Canonicalizing search queries to natural language questions

1. A method implemented by one or more processors, the method comprising:

determining, based on historical data indicating proximities of query submissions, a related search query for a given search query;

generating a well-formed variant, of the related search query, by processing the related search query using a trained canonicalization model, wherein the well-formed variant differs from the related search query;

defining a mapping between the given search query and the well-formed variant generated by processing the related search query using the trained canonicalization model;

subsequent to defining the mapping, and in response to a submission of the given search query via a client device:

causing, based on the mapping previously defined between the given search query and the well-formed variant, a selectable version of the well-formed variant to be rendered by the client device in response to the submission; and

in response to selection, via the client device, of the selectable version of the well-formed variant, providing the related search query to a search system to generate one or more corresponding search results for the related search query.

2. The method of claim 1 , further comprising:

determining that the related search query is not well-formed;

wherein generating the well-formed variant of the related search query and defining the mapping between the given search query and the well-formed variant are performed in response to determining that the related search query is not well-formed.

3. The method of claim 2 , wherein the related search query is not grammatically correct and the well-formed variant is grammatically correct.

4. The method of claim 2 , wherein the related search query is not a question and the well-formed variant is a question.

5. The method of claim 1 , wherein the trained canonicalization model is a sequence to sequence model.

6. The method of claim 1 , wherein the trained canonicalization model includes an encoder portion that is a first recurrent neural network and includes a decoder portion that is a second recurrent neural network.

7. A method implemented by one or more processors, the method comprising:

receiving a search query that is a natural language search query generated at a client device responsive to user interface input received at the client device;

prior to attempting to generate any well-formed variant for the search query, and prior to a search being performed for the search query:

processing features of the search query, using a trained classification machine learning model, to generate output comprising a probability indicating an extent to which the search query conforms to one or more grammar rules; and

determining, based on a magnitude of the probability of the output, whether the search query is well-formed;

in response to determining the search query is not well-formed:

generating a well-formed variant of the search query by processing features of the search query using a trained canonicalization machine learning model;

providing the well-formed variant to a search system to generate one or more search results corresponding to the well-formed variant; and

causing, responsive to receiving the search query, the one or more search results, that correspond to the well-formed variant, to be rendered via the client device;

in response to determining the search query is well-formed:

providing the search query to the search system to generate one or more other search results corresponding to the search query; and

causing, responsive to receiving the search query, the one or more other search results to be rendered via the client device.

8. The method of claim 7 , wherein the features of the search query, processed using the trained classification machine learning model to generate the output, comprise one or more of: one or more characters in the search query, one or more words in the search query, or one or more parts of speech in the search query.

9. The method of claim 7 , wherein the features of the search query, processed using the trained classification machine learning model to generate the output, comprise one or more of: one or more character n-grams, one or more word n-grams, or one or more part of speech n-grams.

10. The method of claim 7 , wherein the probability, of the output, is a value between zero and one.

11. The method of claim 7 , wherein the trained canonicalization machine learning model is a sequence to sequence model.

12. The method of claim 7 , wherein the trained canonicalization machine learning model is trained by:

training the trained canonicalization machine learning model based on a plurality of canonicalization training instances that each includes a corresponding first query which is not well-formed and a corresponding second query which is well-formed.

13. The method of claim 7 , wherein the trained classification machine learning model is trained by:

training the trained classification machine learning model on a plurality of classification training instances that each includes a corresponding input query and a corresponding indication of whether the corresponding input query is well-formed.

14. The method of claim 7 , wherein the search system is remote from the client device and providing the well-formed variant to the search system to generate the one or more search results corresponding to the well-formed variant comprises:

transmitting the well-formed variant to the search system remote from the client device; and

receiving the one or more search results from the search system remote from the client device.

15. A method implemented by one or more processors, the method comprising:

receiving a search query that is a natural language search query generated at a client device responsive to user interface input received at the client device;

prior to attempting to generate any well-formed variant for the search query, and prior to a search being performed for the search query:

processing features of the search query using a trained classification machine learning model, to generate output comprising a probability indicating an extent to which the search query conforms to one or more grammar rules; and

determining, based on a magnitude of the probability of the output, whether the search query is well-formed;

in response to determining the search query is not-well formed:

generating a well-formed variant of the search query by processing features of the search query using a trained canonicalization machine learning model;

causing, responsive to receiving the search query, the client device to render:

an indication the search query is not well-formed, and

the well-formed variant;

in response to determining the search query is well formed:

providing the search query to the search system to generate one or more search results corresponding to the search query; and

causing, responsive to receiving the search query, the one or more search results to be rendered via the client device.

16. The method of claim 15 , wherein the features of the search query, processed using the trained classification machine learning model to generate the output, comprise one or more words in the search query and one or more parts of speech in the search query.

17. The method of claim 15 , wherein the probability, of the output, is a value between zero and one.

18. The method of claim 15 , wherein the trained canonicalization machine learning model is a sequence to sequence model.

19. The method of claim 15 , wherein the trained classification machine learning model is trained by:

training the trained classification machine learning model on a plurality of classification training instances that each includes a corresponding input query and a corresponding indication of whether the corresponding input query is well-formed.