# Zero-Shot Cross-Lingual Summarization via Large Language Models

**Authors:** Jiaan Wang, Yunlong Liang, Fandong Meng, Beiqi Zou, Zhixu Li, Jianfeng Qu, Jie Zhou  
**arXiv:** [2302.14229](https://arxiv.org/abs/2302.14229)  
**Date:** 2023-02-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.14229v3/

## Key points

- Large Language Models (LLMs) like GPT-3.5, ChatGPT, and GPT-4 are being used in cross-lingual summarization (CLS)
- Performance of LLMs in CLS tasks is not well understood
- Empirical exploration of LLM capabilities in zero-shot CLS using different prompts and paradigms
- ChatGPT and GPT-4 generate lengthy summaries with detailed information
- Interactive prompts help achieve a better balance between informativeness and conciseness
- GPT-4 achieves state-of-the-art zero-shot CLS performance comparable to fine-tuned mBART 50 on three widely used CLS datasets
- Some multi lingual and bilingual LLMs have limited zero-shot CLS ability
- CLS poses a challenge for LLMs due to its composite nature requiring simultaneous summarization and translation
- Future LLM research should utilize CLS as a testbed for evaluation purposes.

## Summary

In the field of cross-lingual summarization (CLS), the use of Large Language Models (LLMs) like GPT-3.5, ChatGPT, and GPT-4 has gained significant attention. However, their performance in CLS tasks is not well understood. To address this gap, the authors of this report empirically explore the capabilities of LLMs in zero-shot CLS using different prompts and paradigms (end-to-end and pipeline). They find that ChatGPT and GPT-4 tend to generate lengthy summaries with detailed information but can achieve a better balance between informativeness and conciseness with interactive prompts. Experimental results on three widely used CLS datasets show that GPT-4 achieves state-of-the-art zero-shot CLS performance comparable to fine tuned mBART 50. Additionally, some multi lingual and bilingual LLMs exhibit limited zero shot CLS ability. The authors emphasize that CLS poses a challenge for LLMs due to its composite nature requiring simultaneous summarization and translation. They recommend future LLM research to utilize CLS as a testbed for evaluation purposes.