Reproduce “Task-Oriented Dialogue as Dataflow Synthesis” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Semantic Machines, Jacob Andreas, John Bufe, David Burkett, Charles Chen, Josh Clausman, Jean Crawford, Kate Crim and 38 more
Researchers from Microsoft Semantic Machines introduce a framework representing task-oriented dialogue as dataflow synthesis, where dialogue state is a dynamic graph updated by compositional programs and explicit metacomputation operators. This approach enables greater expressiveness and improves accuracy and robustness in handling complex conversational phenomena like reference and revision.
Key insights
- Explicitly modeling metacomputation operators (reference and revision) as structured operations on a dataflow graph significantly simplifies the learning problem for neural models, making complex dialogue phenomena more predictable and accurate.
- Representing dialogue state as an executable dataflow graph provides a highly expressive and structured foundation that bridges the flexibility of neural models with the interpretability and learnability of symbolic systems.
- An exception-driven approach, where program evaluation failures trigger agent initiative, offers a robust and flexible alternative to rigid dialogue policies for managing conversational breakdowns and guiding users.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.