Rachael Tatman — Conversational AI and Linguistics

AI Challenges

Creating effective conversational AI systems presents significant technical challenges. Rachael Tatman highlights the difficulty of managing conversational flows, especially when users deviate from expected paths. Traditional state machine approaches struggle with this, but Rasa's attention model offers a more flexible solution by ranking possible responses and adapting to interleaved conversational structures 1. She emphasizes the importance of achieving fluent conversational interactions, which remains a major engineering and machine learning challenge 2.

Being able to achieve that really fluent level of conversational interaction is a really large engineering challenge and a really large machine learning challenge.

--- Rachael Tatman

Tatman is excited about the potential of conversational AI to provide natural interactions, especially for users with diverse abilities and backgrounds.

   

RASA Framework

Rasa's framework for conversational AI is built on a foundation of probabilistic policies rather than rigid rules. Rachael Tatman explains that Rasa uses a combination of intents and example entities to train models, allowing for more adaptable and human-like interactions 3. This approach enables the creation of minimally viable assistants that can be iteratively improved through user interactions and feedback.

You build a minimally viable assistant and then you deploy it and have people make test conversations with it.

--- Rachael Tatman

Tatman finds excitement in the evolving design of conversational AI, which now allows for more intuitive computational tasks and broader accessibility 4.

   

NLP Advances

Recent advancements in NLP have significantly impacted conversational AI systems. Rachael Tatman discusses the integration of transformer architectures and contextual embeddings in Rasa's framework, which enhance the system's ability to understand and generate human-like responses 5. However, she cautions against relying solely on neural natural language generation due to potential inaccuracies and the risk of generating inappropriate content.

I would not be comfortable doing a completely neural natural language generation conversational assistant.

--- Rachael Tatman

Tatman underscores the importance of maintaining control over utterances to ensure factual and safe interactions, highlighting the ongoing need for careful design and implementation in conversational AI 4.