EROS: Entity-Driven Controlled Policy document Summarization

Anonymous

Abstract:

Privacy policy documents have a crucial role
in informing individuals about the collection,
usage and protection of user’s personal data
by organizations. Policy documents are no-
torious for their complex and convoluted lan-
guage, posing significant challenges to users
who attempt to comprehend their content. In
this paper, we propose an innovative approach
to enhance the interpretability and readability
of policy documents by using controlled ab-
stractive summarization by enforcing critical
entities using reinforcement learning.
Due to legal jargon, lengthy sentences, and
intricate syntactic structures of privacy policy
documents our approach first identifies critical
information necessary to the user using span-
based entity extraction model (EEPD, and use
these entities to enhance the summary of the
document. Our model EROS uses proximal
policy optimization (PPO) to control the infor-
mation and symantic structure of the generated
summary. Our model shows massive improve-
ment over base summarization techniques

Paper Type: long

Research Area: Summarization