Grammar-Preserving Text Saliency Modulation (GP-TSM)

An AI-Resilient Text Rendering Technique for Reading and Skimming Documents
Ziwei Gu, Ian Arawjo, Kenneth Li, Jonathan K. Kummerfeld, Elena L. Glassman
In the 2024 ACM CHI conference on Human Factors in Computing Systems
CHI ’24, May 11–16, 2024, Honolulu, HI, USA

GP-TSM is an LLM-powered text rendering technique that supports reading and skimming by reifying recursive sentence compression in text saliency. Readers can skip over de-emphasized segments without compromising their reading flow/ comprehension of the text, while still being able to notice and recover from AI suggestions they disagree with.

Read the full paper.
Try the live demo here.
For more information, see our lab website.

Citation

@inproceedings{10.1145/3613904.3642699,
author = {Gu, Ziwei and Arawjo, Ian and Li, Kenneth and Kummerfeld, Jonathan K. and Glassman, Elena L.},
title = {An AI-Resilient Text Rendering Technique for Reading and Skimming Documents},
year = {2024},
isbn = {9798400703300},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3613904.3642699},
doi = {10.1145/3613904.3642699},
booktitle = {Proceedings of the CHI Conference on Human Factors in Computing Systems},
articleno = {898},
numpages = {22},
keywords = {human-AI interaction, natural language processing, text visualization},
location = {<conf-loc>, <city>Honolulu</city>, <state>HI</state>, <country>USA</country>, </conf-loc>},
series = {CHI '24}
}

In this example we show what it looks like when GP-TSM is applied to two paragraphs of text from GRE (The Graduate Record Examinations) Practice Exams. GP-TSM uses an LLM-based recursive sentence compression method to identify successive levels of detail beyond the core meaning of a passage, which are de-emphasized by rendering words with successively lighter but still legible gray text.

Installation

Clone the repository:

git clone https://github.com/ZiweiGu/GP-TSM

Then, set up the virtual environment (called venv) using virtualenv (installation here):


virtualenv -p python3 venv

Activate the virtual environment:


source venv/bin/activate

Install necessary packages:


pip install -r requirements.txt

Usage

Run the app (in development mode) with:

python3 app.py

For backend only, use the get_shortened_paragraph(orig_paragraph, k) function in llm.py or gptsm-lite.py. The latter is an alternative of the original GP-TSM algorithm that runs faster, and is designed for applications that require a high level of responsiveness or interactivity. It achieves higher speed by using smaller values for N and MAX_DEPTH and removing grammaticality from evaluation, which is a time-consuming metric to compute. However, this may mean that the key grammar-preserving feature can be violated at times. To achieve the best output quality, please use the original version in llm.py.

Running API server through Docker (for chrome extension)

docker compose build
docker compose up -d

Go to chrome://extensions/ in Chrome.
Enable Developer mode (top right corner).
Click Load unpacked.
Select the folder chrome_extension_unpacked to install the extension.

Researchers

Name	Affiliation
Ziwei Gu	Harvard University
Ian Arawjo	Harvard University
Kenneth Li	Harvard University
Jonathan K. Kummerfeld	University of Sydney
Elena L. Glassman	Harvard University

License

See LICENSE.md.

0xrushi / GP-TSM