Rasa overview: rules, neural models and chatbot design

Rasa is presented as a framework for building task-oriented dialogue systems: assistants that help users complete a specific action through conversation. Dr. Rachael Tatman explains how developers supply examples of user wording and conversation patterns, then combine machine learning with explicit rules.

The video was created in 2021 and uploaded on July 30, 2021. It covers Rasa's legacy open source project without naming a release version. The supplied description says Rasa no longer actively maintains that project, so the architecture discussion belongs to that historical context.

Tatman separates an assistant into natural language understanding and a dialogue policy. Regular expressions can extract predictable text such as email addresses, while the transformer-based DIET model learns intent classification from examples. She describes rules as lightweight but limited by the patterns they encode; neural methods require training data and can generalize to unfamiliar wording.

For dialogue control, the talk compares state machines with TED, which selects a next action based on the conversation and developer-provided options, usually prewritten responses. Tatman recommends combining rules and neural policies to handle digressions. Her improvement process is concrete: review conversations, annotate errors, add corrections to training data, retrain and redeploy. This is a conceptual introduction rather than an installation guide.