Algorithms for Chatbot

DIET Architecture

Rasa uses Dual Intent and Entity Transformer (DIET) architecture to detect both Intent and Entity using Transformer.

DIET Architecture

The DIET architecture can be customized to a specific domain. The architecture can be thought of as lego blocks. We can customize and use the blocks we want for our use case. Depending on the accuracy, latency and memory usage of the use case we can decide which components to use.

TED Architecture

TED is used to decide what next action to be taken by the bot given the conversation with the user.

Creating Features for TED using Intent, Entity, Slots and Previous actions

Training Transformers for next action prediction using features created and action to be taken at that particular time step to calculate loss

Predicting next action

TED is not trained on sequence of events which happen in the dialog. In dialog lot of context switching will happen. Using Transformers here will be useful because transformers can pay attention to dialog events which occured in the past.

An example of TED paying attention to past events to decide next action

Response selection

Using DIET backend for response selection

Reference