The proliferation of mobile phones has made reading on these devices increasingly common, but it has also exposed readers to a plethora of distractions, both internal (e.g., incoming SMS or phone calls) and external (e.g., navigating a footpath). Resuming a reading task after a distraction can be a daunting and time-consuming endeavour. To address this challenge, our research explores using automated text summaries and visualizations, powered by Large Language Models (LLMs), to aid readers in distractive environments. These summaries and visualizations can be triggered both before and after distractions occur, ofering a flexible approach to mitigating interruptions. This paper delves into the concept of generative primes - text summaries and visualizations presented before a distraction happens - and post hoc text summaries and visualizations. We evaluate the efectiveness of these approaches in terms of their ability to faithfully represent the content and how efectively they assist readers in comprehending and recalling the text. Large Language Models open up new possibilities for developing reading interfaces that empower readers to seamlessly handle distractions and efortlessly pick up where they left of in their reading sessions as well as provide opportunities for memory consolidation. Our research sheds light on how these advances can positively impact the reading experience in today's increasingly distractive digital landscape.
CEUR ceur-ws.org
The advent of mobile technology has transformed the way we engage with written content.
From e-books and digital articles to social media feeds and messaging apps, our mobile devices
have become the primary medium through which we access and consume written information.
While this shift ofers unprecedented convenience and accessibility, it has also introduced new
challenges, particularly in environments where distractions abound [
As reading is increasingly becoming an activity that is performed on the go, it is vital to
address the impact of distractive environments. Recent research highlights the prevalence of
mobile phone usage for media consumption. According to a report by the World Economic
Forum, people now spend over four hours a day reading, listening, and watching on their mobile
phones [
The interruptions experienced while reading on mobile devices are twofold. First, there are
internal disruptions, such as incoming text messages, phone calls, or app notifications, which
divert the reader’s attention away from the text [
At the heart of interruptions lies the concept of switching costs, i.e., the time and mental
resources it takes to shift focus from one activity to another (and potentially back). Workers have
been shown to struggle to shift and resume their tasks after interruptions [
In recent studies, we have further focused on whether such text primes should be administered after an interruption occurred or beforehand (e.g., by delaying an incoming call). The design of these cues, however, has so far been labour-intensive as interruptions can occur at arbitrary points during text reading and cues hence need to be generated for any number of text passages.
In our era, where Large Language Models (LLMs) have become increasingly sophisticated, these summaries and visualizations can be created in an automated fashion, however, and customized to the reader’s preferences. LLMs have significantly advanced the field of Natural Language Processing (NLP) and have been instrumental in tasks like text summarization, language translation, and information retrieval [11, 12]. GPT-3, for example, has demonstrated remarkable capabilities in understanding and generating human-like text, sparking interest in leveraging such models to enhance the reading experience [Brown et al., 2020]. These LLMs can be harnessed to power a new generation of reading interfaces that mitigate the negative efects of distractions. By providing readers with informative summaries and visualizations, these interfaces have the potential to facilitate smoother transitions between reading and interruptions.
This position paper delves into the concept of generative primes, i.e., content cues in the form of text, graphical and image visualizations that help readers preview and review text passages and mitigate the efects of interruptions. We investigate the benefits of presenting such primes before a distraction occurs and explore post-hoc summaries and visualizations, which help readers re-orient themselves after being interrupted. Our research evaluates how well these generative primes represent the content and how efectively they assist readers in comprehending and recalling the text.
Our research aims to contribute to a deeper understanding of how generative primes, empowered by LLMs, can revolutionize how we read in distractive environments. By ofering insights into their efectiveness and applications, we hope to pave the way for a more seamless and enjoyable reading experience in the era of constant digital distractions. 2. The Concept of Generative Primes The concept of ”generative primes” represents an approach to address the challenges of reading in distractive environments. Generative primes are cues that provide readers with text summaries and visualizations either before a distraction occurs or after it has disrupted their reading. These elements act as cognitive aids, preparing the reader for the upcoming content or assisting in the resumption of the reading task after an interruption. By combining textual summaries and visualizations, generative primes create a holistic reading experience that can enhance comprehension, recall, and engagement.
In the past, we utilized the priming efect to facilitate text comprehension [ 13] and tested
diferent visualizations, including text highlights, structured mind maps, and image galleries [
To implement generative primes, Large Language Models (LLMs) like ChatGPT are invaluable for generating text summaries. ChatGPT is trained to understand and generate human-like text, making it ideal for creating concise, coherent, and informative textual overviews of longer passages. These summaries can serve as previews or reorientation tools, ensuring that readers stay connected to the content despite distractions.
Visualizations are essential components of generative primes that complement text summaries. Software tools like Midjourney and DALL-E have been revolutionary in this regard. Midjourney can create visualizations through image stories, distilling complex textual information into visually appealing and easily comprehensible narratives. Meanwhile, DALL-E can generate innovative visualizations, such as mind maps that visually represent the structure and connections within the text.
The synergy of text summaries from LLMs like ChatGPT and visualizations from software like
Midjourney and DALL-E enables the creation of generative primes that cater to diverse learning
and reading styles. These primes hold the potential to revolutionize the way we interact with
written content in distractive environments, making the reading experience more engaging,
informative, and resilient to interruptions.
3. Research Plan and Outlook
In our recent work, we explored the impact of interruptions on reading comprehension and
how these efects can be mitigated using reviews (summaries of already read content) and
previews (summaries of upcoming content) [
We found that previews shown after interruptions significantly enhanced reading comprehension compared to not showing any summaries. Reviews also helped, but to a lesser extent than previews. The efectiveness of these priming cues varied based on the type of reading material and the timing of the cue presentation.
Despite the efectiveness of previews in aiding comprehension, most participants expressed a preference for reviews. They found reviews more helpful in summarizing key points and aiding in answering comprehension questions. A significant number of participants (79.3%), however, preferred having some form of summary (either review or preview) compared to none.
These studies suggest that smart reading applications could incorporate text summaries, especially previews after interruptions, to facilitate better comprehension and re-engagement with reading tasks. Hence, we are now in the process of utilizing LLMs to generate text summaries and visualizations on the fly to create text primes. Figure 2 shows the process of creating three types of primes—text summary, image story, and a mindmap—from prompting to Image-Priming Generation
Please write a brief and concise summary of the following article. It should provide an overview of the main points, key arguments, and ndings, but should not exceed 100 words. Ensure the summary accurately re ects the essence of the article, giving the reader an understanding of the content without having to read the entire piece. Remember, this summary should serve as a preview to the main article. <story goes here. .> Read the fol owing story and identify 4 key moments in the story. I want to generate 4 images based on the key moments. Describe each key moment in adjectives and verbs so that I can describe it to an il ustrator. Al the pictures should be described in a manner that they are independent. Replace Chan with Yui Aragaki, a famous Japanese actor.
<story goes here. .> Mindmap-Priming Generation
Read the fol owing passage and create a succinct mind map in Markdown format for XMind compatibility, suitable for mobile phone display with a tree table structure. <story goes here. .> promptTheprompt oimagine /imagine prompt pave the way for a more seamless and enjoyable reading experience in the
digital age.
Our work defines and highlights the potential of generative primes, which include text summaries and visualizations, to mitigate the challenges of reading in distractive environments.
By providing readers with informative cues before and after distractions occur, these generative primes have the ability to enhance comprehension, recall, and engagement.
Further research may focus on the custo mization options and evaluate the efectiveness of these primes in diferent reading contexts.
Ultimately, we aim to contribute to the development of reading interfaces that empower readers to navigate distractions and sea mlessly continue their reading tasks. We thank the participants of our studies as well as the continuous support by Adobe Research and the team at the Documents Intelligence Lab, foremost Rajiv Jain and Jennifer Healey. [11] OpenAI, Chatgpt: Optimizing language models for dialogue, 2020. [12] OpenAI, Gpt-4 technical report, 2023. arXiv:2303.08774. [13] K. Angerbauer, T. Dingler, D. Kern, A. Schmidt, Utilizing the efects of priming to facilitate text comprehension, in: Proceedings of the 33rd Annual ACM Conference Extended Abstracts on Human Factors in Computing Systems, CHI EA ’15, Association for Computing Machinery, New York, NY, USA, 2015, p. 1043–1048. URL: https: //doi.org/10.1145/2702613.2732914. doi:10.1145/2702613.2732914.