Research at Rasa
Our research team enables developers to build conversational AI that couldn't be built today.
We make Rasa the best tool developers could use to build conversational AI. We engage and share ideas with the broader research community. And we attract the best people in the field to come work with us.
We've added support for our model pipelines to support incremental training. This allows you to fine-tune an existing model after adding new training examples instead of training a new model from scratch. It's an ongoing area of research on how to do this effectively, but our first results look promising
"Whatlies" in Word Embeddings
Whatlies is an open source toolkit for visually inspecting word and sentence embeddings. The project offers a unified and extensible API with current support for a range of popular embedding backends including spaCy, tfhub, huggingface transformers, gensim, fastText and BytePair embeddings.
Dual Intent and Entity Transformer (DIET)
DIET is a new state of the art NLU architecture that jointly predicts intents and entities. It outperforms fine-tuning BERT and is 6x faster to train. You can use DIET together with BERT and other pre-trained language models in a plug-and-play fashion.Explainer Video
What's the best data structure for dialogue memory - a stack? a graph? a flat list? Self-attention gives you great flexibility without a complex memory structure.Read the paper
Transfer Learning Across Dialogue Tasks
You've built an assistant and it can already help users with a few things. Now you're adding new functionality. How can your assistant re-use the dialogue elements it already knows about in this new context?
Compressing Transformer Language Models
Large-scale language models like BERT, GPT-2, and XLNet show excellent performance on a number of NLU tasks but are very resource intensive. Can we compress these models to get something that's almost as accurate but much faster?Read about Quantizing BERT
Entity Resolution using Knowledge Bases
Combining a dialogue system with a knowledge base allows developers to encode domain knowledge in a scalable way and integrate it with statistical NLU and dialogue models. It also helps your assistant understand messages like the second one or which of those is cheaper? .
Supervised Word Embeddings
Pre-trained word embeddings like word2vec and GloVe are a great way to build a simple text classifier. But learning supervised embeddings for your specific task helps you deal with jargon and out-of-vocabulary words. This is now our default intent classification model.
Mixing Single and Multi-turn Dialogue
Dialogue elements like small talk and FAQs are single-turn interactions. New retrieval-based models in Rasa can handle all of these simple responses in a single action. This means your dialogue policy becomes much simpler and you need fewer training stories.Read the blog post
Most language models and word embeddings are trained on prose and don't know anything about the rules of conversation. How can we build embeddings that understand the difference between purposeful dialogue and chit-chat, and can detect non-sequiturs?
Talks and Meetups
We regularly host external speakers at our #botsBerlin meetup to talk about their research.
We’re a partner of the UKRI Centre for Doctoral Training in Natural Language Processing at the University of Edinburgh. We are also sponsors of SigDIAL 2020.
Every year we supervise a few MSc students and take on some interns to work on research. If you are using Rasa in a course, get in touch and we can share materials for use in lectures and group projects.