System and method for providing multi-core and multi-level topical organization in social indexes

• US 9,031,944 B2
• Filed: 04/30/2010
• Issued: 05/12/2015
• Est. Priority Date: 04/30/2010
• Status: Expired due to Fees

• Current Assignee: Palo Alto Research Center, Inc. (Xerox Holdings Corp.)
• Original Assignee: Palo Alto Research Center, Inc. (Xerox Holdings Corp.)
• Inventors: Stefik, Mark Jeffrey, Good, Lance E., Mittal, Sanjay
• Primary Examiner(s): Pulliam, Christyann
• Assistant Examiner(s): Ohba, Mellissa M
• Application Number:
  US12/771,816
• Publication Number:
  US 20110270830A1
• Time in Patent Office: 1,838 Days
• US Class Current: 707/731
• CPC Class Codes: G06F 16/353 into predefined classes

Abstract

A computer-implemented method affords multi-core and multi-level topical organization in social indexes. A corpus of articles is accessed. Each article includes online textual materials. A finite state pattern for a topic filters the articles as candidate articles, which are potentially on-topic. Similarity-based representations are formed for on-topic and off-topic core meanings of the topic. An aggregate score for each of the candidate articles is determined using the similarity-based representations to indicate whether the candidate article is sufficiently on-topic. The candidate articles are presented ordered by their aggregate scores. In a further embodiment, a hierarchy of topics is used to guide the presentation of articles from subtopics, with considerations of fairness of subtopic coverage, elimination of similarity-duplicates in articles, and article freshness.

Claims

1. A computer-implemented system for providing multi-core topic indexing in electronically-stored social indexes, comprising:

a storage device comprising;

a corpus of articles each comprised of online textual materials and a topics;

a finite state pattern for each topic, each finite state pattern defining a fine-grained topic model that is used to identify the articles that are potentially on-topic; and

on-topic training examples and off-topic training examples from the articles for each topic;

one or more of distinct core meanings for the topic by assigning at least one of the on-topic training examples and the off-topic training examples;

a set of average on-topic articles, comprising;

a training module configured to provide a set of random training examples from the corpus;

a match module configured to match the set of random training examples to the finite state pattern for the topic;

an off-topic elimination module configured to eliminate an article that is similar to the off-topic training examples; and

an on-topic addition module configured to add the on-topic training examples into the set of the random training examples; and

an average on-topic core meaning based on the set of the average on-topic articles;

a social indexing system, comprising;

a characteristic words selector configured to specify characteristic words for each of the on-topic training examples, the off-topic training examples, and the set of average on-topic articles, and to assign scores to the characteristic words that were specified for the on-topic training examples, off-topic training examples, and the set of average on-topic articles;

a characteristic words organizer configured to specify on-topic characteristic word term vectors, each on-topic characteristic word term vector comprising the scores of the characteristic words that were specified for each topic for each of the on-topic training examples;

a characteristic words scorer configured to specify off-topic characteristic word term vectors, each off-topic characteristic word term vector comprising the scores of the characteristic words that were specified for each topic for each of the off-topic training examples;

a characteristic words specifier configured to specify average on-topic characteristic word term vectors, each average on-topic characteristic word term vector comprising the scores of the characteristic words that were specified for each topic for the set of average on-topic articles;

an information collector configured to obtain a new article;

a finite state pattern matcher configured to match the new article to the finite state pattern of each of the topics to designate the new article as a candidate article for each topic to which the finite state pattern was matched;

a candidate article characteristic words selector configured to specify characteristic words extracted from the candidate article;

a candidate article characteristic words scorer configured to assign candidate article scores to the characteristic words of the candidate article;

a topic comparer configured to compare the candidate article scores to the off-topic characteristic word term vectors of each topic and to form an off-topic score for each topic, and to discard the candidate article as off-topic for each topic in which the off-topic score for that topic exceeds an off-topic threshold; and

a similarity score comparer configured to compare the candidate article scores to the on-topic characteristic word term vectors and the average on-topic characteristic word term vectors of each topic and to form an on-topic score for each topic and configured to select only the candidate articles as candidate on-topic articles which the on-topic score for that topic exceeds an on-topic threshold; and

a display configured to present the candidate on-topic articles.

2. A computer-implemented method for providing multi-core topic indexing in electronically-stored social indexes, comprising:

accessing a corpus of articles each comprised of online textual materials and a tree of topics;

providing a finite state pattern for each topic, each finite state pattern defining a fine-grained topic model that is used to identify the articles that are potentially on-topic;

providing on-topic training examples and off-topic training examples from the articles for each topic;

defining one or more of distinct core meanings for the topic by assigning at least one of the on-topic training examples and the off-topic training examples;

obtaining a set of an average on-topic articles, comprising;

providing a set of random training examples from the corpus;

matching the set of random training examples to the finite state pattern for the topic;

eliminating an article in the set of random training examples that is similar to the off-topic training examples; and

adding the on-topic training examples into the set of the random training examples;

defining an average on-topic core meaning based on the set of average on-topic articles;

specifying characteristic words for each of the on-topic training examples, off-topic training examples, and the set of average on-topic articles;

assigning scores to the characteristic words that were specified for the on-topic training examples, off-topic training examples, and the set of average on-topic articles;

specifying on-topic characteristic word term vectors, each on-topic characteristic word term vector comprising the scores of the characteristic words that were specified for each topic for each of the on-topic training examples;

specifying off-topic characteristic word term vectors, each off-topic characteristic word term vector comprising the scores of the characteristic words that were specified for each topic for each of the off-topic training examples;

specifying average on-topic characteristic word term vectors, each average on-topic characteristic word term vector comprising the scores of the characteristic words that were specified for each topic for the set of average on-topic articles;

obtaining a new article;

matching the new article to the finite state pattern of each of the topics and designating the new article as a candidate article for each topic to which the finite state pattern was matched;

specifying characteristic words extracted from the candidate article;

assigning candidate article scores to the characteristic words of the candidate article;

comparing the candidate article scores to the off-topic characteristic word term vectors of each topic and forming an off-topic score for each topic;

discarding the candidate article as off-topic for each topic in which the off-topic score for that topic exceeds an off-topic threshold;

comparing the candidate article scores to the on-topic characteristic word term vectors and the average on-topic characteristic word term vectors of each topic and forming an on-topic score for each topic and selecting only the candidate articles as candidate on-topic articles which the on-topic score for that topic exceeds an on-topic threshold; and

presenting the candidate on-topic articles.

3. A computer-implemented method according to claim 2, further comprising at least one of:

automatically generating the finite state pattern based on the on-topic and the off-topic training examples;

generating the finite state pattern through evaluation of default patterns;

suggesting the finite state pattern to the user based on an evaluation of the articles in the topic; and

designating the finite state pattern as a discrete Boolean query specified by a user.

4. A computer-implemented method according to claim 3, further comprising:

selecting the on-topic training examples from the corpus;

selecting the off-topic training examples from the corpus; and

defining the finite state pattern to match a maximum number of the on-topic training examples and to not match a maximum number of the off-topic training examples.

5. A computer-implemented method according to claim 2, further comprising:

specifying a label for each of the topics, each label comprising one or more words;

identifying constraints inherent in the literal structure of the topic tree;

for each topic in the topic tree, creating a candidate topic model that comprises at least one term derived from the words in at least one of the labels;

evaluating the candidate topic models for the topic tree against the constraints; and

identifying the candidate topic model, which best satisfy the constraints, as the finite state pattern for the topic.

6. A computer-implemented method according to claim 2, further comprising:

aggregating the scores of the on-topic characteristic word term vectors for each topic into a set of on-topic scores at a topic center;

finding topic center similarity scores by comparing the candidate article scores to the set of the on-topic scores at the topic center of each topic;

weighting one or more of the on-topic training examples and the off-topic training examples; and

proportioning the topic center similarity scores based on the respective weightings of the on-topic training examples and the off-topic training examples.

7. A computer-implemented method according to claim 2, further comprising:

computing the topic center similarity score by applying a cosine function to weights assigned to the characteristic words identified in each of the candidate articles, the on-topic training examples, and the off-topic training examples.

8. A computer-implemented method according to claim 2, further comprising at least one of:

presenting the candidate articles without filtering through the plurality of the finite state patterns of the topics; and

presenting only the candidate articles comprised of an aggregate score meeting a minimum threshold.