Methods and apparatus for determining original electronic messages that contain asks

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

identifying a corpus of electronic communications stored in one or more databases, wherein a plurality of the electronic communications of the corpus each include a reply message and an original message to which the reply message is responsive;

identifying, from the corpus of electronic communications, a reply group of reply messages based on the reply messages of the reply group each being determined to include a commit in a body of the reply message;

selecting, from the corpus of electronic communications, an original group of original messages based on the reply messages of the reply group being responsive to the original messages of the original group;

generating positive training examples based on the original messages of the original group, each of the positive training examples including training example input based on original message features of a corresponding one of the original messages of the original group and including training example output that indicates presence of the commit;

training a machine learning model based on the positive training examples;

identifying a given electronic communication of the electronic communications of the corpus based on the original message of the given electronic communication not being included in the original group of the original messages;

applying, as input to the trained machine learning model, given message features of the original message of the given electronic communication;

determining, based on output generated over the trained machine learning model based on the input, that the commit is present in the reply message of the given electronic communication that is responsive to the original message of the given electronic communication;

based on determining that the commit is present in the reply message of the given electronic communication, using reply message features of the reply message to select an additional reply group of reply messages that each include the commit;

selecting an additional original group of the original messages based on the reply messages of the additional reply group being responsive to the original messages of the additional original group;

generating additional positive training examples based on the original messages of the additional original group; and

further training the machine learning model based on the additional positive training examples.

2. The method of claim 1 , further comprising:

identifying a new original message sent to a user;

applying, as input to the further trained machine learning model, new original message features of the new original message; and

determining, based on output generated over the further trained machine learning model based on the input, that the commit is likely to be present in a yet to be formulated new reply message that is responsive to the new original message.

3. The method of claim 2 , further comprising:

in response to determining that the commit is likely to be present in the yet to be formulated new reply message that is responsive to the new original message:

providing, for presentation to the user via a computing device of the user, an indication that the commit is likely to be present in the new reply message.

4. The method of claim 3 , wherein the indication is a visual indication applied to a graphical representation of the new original message.

5. The method of claim 4 , wherein the visual indication is applied to the graphical representation of the new original message when the new original message is presented along with graphical representations of one or more additional messages, and wherein the graphical representation visually distinguishes the new original message from the one or more additional messages.

6. The method of claim 3 , wherein the indication is provided for presentation to the user in response to the user interacting with the new original message via the computing device of the user.

7. The method of claim 1 , wherein each of the reply messages of the reply group is determined to include the commit in the body of the reply message based on:

applying each of the reply messages of the reply group to a trained commit classifier.

8. The method of claim 7 , wherein using the reply message features of the reply message to select the additional reply group of the reply messages that each include the commit comprises:

generating a commit classifier training example based on the reply message features;

further training the commit classifier based on the commit classifier training example; and

using the further trained commit classifier to select the additional reply group of the reply messages.

9. The method of claim 1 , further comprising:

determining that the commit is present in a threshold quantity of the reply messages that have the reply message features, the threshold quantity including the reply message;

wherein using the reply message features of the reply message to select the additional reply group of reply messages that each include the commit is based on determining that the commit is present in the threshold quantity of the reply messages that have the reply message features.

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

identifying a corpus of electronic communications stored in one or more databases, wherein a plurality of the electronic communications of the corpus each include a reply message and an original message to which the reply message is responsive;

selecting a group of the electronic communications from the corpus of electronic communications based on determining that the electronic communications of the group each have a corresponding original message that satisfies criteria of a model;

determining reply message features based on their occurrence in a threshold quantity of the electronic communications of the group;

using the reply message features to select an additional group of the electronic communications based on the electronic communications of the additional group each having a corresponding reply message that includes the reply message features;

generating refined criteria for the model using original message features from original messages of the additional group of the electronic communications;

identifying a new original message sent to a user;

determining that the new original message satisfies the refined criteria for the model; and

in response to determining that the new original message satisfies the refined criteria for the model:

performing one or more further actions directed to the new original message, wherein performing the one or more further actions comprises:

providing, for presentation to the user via a computing device of the user, an indication that sensorially distinguishes the new original message from other messages of the user.

11. The method of claim 10 , wherein the indication is a visual indication applied to a graphical representation of the new original message.

12. The method of claim 11 , wherein the visual indication is applied to the graphical representation of the new original message when the new original message is presented along with graphical representations of one or more additional messages.

13. The method of claim 10 , wherein the indication is provided for presentation to the user in response to the user interacting with the new original message via the computing device of the user.

14. The method of claim 10 , wherein the model is a machine learning model.

15. The method of claim 14 , wherein generating the refined criteria for the model using original message features from the original messages of the additional group of the electronic communications comprises:

generating a plurality of positive training examples based on the original message features; and

generating the refined criteria for the model by training the machine learning model based on the positive training examples.

16. The method of claim 10 , wherein using the reply message features to select the additional group of the electronic communications based on the electronic communications of the additional group each having a corresponding reply message that includes the reply message features comprises:

refining a reply message model based on the reply message features; and

using the refined reply message model to select the additional group of the electronic communications.

17. The method of claim 16 , wherein refining the reply message model comprises:

generating a plurality of positive training examples based on the reply message features; and

training the reply message model based on the positive training examples.

18. A system, comprising:

at least one processor;

memory coupled to the processor, wherein the memory stores instructions that, when executed by the at least one processor cause the at least one processor to:

identify a corpus of electronic communications stored in one or more databases, wherein a plurality of the electronic communications of the corpus each include a reply message and an original message to which the reply message is responsive;

select a group of the electronic communications from the corpus of electronic communications based on determining that the electronic communications of the group each have a corresponding original message that satisfies criteria of a model;

determine reply message features based on their occurrence in a threshold quantity of the electronic communications of the group;

use the reply message features to select an additional group of the electronic communications based on the electronic communications of the additional group each having a corresponding reply message that includes the reply message features;

generate refined criteria for the model using original message features from original messages of the additional group of the electronic communications;

identify a new original message sent to a user;

determine that the new original message satisfies the refined criteria for the model; and

in response to determining that the new original message satisfies the refined criteria for the model:

perform one or more further actions directed to the new original message, wherein the instructions to perform the one or more further actions comprise instructions to:

provide, for presentation to the user via a computing device of the user, an indication that sensorially distinguishes the new original message from other messages of the user.