The hand-written UML class diagram data set

The hand-written class diagram data set comes from Song Yang.

Inspirations include the yuml examples page.

The project report is available in this repo's PDF, called Extraction of UML class diagrams from English.pdf. Check it out! It's written like a research publication.

Description

The train-test-validation split is 60-20-20. The data comes in pairs: .txt in English and class diagrams in yuml in .yuml. yUML is a textual representation of UML diagrams, e.g. the LaTex of UML.

Sample: train/inheritance_person

The text is written in English.

An employee is a person. A customer is a person.

There is the associated .yuml.

[Person]^-[Employee]
[Person]^-[Customer]

To visualize the .yuml files, you need the yuml command-line tool on GitHub. After installing, you can simply call the render.sh script.

./render.sh train/inheritance_person.yml # generates last_render.png

Note: Internet connection is required to generate pictures with yUML.

Installation

The data set itself does not require installation. However, the processing tools do.

YUML visualizer

As mentioned earlier, YUML is a UML visualization tool that produces images from text. Visit and install the yuml command-line tool on its GitHub.

Tokenizer

The YUML textual data has a structure similar to LaTex in concept. It is meant to compile symbols into shapes on an image. Because YUML is a domain specific language and not prose (English sentences), it requires a special tokenizer in preparation for NLP tasks.

The tokenizer is written in C and Flex/Bison. In order to build the tokenizer, you need to install Flex and Bison on your system. Naturally, this requires a C compiler and CMake.

Build script

The build.sh script inside tokenizer will create a build directory and run Make and CMake.

cd tokenizer
./build.sh

Alternatively, you can use the Makefile inside the tokenizer directory to generate all the C files from Flex/Bison. And launch CMake from a directory of your choice to compile C.

Execution

Rendering a YUML file

A provided bash script called render.sh helps with rendering YUML, described in Sample: train/inheritance_person.

Running the tokenizer

After compilation, the tokenizer is located inside the tokenizer/build folder. The executable is named yuml_tokenizer.exe. The extension may vary depending on your operating system.

ATTENTION: The executable relies on dynamically linked libraries (Flex/Bison), that differs from platform to platform. If you compiled it with Cygwin, you can only run it on Cygwin.

As with any Flex/Bison executable, its default behavior is to take inputs from either STDIN or user prompts.

./yuml_tokenizer.exe < example.yuml # example.yuml contains your YUML
echo "[Student]" | ./yuml_tokenizer.exe

The echo command chain yields [ Student ]. The tokenizer is spacing out the different YUML syntax and standardizing alternate constructions.

Tokenize script

To simplify things, tokenize.sh calls the executable for you, assuming an .exe extension. All the data set will be tokenized and stored at the indicated folder.

./tokenize.sh my-tokenized-data-folder/

Testing, quality assurance

Conveniently, the accompanying data set serves also as a test suite for the tokenizer. The test.sh comprehensively runs the tokenizer over the entire data set and checks whether it is working properly.

Tests include grammar, spacing out YUML and semantics. The latter makes running the tests take at least 10 minutes and requires Internet connection, since it is using yuml. As such, semantic testing is in a separate script called test-semantics.sh. You must run test.sh successfully before doing semantic testing.

Known bugs

Class with an interface

When a class name contains an interface, the semi-colon separating the brackets from the name of the class may hide the name of the class.

[ << IResizeable >> ; Shape ]

This will produce only a box with IResizeable. But if you remove the semi-colon, the box will contain both words.