How to Do Monolingual, Multilingual, and Cross-lingual Text Classification in April, 2024

For monolingual text classification, fine-tuning language models remains a strong baseline solution compared to using prompting with LLMs

Cross-lingual classification can be achieved through multiple approaches:

• Translating training data to target language
• Using multilingual language models with adapters
• Back-translation techniques
• Prompting multilingual LLMs

Model selection considerations:

• Open source options (BERT, T5, Mistral) vs closed source (GPT)
• Model size vs computational resources
• Language coverage in pre-training data
• Task-specific requirements

Adapter-based fine-tuning provides an efficient approach for multilingual tasks:

• Requires less computational resources
• Allows freezing base model layers
• Enables language-specific adaptations

Translation quality significantly impacts cross-lingual performance:

• DeepL showed strong results for translation tasks
• Consider language family relationships when choosing translation paths
• Multiple translation steps can compound errors

Data considerations:

• Balance datasets across languages and classes
• Account for morphological differences between languages
• Consider domain-specific vocabulary and context
• Leverage existing multilingual datasets when available

LLM prompting strategies:

• Use clear, consistent prompt formats
• Include task-specific instructions
• Consider few-shot examples
• Restrict output format for classification tasks

Evaluation metrics and benchmarks:

• Cross-check model performance across languages
• Compare against monolingual baselines
• Use standardized benchmarks like LLM leaderboards
• Consider practical deployment constraints