Ji's Motivation is 'Quite Clear' as She Produces a Fully Automated Decision-Making Assistant for Experts

5/8/2023 Aaron Seidlitz, Illinois CS

The Illinois CS professor is beginning a DARPA funded five-year research project with collaborators from Kitware, the University of Washington, and more to develop algorithmic decision-makers for users under difficult scenarios.

Written by Aaron Seidlitz, Illinois CS

A new research project – funded with $12 million over the next five years by the Defense Advanced Research Projects Agency’s (DARPA) In the Moment program – piqued Illinois Computer Science professor Heng Ji’s interest, due to its very clear potential to aid people in difficult decision making at the height of inopportune moments.

Insert caption here
Heng Ji

The project, entitled “ALIGN: Aligned Moral Language Models with Interpretable and Graph-based Knowledge,” follows the lead of a company called Kitware as the principal investigator.

Ji’s expertise in natural language processing, especially on information extraction and knowledge base population, plays well with fellow collaborators from Kitware, the University of Washington, the University of California, Berkeley, and more.

Together, the group’s primary goal “is to develop algorithmic decision-makers, which can be dynamically aligned to a trusted group or individual human decision-makers under difficult scenarios, while learning to make use of the provided decision-maker attributes.”

“The motivation for such a project is quite clear to me,” Ji said. “A lot of important decision making is done by humans, and it tends to be pretty subjective – as humans have certain biases. Currently, most of the decision-making algorithms are also very subjective.

“The goal of this project is to provide a fully automated assistant for experts. But the decisions need to be made in a very transparent way, so we must have strong evidence and make the reasoning explainable.”

This, she said, will provide crucial information in times of need for decision makers who oftentimes have difficult moments facing them.

Considering that DARPA’s dedication focuses on breakthrough technologies and capabilities for national security, these instances often take place when the health of a person or multiple people is in jeopardy.

In the past, human decision makers had processes in place to manage such responsibilities, but, as Ji noted, their opinion can be subjective.

Then, decision-making algorithms attempted to improve the process, but often included subjective information.

This project, Ji believes, has the potential to offset human bias with decision-making algorithms – but also the upfront information needed so that human decision makers can trust the information being presented. Or they can quickly glance through the protocols leading to the decision to ensure the right outcome was produced.

“So, the idea is that we have a new scenario, and we have options. We have two patients or three patients, but which one will be rescued first? Decision makers have their profiles and the patient description,” Ji explained. “We will use our natural language processing and knowledge discovery algorithms to construct a knowledge graph for our clients.”

The group’s efforts focus on converting this natural language guideline into a structured neural symbolic program.

“An important aspect comes down to the idea that users can trace the decision back through the program to find out how the decision was made in a step-by-step fashion – offsetting any distrust the expert does find in a recommendation. Then, if the program matching made a mistake, the expert can see that, and they can adjust their decision,” Ji said.

Beyond the work itself and what this project aims to produce, she also appreciates the opportunity work with this specific collaboration.

Ji said that KitWare’s presence is familiar to her after working on projects together in the past. This includes an effort that began earlier this year to map social media activity – conducted with fellow Illinois CS professor Tarek Abdelzaher and funded through DARPA – and a 2021 disinformation grant that was funded through DARPA’s Semantic Forensics (SemaFor) program.

Additionally, the other academic collaborators have already built up a rapport with Ji, who was ecstatic to find out they had been included in this effort.


Share this story

This story was published May 8, 2023.