[Paper Review] Rasa: Open Source Language Understanding and Dialogue Management

Tom Bocklisch, Joey Faulkner+2 more|arXiv (Cornell University)|Dec 14, 2017

Topic ModelingComputer Science14 references139 citations

TL;DR

Rasa introduces open-source Python tools NLU and Core for language understanding and dialogue management, designed for non-specialist developers with decoupled components and machine teaching workflows.

ABSTRACT

We introduce a pair of tools, Rasa NLU and Rasa Core, which are open source python libraries for building conversational software. Their purpose is to make machine-learning based dialogue management and language understanding accessible to non-specialist software developers. In terms of design philosophy, we aim for ease of use, and bootstrapping from minimal (or no) initial training data. Both packages are extensively documented and ship with a comprehensive suite of tests. The code is available at https://github.com/RasaHQ/

Motivation & Objective

• Make machine-learning based dialogue management and language understanding accessible to non-specialist developers.
• Bootstraps from minimal or no initial training data to build conversational systems.
• Provide a modular, decoupled architecture that supports reuse across languages and easy deployment.
• Offer comprehensive documentation, testing, and production-ready deployment workflows.

Proposed method

• Modular architecture with a tracker for conversation state and a policy that selects the next action.
• NLU pipeline combining spaCy for tokenization/POS, GloVe embeddings, and scikit-learn classifiers; optional CRF for entities.
• Predefined pipelines (e.g., spacy_sklearn) with swap‑in components and configurable backends.
• Training data formats: intents/entities in JSON/Markdown for NLU; stories for Dialogues in Markdown.
• Machine teaching workflow to iteratively correct system actions and expand training data.
• Visualization of dialogue graphs to interpret training data and policy behavior.
• Docker-based deployment for reproducible production environments.

Experimental results

Research questions

• RQ1 Can Rasa NLU and Core be flexibly used independently and still interoperate effectively?
• RQ2 How does the system perform with minimal or incremental training data and how does it handle non-linear conversations?
• RQ3 What is the effectiveness of machine teaching for improving the dialogue policy in practice?
• RQ4 How can production deployment be streamlined and reproducible across environments?

Key findings

• The architecture is fully decoupled, enabling reuse of trained models across languages and straightforward production deployment.
• Rasa NLU uses a robust pipeline with spaCy, GloVe, and a multiclass classifier; entities are detected with a CRF component.
• Rasa Core treats dialogue management as a classification problem over predefined actions using a feature-based policy.
• A max_history parameter (default typically 3–6) shapes the featurization of dialogue state for sequence models.
• Machine teaching provides a practical workflow to generate training data by correcting predicted actions and partially training the policy.
• Visualizations of story graphs aid interpretation and debugging of training data and policy decisions.