Conversational AI 20-NEURAL NETWORK APPROACHES TO DIALOGUE MODELING
The neural approach to dialogue systems development is based on an end-to-end architecture in which the various components of the traditional dialogue systems architecture are not required. It involves learning mappings between input and output utterances. At present, this approach has been applied mainly to text-based dialogues. In a spoken dialogue application the speech components operate separately from the end-to-end architecture, so that the results of speech recognition are fed in text form into the architecture and the output response is passed to a TTS component to produce a spoken utterance. Given such an end-to-end architecture, there are two main tasks to accomplish.
- Process and represent the input: this is known as encoding.
- Generate the output: this is known as decoding.
An end-to-end architecture provides certain advantages over the traditional pipelined architecture.