PolicyPal aims to revolutionise the way you interact with website policies, utilising artificial intelligence to simplify the often convoluted, verbose terms and conditions or privacy policies of any website into a set of succinct and concise bullet points using Natural Language Processing (NLP).
It utilises two custom-trained natural language processing models based on the Longformer Encoder Decoder Architecture by Beltagy et al which were trained through transfer learning using Hugging Face on the task of summarising terms and conditions and privacy policies.
Produced summaries are then passed to a custom fine-tuned BERT model which will then provide each produced summary with a rating (Positive, Neutral, Negative), depending on the degree to which the terms and conditions or privacy policy respects your privacy
One method is to clone the project and compile it as approval has not been sought from the Google Chrome Store
A version is also available in the the releases tab however, the server has been closed due to high cost ($45/m) but it's capable of running locally as long as you have decent hardware.
Overall, the models show major improvements in the field of summarisation of terms and conditions and privacy policies compared to existing models:
These models were evaluated on the ROUGE Metric (Lin, 2003)
The classification model is evaluated with Precision and Recall
PolicyPal was evaluated by 30 participants, with all participants inclined to continue using PolicyPal and all participants stating they felt more informed while using PolicyPal than when without.
Some further highlights include:
