I am learning to create a dialogue system. The various parts of such a system are Intent classifier, slot filling, Dialogue state tracking (DST), dialogue policy optimization and NLG.

While reading this paper on DST, I found out that a discriminative sequence model of DST can identify goal constraints, fill slots and maintain state of the conversation.

Does this mean that now I dont need to create an intent classifier and slot filling models separately as the tasks are already being done by this DST? Or I am misunderstanding both the things and they are separate?

  • $\begingroup$ The question asks for the overall design of a chatbot framework. A possible paradigm out there is a blackboard architecture. Is your chatbot utilizing a blackboard which is equal to a multi-agent architecture or is another kind of framework desired, e.g. a simple neural network? $\endgroup$ – Manuel Rodriguez Nov 26 '18 at 14:12
  • $\begingroup$ I need to build a multi-agent architecture with the framework involving different models for different requirements which will be merged into one big pipeline for creation of the dialogue system. It will be more complex than a single simple neural network. $\endgroup$ – mayank agrawal Nov 26 '18 at 14:18
  • $\begingroup$ I've heard about some kind of “Dialogue State Tracking Challenge” in which the details are discussed, but perhaps @OliverMason knows more about it. $\endgroup$ – Manuel Rodriguez Nov 26 '18 at 14:28
  • $\begingroup$ Yes! It was a research challenge focused on improving the state of the art in tracking the state of spoken dialog systems. What I want to know is if I create a very good DST model, will it perform intent classification and slot filling as well. $\endgroup$ – mayank agrawal Nov 26 '18 at 14:32

According to literature only neural dialoque state tracking systems were implemented in the past. The idea is to provide a dataset, for example from “Dialogue State Tracking Challenge” (DSTC), and use neural networks for memorizing the corpus. That means, the network has to create a hypothesis about the dialoque state, or to be more specific, it has to reproduce the known state given by the dataset. In theory, it is possible to extend the task by leaving out the neural network. Creating a Natural language processing pipeline doesn't need deeplearning, it can be done with normal expert systems and blackboard architectures too. The reason why this is seldom discussed in the literature, has to do with that such a system contains more than a single algorithm, but it's an integrated architecture. Similar to Question & Answering systems like “Apache UIMA” it contains of many submodules connected together. The direct answer to the OP is, that a dialogue tracking system only extends other modules like slot filling, it's not able to replace them. Slot filling means, to ground a spoken dialogue on the fly. While parsing the input stream, the RDF-triples are extracted for building the semantic model.


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.