Overview
The CAS2UML dataset provides a collection of hand-drawn Unified Modeling Language (UML) diagrams paired with executable PlantUML code. This resource is intended to facilitate the reproducible evaluation of systems designed for automated UML generation from sketch inputs, a domain experiencing renewed interest with advancements in large language models and multimodal AI. The dataset includes a validation tool and reusable scripts to verify the syntactic correctness and renderability of generated UML outputs.
Research Context
Automated generation of UML diagrams from sketches and images is an area of study that has recently garnered increased attention, largely attributed to developments in large language models and multimodal artificial intelligence. A persistent challenge in this field has been the absence of publicly accessible datasets that include executable groundtruth models, hindering reproducible evaluation efforts. This gap makes it difficult to consistently benchmark the performance of different sketch-to-UML approaches.
Approach
The development of CAS2UML involved the creation of a public dataset consisting of 557 hand-drawn UML diagrams. These diagrams are specifically categorized into two types: 271 class diagrams and 286 activity diagrams. Each hand-drawn diagram in the dataset is associated with corresponding PlantUML code that has undergone manual validation. To further support reproducible evaluation, the project also delivers a PlantUML-based validation tool and a set of reusable scripts. These tools are designed to check the syntactic correctness and renderability of UML artifacts that are generated by various sketch-to-UML systems, thereby enabling standardized benchmarking.
Findings
- The CAS2UML dataset consists of 557 hand-drawn UML diagrams.
- Within this dataset, there are 271 class diagrams.
- The dataset also contains 286 activity diagrams.
- Each diagram in the dataset is paired with manually validated PlantUML code.
- A PlantUML-based validation tool is provided.
- Reusable scripts are included for checking syntactic correctness and renderability of generated UML artifacts.
Why This Matters
The provision of CAS2UML addresses a challenge in the automated UML generation field by offering a public dataset with executable groundtruth models. This dataset, along with its associated validation tool and scripts, is designed to enable reproducible benchmarking of sketch-to-UML approaches. Such resources are relevant for researchers and developers working on systems that leverage large language models and multimodal AI to interpret and generate UML diagrams from visual inputs, potentially fostering more consistent and comparable evaluations.