Neural typographical error modeling via generative adversarial networks

1. A non-transitory computer-readable storage medium storing one or more programs for detecting a typographical error in a user input, the one or more programs comprising instruction, which when executed by one or more processors of an electronic device, cause the electronic device to:

receive the user input including one or more words;

display the user input;

determine, using a trained first learning network, whether the user input includes a typographical error,

wherein the first learning network is trained in a supervised manner based on a data set generated by a second learning network, and

wherein the second learning network is trained in an unsupervised manner based on input words and training data that includes a collection of words having typographical errors, wherein the second learning network includes a neural network; and

in accordance with a determination that the user input includes a typographical error, correct the displayed user input.

2. The non-transitory computer-readable storage medium of claim 1 , wherein the data set generated by the second learning network is generated using one or more programs comprising instruction, which when executed by one or more processors of a second electronic device, cause the second electronic device to:

receive one or more input words;

extract, based on the one or more input words, seed data for unsupervised training of the second learning network;

obtain the training data that includes a collection of words having typographical errors;

determine, using the second learning network and based on the seed data and the training data, one or more output words having a probability distribution corresponding to a probability distribution of the training data, wherein the one or more output words include typographical errors; and

generate, based on the determined one or more output words, the data set for supervised training of the first learning network.

3. The non-transitory computer-readable storage medium of claim 2 , wherein extracting, based on the one or more input words, the seed data for unsupervised training of the second learning network comprises:

obtaining an input character sequence corresponding to each of the one or more input words;

encoding the input character sequence; and

determining, based on the encoded input character sequence, a vector representing at least a portion of the seed data, wherein the vector is encoded with contextual data associated with the one or more input words.

4. The non-transitory computer-readable storage medium of claim 2 , wherein determining the one or more output words having a probability distribution corresponding to the probability distribution of the training data comprises:

performing an unsupervised training of the second learning network based on the seed data and the training data, wherein the second learning network is a generative adversarial network comprising a generator and a discriminator; and

determining based on the training results of the second learning network, the one or more output words having a probability distribution corresponding to the probability distribution of the training data.

5. The non-transitory computer-readable storage medium of claim 4 , wherein performing the unsupervised training of the second learning network based on the seed data and the training data comprises:

determining a first probability distribution associated with the training data; and

in each iteration of the unsupervised training,

generating, by the generator, representations of a generated character sequence representing a word having one or more typographical errors;

determining a second probability distribution associated with the representations of a generated character sequence; and

determining whether the unsupervised training is completed based on the first and second probability distributions.

6. The non-transitory computer-readable storage medium of claim 1 , wherein the training data that includes a collection of words having typographical errors comprises a plurality of words collected from a plurality of users, each word of the plurality of words including a typographical error made by one of the plurality of users.

7. The non-transitory computer-readable storage medium of claim 1 , wherein a typographical error includes at least one of a non-atomic typographical error or an atomic typographical error, wherein a word having a non-atomic typographical error is lexically incorrect and a word having an atomic typographical error is lexically correct but contextually incorrect.

8. The non-transitory computer-readable storage medium of claim 7 , wherein determining whether the user input includes a typographical error comprises:

determining, based on at least one of the user input or a context of the user input, whether the typographical error is a non-atomic typographical error or an atomic typographical error.

9. The non-transitory computer-readable storage medium of claim 8 , wherein the context of the user input comprises at least one of words preceding a current word of the user input or words following the current word of the user input.

10. The non-transitory computer-readable storage medium of claim 1 , wherein correcting the displayed user input comprises:

determining, using a typographical error model associated with the trained first learning network, a plurality of candidate words absent typographical errors;

providing the plurality of candidate words absent typographical errors to the user; receiving, from the user, a selection of one candidate word of the plurality of candidate words; and

in response to receiving the user selection, displaying a corrected user input.

11. The non-transitory computer-readable storage medium of claim 10 , wherein displaying the corrected user input comprises:

deleting the displayed user input; and

displaying the selected candidate word.

12. The non-transitory computer-readable storage medium of claim 10 , wherein determining whether the user input includes a typographical error comprises determining, using the typographical error model associated with the trained first learning network, one or more incorrect characters of a word of the user input.

13. The non-transitory computer-readable storage medium of claim 12 , wherein displaying the corrected user input comprises:

deleting the one or more incorrect characters of the word of the user input; and

inserting one or more correct characters into the word of the user input.

14. The non-transitory computer-readable storage medium of claim 1 , wherein correcting the displayed user input comprises:

determining, using a typographical error model associated with the trained first learning network, a plurality of candidate words absent typographical errors;

determining a ranking of the plurality of candidate words absent typographical errors; and

correcting the displayed user input with the highest ranked candidate word.

15. The non-transitory computer-readable storage medium of claim 14 , wherein the ranking of the plurality of candidate words is based on the popularity of the candidate words.

16. The non-transitory computer-readable storage medium of claim 14 , wherein the ranking of the plurality of candidate words is based on the context associated with the user input.

17. The non-transitory computer-readable storage medium of claim 14 , wherein correcting the displayed user input with the highest ranked candidate word comprises:

deleting at least portion of the displayed user input; and

displaying the highest ranked candidate word.

18. The non-transitory computer-readable storage medium of claim 14 , wherein determining whether the user input includes a typographical error comprises determining, using a typographical error model associated with the trained first learning network, one or more incorrect characters of a word of the user input.

19. The non-transitory computer-readable storage medium of claim 18 , wherein correcting the displayed user input with the highest ranked candidate word comprises:

deleting the one or more incorrect characters of the word of the user input; and

inserting one or more correct characters into the word of the user input.

20. A method for detecting a typographical error in a user input, comprising:

at an electronic device with one or more processors and memory:

receiving the user input including one or more words;

displaying the user input;

determining, using a trained first learning network, whether the user input includes a typographical error,

wherein the first learning network is trained in a supervised manner based on a data set generated by a second learning network, and

wherein the second learning network is trained in an unsupervised manner based on input words and training data that includes a collection of words having typographical errors, wherein the second learning network includes a neural network; and

in accordance with a determination that the user input includes a typographical error, correcting the displayed user input.

21. An electronic device, comprising:

one or more processors;

memory; and

one or more programs stored in memory, the one or more programs including instructions for:

receiving a user input including one or more words;

displaying the user input;

determining, using a trained first learning network, whether the user input includes a typographical error,

wherein the first learning network is trained in a supervised manner based on a data set generated by a second learning network, and

wherein the second learning network is trained in an unsupervised manner based on input words and training data that includes a collection of words having typographical errors, wherein the second learning network includes a neural network; and

in accordance with a determination that the user input includes a typographical error, correcting the displayed user input.