I’m just getting started with neural networks.

I want to make the bot an expert in one very simple subject. I want to teach the user a simple process so that there’s not a huge breadth of replies that requires a lot of knowledge.

In the long term, I would like to evolve it into a full-blown help desk/customer service bot that would be able to identify the user’s problems and troubleshoot basic ones..  Or recommend a solution if it can’t help.

I am currently pursuing the machine learning code at coursera.
so if I code a neural network and plug these in as inputs:

Length of user inputs
# of replies (to know how far in the process we are)
List of keywords
?

and feed it lots of user conversations, would that bring something usable in the medium term or am I not ‘getting it”? Are there features that I should add that are missing from the list? Anyone got tips?:D

Thanks

Guanzhou Li - Jul 25, 2013:

In the long term, I would like to evolve it into a full-blown help desk/customer service bot that would be able to identify the user’s problems and troubleshoot basic ones..  Or recommend a solution if it can’t help.

First, I’m glad to see somebody wise enough to post their goal when asking a question. That helps tremendously.

My reaction is that you are off on the wrong track, using the wrong tools for the job. Neural networks are typically used for statistical pattern matching, especially for partial matches / associative memory at a subsymbolic level (i.e., below the level of text). While it’s true that such matching could be of some use in a help desk due to associations, you could do the same thing with standard textual keywords, and probably far better. The inputs you listed are going to be close to useless for machine understanding, especially for specific, detailed, technical knowledge at a help desk. To be useful, I think the system would have to be trained on concepts, probably implemented as keywords, which is not only problematic in itself due to synonyms and volume of training data, but neural networks simply don’t “understand” the types of problems you’re describing because of the Binding Problem. At least regular text and computer code can do binding, and there are many tricks that can be used for handling synonyms, associations, and so on without using neural networks. The idea of feeding a neural network a lot of former user conversations in an attempt to have it learn or understand the domain makes me cringe: it will just turn into a nightmare that could never work as you want.

If you insist on using known AI techniques, I’d go for an expert system instead, and I’m not even sure that would be the best approach.

I tend to agree with Mark. I took the machine learning courses at Coursera and believe there are more efficient ways to code a help desk chatbot than neural nets. A curated corpus may give you the best start. In many cases your goal is to give a specific response, and it is hard to train a net to do this reliably. It might be a fallback from other techniques.

Hi thanks for the responses.

I will thus not waste time on neural networks.

This page lists publications on the most advanced research that has been done so far with regard to natural language understanding. Unfortunately the amount of effort and expertise that is required to produce systems like these is beyond the present capabilities of most of us. Nevertheless you may find some useful pointers among all the published papers.

http://www.cs.rochester.edu/~james

This page contains videos demonstrating one of these systems in action.

http://www.cs.rochester.edu/research/cisd/projects/plow/