SYSTEM AND METHOD FOR ASSISTED DOCUMENT REVIEW

• US 20100312725A1
• Filed: 06/08/2009
• Published: 12/09/2010
• Est. Priority Date: 06/08/2009
• Status: Active Grant

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Hoppenot, Yves, PRIVAULT, Caroline, Ciriza, Victor, O'Neill, Jacki, Renders, Jean-Michel
• Granted Patent: US 8,165,974 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning
• CPC Class Codes: G06N 3/126 Evolutionary algorithms, e.g. genetic algorithms or genetic programming
• CPC Class Codes: G06N 5/043 Distributed expert systems; Blackboards
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks
• CPC Class Codes: G06Q 10/10 Office automation; Time management
• CPC Class Codes: G06Q 50/18 Legal services

Abstract

A system and method for reviewing documents are provided. A collection of documents is portioned into sets of documents for review by a plurality of reviewers. For each set, documents in the set are displayed on a display device for review by a reviewer and temporarily organized through grouping and sorting. The reviewer'"'"'s labels for the displayed documents are received. Based on the reviewer'"'"'s labels, a class from a plurality of classes is assigned to each of the reviewed documents. A classifier model stored in computer memory is progressively trained, based on features extracted from the reviewed documents in the set and their assigned classes. Prior to review of all documents in the set, a calculated subset of documents for which the classifier model assigns a class different from the one assigned based on the reviewer'"'"'s label is returned for a second review by a reviewer. Models generated from one or more other document sets can be used to assess the review of a first of the sets.

Claims

1. A method of reviewing documents comprising:

partitioning a collection of documents into sets of documents for review by a plurality of reviewers;

for each set;

organizing the documents in the set into a plurality of groups,

displaying documents in the set on a display device for review by a reviewer,

receiving the reviewer'"'"'s labels for the displayed documents,

based on the reviewer'"'"'s labels, assigning a class from a plurality of classes to each of the reviewed documents,

progressively training a classifier model stored in computer memory based on features extracted from the reviewed documents in the set and their assigned classes; and

prior to review of all documents in the set, identifying documents in the set for which the classifier model assigns a class different from the one assigned based on the reviewer'"'"'s label and returning a subset of the identified documents for a second review by a reviewer.

2. The method of claim 1, further comprising classifying the documents in a first of the sets of documents with a second classifier model trained on features of reviewed documents in a second of the sets of documents and their assigned classes.

3. The method of claim 2, further comprising identifying documents in the first set which the trained second classifier model assigns a class different from the one assigned based on the reviewer'"'"'s label.

4. The method of claim 3, further comprising measuring a quality of the review of the first set of documents based on the identified documents.

5. The method of claim 2, further comprising, for each set of documents, measuring a quality of the review of the documents in the set, the measure of quality of the review of each set being a function of at least one of:

a number of documents for which the classifier model trained on that set outputs a different class from the class assigned based on the reviewer'"'"'s label; and

a number of documents for which at least one classifier model trained on another of the sets outputs a different class from the class assigned based on the reviewer'"'"'s label.

6. The method of claim 1, wherein the returning of the subset of the documents for a second review by a reviewer includes applying a heuristic to the documents for which the classifier model assigns a class different from the one assigned based on the reviewer'"'"'s label to identify mislabeled documents to be returned.

7. The method of claim 1, further comprising classifying unlabeled documents in the set with the classifier model trained on already labeled documents in the set.

8. The method of claim 7, further comprising suggesting an unlabeled document for review by the reviewer which is determined to be similar to an already labeled document in the set.

9. The method of claim 7, further comprising, separating the unlabeled documents in the set into piles, one pile for each class, the separation being based on the classes assigned to the unlabeled documents by the trained classifier model.

10. The method of claim 9, further comprising graphically representing the piles on the display.

11. The method of claim 7, further comprising, suggesting a label for an unlabeled document being reviewed by the reviewer based on the trained classifier model.

12. The method of claim 1, wherein the organizing of the documents into groups comprises automatically clustering the documents in the set into a plurality of clusters based on similarity of extracted features and displaying graphical representations of the clusters on the display.

13. The method of claim 12, further comprising displaying, for each cluster, at least one feature of the documents which resulted in documents being clustered in the cluster.

14. The method of claim 12, further comprising displaying a selector for enabling the reviewer to select how many clusters are generated.

15. The method of claim 1, wherein the sets of documents are generated by automated clustering of the collection of documents.

16. The method of claim 15, wherein the clustering is one of unsupervised clustering or partially supervised clustering in which a portion of the documents are manually assigned to clusters.

17. The method of claim 1, wherein the display comprises a multi-touch interface.

18. The method of claim 17, wherein the multi-touch interface enables contemporaneous review and interaction with the displayed documents by a team of reviewers.

19. The method of claim 1, wherein there are at least three sets of documents and wherein at least three classifier models are trained, one classifier model for each of the sets of documents, the method including comparing the classes assigned to the documents in the first set based on the labels of the reviewer with classes assigned by at least two classifiers to the first set of documents, which classifiers have not been trained on the first set of the documents.

20. A computer program product encoding instructions which, when executed by a computer, performs the method of claim 1.

21. A reviewing system comprising:

memory which stores a set of documents for review;

a display which displays documents in the set;

a user input device for receiving information from which class labels applied to the documents by a reviewer are determined;

a classifier model which is progressively trained based on the applied class labels;

a document reviewing application stored in memory which, prior to all documents in the set being labeled, compares a class assigned by the classifier for a labeled document to the label applied by the reviewer and identifies labeled documents where the class label applied by the reviewer does not match the class assigned by the reviewer and returns, for a second review, a subset of the identified labeled documents.

22. The reviewing system of claim 21, further comprising:

an assessment component for evaluating the review which compares class labels assigned to the documents in a first of the sets of documents with classes assigned by a second classifier model trained on features of reviewed documents in a second of the sets of documents and their assigned classes.

23. A method of reviewing documents comprising:

partitioning a collection of documents into sets of documents;

for a first of the sets of documents, generating a first classifier model based on reviewer-applied labels for documents in the first set;

for a second of the sets of documents, generating a second classifier model based on reviewer-applied labels for documents in the second set; and

assessing a quality of the labels applied to the first set of documents including;

assigning classes to the documents from the first set of documents with the second classifier model, and

comparing the reviewer-applied labels of the documents in the first set with the classes assigned by the second classifier model to identify documents for which the reviewer-applied labels do not match the assigned classes.