Generating training data for machine learning

Patent metadata

Field	Value
Publication number	US-10679144-B2
Application number	US-201615207918-A
Country	US
Kind code	B2
Filing date	Jul 12, 2016
Priority date	Jul 12, 2016
Publication date	Jun 9, 2020
Grant date	Jun 9, 2020

Abstract

A computer-implemented method includes receiving a rule, wherein the rule includes at least one token, and receiving at least two dictionaries, wherein the at least two dictionaries include at least one general language dictionary and at least one domain-specific dictionary for a domain. The computer-implemented method further includes, for each of the at least one token, selecting at least one word at random from at least one of the at least two dictionaries and adding the at least one word to a test data line, such that the test data line includes a candidate statement conforming to the rule. The computer-implemented method further includes filtering the candidate statement based on a domain-specific model for the domain and including the candidate statement in training data provided to a machine learning model. A corresponding computer program product and computer system are also disclosed.

First claim

What is claimed is:

1. A computer program product comprising one or more computer readable storage media and program instructions stored on said one or more computer readable storage media, said program instructions comprising instructions to:
receive a rule, wherein said rule comprises at least one token;
receive at least two dictionaries, wherein said at least two dictionaries comprise at least one general language dictionary and at least one domain-specific dictionary for a domain;
for each of said at least one token, select at least one word at random from at least one of said at least two dictionaries and add said at least one word to a test data line, such that said test data line comprises a candidate statement conforming to said rule;
filter said candidate statement based on a domain-specific model for said domain; and
include said candidate statement in training data provided to a machine learning model.

2. The computer program product of claim 1 , wherein said program instructions further comprise instructions to insert at least one additional word randomly selected from at least one of said at least two dictionaries into said test data line.

3. The computer program product of claim 1 , wherein said instructions to filter said candidate statement comprise instructions to discard said candidate statement, if said candidate statement fails to meet a definition of semantically correct candidate statements for said domain, according to said domain-specific model.

4. The computer program product of claim 3 , wherein said domain-specific model is based on a general corpus within said domain.

5. The computer program product of claim 1 , wherein said rule is expressed using regular expressions.

6. The computer program product of claim 1 , wherein said rule is expressed as a state machine.

7. The computer program product of claim 1 , wherein said rule encodes engineered knowledge of a human expert.

8. A computer system comprising:
one or more processors;
one or more computer readable storage media;
computer program instructions;
said computer program instructions being stored on said one or more computer readable storage media;
said computer program instructions comprising instructions to:
receive a rule, wherein said rule comprises at least one token;
receive at least two dictionaries, wherein said at least two dictionaries comprise at least one general language dictionary and at least one domain-specific dictionary for a domain;
for each of said at least one token, select at least one word at random from at least one of said at least two dictionaries and add said at least one word to a test data line, such that said test data line comprises a candidate statement conforming to said rule;
filter said candidate statement based on a domain-specific model for said domain; and
include said candidate statement in training data provided to a machine learning model.

9. The computer system of claim 8 , wherein said program instructions further comprise instructions to insert at least one additional word randomly selected from at least one of said at least two dictionaries into said test data line.

10. The computer system of claim 8 , wherein said instructions to filter said candidate statement comprise instructions to discard said candidate statement, if said candidate statement fails to meet a definition of semantically correct candidate statements for said domain, according to said domain-specific model.

11. The computer system of claim 10 , wherein said domain-specific model is based on a general corpus within said domain.

12. The computer system of claim 10 , wherein said domain-specific model is based on a corpus that excludes user-specific information.

13. The computer system of claim 10 , wherein said domain-specific model is an n-gram model of domain-specific statements.

14. The computer system of claim 10 , wherein said domain is medical diagnosis.

15. The computer system of claim 14 , wherein said domain-specific model is based on a general medical corpus.

16. The computer system of claim 14 , wherein said domain-specific model is based on a corpus that excludes medical patient records.

17. The computer system of claim 14 , wherein said domain-specific model is an n-gram model of medical diagnosis statements.

18. The computer system of claim 8 , wherein said rule is expressed using regular expressions.

19. The computer system of claim 8 , wherein said rule is expressed as a state machine.

20. The computer system of claim 8 , wherein said rule encodes engineered knowledge of a human expert.

Assignees

• IBM

Inventors

• Fink Patrick W
• Mcneil Kristin E
• Parker Philip E
• Werts David B

Classifications

• G06F40/242

  Dictionaries · CPC title

• G06N5/022

  Knowledge engineering; Knowledge acquisition · CPC title

• G06N20/00 Primary

  Machine learning · CPC title