# Marvista: Exploring the Design of a Human-AI Collaborative News Reading Tool

**Authors:** Xiang 'Anthony' Chen, Chien-Sheng Wu, Lidiya Murakhovs'ka, Philippe Laban, Tong Niu, Wenhao Liu, Caiming Xiong  
**arXiv:** [2207.08401](https://arxiv.org/abs/2207.08401)  
**Date:** 2022-07-18  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2207.08401v2/

## Key points

- Marvista is a human-AI collaborative tool designed for the review step of Robinson's SQ3R technique for online news articles
- Key feature: Generates abstractive summaries to aid in recalling important information and skip re-reading for returning readers
- Implementation uses Transformers encoder-decoder architecture with RoBERTa as the decoder and Longformer for handling long articles effectively
- Demonstrated reasonable summarization results on the CNN/Daily Mail dataset, showing efficacy as a proof-of-concept solution
- Future work may involve incorporating OpenAI's InstructGPT and ChatGPT for improved abstractive summarization
- Utilizes personalized user data from reading sessions to generate summaries, incorporating both implicitly-generated data (e.g., dwelling time on paragraphs) and explicitly-generated data (e.g., user notes and highlighted text)
- Participants find the personalized approach valuable in reflecting their interests and controlling the focus of AI-generated summaries
- Emphasizes content-specific guidance throughout the reading process to enhance user engagement and comprehension of online news articles

## Summary

Marvista is a human-AI collaborative tool designed to assist readers in the review step of Robinson's SQ3R technique when reading online news articles. Its key feature is the generation of an abstractive summary that not only helps users recall important information but also allows returning readers to skip re-reading and directly access the summary page. While currently provided at the end of reading, future iterations may prompt users with summative questions to encourage reflection on the entire article. The implementation of Marvista utilizes standard Transformers encoder-decoder architecture, with RoBERTa as the decoder and Longformer for handling long articles more effectively than BERT. This model has shown reasonable summarization results on the CNN/Daily Mail dataset, demonstrating its efficacy as a proof-of-concept solution. Future work may incorporate OpenAI's InstructGPT and ChatGPT for improved abstractive summarization. One unique aspect of Marvista is its use of personalized user data from reading sessions to generate summaries. Implicitly-generated data such as dwelling time on paragraphs influences paragraph importance in the summary, while explicitly-generated data like user notes and highlighted text are also incorporated into the process. Participants have found this personalized approach valuable in reflecting their interests and controlling the focus of the AI-generated summary. Overall, Marvista's design emphasizes content-specific guidance throughout the reading process, enhancing user engagement and comprehension of online news articles. By leveraging natural language processing models and user-generated data, Marvista provides tailored support for readers seeking efficient and effective ways to engage with digital content.