26.3 GUS: Simple Frame-based Dialogue Systems

What are the three things you want to extract from user’s utterance?
  1. Domain classification – what’s the user talking about?

  2. Intent determination – what’s the task or goal of the user?

  3. Slot filling – extract the slots and fillers the user wants the system to understand

Below showcase how we convert a user utterance into a format the dialogue system would understand.

What are other components of the frame-based dialogue?
  1. Automatic Speech Recognition

  2. NLG module using template-based generation and prompts. The template means most of the words in the response has been pre-determined by the dialogue designer. Templates can be fixed or include some variables that the users need to fill in

26.4 The Dialogue-State Architecture

What are the six components of a typical dialogue-state system?
  1. Automatic Speech Recognition

  2. Spoken Language Understanding

  3. Dialog State Tracker (DST) – maintains the current state of the dialogue

  4. Dialog Policy – decides what the system should do or say next

  5. Natural Language Generation (NLG)

  6. Text to Speech (TTS)

What are dialogue acts?

Dialogue acts combine both speech acts and grounding to create a single representation of the turn or sentence. Figure below showcase the utterance and its respective dialogue act.

What’s semantic parsing?

Semantic parsing is the task of associating each sentence with the correct set of slots, domain, and intent.

What’s a simple method to perform slot filling?

Train a sequence model to map input words representation to slot fillers, domain, and intent. This can be done by converting input sentence into embedding using contextual embeddings and feed the sentence embeddings into a feedforward layer to determine the domain and intent. The figure below showcase this simple architecture.

Ryan

Ryan

Data Scientist

Leave a Reply