Training a question-answer dialog system to avoid adversarial attacks

1. A method for detecting an information operations campaign, the method comprising:

A method comprising:

generating, by a computing device, a first adversarial statement via:

receiving a question and an original context for the question;

converting the question into a statement with a placeholder answer;

picking randomly an answer entity from a training text corpus;

replacing the placeholder answer with the randomly picked answer entity; and

leaving a correct question entity in the statement;

inserting, by the computing device, the first adversarial statement into the original context to form a first adversarial context;

inputting, by the computing device, the question and the first adversarial context as a first pair and the question and the original context as a second pair into a question-answer dialog system to train the question-answer dialog system; and

using the trained question-answer dialog system to automatically respond to a further question that is received.

2. The method of claim 1 , further comprising determining, by the computing device, an entity type of an answer to the question, wherein the picking of the answer entity for the adversarial statement from the training text corpus occurs from entities in the training text corpus having the determined entity type.

3. The method of claim 2 , wherein the determining of the entity type comprises the computing device predicting the entity type in a non-adversarial setting.

4. The method of claim 2 , wherein the determined entity type is a type of entity of a correct answer to the question.

5. The method of claim 1 , wherein the original context includes a correct answer to the question, the correct answer is disposed in an answer position within the original context, and the answer position is input into the question-answer dialog system to train the question-answer dialog system.

6. The method of claim 5 , wherein the answer position is input into the question-answer dialog system as a group together with the second pair.

7. The method of claim 1 , wherein the placeholder answer is disposed in a candidate answer position within the first adversarial context, and the candidate answer position is input into the question-answer dialog system to train the question-answer dialog system.

8. The method of claim 7 , wherein the candidate answer position is input into the question-answer dialog system as a group together with the first pair.

9. The method of claim 1 , wherein the first adversarial context includes a correct answer to the question.

10. The method of claim 1 , wherein in response to receiving a new question and a new context for answering the new question, the trained question-answer dialog system provides a new answer for the new question and provides a predicted position of the new answer within the new context.

11. The method of claim 1 , wherein the original context is in a first language, and the first adversarial statement is in a second language that is different from the first language.

12. A computer program product comprising a computer readable storage medium having program code embodied therewith, wherein the computer readable storage medium is not a transitory signal per se, wherein the program code is readable and executable by a processor to perform a method of training a question-answer dialog system to avoid adversarial attacks, and wherein the method comprises:

generating a first adversarial statement via:

receiving a question and an original context for the question;

converting the question into a statement with a placeholder answer;

picking randomly an answer entity from a training text corpus;

replacing the placeholder answer with the randomly picked answer entity; and

leaving a correct question entity in the statement;

inserting the first adversarial statement into the original context to form a first adversarial context;

inputting the question and the first adversarial context as a first pair and the question and the original context as a second pair into a question-answer dialog system to train the question-answer dialog system; and

using the trained question-answer dialog system to automatically respond to a further question that is received.

13. The computer program product of claim 12 , wherein the method further comprises determining an entity type of an answer to the question, wherein the picking of the answer entity for the adversarial statement from the training text corpus occurs from entities in the training text corpus having the determined entity type.

14. The computer program product of claim 12 , wherein the original context includes a correct answer to the question, the correct answer is disposed in an answer position within the original context, and the answer position is input into the question-answer dialog system to train the question-answer dialog system.

15. The computer program product of claim 12 , wherein the placeholder answer is disposed in a candidate answer position within the first adversarial context, and the candidate answer position is input into the question-answer dialog system to train the question-answer dialog system.

16. The computer program product of claim 12 , wherein in response to receiving a new question and a new context for answering the new question, the trained question-answer dialog system provides a new answer for the new question and provides a predicted position of the new answer within the new context.

17. A computer system comprising one or more processors, one or more computer readable memories, and one or more computer readable non-transitory storage mediums, and program instructions stored on at least one of the one or more computer readable non-transitory storage mediums for execution by at least one of the one or more processors via at least one of the one or more computer readable memories, the stored program instructions executed to perform a method comprising:

generating a first adversarial statement via:

receiving a question and an original context for the question;

converting the question into a statement with a placeholder answer;

picking randomly an answer entity from a training text corpus;

replacing the placeholder answer with the randomly picked answer entity; and

leaving a correct question entity in the statement;

inserting the first adversarial statement into the original context to form a first adversarial context;

inputting the question and the first adversarial context as a first pair and the question and the original context as a second pair into a question-answer dialog system to train the question-answer dialog system; and

using the trained question-answer dialog system to automatically respond to a further question that is received.

18. The computer system of claim 17 , wherein the method further comprises determining an entity type of an answer to the question, wherein the picking of the answer entity for the adversarial statement from the training text corpus occurs from entities in the training text corpus having the determined entity type.

19. The computer system of claim 17 , wherein the original context includes a correct answer to the question, the correct answer is disposed in an answer position within the original context, and the answer position is input into the question-answer dialog system to train the question-answer dialog system.

20. The computer system of claim 17 , wherein in response to receiving a new question and a new context for answering the new question, the trained question-answer dialog system provides a new answer for the new question and provides a predicted position of the new answer within the new context.