Training natural language system with generated dialogues

1. A method for generating training data for training a natural language processing system, comprising:

loading, into a computer memory, a computer-readable transcript representing an ordered sequence of one or more dialogue events, wherein a dialogue event includes a client utterance or a computerized assistant response;

acquiring a computer-readable command parametrized by a seed semantic parameter and describing an exemplary ordered subsequence of one or more dialogue events from the computer-readable transcript, the seed semantic parameter indicating a semantic situation described in the computer-readable transcript;

acquiring an alternative semantic parameter differing from the seed semantic parameter, the alternative semantic parameter indicating a different semantic situation not described in the computer-readable transcript, and re-parametrizing the computer-readable command by replacing the seed semantic parameter with the alternative semantic parameter;

generating an alternative ordered subsequence of one or more dialogue events based on the computer-readable command and the alternative semantic parameter, the alternative ordered subsequence of one or more dialogue events differing from the exemplary ordered subsequence of one or more dialogue events; and

outputting, to a data store, an alternative computer-readable transcript including the alternative ordered subsequence of one or more dialogue events, the alternative computer-readable transcript having a predetermined format usable to train the natural language processing system to respond to the different semantic situation.

2. The method of claim 1 , wherein the computer-readable command includes one or more sub-commands, each sub-command corresponding to a portion of the ordered sequence of one or more dialogue events.

3. The method of claim 2 , wherein re-parametrizing the computer-readable command further includes, for a focal sub-command of the one or more sub-commands, replacing a semantic parameter associated with the focal sub-command with a different, synthetic semantic parameter.

4. The method of claim 1 , wherein acquiring the computer-readable command includes:

graphically displaying the computer-readable transcript; and

receiving one or more computer inputs selecting the computer-readable command.

5. The method of claim 4 , further comprising graphically displaying a hierarchical menu including a plurality of candidate commands, wherein the one or more computer inputs selecting the computer-readable command indicate one of the candidate commands.

6. The method of claim 1 , wherein acquiring the alternative semantic parameter includes:

graphically displaying a portion of the computer-readable transcript corresponding to the seed semantic parameter; and

receiving one or more computer inputs indicating the alternative semantic parameter.

7. The method of claim 1 , wherein acquiring the alternative semantic parameter includes operating a computer model linking a computer-readable description of the seed semantic parameter to one or more candidate alternative semantic parameters.

8. The method of claim 1 , further comprising outputting, to a semantic parameter data store, a computer-readable representation of the alternative semantic parameter, addressable based on the seed semantic parameter.

9. The method of claim 8 , wherein acquiring the alternative semantic parameter includes looking up one or more candidate alternative semantic parameters in the semantic parameter data store by locating an address defined by the seed semantic parameter.

10. The method of claim 1 , wherein generating the alternative ordered subsequence of one or more dialogue events includes:

outputting an initial ordered subsequence of one or more dialogue events based on the computer-readable command and the alternative semantic parameter; and

paraphrasing the initial ordered subsequence to generate the alternative ordered subsequence.

11. The method of claim 10 , wherein paraphrasing the initial ordered subsequence includes, for a portion of the initial ordered subsequence, operating a computer model to select a candidate paraphrase of the portion.

12. The method of claim 11 , wherein the computer model includes a natural language model configured to recognize a semantic relationship between a phrase and a candidate paraphrase for the phrase.

13. The method of claim 10 , wherein paraphrasing the initial ordered subsequence includes:

graphically displaying a portion of the initial ordered subsequence; and

receiving one or more computer inputs defining a candidate paraphrase of the portion.

14. The method of claim 10 , further comprising outputting, to a paraphrase data store, a computer-readable representation of a candidate paraphrase for a portion of the initial ordered subsequence, addressable based on the portion of the initial ordered subsequence.

15. The method of claim 14 , wherein paraphrasing the initial ordered subsequence includes, for a portion of the initial ordered subsequence, looking up one or more candidate paraphrases in the paraphrase data store by locating an address defined by the portion.

16. A pipeline for generating training data for training a natural language processing system, comprising:

a seed dialogue acquisition machine configured to load, from a computer memory, a computer-readable transcript representing an ordered sequence of one or more dialogue events, wherein a dialogue event includes a client utterance or a computerized assistant response;

an annotation acquisition machine configured to acquire a computer-readable command parametrized by a seed semantic parameter and describing an exemplary ordered subsequence of one or more dialogue events from the computer-readable transcript, the seed semantic parameter indicating a semantic situation described in the computer-readable transcript;

a synthetic data generation machine configured to:

generate an alternative semantic parameter differing from the seed semantic parameter, the alternative semantic parameter indicating a different semantic situation not described in the computer-readable transcript, and to re-parametrize the computer-readable command by replacing the seed semantic parameter with the alternative semantic parameter;

generate an alternative ordered subsequence of one or more dialogue events based on the computer-readable command and the alternative semantic parameter, the alternative ordered subsequence of one or more dialogue events differing from the exemplary ordered subsequence of one or more dialogue events; and

output, to a data store, an alternative computer-readable transcript including the alternative ordered subsequence of one or more dialogue events, the alternative computer-readable transcript having a predetermined format usable to train the natural language processing system to respond to the different semantic situation.

17. The training pipeline of claim 16 , further comprising a deployed computerized assistant machine configured to interact with a client and to store, into a historical agent log store of the computer memory, a computer-readable transcript describing the interaction.

18. The training pipeline of claim 17 , further comprising a training machine configured to train the deployed computerized assistant machine based at least on the alternative computer-readable transcript including the alternative ordered sequence of one or more dialogue events, the one or more dialogue events including alternative client utterances and corresponding alternative computerized assistant responses.

19. The training pipeline of claim 16 , wherein outputting the alternative ordered subsequence in the synthetic data generation machine includes:

outputting an initial ordered subsequence of one or more dialogue events based on the computer-readable command and the alternative semantic parameter; and

paraphrasing the initial ordered subsequence to generate the alternative ordered subsequence.

20. A computer system, comprising:

a logic device; and

a storage device configured to hold instructions executable by the logic device to:

load, into a computer memory, a computer-readable transcript representing an ordered sequence of one or more dialogue events, wherein a dialogue event includes a client utterance or a computerized assistant response;

acquire a computer-readable command parametrized by a seed semantic parameter and describing an exemplary ordered subsequence of one or more dialogue events from the computer-readable transcript, the seed semantic parameter indicating a semantic situation described in the computer-readable transcript;

acquire an alternative semantic parameter differing from the seed semantic parameter, the alternative semantic parameter indicating a different semantic situation not described in the computer-readable transcript, and re-parametrize the computer-readable command by replacing the seed semantic parameter with the alternative semantic parameter;

generate an alternative ordered subsequence of one or more dialogue events based on the computer-readable command and the alternative semantic parameter, the alternative ordered subsequence of one or more dialogue events differing from the exemplary ordered subsequence of one or more dialogue events; and

output, to a data store, an alternative computer-readable transcript including the alternative ordered subsequence of one or more dialogue events, the alternative computer-readable transcript having a predetermined format usable to train the natural language processing system to respond to the different semantic situation.