PatentDe  


Dokumentenidentifikation EP1840772 15.11.2007
EP-Veröffentlichungsnummer 0001840772
Titel System und Verfahren zur Vermeidung der Übermittlung unerwünschter Informationen
Anmelder Xerox Corp., Rochester, N.Y., US
Erfinder Guerraz, Agnes, 38450, Le Gua, FR;
Privault, Caroline, 38330, Montbonnot, FR;
Goutte, Cyril, Ottawa Ontario K1S 0P9, CA;
Gaussier, Eric, 38320, Eybens, FR;
Pacull, Francois, 38920, Crolles, FR;
Renders, Jean-Michel, 38330, Saint-nazaire-les-eymes, FR
Vertreter derzeit kein Vertreter bestellt
Vertragsstaaten 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
Sprache des Dokument EN
EP-Anmeldetag 27.03.2007
EP-Aktenzeichen 071050116
EP-Offenlegungsdatum 03.10.2007
Veröffentlichungstag im Patentblatt 15.11.2007
IPC-Hauptklasse G06F 17/30(2006.01)A, F, I, 20070904, B, H, EP

Beschreibung[en]
BACKGROUND

The following relates to the document processing arts. It is described with example reference to embodiments employing probabilistic hierarchical clustering in which documents are represented by a bag-of-words format. However, the following is also applicable to non-hierarchical probabilistic clustering, to other types of clustering, and so forth.

In typical clustering systems, a set of documents is processed by a training algorithm that classifies the documents into various classes based on document similarities and differences. For example, in one approach the documents are represented by a bag-of-words format in which counts are stored for keywords or for words other than certain frequent and typically semantically uninteresting stop words (such as "the", "an", "and", or so forth). Document similarities and differences are measured in terms of the word counts, ratios, or frequencies, and the training partitions documents into various classes based on such similarities and differences. The training further generates probabilistic model parameters indicative of word counts, ratios, or frequencies characterizing the classes. For example, a ratio of the count of each word in the documents of a class respective to the total count of words in the documents of the class provides a word probability or word frequency modeling parameter. Optionally, the classes are organized into a hierarchy of classes, in which the documents are associated with leaf classes and ancestor classes identify or associate semantically or logically related groupings of leaf classes. Once the training is complete, the clustering system can be used to provide a convenient and intuitive interface for user access to the clustered documents.

A problem arises, however, in that the classification system generated by the initial cluster training is generally static. The probabilistic modeling parameters are computed during the initial training based on counting numbers of words in documents and classes. If a document in the clustering system is moved from one class to another class, or if a class is split or existing classes are merged, or so forth, then the probabilistic modeling parameters computed during the training are no longer accurate.

To maintain up-to-date probabilistic modeling parameters, the clustering system can be retrained after each update (such as after each document or class move, after each class split or merge, or so forth). However, a large clustering system may contain tens or hundreds of thousands of documents, or more, with each document containing thousands, tens of thousands, or more words. Accordingly, re-training of the clustering system is typically a relatively slow proposition. For document bases of tens of hundreds of documents each including thousands or tens of thousands of words, retraining can take several minutes or longer. Such long time frames are not conducive to performing real-time updates of the hierarchy of classes. Moreover, the effect of such retraining will generally not be localized to the moved documents or classes that have been moved, merged, split, or otherwise updated. Rather, a retraining of the clustering system to account for updating of one region of the hierarchy of classes may have unintended consequences on other regions that may be far away from the updated region.

BRIEF DESCRIPTION

According to certain aspects illustrated herein, there is provided a method for updating a probabilistic clustering system defined at least in part by probabilistic model parameters indicative of word counts, ratios, or frequencies characterizing classes of the clustering system. An association of one or more documents is changed from one or more source classes to one or more destination classes. Probabilistic model parameters characterizing classes affected by the changed association are locally updated without updating probabilistic model parameters characterizing classes not affected by the changed association.

In a further embodiment the probabilistic clustering system includes a hierarchy of classes with documents assigned to leaf classes of the hierarchy, and the changing of the association comprises:

  • creating a merged class;
  • changing the association of one or more documents or descendant classes from two or more source classes having a common immediate ancestor class to the merged class; and
  • replacing the two or more source classes with the merged class in the hierarchy of classes.

In a further embodiment the local updating comprises:

  • algebraically combining probabilistic model parameters characterizing the two or more source classes to generate probabilistic model parameters characterizing the merged class.

In a further embodiment the probabilistic clustering system includes a hierarchy of classes with documents assigned to leaf classes of the hierarchy, and the changing of the association comprises:

  • creating a merging ancestor class; and
  • inserting the merging ancestor class into the hierarchy of classes between (i) two or more classes each having a common immediate ancestor class and (ii) the common immediate ancestor class.

In a further embodiment the common immediate ancestor class is a root node of the hierarchy of classes.

In a further embodiment the local updating comprises:

  • algebraically combining probabilistic model parameters characterizing the two or more classes to generate probabilistic model parameters characterizing the merging ancestor class.

In a further embodiment the probabilistic clustering system includes a hierarchy of classes with documents assigned to leaf classes of the hierarchy, and the changing of the association comprises:

  • creating two or more split leaf classes; and
  • performing clustering training to associate each document of a pre-existing leaf class with one of the two or more split leaf classes, the cluster training being limited to documents associated with the pre-existing leaf class, the cluster training generating local probabilistic model parameters for characterizing the split leaf classes respective to the documents associated with the pre-existing leaf class.

In a further embodiment the changing of the association further comprises:

  • replacing the pre-existing leaf class with the split leaf classes in the hierarchy of classes.

In a further embodiment the changing of the association further comprises:

  • associating each split leaf class with the pre-existing leaf class such that the pre-existing leaf class is the immediate ancestor of each split leaf class; and
  • removing all documents from the pre-existing leaf class, said documents being associated with the two or more split leaf classes by the cluster training.

In a further embodiment the local updating comprises:

  • algebraically combining the local probabilistic model parameters of each split leaf class and the probabilistic parameters of the pre-existing leaf class to produce probabilistic model parameters characterizing the two or more split leaf classes respective to the documents of the probabilistic clustering system.

In a further embodiment the changing of the association further comprises:

  1. (i) defining a plurality of documents as a document group that is to be kept together, each document of the document group being represented as a bag-of-words and being associated with the pre-existing leaf class;
  2. (ii) prior to the performing of cluster training, replacing the plurality of documents of the document group with a temporary document having a bag-of-words representation that combines word counts of the documents of the document group such that the subsequent cluster training associates the temporary document with one of the split leaf classes;
  3. (iii) after the performing of the cluster training, replacing the temporary document with the plurality of documents of the document group, each of the plurality of documents being associated with the same split leaf class with which the temporary document was associated by the cluster training.

In a further embodiment the operations (i), (ii), and (iii) are repeated for two or more document groups, and the performing of cluster training further comprises:

  • associating the temporary document replacing each document group with a different split leaf class.

According to certain aspects illustrated herein, there is provided a method, for use in a clustering system in which each document is represented as a bag-of-words, for splitting a pre-existing class into two or more split leaf classes. A plurality of documents associated with the pre-existing class are defined as a document group that is to be kept together. The plurality of documents of the document group are replaced with a temporary document having a bag-of-words representation that combines word counts of the documents of the document group. Clustering is performed to associate each document of the pre-existing leaf class, including the temporary document, with one of the two or more split leaf classes. The clustering is limited to documents associated with the pre-existing leaf class. After the clustering, the temporary document is replaced with the plurality of documents of the document group, each of the plurality of documents being associated with the same split leaf class with which the temporary document was associated by the clustering.

In a further embodiment the clustering system includes a hierarchy of classes with documents assigned to leaf classes of the hierarchy, the method further comprising:

  • associating each split leaf class with the pre-existing leaf class such that the pre-existing leaf class is the immediate ancestor of each split leaf class; and
  • removing all documents from the pre-existing leaf class, said documents being associated with the two or more split leaf classes by the clustering.

In a further embodiment the defining of a plurality of documents is repeated to define two or more document groups each of which is replaced by a temporary document prior to the clustering each temporary document being replaced by the plurality of documents of the corresponding document group after the clustering.

In a further embodiment the performing of clustering comprises:

  • associating the temporary document replacing each document group with a different split leaf class.

According to certain aspects illustrated herein, there is provided a probabilistic clustering system operating in conjunction with documents grouped into classes characterized by probabilistic model parameters indicative of word counts, ratios, or frequencies. A user interface is configured to receive a user selection of a change of association of one or more documents from one or more source classes to one or more destination classes. A processor is configured to locally update probabilistic model parameters characterizing classes affected by the changed association without updating probabilistic model parameters characterizing classes not affected by the changed association.

In a further embodiment the user interface is configured to receive a user selection to move a single document from a source class to a destination class, and the processor is configured to algebraically update probabilistic model parameters characterizing the source class and the destination class based on word counts of the moved document to reflect removal of the document from the source class and addition of the document to the destination class, respectively.

In a further embodiment the classes are organized into a hierarchy of classes with documents assigned to leaf classes of the hierarchy, the user interface is configured to receive a user selection to move a document or class from a source class or ancestor class to a destination class or ancestor class, and the processor is configured to (i) identify a common parent class that is a common ancestor for both the source class or ancestor class and the destination class or ancestor class, (ii) identify at least one intermediate ancestor class disposed in the hierarchy between the source class or ancestor class and the common parent class or disposed in the hierarchy between the destination class or ancestor class and the common parent class, and (iii) algebraically update probabilistic model parameters characterizing each intermediate ancestor class based on word counts of the moved document or class to reflect the moving of the document or class from the source class or ancestor class to the destination class or ancestor class.

In a further embodiment the classes are organized into a hierarchy of classes with documents assigned to leaf classes of the hierarchy, the user interface is configured to receive a user selection to merge two or more classes to be merged, said two or more classes having a common immediate ancestor class, and the processor is configured to algebraically combine probabilistic model parameters characterizing the two or more classes to be merged to generate probabilistic model parameters characterizing a merged class.

In a further embodiment the merged class is one of (i) a substitute class that replaces the two or more classes to be merged and (ii) a new ancestor class that is inserted between the two or more classes to be merged and a common ancestor of the two or more classes to be merged.

In a further embodiment the classes are organized into a hierarchy of classes with documents assigned to leaf classes of the hierarchy, the user interface is configured to receive a user selection to split a pre-existing leaf class into two or more split leaf classes, and the processor is configured to (i) perform clustering training to associate each document of the pre-existing leaf class with one of the two or more split leaf classes, the cluster training being limited to documents associated with the pre-existing leaf class and generating local probabilistic model parameters for characterizing the split leaf classes respective to the documents associated with the pre-existing leaf class and (ii) algebraically combine the local probabilistic model parameters of each split leaf class and the probabilistic parameters of the pre-existing leaf class to produce probabilistic model parameters characterizing the two or more split leaf classes respective to the documents of the probabilistic clustering system.

In a further embodiment the two or more split leaf classes one of (i) replace the pre-existing leaf class, and (ii) are arranged as immediate descendant classes of the pre-existing leaf class, the documents being removed from the pre-existing leaf class and associated with the two or more split leaf classes by the cluster training.

BRIEF DESCRIPTION OF THE DRAWINGS

FIGURE 1 diagrammatically shows a clustering system including real-time hierarchy updating.

FIGURE 2 diagrammatically shows an update in which a document is moved from a source class to a destination class.

FIGURE 3 diagrammatically shows an update in which a leaf class is moved from a source ancestor class to a destination ancestor class.

FIGURE 4 diagrammatically shows an update in which two leaf classes having a common immediate ancestor class are merged.

FIGURE 5 diagrammatically shows an update in which two leaf classes having a common immediate ancestor class are merged by creating a common immediate ancestor class.

FIGURE 6 diagrammatically shows an update in which a pre-existing leaf class is split into two split leaf classes that replace the pre-existing leaf class.

FIGURE 7 diagrammatically shows an update in which a pre-existing leaf class is split into two split leaf classes by emptying the documents of the pre-existing leaf class into two new split leaf classes having the pre-existing leaf class as a common immediate ancestor class.

FIGURE 8 diagrammatically shows an update in which a non-leaf class and a leaf class having a common immediate ancestor class are merged by creating a common immediate ancestor class.

DETAILED DESCRIPTION

With reference to FIGURE 1, a clustering system is described. The clustering system is a probabilistic clustering system that organizes set of documents 10 to generate a hierarchy of classes 12. Alternatively, in some embodiments a flat set of classes is generated. The hierarchy of classes 12 include document/leaf class associations 14 indicating which of the documents 10 is associated with which leaf classes of the hierarchy of classes 12. (In the art the term "class" is used, or an alternative but substantive equivalent term such as "cluster" or "category" is used. The term "class" as used herein is intended to be broadly construed to encompass the terms "cluster" or "category" as well). In some embodiments, the document/leaf class associations 14 are annotated to the documents. Each document of the set of documents 10 is suitably represented in a bag-of-words format. In the bag-of-words format, the position of words in the document is not utilized. Rather, a word-frequency vector or similar format is used, in which the document is represented by counts of keywords or of words other than certain frequent and typically semantically uninteresting stop words (such as "the", "an", "and", or so forth). The specific format used can vary - for example, the bag-of-words representation optionally has the counts normalized by the total number of words in the document, or the total number of non-stopwords in the document, or so forth.

The hierarchy of classes 12 also includes a set of probabilistic modeling parameters 16 indicative of word counts, ratios, or frequencies that characterize the classes. The example illustrated clustering system employs a probabilistic model. The hierarchy of classes 12 includes leaf classes associated directly with documents. Optionally, the hierarchy further includes one or more ancestor classes which do not themselves contain any documents but which are ancestral to leaf documents that contain documents. The hierarchy of classes 12 optionally includes more than one level of ancestor classes. In an example embodiment, the probabilistic modeling parameters 16 that characterize the classes include: a class probability P(C) for each class C that is indicative of a ratio of the count of words in documents of the class C respective to a total count of words in the set of documents 10; a document probability P(d|C) that is indicative of a ratio of the count of words in the document d in the class C respective to the total count of words in the class C (for hard partitioning, P(d|C)=0 if the document d is not associated with the class C, and P(d|C) is indicative of the influence of document d on class C if document d is associated with class C); a word probability P(w|C) that is indicative of a ratio of the count of word w in the class C respective to the total count of words in class C. In the illustrated embodiments, hard partitioning is used. However, it is also contemplated to employ soft partitioning in which a given document may have fractional associations with more than one class.

The example probabilistic model parameters P(C), P(d|C), and P(w|C) are illustrative and non-limiting. Other suitable probabilistic model parameters may be defined instead of or in addition to these. Moreover, in some embodiments, clustering may be performed using a non-probabilistic clustering technique such as k-means, latent semantic indexing or hierarchical agglomerative methods.

The hierarchy of classes 12 is typically initially generated using a suitable training approach. In some embodiments, a maximum likelihood approach such as expectation-maximization (EM) or gradient descent is used to perform the training of the probabilistic clustering model. The training generates the probabilistic modeling parameters 16 based on the document/leaf class associations 14 obtained during a training phase. Leaf classes of the hierarchy are directly associated with documents. Intermediate classes from which leaf classes depend either directly or indirectly also have associated probabilistic modeling parameters such as the example probabilities P(c), P(d|c) and P(w|c) (where lower-case "c" here indicates a non-leaf class). In one suitable approach for non-leaf classes, P(c)=0 for all non-leaf classes c, indicating that no documents are directly contained by the non-leaf class c, and P(d|c) and P(w|c) are the weighted averages of the corresponding parameters of its children or descendents, that is: P w | c = C c P w | C P C C c P C , where C c indicates C is a descendent of c and P d | c = C c P d | C P C C c P C , where C c indicates C is a descendent of c

Once the clustering system is trained, a user interface 20 can be used by a user to communicate with a processor 22 that implements a clusters navigator 24 so as to navigate the clustering system. In some embodiments, the user interface is a graphical user interface (GUI) providing, for example: a window showing a tree representation of the hierarchy of classes, optionally configured so that various branches can be expanded or collapsed to focus in on regions of interest; a document window for displaying a selected document; and so forth. Optionally, various windows can be opened and closed using suitable menu options, mouse selections, or other user input mechanisms of the graphical user interface. As long as the user is merely viewing documents, searching for documents, or performing other read-only cluster system navigation operations, the clustering system is static and the hierarchy of classes 12 is unchanged.

However, the user may decide to change the association of one or more documents from one or more source classes to one or more destination classes. In one example association change, the user chooses to move a single document from one class into another class. (See example FIGURE 2). In another example association change, the user moves a leaf class from one ancestor class to another ancestor class, and in the process moves the documents associated with the moved leaf class. (See example FIGURE 3). In another example association change, the user merges two or more leaf classes to form a merged leaf class, and in the process moves the documents associated with the original leaf class into the merged leaf class. (See example FIGURE 4). In another example association change, the user merges two or more leaf classes by inserting a new ancestor class so as to be the immediate ancestor to the two or more leaf classes to be merged, and in the process moves the documents to be indirectly off of the new ancestor class. (See example FIGURE 5). In another example association change, the user splits a pre-existing leaf class into two or more split leaf classes that replace the pre-existing leaf class, and in the process moves the documents associated with the pre-existing leaf class into the split leaf classes. (See example FIGURE 6). In another example association change, the user performs a split by transferring documents of a pre-existing leaf classes into two or more split leaf classes that are created immediately descendant to the pre-existing leaf class, and in the process moves the documents associated with the pre-existing leaf class into the new split leaf classes. (See example FIGURE 7).

Each of these example association changes impacts the hierarchy of classes 12 including the probabilistic modeling parameters 16. If no new classes are created (for example, in the move operations of FIGURES 2 and 3), the effect is to change values of certain probabilistic modeling parameters 16. When new or different classes are created (for example, in the merge and split operations of FIGURES 4-7), values of existing probabilistic modeling parameters 16 may change and additionally new probabilistic modeling parameters are created and/or deleted. Advantageously, however, the clustering system of FIGURE 1 includes the processor 22 that, in addition to implementing or including components for the clusters navigator 24, also implements or includes components for locally updating probabilistic model parameters characterizing those classes that are affected by the changed association without updating probabilistic model parameters that characterize classes not affected by the changed association. As will be described herein, in move and merge operations an algebraic probabilistic modeling parameters updater 30 implemented by the processor 22 can locally update the probabilistic model parameters 12 algebraically, without invoking cluster training. For split operations, a clustering builder 32 implemented by the processor 22 is used to split documents amongst the split leaf classes. The clustering operates only on the documents associated with the pre-existing class that is being split. The clustering builder 32 may employ a maximum likelihood probabilistic clustering algorithm such as an expectancy maximization (EM) algorithm, a gradient descent algorithm, or so forth, to perform the limited cluster training. Typically, the clustering builder 32 is the same algorithm or component used to perform the initial cluster training on the complete set of documents 10 to create the clustering system; however, during a subsequent split operation the clustering builder 32 is applied only to the documents associated with the pre-existing class that is being split.

An advantage of performing local updating using the components 30, 32 is that these updates can be performed in real-time, that is, over a processing time that is acceptable to the user, such as preferably a few minutes or less and more preferably a few seconds or less. Algebraic updating 30 is fast because it performs local updating without clustering (or re-clustering) documents. Application of the clustering builder 32 during a split operation is also typically fast because the number of documents in the pre-existing cluster that is being split is typically a small sub-set of the total number of documents 10 that are organized by the clustering system. For example, the application of the clustering builder 32 during a split operation has an execution time of typically a few seconds or less for pre-existing clusters including a tens, hundreds, or a few thousand documents. In contrast, applying the clustering builder 32 to perform a complete retraining of the cluster system involves clustering typically tens of thousand or more documents in the set of documents 10 and hence takes substantially longer. Such a complete retraining typically is not practical to perform in real-time.

Having described the overall clustering system including real-time updating with reference to FIGURE 1 and with brief references to FIGURES 2-8, the processing for example association changes including document move, class move, and several types of merge and split operations is next described in greater detail with detailed reference to FIGURES 2-8 in turn. It is to be appreciated that these various operations can be combined in various ways to perform complex real-time updating of the clustering system. As just one example, a class can be moved to another part of the hierarchy, then split, and then one or more documents moved from one of the generated split classes to another of the generated split classes.

With reference to FIGURE 2, a document selected from the set of documents 10 organized by the clustering system is moved from a leaf class "L1" to a leaf class "LN". In such an operation, the hierarchy of classes may be flat (that is, including only leaf classes or only leaf classes off of a common root class), or it may be more complex, as indicated by the optional higher level class structure shown by short dashed lines in FIGURE 2 and including a higher level class "H1" and a root class. The document move operation involves removal of the document from the leaf class "L1" (removal indicated by the "Ø" symbol in FIGURE 2) and addition of the document to the leaf class "LN" (addition indicated by the dashed connector with the "+" symbol in FIGURE 2). In order to account for the move of the document from source class "L1" to destination class "LN", a complete re-build of the hierarchy of classes 12 would be performed - however, such a complete rebuild is typically not feasible for real-time processing.

Accordingly, the algebraic probabilistic modeling parameters updater 30 of the processor 22 makes a local update of the probabilistic modeling parameters, which affects the source and destination leaf classes "L1" and "LN" as well as the intermediate class "H1" that is between the source leaf class "L1" and the common parent class "root" that is a common ancestor class for both the source leaf class "L1" and the destination leaf class "LN". More generally, the algebraic probabilistic modeling parameters updater 30 updates probabilistic modeling parameters characterizing the source and destination leaf classes as well as any intermediate ancestor class or classes disposed in the hierarchy between the source leaf class and the common parent class or disposed in the hierarchy between the destination leaf class and the common parent class. The term "common parent class" is to be understood as referring to the common parent class furthest down the hierarchy, that is, closest in the hierarchy to the source and destination leaf classes. The common parent class may itself have parent classes, which are also common to both the source and destination leaf classes.

The algebraic updating for the move operation is suitably performed as follows. Let the source leaf class be denoted by Cs and the destination or target leaf class be denoted by Ct. Further let D denote the document to be moved, and let d denote any other document in the source class Cs or in the target class Ct. The updating is suitably performed in four operations: removing the document D from the source class Cs and updating the probabilistic modeling parameters for source class Cs; propagating the updating to any intermediate classes disposed in the hierarchy between source class Cs and the common parent class of source and target classes Cs and Ct; adding the document D to the destination or target cluster Ct and updating the probabilistic modeling parameters for target class Ct; and propagating the updating to any intermediate classes disposed in the hierarchy between target class Ct and the common parent class of source and target classes Cs and Ct. Only probabilistic modeling parameters for the local classes, namely Cs, Ct, and any intermediate classes disposed between the source or target classes and the common parent class, are modified.

In FIGURE 2, source class Cs is the class "L1", target class Ct is the class "LN", and there is one intermediate ancestor class "H1" between class "L1" and the common parent class (in the illustrated case, the root class) of classes "L1" and "LN". Other classes are not local, and hence are not updated. In FIGURE 2, these non-local classes include leaf class "L2" and the "root" class. The move operation including algebraic updating of the probabilistic modeling parameters is as follows, where |Cs| denotes the total word count of the source class Cs, |Ct| denotes the total word count of the target class Ct, |D| denotes the total word count of the document D, and N denotes the total word count for the collection of documents 10. Removal of the document D from the source class Cs is accounted for by the following algebraic adjustments (where P() denotes the probabilistic modeling parameter before adjustment, and P̂() denotes the probabilistic modeling parameter after adjustment by the algebraic probabilistic modeling parameters updater 30 to account for the moving of document D): P ^ D | C s = 0 P ^ d | C s = P d | C s 1 - D / C s , d C s \ D P ^ C s = P C s - | D | / N and P ^ w | C s = P w | C s - N wD / C s 1 - D / C s , w where in Equation (6) NwD is the word count of word w in document D. Equations (3)-(6) algebraically account for the effect of the removal of document D from the source class Cs. Additionally, if there are any intermediate ancestor classes disposed in the hierarchy between source class Cs and the common parent class of source and target classes Cs and Ct, the probabilistic modeling parameters characterizing these classes are algebraically adjusted as follows. Let Cp denote the common parent class of source and target classes Cs and Ct. Then the following iterative procedure is employed:

Addition of the document D to the target class Ct is accounted for as follows: P ^ D | C t = D D + C t P ^ d | C t = P d | C t 1 - D / C t , d C t \ D P ^ C t = P C t + | D | / N and P ^ w | C t = P w | C t + N wD / C t 1 - D / C t , w Equations (7)-(10) algebraically account for the effect of the addition of document D to the destination or target class Ct. Additionally, if there are any intermediate ancestor classes disposed in the hierarchy between target class Ct and the common parent class Cp of source and target classes Cs and Ct, the probabilistic modeling parameters characterizing these classes are algebraically adjusted as follows:

With reference to FIGURE 3, a class move operation is next described. Leaf class "L2" is moved from class "H1" to the root class. The leaf class move involves removal of the leaf class "L2" from the source immediate ancestor class "H1" (removal indicated by the "Ø" symbol in FIGURE 3) and addition of the leaf class "L2" to the destination or target immediate ancestor class "root" (addition indicated by the dashed connector with the "+" symbol in FIGURE 3). In order to account for the move of the leaf class "L2" from source immediate ancestor class "H1" to destination immediate ancestor class "root", a complete re-build of the hierarchy of classes 12 would be performed - however, such a complete rebuild is typically not feasible for real-time processing.

Accordingly, the algebraic probabilistic modeling parameters updater 30 of the processor 22 makes a local update of the probabilistic modeling parameters. The local update affects the source and destination immediate ancestor classes as well as any intermediate ancestor class or classes disposed in the hierarchy between the immediate source ancestor class and the common parent class or disposed in the hierarchy between the immediate destination ancestor class and the common parent class. The algebraic updating is substantially analogous to the case for the document move operation. Removal of the leaf class, denoted C, from the source ancestor class, is reflected in the probabilistic modeling parameters characterizing the immediate ancestor class, denoted Cp s, is as follows (where |C| denotes the total word count for the moved leaf class C): P ^ d | C s p = 0 , d C P ^ d | C s p = P d | C s p 1 - C / C s p , d C s p \ C and P ^ w | C s p = P w | C s p - P w | C C / C s p 1 - C / C s p , w where P() denotes the probabilistic modeling parameter before adjustment, and P̂() denotes the probabilistic modeling parameter after adjustment by the algebraic probabilistic modeling parameters updater 30 to account for the moving of the leaf cluster C. Additionally, if there are any intermediate ancestor classes disposed in the hierarchy between immediate source ancestor class C s p and the common parent class Cp of immediate source ancestor class C s p and the immediate destination or target ancestor class, denoted C t p , the probabilistic modeling parameters characterizing these classes are algebraically adjusted as follows:

Addition of the leaf class C to the immediate target ancestor class C t p is accounted for as follows: P ^ d | C t p = d C + C t p , d C where |d| denotes the total word count for document d, and P ^ d | C t p = P d | C t p 1 - C / C t p , d C t p and P ^ w | C t p = P w | C t p - P w | C C / C t p 1 - C / C t p , w

Additionally, if there are any intermediate ancestor classes disposed in the hierarchy between immediate target ancestor class C t p and the common parent class Cp of immediate source ancestor class C s p and the immediate target ancestor class C t p , the probabilistic modeling parameters characterizing these classes are algebraically adjusted as follows:

With reference to FIGURE 4, a flat merge operation is described. Leaf classes "L1" and "L2" having a common immediate ancestor class "root" are merged to form a merged leaf class "L1&2" that replaces the leaf classes "L1" and "L2". In other words, the association of documents of the leaf classes "L1" and "L2" are changed to the newly created merged class "L1&2", which replaces the classes "L1" and "L2" in the hierarchy of classes. The hierarchy of classes in FIGURE 4 is a flat hierarchy; however, a more complex hierarchy can be used. For example, rather than being off the root class, the classes to be merged can have a common ancestor class that is intermediate between the root and the leaf classes to be merged. For such a merge operation to be meaningful, the common ancestor class should have at least one additional descendant leaf class that is not included in the merge operation.

The algebraic probabilistic modeling parameters updater 30 of the processor 22 accounts for a flat merge in the probabilistic modeling parameters as follows. Because the merged class replaces the classes to be merged in the hierarchy of classes, and because the classes to be merged have a common immediate ancestor class, the effect of the merge on the probabilistic modeling parameters is highly localized, and affects only the created merged class. The probabilistic modeling parameters of the merged class are suitably computed as follows, where L denotes the number of classes to be merged (L>1), Cn denotes a leaf class to be merged (leaf class "L1" or "L2" in the example of FIGURE 4), the index n runs over the two or more classes to be merged (in the following, a base-one index in which n∈{1,...,L} is used, but a base-zero index in which n∈{0,...,L-1}, or other indexing arrangement could also be used), and C denotes the merged class (merged class "L1&2" in the example of FIGURE 4): P C = n - 1 L P C n P w | C = n = 1 L P w | C n P C n P C and P d | C = n = 1 L P d | C n P C n P C = d C

where |d| and |C| are the total word counts for document d and merged class C, respectively. The original leaf classes Cn are removed from the hierarchy of classes and replaced by the merged class C, with all documents previously associated with the leaf classes Cn now associated with the merged class C.

With reference to FIGURE 5, a hierarchal merge is described. Leaf classes "L1" and "L2" having a common immediate ancestor class "root" are merged by inserting a merged leaf class "H1&2" between the leaf classes "L1" and "L2" and the immediate ancestor class "root". The inserted merged leaf class "H1&2" is therefore indicative of the combination or merger of leaf classes "L1" and "L2", but, unlike in the flat merge, here the original leaf classes "L1" and "L2" are retained. The effect of the hierarchal merge on the probabilistic modeling parameters is again highly localized, and affects only the inserted merged class and the position in the hierarchy of the leaf classes to be merged. The probabilistic modeling parameters of the merged class are suitably computed as follows, where L denotes the number of classes to be merged (L>1), Cn denotes a class to be merged (classes "L1" and "L2" in the example of FIGURE 4), the index n runs over the two or more classes to be merged (again using the conventional base-one index in which n∈{1,...,L}), and C denotes the merged class (inserted merged class "L1&2" in the example of FIGURE 4): P C = 0 (indicating that the inserted merged class C is a non-leaf class that itself contains no documents), P w | C = n = 1 L P w | C n P C n n = 1 L P C n = N wC C and P d | C = n = 1 L P d | C n P C n n = 1 L P C n = d C

where |d| and |C| are the total word counts for document d and merged class C, respectively, and NwC is the total count of occurrences of word w in the merged class C. The original leaf classes Cn (leaf classes "L1" and "L2" in the example of FIGURE 5) are retained in the case of a hierarchal merge, and the newly created merged class C (class "H1&2" in the example of FIGURE 5) is inserted in the hierarchy of classes between the original leaf classes Cn and the immediate common ancestor class of the original leaf classes Cn. All documents remain associated with their respective leaf classes Cn, but now also indirectly depend from the inserted merged class C.

With reference to FIGURE 6, a flat split operation is described. A pre-existing leaf class "L2" is split into two or more split leaf classes, in FIGURE 6 two split leaf classes "L2A and L2B", that replace the pre-existing leaf class in the hierarchy of classes. Before the split the pre-existing leaf class "L2" has class "H1" as its immediate ancestor class, and after the split the replacement split leaf classes "L2A and L2B" have their common immediate ancestor class "H1". While in the example of FIGURE 6 the common immediate ancestor class "H1" is an intermediate class, in other embodiments the hierarchy of classes may be flat, and the immediate ancestor class for the pre-existing leaf class and the replacement split leaf classes may be the root class.

To divide documents associated with the pre-existing leaf class amongst the two or more replacement split leaf classes, cluster training is performed. However, in order to promote rapid processing conducive to real-time operations, the clustering builder 32 of the processor 22 is limited to processing documents associated with the pre-existing leaf class (example pre-existing leaf class "L2" in FIGURE 6). The limited cluster training generates local probabilistic model parameters for characterizing the split leaf classes (example split leaf classes "L2A" and "L2B" in FIGURE 6) respective to the documents associated with the pre-existing leaf class. Because the number of documents associated with the single pre-existing class to be split is typically substantially smaller than the total number of documents organized by the clustering system, this limited and local cluster training can be performed substantially faster than a retraining of the clustering system as a whole. Additionally, since the limited cluster training is local to documents of the pre-existing leaf class, its effect is localized.

Suitable processing for performing the flat split is as follows. The pre-existing class is denoted by C, while the set of documents Dc denotes those documents associated with the pre-existing leaf class C. For hard partitioning, P(d|C)=0 for all documents that are not members of Dc. The clustering builder 32 is applied to the set of documents Dc with L denoting the number of split leaf classes. The index n runs over the split leaf classes, and Cn denotes the split leaf classes. The local cluster training produces a set of local probabilistic modeling parameters for the split leaf classes Cn, such as P̂(Cn), P̂(d|Cn), P̂(w|Cn), or so forth. These local probabilistic modeling parameters are computed for the set of documents Dc without accounting for other documents of the clustering system. Thus, for example: n = 1 L P ^ C n = 1 indicating that every document in the set of documents Dc is associated with one of the split leaf classes Cn. On the other hand, in the context of the global clustering system: n = 1 L P C n = P C since the set of documents Dc is the set of documents associated with the pre-existing leaf class C in the global clustering system. To convert the local probabilistic modeling parameters P̂(Cn), P̂(d|Cn), P̂(w|Cn) into corresponding probabilistic modeling parameters P(Cn), P(d|Cn), and P(W|Cn) suitable for use in the context of the clustering system as a whole, the algebraic probabilistic modeling parameters updater 30 is applied as follows: P C n = P ^ C n P C P d | C n = P ^ d | C n P w | C n = P ^ w | C n Equations (26) and (27) are identity equations, as the values for these probabilistic modeling parameters are the same in the local context as in the global clustering system context. With the global probabilistic modeling parameters determined, the flat split is completed by replacing the pre-existing leaf class (leaf class "L2" in example FIGURE 6) with the split leaf classes (leaf classes "L2A" and "L2B" in example FIGURE 6) with the documents of the pre-existing leaf class being associated with the various split leaf classes in accordance with the results of the local cluster training.

It will be noted that the flat merge is the reverse operation to the flat split in some sense. Splitting a pre-existing leaf class and then merging the resulting split leaf classes will result in obtaining identically the original pre-existing leaf class. The reverse is not true - merging several leaf classes and then splitting the merged class may produce different leaf classes. That is, for example, merging five leaf classes and then splitting the merged leaf class back into five leaf classes will in general not result in identically the same original five leaf classes.

With reference to FIGURE 7, a hierarchal split is described. In this case, the pre-existing leaf class is not replaced - rather, the split leaf classes Cn are added and placed off of the pre-existing leaf class C, and the document contents Dc of the pre-existing leaf class are emptied into the added split leaf classes. FIGURE 7 depicts an example, in which document contents of the pre-existing leaf class "L2" are emptied into added split leaf classes "L2A" and "L2B". After the hierarchal split, the pre-existing leaf class is no longer a leaf class, but rather is the immediate ancestor class of the split leaf classes. Accordingly, after the hierarchal split, the pre-existing class itself contains no documents. The division of the set of documents Dc amongst the added split leaf classes Cn are suitably determined using the clustering builder 32 of the processor 22, as in the flat split, and the clustering produces local probabilistic modeling parameters such as P̂(Cn), P̂(d|Cn), and P̂(w|Cn). To convert the local probabilistic modeling parameters P̂(Cn), P̂(d|Cn), P̂(w|Cn) into corresponding probabilistic modeling parameters P(Cn), P(d|Cn), and P(w|Cn) suitable for use in the context of the clustering system as a whole, the algebraic probabilistic modeling parameters updater 30 is again applied in accordance with Equations (25)-(27). Additionally, for the pre-existing class C which is retained in the case of hierarchal clustering: P C = 0 and P(w|C) and P(d|C) are unchanged. The pre-existing class C is retained and serves as the immediate ancestor class for the added split leaf clusters, and the documents that were associated with the pre-existing class C are associated with the various added split leaf clusters in accordance with the local cluster training.

In the case of a flat split followed by a flat merge of the split leaf classes, the original leaf class is obtained. In contrast, the hierarchical merge is not an inverse operation to the hierarchical split. This is because both the hierarchical merge and the hierarchical split create additional hierarchy.

In the foregoing example flat and hierarchical split operations, it was assumed that the user has no preference as to how the split is performed. In other words, the user relies entirely upon the mechanics of the clustering builder 32 which decides which documents are grouped into which split leaf classes. In some cases, however, the user may want to keep a certain sub-set of documents together during the split operation. This is suitably referred to as a constrained split. Accordingly, a technique is described for performing a split (either flat or hierarchal) in a clustering system in which documents are represented in a bag-of-words format, in which the document is represented by a word-frequency vector or other format that stores counts of keywords or of words other than certain frequent and typically semantically uninteresting stop words (such as "the", "an", "and", or so forth). The described technique for keeping a selected sub-set of documents together during the split can be applied in the context of substantially any split operation that employs a clustering technique operating on documents in a bag-of-words format to distribute documents amongst the split classes. The technique is not limited to probabilistic clustering systems. The example notation used herein is as follows: C denotes the class to be split; L denotes the number of split classes to be generated by the split (the split leaf clusters may replace the class C in the case of a flat split, or may be appended to the class C in the case of a hierarchal split); and G={d1,...dn} denotes a sub-set of n documents to be grouped together, where n>1.

Prior to performing the split, the sub-set of documents G is replaced by a single temporary document g=d1*d2*...*dn where the operator "*" denotes additive combination of the word counts of the bag-of-words-formatted documents. Table 1 shows an illustrative example of generation of the temporary document g=d1*d2*d3 for the case of a sub-set of three documents G={d1, d2, d3} with the bag-of-words representation indicated in Table 1. For example, Table 1 shows that for the word "computer", document d1 has a word count of 53, document d2 has a word count of 420, and document d3 has a word count of 232. Accordingly, the temporary document g=d1*d2*d3 has a word count for the word "computer" of 53+420+232=705. The same approach is used for each word of the bag-of-words, such as for "hard drive", "RAM", and so forth. The document d1 has a total word count (optionally excluding selected stopwords) of 2245 words; document d2 has a total word count of 11,025 words; and document d3 has a total word count of 6346 words. Accordingly, the temporary document g=d1*d2*d3 has a total word count of 2245+11,025+6346=19,616 words. Note that the term "floppy" does not occur in document d3. In such a case, where for example "floppy" is not in the bag-of-words representing document d3, the count for term "floppy" is taken as zero for document d3. More generally, the temporary document g includes all words that are present in any of the bag-of-words representations of any of the documents of the sub-set of documents G. Table 1 - Example of generating temporary document g for G={d<sub>1</sub>, d<sub>2</sub>, d<sub>3</sub>} Word d1 word count d2 word count d3 word count g word count "computer" 53 420 232 705 "hard drive" 23 321 111 455 "RAM" 10 43 120 173 "floppy" 32 23 --- 55 ... ... ... ... ... Total 2245 11025 6346 19616

The temporary document g is substituted for the sub-set of documents G in the pre-existing class to be split. After this substitution, the clustering algorithm is performed on the documents of the class C with the number of split classes set to L. As the temporary document g is a single document, it will be clustered as a unit. In the case of hard partitioning, the temporary document g will be associated with a single one of the split classes by the clustering. The clustering can use probabilistic clustering or substantially any other clustering algorithm, including a non-probabilistic clustering technique such as k-means, latent semantic indexing or hierarchical agglomerative methods. For probabilistic clustering, a maximum likelihood probabilistic clustering algorithm may be employed, such as an expectancy maximization (EM) algorithm, a gradient descent algorithm, or so forth. Once the clustering is complete, the temporary document g is replaced by the sub-set of documents G={d1,...dn}, with each document d1, ..., dn being associated with the same split class with which the clustering associated the temporary document g. For example, if L=7 and the clustering associated the temporary document g with a particular split class Ck of the seven split classes, then every document d1, ..., dn is associated with the same split class Ck. If probabilistic clustering is employed, then the document probabilities P(d|Ck)for the documents d1, ..., dn of the sub-set G, and for split class k with which the temporary document g is associated, are computed based on the document probability P̂(g|Ck) generated by the clustering, as follows: P d i | C k = d i d j G d j P ^ g | C k , d i G where |di| denotes the total word count for document di and |dj| denotes the total word count for document dj. Probabilities P(Ck) and P(w|Ck) are unchanged by the substitution of the documents d1, ..., dn for the temporary document g in the class Ck.

If the user identifies more than one sub-set of documents, each of which is to be kept together, the above approach can continue to be used. For example, if the user identifies N sub-sets of documents, with each sub-set of documents denoted as G(p)={d1 (p), d2 (p), ..., dn (p)} where the index p runs over the N sub-sets of documents and each sub-set G(p) has no intersection with (that is, no documents in common with) any other user-identified sub-set of documents, then for each sub-set of documents G(p) a corresponding temporary document g(p)=d1 (p) *d2 (p) * ... *dn (p) is created to replace the set of documents G(p) in the class to be split. The clustering is then performed as before, so that each temporary document g(p) is assigned to a split class, and each temporary document g(p) is expanded as before, including application of Equation (29) in the case of probabilistic clustering.

In some cases, not only will the user identify more than one sub-set G(p) of documents, each of which is to be kept together, but the user will also indicate that each sub-set of documents G(p) is to be associated with a different split class. This is suitably termed an "exclusive group constraint" or a "disjunction constraint", and is readily accomplished by manipulating the clustering to ensure that each corresponding temporary document g(p) is assigned to a different class during the clustering process. For example, in the case of probabilistic clustering employing the expectation maximization (EM) algorithm, the following approach is suitable. The EM clustering algorithm is initialized with special values for the document probability (that is, P(d|c)) parameters to force the disjunction between the groups. This initialization implements a simple and reasonable assumption: user-specified exclusive constraints are transitive. That is, if the user specifies that classes G(1) and G(2) are disjoint and that classes G(2) and G(3) are disjoint, then it is assumed that G(1) and G(3) are also disjoint. In other words, it is assumed that every sub-set of documents G(p) is to be disjoint from (that is, assigned to a different class from) every other sub-set of documents. The initialization implements a default assignment of the temporary documents g(p) to the split classes, denoted ck, which ensures that the sub-sets G(p) will be disjoint in the resulting split structure. In a suitable approach using EM clustering, the initialization suitably assigns P(ck|g(p))=1 for k=p and P(ck|g(p))=0 for k≠p, and assigns P(ck|di)=1/L for di∉G(1),...,G(n) (that is, the documents not in any of the sub-sets G(p) are initially assigned a uniform distribution amongst the split classes ck). The probability P(d)=|d|/|C| is initially assigned for any document d that is in the class to be split, denoted C. Finally, the initial word probabilities are assigned as follows: P w | c k = N wC C + &egr;

where NwC is the number of occurrences of word w in the class C to be split, and s is a small perturbation. The EM clustering is then performed. The EM clustering iterative process ensures that after the above initialization, each next iteration will keep the same binary values for P(ck|g(p)). The initial state respects the disjunctions between the sub-sets of documents G(p) represented by temporary documents g(p). Accordingly, after several iterations of the EM algorithm a locally optimal clustering of the documents is reached, initially contained in cluster C that respects the exclusive constraints. After the EM clustering, for each split class ck, any temporary document g(p) associated therewith is expanded back to the constituent documents G(p)={d1 (p), d2 (p), ..., dn (p)} including application of Equation (29).

With returning reference to FIGURE 1, the document move, leaf class move, leaf classes merge, and leaf class split operations described herein are suitably implemented in real-time in conjunction with user inputs supplied via the user interface 20. By the term "real-time", it is meant that the operation can be performed in a time-frame that is acceptable to a user interacting with the user interface 20. Typically, such a time-frame is preferably a few minutes or less, and more preferably a few seconds or less. In some embodiments, the user interface 20 is a graphical user interface (GUI) in which the hierarchy of classes 12 is suitably represented in a tree or tree-like structure, optionally including expandable and collapsible branches, in which the user can select classes, documents, and so forth using a mouse or other pointing device. For example, in one approach the user performs a move operation on a document or leaf class by right-clicking on the document or leaf class to bring up a context-sensitive menu including a selection for the move operation. The user selects the move operation off of the context-sensitive menu, which then brings up a dialog asking for the destination class. The user selects the destination class, for example by clicking on it using the mouse point, the document or leaf class move is performed as described herein, and an updated tree representation of the hierarchy of classes 12 is displayed. The GUI preferably includes suitable checks on the operation, such as displaying the move operation selection option in the context-sensitive menu only if the user right-clicks on a document or leaf class. In some embodiments, the user can select several documents or leaf classes to move by, for example, holding down the <CONTROL> or <ALT> key while left-clicking on the documents, and then right-clicking on the selected group to bring up the context-sensitive menu including the move operation selection option.

To execute a merge operation, the user suitably selects multiple leaf classes for merging by, for example, holding down the <CONTROL> or <ALT> key while left-clicking on the documents, and then right-clicking on the selected group to bring up the context-sensitive menu. If the GUI determines that the selections are all leaf classes having a common immediate ancestor class (which may be the root class), then the GUI includes a merge option in the context-sensitive menu. If the user selects the merge option, then the GUI responds by asking whether a flat or hierarchal merge is intended. (Alternatively, separate "flat merge" and "hierarchal merge" options can be included on the context-sensitive menu). The selected leaf classes are then merged using the selected flat or hierarchal merge as described herein, and an updated tree representation of the hierarchy of classes 12 is displayed.

To execute a split operation, the user selects a leaf class to be split, and right-clicks to bring up the context-sensitive menu. If the GUI determines that the selection is a leaf class, then it includes the split option on the context sensitive menu. If the user selects the split option off the context-sensitive menu, then the GUI responds by asking whether a flat or hierarchal split is intended. (Alternatively, separate "flat split" and "hierarchal split" options can be included on the context-sensitive menu). Additionally, the GUI asks for a value indicating the number of split classes to create. Preferably, the GUI includes a check to ensure that the number of split classes supplied by the user is greater than one and is less than a total number of documents in the class to be split. The selected leaf class is then split using the selected flat or hierarchal split as described herein, and an updated tree representation of the hierarchy of classes 12 is displayed.

To implement a constrained split in which one or more sub-sets of documents G(p) are kept together, the user opens the leaf class, for example by double-clicking, and selects a sub-set by holding the <CONTROL>, <ALT>, or other grouping key down while clicking on the documents that are to be included in the sub-set of documents. This process is optionally repeatable to define more than one sub-set of documents to be kept together. The user then selects the split option as before, and the GUI recalls the selected groups and keeps them together using substituted temporary documents as described herein. Optionally, before performing the split the GUI asks the user whether the groups are to be clustered into separate split leaf classes (that is, whether the clustering should employ an "exclusive group constraint" in the clustering), and performs the constrained split in accordance with the user's response.

The described GUI operations are illustrative examples which use mouse manipulations familiar to users of typical GUI systems such as Microsoft Windows® computer operating systems. Other GUI implementations can be employed to facilitate execution of real-time document move, leaf class move, classes merge and class split operations as described herein. Moreover, it is contemplated to implement these operations, either in real-time or as processes queued for later processing, using non-graphical interfacing such as a batch processing script, command-line user interface, or so forth.

Heretofore, the class move, split, and merge operations have been described respective to moving, splitting, or merging a leaf cluster. However, these operations can be extended to moving, splitting, or merging larger branches of the hierarchy of classes.

For example, when a non-leaf class is moved along with all descendant classes, the algebraic updating is suitably as follows. The non-leaf class that is moved is denoted c, and its immediate ancestor class (that is, the source ancestor class) is denoted C s p . The symbol C s p denotes the number of word occurrences in the source ancestor class C s p before non-leaf class c is moved, and |c| denotes the number of word occurrences in all leaf classes that depend from non-leaf class c. The immediate destination or target ancestor class to which class c is moved is similarly denoted C t p , and C t p denotes the number of word occurrences in the target ancestor class C t p before non-leaf class c is moved. The probabilistic modeling parameters of the source ancestor class C s p are suitably updated in a manner analogous to that of Equations (11)-(13) as follows: P ^ d | C s p = 0 , d C , C c P ^ d | C s p = P d | C s p 1 - c / C s p , d C s p \ C and P ^ w | C s p = P w | C s p - P w | c c / C s p 1 - c / C s p , w

The changes are propagated to any intermediate ancestor classes between source ancestor class C s p and the common parent class Cp in the same way as in the case of a leaf class move operation. The effect on the probabilistic modeling parameters characterizing C t p of moving the non-leaf class c onto the immediate target ancestor class C t p is accounted for analogously to Equations (14)-(16): P ^ d | C t p = d c + C t p , d C , C c where |d| denotes the total word count for document d, and P ^ d | C t p = P d | C t p 1 - c / C t p , d C t p and P ^ w | C t p = P w | C t p - P w | c c / C t p 1 - c / C t p , w

The changes are again propagated to any intermediate ancestor classes between target ancestor class C t p and the common parent class Cp in the same way as in the case of a leaf class move operation.

In similar fashion, a flat merge into a single non-leaf class c of non-leaf classes cn having a common immediate parent class, where index n runs from 1 to L, is performed as follows. The number of words in documents associated with the leaf classes descending from non-leaf class cn is denoted |cn|, and the number of words in documents associated with leaf classes descending from merged non-leaf class c is denoted |c|. Thus, c = n = 1 L c n

The algebraic probabilistic model parameters updater 30 computes the parameters for the merged class c in a manner analogous to Equations (17)-(19) as follows: P c = n - 1 L P c n = 0 where equality with zero results because a non-leaf class itself contains no documents, hence P(cn)=0. The word probabilities for merged class c are: P w | c = n = 1 L P w | c n c n n = 1 L | c n | and the document probabilities are: P d | c = n = 1 L P d | c n c n n = 1 L c n = d c

where |d| is the number of words in document d. To complete the merge, the non-leaf classes cn are replaced by the merged class c which descends from the common parent of the classes cn and which contains the descendent classes that descended from the non-leaf classes cn prior to the flat merge.

A hierarchical merge of leaf classes has been described with reference to FIGURE 5 and Equations (20)-(22). More generally, any set of classes, either leaf classes or non-leaf classes, having the same immediate ancestor (i.e., parent) class Cp can be merged.

With reference to FIGURE 8, for example, a non-leaf class "H2" having immediate ancestor class "H1" and having descendant leaf classes "L2" and "L3" is hierarchically merged with a leaf class "L4" also having the immediate ancestor class "H1". The hierarchical merge is performed by inserting a new merged non-leaf class "c" between the immediate ancestor class "H1" and the merged classes "H2" and "L4" as shown in FIGURE8.

The algebraic probabilistic model parameters updater 30 computes the parameters for the merged class c in a manner analogous to Equations (20)-(22) as follows. Let cn represent the classes (leaf or non-leaf) to be merged, with index n running from 1 to L. The common immediate ancestor class to the classes Cn is denoted Cp. The notation |cn| denotes the number of words in documents associated with class cn if class cn is a leaf class, or denotes the number of words in documents associated with leaf classes descending from the class cn if class cn is a non-leaf class. The latter is suitably expressed as: c = C c n C

Then the number of words *|c| in the inserted merged class c is given by: c = n = 1 L c n

The algebraic probabilistic model parameters updater 30 computes the parameters for the inserted merged class c in a manner analogous to Equations (20)-(22) as follows: P c = 0 (indicating that the inserted merged class c is a non-leaf class that itself contains no documents), P w | c = n = 1 L P w | c n c n n = 1 L | c n | = N wc c and P d | c = n = 1 L P d | c n c n n = 1 L c n = d c

where Nwc is the total count of occurrences of word w in documents associated with leaf classes descending from the merged class c. To complete the merge, the class c is inserted in the hierarchical structure under the common parent Cp, and the merged classes cn are arranged to descend from inserted class c.


Anspruch[en]
A method for updating a probabilistic clustering system defined at least in part by probabilistic model parameters indicative of word counts, ratios, or frequencies characterizing classes of the clustering system, the method comprising: changing an association of one or more documents from one or more source classes to one or more destination classes; and locally updating probabilistic model parameters characterizing classes affected by the changed association without updating probabilistic model parameters characterizing classes not affected by the changed association. The method as set forth in claim 1, wherein the one or more documents are a single document, and the changing of the association comprises: changing the association of the document from a source class to a destination class to move the document from the source class to the destination class. The method as set forth in claim 2, wherein the local updating comprises: algebraically updating probabilistic model parameters characterizing the source class and the destination class based on word counts of the moved document to reflect removal of the document from the source class and addition of the document to the destination class, respectively. The method as set forth in claim 3, wherein the probabilistic clustering system includes a hierarchy of classes with documents assigned to leaf classes of the hierarchy, the local updating further comprising: identifying a common parent class that is a common ancestor for both the source class and the destination class; identifying at least one intermediate ancestor class disposed in the hierarchy between the source class and the common parent class or disposed in the hierarchy between the destination class and the common parent class; and algebraically updating probabilistic model parameters characterizing each intermediate ancestor class based on word counts of the moved document to reflect the moving of the document from the source class to the destination class. The method as set forth in claim 2, further comprising: repeating the association changing and local updating for a plurality of documents so as to move the plurality of documents from the source class to the destination class. The method as set forth in claim 1, wherein the probabilistic clustering system includes a hierarchy of classes with documents assigned to leaf classes of the hierarchy, and the changing of the association comprises: moving a class directly or indirectly containing the one or more documents from a source ancestor class in the hierarchy to a destination ancestor class in the hierarchy. The method as set forth in claim 6, wherein the local updating comprises: identifying a common parent class that is a common ancestor for both the source ancestor class and the destination ancestor class; identifying at least one intermediate ancestor class disposed in the hierarchy between the source ancestor class and the common parent class or disposed in the hierarchy between the destination ancestor class and the common parent class; and algebraically updating probabilistic model parameters characterizing each intermediate ancestor class based on word counts of the moved class to reflect the moving of the class from the source ancestor class to the destination ancestor class. In a clustering system in which each document is represented as a bag-of-words, a method for splitting a pre-existing class into two or more split leaf classes, the method comprising: defining a plurality of documents associated with the pre-existing class as a document group that is to be kept together; replacing the plurality of documents of the document group with a temporary document having a bag-of-words representation that combines word counts of the documents of the document group; performing clustering to associate each document of the pre-existing leaf class, including the temporary document, with one of the two or more split leaf classes, the clustering being limited to documents associated with the pre-existing leaf class; after the clustering, replacing the temporary document with the plurality of documents of the document group, each of the plurality of documents being associated with the same split leaf class with which the temporary document was associated by the clustering. The method as set forth in claim 8, further comprising: replacing the pre-existing leaf class with the split leaf classes. A probabilistic clustering system operating in conjunction with documents grouped into classes characterized by probabilistic model parameters indicative of word counts, ratios, or frequencies, the probabilistic clustering system comprising: a user interface configured to receive a user selection of a change of association of one or more documents from one or more source classes to one or more destination classes; and a processor configured to locally update probabilistic model parameters characterizing classes affected by the changed association without updating probabilistic model parameters characterizing classes not affected by the changed association.






IPC
A Täglicher Lebensbedarf
B Arbeitsverfahren; Transportieren
C Chemie; Hüttenwesen
D Textilien; Papier
E Bauwesen; Erdbohren; Bergbau
F Maschinenbau; Beleuchtung; Heizung; Waffen; Sprengen
G Physik
H Elektrotechnik

Anmelder
Datum

Patentrecherche

Patent Zeichnungen (PDF)

Copyright © 2008 Patent-De Alle Rechte vorbehalten. eMail: info@patent-de.com