Automatic Section Title Generation to Improve the Readability of Privacy Policies

Shomir Wilson,Mysore Gopinath,Norman M. Sadeh,Abhijith Athreya,Vinayshekhar Bannihatti Kumar

Abstract:The text in a typical privacy policy is organized into different topics, either implicitly or explicitly. Few policies with explicit visual organization split the text into various sections with titles that briefly describe the paragraphs that follow. Prior studies prove that the explicit visual organization im-proves the readability of privacy policies. However, policies with implicit organization either lack visual cues or do not contain titles for paragraphs, rendering them difficult to read. To make these policies more readable, we work on the automatic generation of titles for paragraphs. We exploit the explicit structure of web privacy policies to generate a title-paragraph dataset and train supervised generative models that generate a title for a given chunk of the privacy policy text. We score the models on both automatic and human evaluation metrics and observe that the Transformer based approach outperforms se-quential models producing relevant and grammatical outputs for 48% of the test cases.

Computer Science