Leveraging Linguistic Metrics to Improve UML Class Diagram Generation Using Large Language Models: A Cross-Model Study of Rewriting-as-a-Service

Complexity-Levels Experiment for LLM-Based UML Class Diagram Generation This archive contains the code, generated inputs, generated UML diagrams, and results for a study of how rewriting a software system's prose specification along different linguistic-complexity and information-density objectives affects a large language model's ability to generate a correct UML class diagram from it. Starting from 45 case-study specifications (source descriptions plus gold-standard UML class diagrams), each description was rewritten into 18 complexity-level variants — ranging from compact structural notes and narrative simplifications to objectives that directly minimize syntactic (dependency-tree) depth, maximize entity/relationship recall density per token, or combine both under a min-gated scoring rule. All rewrites are constrained to preserve every entity, attribute, relationship, action, and constraint present in the source; only linguistic form or information density changes. Each of the 18 level variants was then used to generate a UML class diagram with three large language models — Claude, GPT-4o-mini, and a locally-run Gemma model — and scored against the gold diagram on class, relationship, attribute, and cardinality F1. Content The archive is organized into: the 45 source cases with all rewritten description variants and gold diagrams (dataset/); every generated UML diagram in raw PlantUML source form (text_output/); the text-complexity metric suite and rewriting pipeline (text/); the experiment orchestration, F1 scoring, and result tables/plots (experiments/levels/); and the underlying generation/evaluation engine (src/). A detailed README.md documents the directory structure, the objective behind each of the 18 levels, commands to reproduce the pipeline, and the schema of the results table (levels_f1.csv).