Reproduce “Generative Query Expansion with Multilingual LLMs for Cross-Lingual Information Retrieval” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Olivia Macmillan-Scott, Roksana Goworek, Eda B. Özyiğit
Researchers from The Alan Turing Institute systematically evaluated generative query expansion methods using multilingual large language models (mLLMs) in cross-lingual information retrieval (CLIR), demonstrating performance improvements across various language pairs and identifying optimal strategies contingent on query length and linguistic characteristics.
Key insights
- Optimal generative query expansion strategies are highly dependent on the input query's length and type; zero-shot prompting is better for short queries, while few-shot prompting excels for longer, natural language questions.
- Generative expansion consistently improves CLIR performance but does not fully overcome absolute performance gaps for typologically distant or lower-resource languages.
- Supervised fine-tuning provides domain-specific performance benefits but can lead to degradation when applied to retrieval tasks with different query styles or domains than the fine-tuning data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.