Deep Learning approach for Signature extraction from emails

Question

Deep Learning (DL) is now days are used in various fields. I may like to know any idea/approach on usin DL for Signature extraction from emails.

There are some papers which uses ML like
http://keg.cs.tsinghua.edu.cn/jietang/publications/f142-tang.pdf http://www.cs.cmu.edu/~vitor/papers/sigFilePaper_finalversion.pdf    there are few more.

There are some open source libraries in Python like talon which uses NLP (Regular expressions and other techniques) and uses ML algorithms ( SVM ) for the same.

I have done a little research but I am not getting any resource or any idea on how to proceed with DL for this task.
I thought of using seq2seq models but I am not sure it can be of much help for this task. I am not sure whether summarization techniques can be useful. Any idea or resources for the same can be helpful.

Answer 1

Detecting a signature in an email is more like a detection problem in an image rather than a natural language processing (NLP) problem. It does not particularly matter what the signature says, you simply want to be able to detect it.

In such a case, detection in images is currently dominated by convolutional neural networks (CNN). This is a very powerful technique which has proved to be the best for a multitude of image-based tasks. However, it requires a very large dataset. So you will need to collect a lot of varied examples and annotate them yourself.

The output of the CNN can provide either a binary output 0/1 determining if a signature exists. Or it can provide an output which is the coordinates of the center of the signature (X, Y coordinates) such that you can know where it is placed in the document.

Answer 2

In this problem, the signature in each email is related to the text that is before the signature that's why I would suggest using recurrent neural networks like LSTMs or GRUs especially if it's a classification task. You can also try the LSTM seq to seq model if your task is to output the signature.