Deep Learning project

Hello everyone.

I work in customer service, and I was thinking about proactively developing an algorithm to assist in the supporting service. In general, my job focuses in a single technology, and our job is to receive “cases” and solve them. The algorithm would receive an input X which would be the case description - “having problems doing abc”, “the connection is very slow”, etc, and would provide an output Y which would be a solution or at least would narrow the possibilities, allowing the supporting team to save time. So, for each problem X, a solution Y would be provided (not to the customer, only to the supporting team), and this would be based on all the historical data the company has, which maps a problem description X to a solution Y.

This is a NLP problem, and it makes sense to have something like this because there’s a finite number of problems a specific technology has, and for a specific problem X there’s a high probability that this problem X has troubled a couple of users in the past; and instead of going through the normal troubleshooting process, it would be nice to have an algorithm that would analyze all the historical data and understand that this problem X someone presented was solved by Y in the past.

I’m still finishing the NLP course so I will probably have it a little more clear on my mind by the end of it, but just wanted to generally discuss it with you guys, a little more experienced in NLP than me, how would you tackle something like this, and if there’s a similar topic from which I can take some inspiration. I see some similarities in machine translation for example, as the input X gets mapped to a solution Y, here the solution being the translated sentenced.


Hello jpedroanascimento,

this is interesting task and I’m working on similar project. I’m almost done with NLP course too and in the Processing with Sequence course in week 4, there is lesson named One shot learning. This can be useful when searching if some “case” has been solved allready.

So basically I would have some prepared “case scenarios”/“how to” and algorithm would just decide from user text input, if there is some “how to” solution and serve that solution to user.

If user is entering text string I would use Entity recognition technique to better understand what should be solved. Based on obtained informations I would measure similarity between user problem and “how to” prepared steps.

This is what came to my mind now.