# Text Classification via Large Language Models

**Authors:** Xiaofei Sun, Xiaoya Li, Jiwei Li, Fei Wu, Shangwei Guo, Tianwei Zhang, Guoyin Wang  
**arXiv:** [2305.08377](https://arxiv.org/abs/2305.08377)  
**Date:** 2023-05-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.08377v1/

## Key points

- CARP enhances the performance of large-scale Language Models (LLMs) in text classification tasks.
- LLMs like GPT-3 underperform compared to fine-tuned models in text classification due to their lack of reasoning ability and limited number of tokens allowed in in-context learning.
- CARP adopts a progressive reasoning strategy by prompting LLMs to identify superficial clues and inducing a diagnostic reasoning process for making final decisions.
- CARP utilizes a fine-tuned model on a supervised dataset for kNN demonstration search during the in-context learning phase.
- CARP achieves new state-of-the-art performances on four out of five widely-used text-classification benchmarks: SST-2, AGNews, R8, and R52. It performs comparably to state-of-the-art models on MR benchmark.
- CARP shows impressive abilities in low-resource and domain-adaptation setups, achieving comparable performances with only 16 examples per class compared to supervised models with 1,024 examples per class.
- The related work section discusses large language models (LLMs) categorized into encoder-only models like BERT and decoder-only models like GPT.
- CARP significantly improves text classification performance by incorporating clue-based prompting and reasoning strategies into LLMs.
- Future work aims to explore CARP in other natural language understanding tasks.

## Summary

, , , , 

This paper presents Clue And Reasoning Prompting (CARP) as a solution to enhance the performance of large-scale Language Models (LLMs) in text classification tasks. While LLMs like GPT-3 have achieved remarkable success, they still underperform compared to fine-tuned models in text classification. This is due to their lack of reasoning ability in addressing complex linguistic phenomena and the limited number of tokens allowed in in-context learning. To overcome these limitations, CARP adopts a progressive reasoning strategy tailored specifically for handling complex linguistic phenomena involved in text classification. It first prompts LLMs to identify superficial clues such as keywords, tones, semantic relations, and references. Based on these clues, CARP induces a diagnostic reasoning process for making final decisions. In addition, CARP utilizes a fine-tuned model on a supervised dataset for kNN demonstration search during the in-context learning phase. This allows the model to leverage both the generalization ability of LLMs and the task-specific evidence provided by the labeled dataset. The results demonstrate that CARP achieves new state-of-the-art performances on four out of five widely-used text-classification benchmarks: SST-2, AGNews, R8, and R52. It also performs comparably to state-of-the-art models on MR benchmark. Notably, CARP shows impressive abilities in low-resource and domain-adaptation setups. With only 16 examples per class, CARP achieves comparable performances to supervised models with 1,024 examples per class. The related work section discusses large language models (LLMs) and categorizes them into encoder-only models like BERT and decoder-only models like GPT. These models follow the pre-training then fine-tuning paradigm for NLP tasks. In conclusion, CARP significantly improves text classification performance by incorporating clue-based prompting and reasoning strategies into large-scale Language Models (LLMs). The approach achieves state-of-the-art results on multiple benchmarks and demonstrates promising capabilities in low-resource and domain-adaptation scenarios. Future work aims to explore CARP in other natural language understanding tasks.