emails/ folder not included for training

[oxlsf]

emails/ folder not included for training

[obukhov-sergey]

Hi @oxlsf we plan to open-source the emails we used for creating the dataset but it will require some work / time since not all of them were originally publicly available like emails from enron public dataset. I.e. we'll need to remove all sensitive information.

Right now we provide just the processed dataset here https://github.com/mailgun/talon/blob/master/talon/signature/data/train.data. Each line represents a line from an email. Each element in a line is either 0 or 1 except for the last one. It's 0 if the corresponding feature from the feature set is false for the line and 1 otherwise. The last element is 1 if the line belongs to a signature line and -1 otherwise. Here's the feature set that we used https://github.com/mailgun/talon/blob/master/talon/signature/learning/featurespace.py#L15

[dichen001]

Hi @obukhov-sergey, how is your progress on opensource your training data? I have the same need with @oxlsf.
By the way, is it possible that you could kindly provide the processed data from Enron (i.e marked with #sig# if that line belongs a signature part)?

[obukhov-sergey]

@oxlsf @dichen001 we recently open-sourced annotated email dataset. It's not the one used by talon but we plan to switched to it once it has over 600 emails (now it's over 190 emails).

The idea is to use cleansed data that doesn't have private or personal information and encourage people to contribute their own emails (with cleansed phone numbers, URLs, etc) to keep the dataset up to date.

Feel free to contribute :) I'll be adding more emails shortly. I also plan to refactor the code that prepares the train data so that it's easy to add more emails to the dataset and test the library.

[itsvivekshetty]

Hi @obukhov-sergey, I have the same need of training the data, is there a way i can do it now? can you please provide steps to train the data.

[obukhov-sergey]

@itsvivekshetty @dichen001 @oxlsf here's the open-sourced dataset https://github.com/mailgun/forge, it's not the one used for training but once it has more data we'll use it instead, PRs are welcomed, I've also added a section in Readme with more info on how to retrain the classifier with your own raw emails.