Agentopic: LLM-Based Workflow for Explainable Topic Modeling
Agentopic introduces an innovative agent-based workflow for explainable topic modeling, utilizing the reasoning strengths of Large Language Models (LLMs). Traditional methods like Latent Dirichlet Allocation (LDA) and BERTopic often fall short in providing clarity on topic assignments. To remedy this, Agentopic employs several agents that work together to identify topics, validate them, group them hierarchically, and provide natural language explanations. This approach allows users to understand the rationale behind topic assignments, thereby improving interpretability while maintaining accuracy. When tested with the British Broadcasting Corporation (BBC) dataset, Agentopic reached an F1-score of 0.95, comparable to GPT-4.1, surpassing LDA's 0.93, and nearing BERTopic's 0.98. Additionally, the unseeded version of Agentopic enriched the BBC dataset with enhanced explanations.
Key facts
- Agentopic is a novel agent-based workflow for explainable topic modeling.
- It leverages the reasoning capabilities of Large Language Models (LLMs).
- Existing approaches like LDA and BERTopic lack transparency.
- Agentopic uses multiple agents for topic identification, validation, hierarchical grouping, and explanation.
- It enables users to trace reasoning behind topic assignments.
- Seeded with BBC dataset, Agentopic achieves F1-score of 0.95.
- This matches GPT-4.1, improves on LDA (0.93), and is close to BERTopic (0.98).
- Unseeded Agentopic generated explanations to augment the BBC dataset.
Entities
Institutions
- British Broadcasting Corporation (BBC)