an audio sequence, a video sequence, and a hybrid document including a combination of text and images.

13. A method of classifying a data object, the method comprising:

selecting a first node of a hierarchically organized classifier having a plurality of nodes;

determining if the first node of said plurality of nodes is the parent of one or more child nodes;

upon determining that said first node is the parent of one or more child nodes, selecting a first of said one or more child nodes and classifying said data object at the first of said one or more child nodes to produce a confidence rating, said data object comprising an electronic document;

recursively selecting each of said one or more child nodes that remain and classifying the data object at each selected one or more child nodes to respectively produce a confidence rating for each selected one or more child nodes; and

assigning the data object to each node of said plurality of nodes having produced an acceptable confidence rating.

14. The method ofclaim 13, wherein the first node is a root node.

15. The method ofclaim 13, wherein said acceptable confidence rating comprises a confidence rating that exceeds a minimum threshold.

16. The method ofclaim 13, further comprising:

assigning the data object to the first node if the first node is the parent of said one or more child nodes and none of said one or more child nodes producing a confidence rating that exceeds the minimum threshold.

17. The method ofclaim 13, wherein if said first node is a root node, then categorizing the data object as undefined.

18. The method ofclaim 13, further comprising:

determining a mean and standard deviation of the confidence ratings of the one or more child nodes.

19. The method ofclaim 18, wherein the data object is assigned to only those of the plurality of nodes having an associated confidence rating that exceeds the mean minus the standard deviation.

20. The method ofclaim 13, further comprising:

determining if at least one of said child nodes producing an acceptable confidence rating is a parent of one or more additional child nodes; and

upon determining that at least one of said child nodes producing an acceptable confidence rating is a parent of said one or more additional child nodes, successively selecting and classifying each of said additional child nodes.

21. (canceled)

22. The method ofclaim 13, wherein said electronic document comprises at least one of a text document, an image file, an audio sequence, a video sequence, and a hybrid document including a combination of text and images.

23.-34. (canceled)

35. An apparatus comprising:

a storage medium having stored therein a plurality of programming instructions designed to implement a plurality of functions of a category name service for providing a category name to a data object, including first one or more functions to

select a first node of a hierarchically organized classifier having a plurality of nodes,

determine if the first node of said plurality of nodes is a parent of one or more child nodes,

select a first of said one or more child nodes and classify said data object at the first of said one or more child nodes to produce a confidence rating if said first node is the parent of one or more child nodes,

select each of said one or more child nodes that remain and classify the data object at each selected one or more child nodes to respectively produce a confidence rating for each selected one or more child nodes,

assign the data object to each node of said plurality of nodes having produced an acceptable confidence rating; and

a processor coupled to the storage medium to execute the programming instructions.

36. The apparatus ofclaim 35, wherein the first node is a root node.

37. The apparatus ofclaim 35, wherein said acceptable confidence rating comprises a confidence rating that exceeds a minimum threshold.

38. The apparatus ofclaim 35, wherein said plurality of programming instructions further comprises instructions to

assign the data object to the first node if the first node is the parent of said one or more child nodes and none of said one or more child nodes produces a confidence rating that exceeds the minimum threshold.

39. The apparatus ofclaim 35, wherein if said first node is a root node, then said data object is categorized as undefined.

40. The apparatus ofclaim 35, wherein said plurality of instructions further determine a mean and standard deviation of the confidence ratings of the one or more child nodes.

41. The apparatus ofclaim 40, wherein the data object is assigned to only those of the plurality of nodes having an associated confidence rating that exceeds the mean minus the standard deviation.

42. The apparatus ofclaim 35, wherein said plurality of programming instructions further comprises instructions to

determine if at least one of said child nodes producing an acceptable confidence rating is a parent of one or more additional child nodes; and

successively select and classify each of said additional child nodes, if it is determined that at least of said child nodes producing an acceptable confidence rating is a parent of said one of more additional child nodes.

43. The apparatus ofclaim 35, wherein said data object comprises an electronic document.

44. The apparatus ofclaim 43, wherein said electronic document comprises at least one of the text document, an image file, an audio sequence, a video sequence, and a hybrid document including a combination of text and images.

Images