OPRA-Vis: Visual Analytics System to Assist Organization-Public Relationship Assessment with Large Language Models

Authors: Sangbong Yoo*, Seongbum Seo*, Chanyoung Yoon, Hyelim Lee, Jeong-Nam Kim, Chansoo Kim, Yun Jang†, and Takanori Fujiwara
(*Equal contribution, †Corresponding author)

This repository contains the source code and materials for OPRA-Vis, a visual analytics system that leverages Large Language Models (LLMs) for Organization-Public Relationship Assessment (OPRA) without requiring extensive labeled datasets.

Abstract

OPRA-Vis integrates LLM prompting with interactive visualizations to help PR experts analyze public opinion data. The system employs few-shot examples and expert-informed clues to guide LLM reasoning, while providing visualizations that reveal the assessment process for expert review and refinement.

Demo Video

OPRA-Vis-compressed.mp4

▶ Watch on YouTube

System Architecture

(a) Concept Labeling: Uses Gemma with Chain-of-Thought prompting for OPRA concept classification
(b) Sentiment Analysis: Employs BERT for sentiment classification and word frequency analysis
(c) Certainty Computation: Calculates Certainty of Concepts (CoC) for uncertainty quantification

Interface Components

OPRA Concept Space View

The scatter plot features an octagonal layout that displays sentence positioning based on OPRA concepts (Trust, Satisfaction, Commitment, Control Mutuality). The visualization uses color-coded certainty levels with a blue-yellow-red scheme and includes tag clouds for sentiment-aware word analysis.

Position Adjustment using Gravity Model

Physics-inspired gravity model that adjusts sentence positions based on Certainty of Concepts (CoC) values, creating intuitive spatial clustering.

Usage

Basic Usage

python app.py --data {dataname} --scale {scale} [--volume {volume}]

Parameters

• --data: Dataset name - Required
  • amazon: Amazon product reviews
  • local: Google Local business reviews
  • imdb: IMDB movie reviews
  • jigsaw: Jigsaw toxic comments dataset
• --scale: Measurement scale - Required
  • opra: Organization-Public Relationship Assessment (4 concepts)
  • toxicity: Toxicity classification (5 dimensions)
• --volume: Number of data samples to load - Optional

Examples

# Run OPRA analysis on Amazon reviews
python app.py --data amazon --scale opra

# Run toxicity analysis on Jigsaw dataset
python app.py --data jigsaw --scale toxicity --volume 10000

# Run OPRA analysis on Google Local reviews
python app.py --data local --scale opra --volume 10000

License

OPRA-Vis is released under the Apache-2.0 license. See the LICENSE file for more details.