I am building a neural network to do intent detection and slot filling. The results I am getting are somewhat poor. Hence, I am looking for an approach to improve my results.
My idea is to use multi-task learning (like in Collobert and Westonäs paper) to take advantage of an additional's task knowledge for improving the results.
So far, I have been thinking about using a question-answering approach (sequence-to-sequence) that would share the intermediate representation with the rest of my architecture.
Do you think that question-answering would be a good task in this case?
The rationale behind this choice is that question-answering would need to learn a representation of the input sentence's intent in order to be able to generate an answer.