MLMU Online #17: Building Chatbots in Practice and Lessons Learned

  Machine learning

The talks will be live on YouTube.

Talk #1: How we use simple models to improve intent matching (Tomáš Vodrážka, Feedyou)

Building a chatbot is not tricky anymore, especially with the Feedyou Platform. But to teach chatbot to communicate to answer even common questions automatically? It is a bit more complicated discipline. At Feedyou, we focus on a hybrid solution that uses common tree structures and NLP models. In this talk, we would like to share the problem of matching the right intent. Can a sophisticated model be the solution to this complex problem? What about using simpler models and focusing on the training data instead?

Tomas in Feedyou develops NLP engine and sets the right vision for future development and involvement of AI and ML in the product.

Talk #2: Lessons learned from using Flowstorm conversational AI platform (Jan Pichl, PromethistAI)

Building a conversational application has been a challenging task. Using the right tool, one can easily create a high-quality application. While creating our successful conversational applications, we quickly discovered that we needed a custom tool that makes the process easier. We were asking the following questions: How to make the intent recognition robust and fast? How to make the application modular and reuse parts in other applications? What about the unexpected utterances? When the unexpected utterance is recognized, how to generate the response properly? The solutions for these questions are now part of the Flowstorm platform, making the applications more robust, coherent, and human-like.

Jan works as NLP engineering lead in PromethistAI and connects the research and production worlds. He was also a team leader of the Alquist team in the Alexa Prize competition 2017, 2018, and 2020.