DIAGRAPH: An Open-Source Graphic Interface for Dialog Flow Design

Dirk Väth, Lindsey Vanderlyn, Ngoc Thang VuACL 2023

TL;DR

DIAGRAPH is a new open-source graphical tool that makes designing dialog systems accessible to everyone, not just coders. It allows users to visually build dialog flows, which are then automatically converted into functional dialog systems, speeding up prototyping and research.

What Problem Does This Paper Address?

Creating sophisticated dialog systems, like chatbots, has traditionally required significant coding expertise. While existing toolkits have helped, they often present a steep learning curve, demanding users to write code to customize models. This creates a barrier for domain experts, researchers from other fields, and students who want to leverage dialog systems without becoming software engineers. Furthermore, the "black-box" nature of some advanced neural models can be problematic in sensitive domains where control and transparency are crucial.

Methods

DIAGRAPH tackles this by providing a visual, drag-and-drop interface for designing dialog flows. Users construct these flows by connecting different types of nodes, each representing a specific dialog action or logic. The core innovation lies in its ability to automatically translate these visual graphs into functional dialog systems.

Key components of the DIAGRAPH system include:

  • Graphical Editor: A web-based interface built with React and React Flow, allowing users to create, edit, and manage dialog graphs.
  • Node Types: Five distinct node types (User Response, Information, Variable, Variable Update, Logic) cater to various dialog design needs, from simple information delivery to complex conditional branching and variable manipulation.
  • Template System: A minimal syntax within nodes allows for dynamic personalization of system output using variables and external data.
  • Backend Integration: The system leverages the ADVISER toolkit for its backend, enabling the conversion of dialog graphs into runnable dialog systems.
  • Debugging View: A parallel window allows designers to test dialogs interactively, inspect variable states, and understand flow execution in real-time.
  • Graph Management Dashboard: A central hub for creating, editing, deleting, and sharing dialog systems, including generating sharable links for embedded chat interfaces.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Democratizing Dialog Design: The most significant insight is that a graphical, intuitive interface can drastically lower the barrier to entry for dialog system creation, opening it up to a much wider audience.
  • Bridging the Gap: DIAGRAPH effectively bridges the gap between domain expertise and technical implementation, allowing subject-matter experts to directly build and iterate on dialog systems relevant to their fields.
  • Rapid Prototyping and Education: The tool's ease of use makes it ideal for rapid prototyping of dialog behaviors for research and for teaching students the fundamentals of dialog systems and programming concepts.
  • Transparency and Control: By using a visual, rule-based approach (though extensible to RL policies), DIAGRAPH offers greater transparency and control over dialog system behavior compared to purely end-to-end neural models, which is vital for sensitive applications.

Key Results

The paper presents strong evidence of DIAGRAPH's usability and effectiveness through user evaluations:

  • Complex Real-World Domain: Subject-area experts from a university travel reimbursement department, with no prior chatbot experience, successfully created a dialog system. This system handled approximately 2000 dialogs with an average SUS score of 82 (significantly above average).
  • Rapid Dialog System Design: In a short 15-30 minute session, 19 participants were able to create and test their own dialog systems, achieving an average SUS score of 75, indicating above-average usability.
  • Educational Tool: A workshop with high school students showed that DIAGRAPH is engaging and effective for teaching dialog systems and programming concepts, with all students successfully creating dialog graphs involving logical operations.

Limitations and Future Work

  • Current Focus: DIAGRAPH is currently designed for text-based dialogs. The authors note that extending it to spoken dialog systems by incorporating ADVISER's text-to-speech and automatic speech recognition modules is possible but not included in the default distribution.
  • Future Enhancements: The authors plan to extend the tool with the ability to query web APIs and allow users to define expected inputs for variable nodes using custom regular expressions, further increasing flexibility.
  • Open-Source Vision: By releasing DIAGRAPH as open-source, the authors aim to make dialog design more accessible and encourage further research into controllable dialog policies.

Summary updated

BibTeX
@inproceedings{vath2023diagraph,
  author = {Dirk Väth and Lindsey Vanderlyn and Ngoc Thang Vu},
  title = {{DIAGRAPH: An Open-Source Graphic Interface for Dialog Flow Design}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure