Innovating AI for humanitarian action in emergencies

Aradhana R, Rajagopal A, Nirmala V, Immanuel Johnraja Jebadurai

Keywords: Large Language Models, Humanitarian action

TL;DR: AI for humanity

Abstract:

While the research community is making significant leaps in Generative AI, millions of children and women are facing humanitarian crisis. UNICEF reports millions of children are experiencing a high burden of unaddressed grief, loss, trauma due to recent war in Middle East and Ukraine-Russia. WHO identifies the challenge of providing timely psychological first aid (PFA) for victims of war as a humanitarian response. Due to limited access to professional counselors inside war torn regions, and the mounting psychological trauma of war on children, the need for research to creatively apply AI to sup-port vulnerable people in emergencies with Humanitarian As-sistance. The challenge is to create an AI that can offer per-sonalized counselling for previously unseen scenarios, given people have experienced different types of scenarios of vio-lence. To create LLM powered AI copilots that can be trust-ed, the paper presents a novel design for explain-ability to as-sist UNRWA staff to offer therapy at refugee camps.
WHO reports 1 in every 5 people exposed to war are impacted by mental health disorder. This is the 2nd challenge we addressed. To provide timely mental health intervention for large populations in inaccessible war zones such Gaza, in-spite of occasional internet availability in warzones, we de-veloped an approach to finetune a GPT-2 model personalized for every victim’s situation. This opens up the opportunity to create & deploy personalized GPT-2 models on smartphone of millions of people to enable personalized healing.
The demos of both the solutions including source code are at presented at https://sites.google.com/view/hadr-ai.

Submission Number: 4