Automated Generation of Impactful Text Formatting Using Large Language Models

Abstract

This master's thesis aims to develop an automated system for generating structured and visually impactful content for Zeiko, a Belgian start-up that offers a software for interactive document creation. The project addresses the need for a solution that helps Zeiko’s users transform long and unstructured textual content into concise, structured, and visually engaging formats compatible with Zeiko's software.

The proposed solution uses large language models to analyze and summarize input text while preserving coherence and key elements. A central component of the system is a fine-tuned Mistral 7B model, trained specifically for structured JSON generation. This model extracts key information such as titles, summaries, and highlights from unstructured text, and returns a structured JSON designed for infographic templates.

The fine-tuning process was managed using custom datasets and its effectiveness was validated through evaluations. The generated outputs are mapped into predefined infographic templates known as z-blocks, a format specific to the company. Additionally, a complementary management platform was developed to allow Zeiko’s design team to deploy new templates with associated metadata.

The entire solution is deployed on Google Cloud using Docker and Kubernetes to ensure scalability and reliability. Evaluations demonstrate that the system significantly improves the speed and consistency of content creation. It also confirms that fine-tuning models like Mistral 7B is an effective method for extracting and structuring information from raw text, providing a practical tool to simplify document production.