1
$\begingroup$

I've been working on a developing a product selection network for my workplace. I work with lots of chemicals and the clients don't always know what they want/need so most of the time I have to ask a bunch of question, collect useful/ignore useless information, then make a selection from there. Discussions take place over the phone.

Ideal situation In terms of information collection/flow: 1. I ask a probing question which specifies which feature I am referring to 2. Client answering question and speech recogniser converts voice to text filling in feature input. 3. Text summariser searches the feature and reduces down to specific key words. - For example: Me: "what kind of application are you looking to perform?" Client: "I'm wanting to adhere two pieces of wood together" Feature summarised to: adhere, wood 4. Once feature vector has enough information the network recommends the most suitable product.

Problems: 1. Clients tend to waffle and give useless information therefore network will need lots of training data. 2. Once a question is asked - client may not directly answer that question and may incidentally answer another feature question.

I would think the logical place to start would be a speech recognition RNN - I have written a weak tensorflow one however I think I'll just look to tap into Google's cloud speech recognition API here. This is where I get stuck - should I just use a simple forward/back propagation network from here and treat it as a classification problem or is there another way to do it?

Any direction pointers would be greatly appreciated.

Kind Regards, Andy

$\endgroup$
1
$\begingroup$

I don't think existing state-of-the-art technology will be adequate to solve this task effectively. The current state of knowledge in NLP and AI probably isn't good enough to build something that will work well in practice. Instead, I think you'll need to use humans. Perhaps you can hire someone and train them -- that will probably be both cheaper and more effective than anything feasible with AI today.

(It seems impractical to obtain the large training sets typically needed for deep learning. Also, your task requires deep understanding of spoken text at a semantic level, which is currently hard. Finally, your task requires domain knowledge about the kinds of application tasks that clients are likely to ask about.)

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.