Granted Patent B2

US 8,356,044 · App. 12/360,825 · Granted Jan 15, 2013

Inventors: Mark J. Stefik (Portola Valley, CA); Sanjay Mittal (Fremont, CA)

Assignee: Palo Alto Research Center Incorporated

Abstract

Claims (63)

1. A computer-implemented method for providing default hierarchical training for social indexing, comprising:

maintaining articles of digital information for social indexing;

specifying a hierarchically-structured tree of topics, which each comprise a label comprising one or more words;

identifying hard constraints based on the labels comprised in the topic tree and the topic tree's hierarchical structure, and defining the hard constraints to include immutable rules comprising at least one of:

requiring that a topic model comprises a single term comprised from a label that is duplicated within the topic tree;

requiring that a topic model includes no term from the label for the topic to which the topic model belongs; and

when the label is duplicated within the topic tree, requiring that a topic model includes no term from the label of a parent topic for the topic to which the topic model belongs;

for each topic in the topic tree, creating a topic model subject to the hard constraints, the topic model comprising a finite state pattern that comprises a pattern evaluable against the articles;

evaluating the topic models for the topic tree against the hard constraints and disfavoring those topic models that violate one or more of the immutable rules; and

identifying for each topic, the topic model, which best satisfies the constraints.

2. A method according to claim 1 , further comprising:

defining soft constraints to include permissive rules; and

discounting those topic models that violate one or more of the permissive rules.

3. A method according to claim 2 , wherein the permissive rules comprise one or more of:

forming one such topic model that includes no term from the label for the topic to which the topic model belongs that is not duplicated in any other label within the topic tree;

when the label is duplicated in another label within the topic tree, forming one such topic model that includes no non-duplicated term from the label of a parent topic for the topic to which the topic model belongs; and

when the label is duplicated in another label within the topic tree, forming one such topic model that includes a term from a label of a common ancestor topic for the topic to which the topic model belongs.

4. A method according to claim 1 , further comprising:

counting the articles that are matched by each of the topic models; and

favoring those topic models proportionately matching a percentage of the articles closest to an ideal percentage.

5. A method according to claim 1 , further comprising:

for each of the topic models, evaluating structural complexity, comprising one or more of:

favoring those topic models that include at least one n-gram;

favoring those topic models that include one or more non-duplicated terms; and

favoring those topic models that include at least one group of the terms in the labels.

6. A method according to claim 1 , further comprising:

identifying the topic models that have been evaluated as equally satisfying the hard constraints; and

favoring the topic model that includes a bonus label comprising one of a matching single term pattern, conjunction of all of the terms in the label for the topic to which the topic model belongs, and an n-gram of all of the terms in the label for the topic to which the topic model belongs appearing in order.

7. A method according to claim 1 , further comprising:

for each of the topic models, disfavoring those topic models with a count of the terms that is less than a count of the terms in the label for the topic to which the topic model belongs.

8. A method according to claim 1 , wherein the at least one term comprises at least one of one of the terms in the labels, an n-gram comprised of a plurality of the terms in the labels, and a group of the terms in the labels.

9. A method according to claim 1 , wherein each topic model comprises one of a conjunction and an n-gram, which are both comprised of the same terms as in the label for the topic to which the topic model belongs.

10. A method according to claim 1 , further comprising:

deriving each term from a stemmed version of a word in one such label.

11. A computer-implemented system for providing default hierarchical training for social indexing, comprising:

an electronic database, comprising:

articles of digital information maintained for social indexing; and

a hierarchically-structured tree of topics, which each comprise a label comprising one or more words;

a processor and memory within which code for execution by the processor is stored, further comprising:

an electronically-stored rules set identifying hard constraints based on the labels comprised in the topic tree and the topic tree's hierarchical structure, wherein the hard constraints are defined to include required immutable rules comprising at least one of:

that a topic model comprises a single term comprised from a label that is duplicated within the topic tree;

that a topic model includes no term from the label for the topic to which the topic model belongs; and

when the label is duplicated within the topic tree, that a topic model includes no term from the label of a parent topic for the topic to which the topic model belongs;

a topic builder module that, for each topic in the topic tree, creates a topic model subject to the hard constraints, the topic model comprising a finite state pattern that comprises a pattern evaluable against the articles; and

an evaluator module evaluating the topic models for the topic tree against the hard constraints, and disfavoring those topic models that violate one or more of the immutable rules; and

a user interface visually identifying for each topic, the topic model, which best satisfies the constraints.

12. A system according to claim 11 , wherein the electronically-stored rules set further identify soft constraints are defined to include permissive rules, and those topic models that violate one or more of the permissive rules are discounted.

13. A system according to claim 12 , wherein the permissive rules comprise one or more of:

one such topic model that includes no term from the label for the topic to which the topic model belongs that is not duplicated in any other label within the topic tree;

one such topic model that includes no non-duplicated term from the label of a parent topic for the topic to which the topic model belongs when the label is duplicated in another label within the topic tree; and

one such topic model that includes a term from a label of a common ancestor topic for the topic to which the topic model belongs when the label is duplicated in another label within the topic tree.

14. A system according to claim 11 , wherein the processor and memory further comprise:

an article matching evaluation module counting the articles that are matched by each of the topic models, and favoring those topic models proportionately matching a percentage of the articles closest to an ideal percentage.

15. A system according to claim 11 , wherein the processor and memory further comprise:

a structural complexity evaluation module evaluating structural complexity for each of the topic models, comprising one or more of favoring those topic models that include at least one n-gram, favoring those topic models that include one or more non-duplicated terms, and favoring those topic models that include at least one group of the terms in the labels.

16. A system according to claim 11 , wherein the processor and memory further comprise:

a label evaluation module identifying the topic models that have been evaluated as equally satisfying the hard constraints, and favoring the topic model that includes a bonus label comprising one of a matching single term pattern, conjunction of all of the terms in the label for the topic to which the topic model belongs, and an n-gram of all of the terms in the label for the topic to which the topic model belongs appearing in order.

17. A system according to claim 11 , wherein the processor and memory further comprise:

a term count evaluation module disfavoring, for each of the topic models, those topic models with a count of the terms that is less than a count of the terms in the label for the topic to which the topic model belongs.

18. A system according to claim 11 , wherein the at least one term comprises at least one of one of the terms in the labels, an n-gram comprised of a plurality of the terms in the labels, and a group of the terms in the labels.

19. A system according to claim 11 , wherein each topic model comprises one of a conjunction and an n-gram, which are both comprised of the same terms as in the label for the topic to which the topic model belongs.

20. A system according to claim 11 , wherein the processor and memory further comprise:

a word stemming module deriving each term from a stemmed version of a word in one such label.

Assignments (8)

Continuity (1)