0
$\begingroup$

I am trying to develop an NLP - CNN algorithm to detect documents with sensitive information such as passport, license and distinguish them from other documents like resume, email, form or advertisements.

I personally consider this as a document classification problem and looked for open source datasets which had documents from different category/classes. I found the RVL-CDIP Dataset and tobacco3482 dataset with classes such as Email, form, letter, news, resume, scientific. However, the dataset collection looks from an old collection (ie samples in these datasets represent how resumes were written in 20000 and invoices collected in early 2000 )These datasets doesn't quite represent recent sample set of documents that we might come across on a daily basis. Is there a more recent document classification dataset available which would be more relevant to the current format of these documents?

EDIT : Looking for a similar dataset to RVL-CDIP Dataset and tobacco3482 dataset but which is more recent created. Use case is to still distinguish between resumes, advertisements, passports, and emails. I am not looking at detecting sensitive information within emails so my usecas is similar to RVL-CDIP Dataset use-case

$\endgroup$
4
  • $\begingroup$ Welcome to DataScienceSE. I think you problem is not very well defined for two related reasons: what would be the specific use case you're targeting, and how do you define "sensitive" in this context? for example some emails could contain sensitive information, resumes contain some personal information, etc. Also the level of sensitivity depends on whether the context is a company protecting their IP or an individual protecting their privacy (and you could even consider a secret service, which would be a whole new level). Once this point is clarified, it should be easier to decide what ... $\endgroup$
    – Erwan
    Feb 18, 2021 at 9:53
  • $\begingroup$ ... kind of data you need. Because if you start by collecting data randomly anywhere, chances are that the system won't do anything useful for anyone since it won't correspond to any clear definition of 'sensitive'. $\endgroup$
    – Erwan
    Feb 18, 2021 at 9:54
  • $\begingroup$ Thankyou for inputs I am looking for a similar dataset to RVL-CDIP Dataset and tobacco3482 dataset but which is more recent created. Use case is to still distinguish between resumes, advertisements, passports, and emails. I am not looking at detecting sensitive information within emails so my usecase is similar to RVL-CDIP Dataset use-case $\endgroup$
    – siya m
    Feb 18, 2021 at 16:38
  • $\begingroup$ Ok so actually you don't really care about sensitivity here. You could try asking on opendata.stackexchange.com. Note that I don't see why it's important to have recent data for this kind of task, since these classes of documents probably look the same nowadays than they did 20 year ago. $\endgroup$
    – Erwan
    Feb 18, 2021 at 18:28

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.