Under Article 24(5) of the EU's Digital Services Act, online platforms must submit every content moderation decision they make, along with a corresponding Statement of Reasons, to the DSA Transparency Database. That requirement generates a large and growing record of how platforms enforce their rules.
The paper, posted on arXiv, argues that this database is only useful if its taxonomy can be connected to the specific language platforms use in their terms and conditions. The authors propose using large language models to perform that mapping, aiming for more granular reasoning than current approaches allow.
The work highlights the potential of LLMs to bridge the gap between a legal framework and the varied policy texts of individual platforms. It is a step toward making the transparency database more amenable to detailed analysis and oversight.