# ChatGraph: Interpretable Text Classification by Converting ChatGPT Knowledge to Graphs

**Authors:** Yucheng Shi, Hehuan Ma, Wenliang Zhong, Gengchen Mai, Xiang Li, Tianming Liu, Junzhou Huang  
**arXiv:** [2305.03513](https://arxiv.org/abs/2305.03513)  
**Date:** 2023-05-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.03513v1/

## Key points

- The paper addresses limitations of ChatGPT, a large language model (LLM)
- Limitations include inflexibility of finetuning on downstream tasks and lack of interpretability in decision-making process
- Authors propose a framework that utilizes ChatGPT for text classification tasks while enhancing interpretability
- Framework involves extracting refined and structural knowledge from raw data using ChatGPT through knowledge graph extraction task
- Rich knowledge is converted into a graph, which is used to train an interpretable linear classifier for predictions
- Method shows significant improvements in performance compared to directly utilizing ChatGPT for text classification tasks
- Approach offers a more transparent decision-making process compared to previous methods
- Paper authored by Yucheng Shi, Hehuan Ma, Wenliang Zhong, Gengchen Mai, Xiang Li, Tianming Liu and Junzhou Huang in the field of computer science with primary categories including cs.CL (Computation and Language), cs.AI (Artificial Intelligence) and cs.LG (Machine Learning)

## Summary

The paper titled "ChatGraph: Interpretable Text Classification by Converting ChatGPT Knowledge to Graphs" addresses the limitations of ChatGPT, a large language model (LLM), which include the inflexibility of finetuning on downstream tasks and the lack of interpretability in its decision-making process. To overcome these limitations, the authors propose a novel framework that utilizes ChatGPT for text classification tasks while enhancing its interpretability. The framework involves extracting refined and structural knowledge from raw data using ChatGPT through a knowledge graph extraction task. This rich knowledge is then converted into a graph, which is used to train an interpretable linear classifier for making predictions. The effectiveness of this method is evaluated through experiments on four datasets, demonstrating significant improvements in performance compared to directly utilizing ChatGPT for text classification tasks. Additionally, this approach offers a more transparent decision-making process compared to previous text classification methods. The paper was authored by Yucheng Shi, Hehuan Ma, Wenliang Zhong, Gengchen Mai, Xiang Li, Tianming Liu and Junzhou Huang in the field of computer science with primary categories including cs.CL (Computation and Language), cs.AI (Artificial Intelligence) and cs.LG (Machine Learning).