An Exploration of Causality in Social Media Data with Knowledge Graphs

Abstract

This study explores the integration of Knowledge Graphs (KGs) with Large Language Models (LLMs) to perform causal analysis on text-based social media data. The objective is to uncover the underlying causes and sentiments driving discussions around emergency management scenarios such as the Israel-Palestine conflict, thereby providing critical insights for decision-making. The research focuses on advanced techniques to effectively represent text as KGs and retrieve the most suitable context of information to analyse. Various methods are evaluated across different datasets. The proposed model, PRAGyan, combines LLMs and KGs under the Retrieval Augmented Generation (RAG) framework. It utilizes the Neo4J Graph Database to handle continuous real time data and GPT-3.5 Turbo LLM for causal reasoning. This yields more accurate results compared to the baseline model (GPT-3.5 Turbo LLM without KG). Quantitative analysis using metrics such as BLEU and cosine similarity show an improvement by 10%.