Joint approach to feature and document labeling

• US 10,055,479 B2
• Filed: 01/12/2015
• Issued: 08/21/2018
• Est. Priority Date: 01/12/2015
• Status: Expired due to Fees

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Clinchant, Stephane, Bouchard, Guillaume
• Primary Examiner(s): Pham, Michael
• Application Number:
  US14/594,622
• Publication Number:
  US 20160203209A1
• Time in Patent Office: 1,317 Days
• Field of Search: 707738
• CPC Class Codes: G06F 16/285 Clustering or classification
• CPC Class Codes: G06F 16/353 into predefined classes
• CPC Class Codes: G06N 20/00 Machine learning

Abstract

Documents of a set of documents are represented by bag-of-words (BOW) vectors. L labeled topics are provided, each labeled with a word list comprising words of a vocabulary that are representative of the labeled topic and possibly a list of relevant documents. Probabilistic classification of the documents generates for each labeled topic a document vector whose elements store scores of the documents for the labeled topic and a word vector whose elements store scores of the words of the vocabulary for the labeled topic. Non-negative matrix factorization (NMF) is performed to generate a document-topic model that clusters the documents into k topics where k>L. NMF factors representing L topics of the k topics are initialized to the document and word vectors for the L labeled topics. In some embodiments the NMF factors representing the L topics initialized to the document and word vectors are frozen, that is, are not updated by the NMF after the initialization.

Claims

1. A system operating on a set of documents comprising text, the system comprising:

a computer;

a user interfacing device including a display and at least one user input device; and

a non-transitory storage medium storing instructions programming the computer to perform operations including;

(i) receiving, via the user interfacing device, L labeled topics each labeled with a word list comprising words representative of the labeled topic;

(ii) performing probabilistic classification of the documents of the set of documents to generate for each labeled topic of the L labeled topics a document vector whose elements store scores of the documents for the labeled topic and a word vector whose elements store scores of words of a vocabulary for the labeled topic;

(iii) performing non-negative matrix factorization (NMF) to generate a NMF model that clusters the set of documents into k topics where k>

L and the performing NMF includes initializing NMF factors representing L topics of the k topics to the document and word vectors for the L labeled topics generated in the operation (ii); and

performing data mining using the NMF model.

2. The system of claim 1 wherein the operation (ii) comprises:

(ii)(a) performing probabilistic classification of the documents of the set of documents based on text content of the documents and the word lists of the L labeled topics and generating a document vector for each labeled topic whose elements store scores of the documents for the labeled topic produced by the probabilistic classification; and

(ii)(b) based on the document vectors for the L labeled topics and the text content of the documents of the set of documents, generating a word vector for each labeled topic whose elements store scores of words of a vocabulary for the labeled topic.

3. The system of claim 2 wherein the NMF factors representing the L topics initialized to the document vectors generated in operation (ii)(a) and the word vectors generated in operation (ii)(b) are not updated by the NMF after the initialization.

4. The system of claim 2 wherein, in the operation (ii)(a), the probabilistic classification of the documents of the set of documents is performed further based on topic labels of a sub-set of the set of documents which are labeled with topic labels.

5. The system of claim 1 wherein the NMF factors representing the L topics initialized to the document and word vectors generated in operation (ii) are not updated by the NMF after the initialization.

6. The system of claim 1 wherein the probabilistic classification of operation (ii) is performed using a Naï

ve Bayes or logistic regression classifier.

7. The system of claim 1 wherein the data mining performed using the NMF model includes:

(iv) assigning at least one label to a document of interest of the set of documents based on the generated NMF model; and

(v) displaying the at least one label assigned to the document of interest on the display.

8. The system of claim 1 wherein the data mining performed using the NMF includes displaying a list of documents of the set of documents indicated by the generated NMF model as belonging to a topic of interest of the k topics.

9. A method comprising:

receiving, at a computer, D documents each represented by a bag-of-words (BOW) vector of length W over a vocabulary;

receiving, at the computer and from a human annotator, L labeled topics wherein each labeled topic is labeled with a word list comprising words of the vocabulary that are representative of the labeled topic;

performing, using the computer, probabilistic classification of the D documents to generate for each labeled topic a document vector of length D and a word vector of length W;

performing, using the computer, non-negative matrix factorization (NMF) to generate a NMF model that clusters the D documents into k topics where k>

L and the performing NMF includes initializing NMF factors representing L topics of the k topics to the document and word vectors for the L labeled topics generated by the probabilistic classification of the D documents; and

performing data mining using the NMF model including at least one of;

displaying a list of documents of the D documents indicated by the NMF model as belonging to a topic of interest of the k topics, and

receiving from a human user a selection of a document of the set of documents and displaying a list of other documents of the set of documents having most similar topic scores indicated by the NMF model.

10. The method of claim 9 wherein performing probabilistic classification comprises:

performing probabilistic classification of each document with respect to each labeled topic by computing an overlap score quantifying overlap of the BOW vector representing the document with the word list of the labeled topic,

generating a document vector of length D for each labeled topic whose elements store the overlap scores of the D documents with the word list of the labeled topic, and

generating a word vector of length W for each labeled topic based on the document vector for the labeled topic and the BOW vectors representing the D documents.

11. The method of claim 10 wherein the NMF factors representing the L topics initialized to the document vectors of length D and word vectors of length W generated by the probabilistic classification are not updated by the NMF after the initialization.

12. The method of claim 10 wherein the probabilistic classification is performed using a Naï

ve Bayes or logistic regression classifier.

13. The method of claim 10 wherein the data mining performed using the NMF model includes:

displaying, on a display device, a label assigned to a document of the D documents based on the NMF model.

14. The method of claim 10 wherein the data mining performed using the NMF model includes:

displaying, on a display device, a list of documents of the D documents indicated by the NMF model as belonging to a topic of interest of the k topics.

15. The method of claim 9 further comprising:

generating a semantic identification for at least one topic of the k topics, which is not one of the L labeled topics, based on highest-ranked words a word vector generated for the topic by the performed NMF.

16. The method of claim 9 wherein the NMF factors representing the L topics initialized to the document vectors of length D and word vectors of length W generated by the probabilistic classification are not updated by the NMF after the initialization.

17. A non-transitory storage medium storing instructions executable by a computer to operate on documents of a set of D documents each represented by a bag-of-words (BOW) vector of length W over a vocabulary by operations including:

(i) receiving L labeled topics each labeled with a word list comprising words of the vocabulary that are representative of the labeled topic;

(ii) performing probabilistic classification of the D documents to generate for each labeled topic of the L labeled topics a document vector of length D whose elements store scores of the D documents for the labeled topic and a word vector of length W whose W elements store scores of the W words of the vocabulary for the labeled topic;

(iii) generating a document-topic model by clustering the D documents into k clusters where k>

L and the clustering includes initializing L clusters of the k clusters to the L labeled topics generated in the operation (ii); and

performing data mining using the document-topic model.

18. The non-transitory storage medium of claim 17 wherein the operation (ii) comprises:

(ii)(a) performing probabilistic classification of the D documents based on the BOW vectors representing the D documents and the word lists of the L labeled topics and generating a document vector of length D for each labeled topic whose D elements store scores of the D documents for the labeled topic produced by the probabilistic classification; and

(ii)(b) generating the word vectors of length W for the L labeled topics based on the document vectors for the L labeled topics and the BOW vectors representing the D documents.

19. The non-transitory storage medium of claim 17 wherein the clustering operation (iii) comprises performing non-negative matrix factorization (NMF).

20. The non-transitory storage medium of claim 17 wherein the L clusters initialized to the L labeled topics generated in operation (ii) are not updated by the clustering after the initialization.