PRISM: Personalizing Reporting With Intelligent Summarization Through Multiple Frames

Abstract

As the challenges of information overload and selection fatigue in the digital news environment persist, natural language processing techniques have become increasingly critical for personalizing information delivery. This paper introduces a novel, frame-specific summarization framework aimed at generating summaries that accurately reflect various frames within news content. Unlike conventional text summarization methodologies that predominantly focus on extracting key frames from a corpus of content, our approach utilizes media framing techniques, coupled with bag-of-words and deep learning model for framing analysis, to extract varied perspectives from a single content and infuse summaries with the relevant contextual semantics. Utilizing a conditional variational autoencoder structure, we created and trained data that embeds multiple frames within a single document, enabling the summary to reflect frame-specific information. Our framework’s effectiveness is supported by its superior performance in terms of summarization accuracy on the benchmark dataset compared to baseline models. The implications of this work aim to support public interest by offering the potential to broaden perspectives on complex societal issues and enhance news personalization techniques.

Keywords

• Media framing
• multi-frame dataset generation
• controllable text summarization
• multi-task learning