Synthetic Data for Identifying Inclusive Language (Case Study: Job Descriptions in Italian)

Using a comprehensive list of job titles, we propose a framework to automatically generate job descriptions in Italian. This synthetic data is then used in a Large Language Model to detect inclusive language in job postings. Finally, we compare the results of this synthetic dataset with real data. Our study demonstrates that the data format and prompting method signif-icantly impact performance. Additionally, we identify limitations and key considerations for unifying synthetic data with real data for fine-tuning purposes. We also propose improvements to the framework and provide guidelines for effectively integrating these two types of data. The novelty of our work is generating and integrating synthetic data due to the scarcity of annotated Italian job descriptions, thereby improving the training of Large Language Models (LLMs) tailored specifically for Italian.