Stefan Schaffer is a Senior Researcher and Group Leader on the Cognitive Assistants department of the German Research Center for Artificial Intelligence (DFKI). His works have resulted in several conversational interfaces for domains reminiscent of mobility, automotive, tax information, customer support, etc. Currently, he’s working on AI chatbots for value chains and hybrid events. Before joining DFKI, Stefan worked as a product manager at Linon Medien. He studied communication science and computer science and did his doctorate on the Technical University of Berlin in the sector of multimodal human-computer interaction.

What initially attracted you to machine learning?

Already during my studies, I had an excellent interest in speech recognition and took courses wherein we built speech recognizers from scratch.

What initially attracted you to speech recognition?

I used to be already fascinated by speech-based human-computer interaction when Captain Picard spoke to “the pc” and received meaningful answers.

Certainly one of your most up-to-date projects was constructing a chatbot interface for a museum that may anticipate what visitors would ask. Could you discuss how your team approached this?

To integrate the query answering functionality into the museum chatbot, we first collected numerous questions and used them to enhance the system’s query answering capabilities. This was done by categorizing the questions in addition to the reply material we received from our project partner Linon Medien, an organization specializing within the production of speech and text content for exhibitions.

Your team also discovered that content type annotations can improve accuracy, what style of accuracy differences are seen from annotations?

The content type annotations improved the general accuracy of the chatbot in natural language understanding. Which means that resulting from the extra annotations, the system was able to present more correct answers.

What are among the core challenges behind constructing a conversational AI?

Without this data, normally one can only offer scripted experiences that mimic conversations between real humans, but are static and thus highly unnatural. One other challenge is that the means of developing a conversational AI interface requires special expertise in the precise area wherein the system will probably be used. Sharing the needed information between conversational design experts and domain experts is typically a difficult process that requires the support of additional experts in user-centered design methods.

What’s your approach for constructing a user friendly chatbot and conversational user interface?

We strictly follow the paradigm of user-centered design. Which means that we engage with our customers and users in early project phases, when a system shouldn’t be yet available. We start with focus groups and data collection and have stakeholders review system variants in early development phases.

What are your views on ChatGPT and GPT-4, is there anything you’d do in a different way?

Currently we use ChatGPT and GPT4 as tools for data generation. Nevertheless, we normally attempt to avoid supporting the closed nature of those products through our use in our research projects. We expect that comparable open-source models will grow to be available within the near future.

You’ll be speaking on the upcoming Way forward for Chatbots & Conversational AI Summit, what is going to you be discussing?

I’ll be speaking about connections between User Experience and conversational AI. I may have a give attention to user-centered design, data-driven user-centered implementation, and evaluation of conversational user interfaces.