# Salute the Classic: Revisiting Challenges of Machine Translation in the Age of Large Language Models

**Authors:** Jianhui Pang, Fanghua Ye, Longyue Wang, Dian Yu, Derek F. Wong, Shuming Shi, Zhaopeng Tu  
**arXiv:** [2401.08350](https://arxiv.org/abs/2401.08350)  
**Date:** 2024-01-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2401.08350v1/

## Key points

- Challenges in Neural Machine Translation (NMT) with Large Language Models (LLMs):
- Domain mismatch
- Amount of parallel data
- Rare word prediction
- Translation of long sentences
- Attention model as word alignment
- Sub-optimal beam search
- LLMs reduce reliance on parallel data during pretraining for major languages.
- LLMs significantly improve translation of long sentences up to 512 words.
- Domain mismatch and rare word prediction.
- Inference efficiency
- Translation of low-resource languages during pretraining
- Human-aligned evaluation
- Datasets and models released for further exploration.
- LLMs excel in translating long sentences and document-level tasks.
- Addressing domain mismatch and predicting rare words.
- Efficiency of inference
- Resource imbalance during pretraining for low-resource languages
- Human-like evaluation issues
- Model interpretability
- Experiments conducted using the Llama2-7b model, limiting generalizability to other LLMs such as GPT-4.
- Future studies should consider a broader range of base models and address potential limitations in experimental designs.

## Summary

This study explores the challenges and advancements in Neural Machine Translation (NMT) with the use of Large Language Models (LLMs). The six core challenges identified in previous research, including domain mismatch, amount of parallel data, rare word prediction, translation of long sentences, attention model as word alignment, and sub-optimal beam search, are revisited to assess their ongoing relevance. The findings reveal that LLMs reduce the reliance on parallel data during pretraining for major languages and significantly improve the translation of long sentences up to 512 words. However, challenges related to domain mismatch and rare word prediction persist. Additionally, three new challenges specific to LLMs in translation tasks are identified: inference efficiency, translation of low-resource languages during pretraining, and human-aligned evaluation. The study also releases datasets and models for further exploration. Further analysis demonstrates that LLMs excel in translating long sentences and document-level tasks but face limitations in addressing domain mismatch and predicting rare words. The efficiency of inference, resource imbalance during pretraining for low-resource languages, human-like evaluation issues, and model interpretability are highlighted as emerging challenges for future research. It is important to note that the experiments were conducted using the Llama2-7b model, which may limit generalizability to other LLMs such as GPT-4. Future studies should consider a broader range of base models and address potential limitations in experimental designs.