CLIRudit: Cross-Lingual Information Retrieval of Scientific Documents
Paper: CLIRudit: Cross-Lingual Information Retrieval of Scientific Documents (arXiv, April 2025) Authors: Francisco Valentini, Diego Kozlowski, Vincent Larivière
Overview
CLIRudit tackles cross-lingual information retrieval (CLIR) for scientific documents: enabling users to find relevant scholarly literature even when the query language differs from the document language. This matters for academic search engines and databases, where a substantial share of high-quality research is published in languages other than English, yet most retrieval tooling and evaluation remain English-centric.
The forum post placing this paper in context frames it within the evolution of information retrieval systems:
- Sparse retrieval (BM25) — efficient, interpretable, but lexical and language-sensitive
- Dense retrieval — bi-encoder embeddings with approximate nearest-neighbor search; strong recall but sensitive to domain shift and long-tail queries
- Cross-encoder re-ranking — higher precision, but document representations cannot be precomputed
- Generative retrieval / RAG / agentic search — LLM-era paradigms where retrieval becomes iterative, verifiable, and plannable, with new emphasis on reasoning budgets and tool-calling action spaces
- Cross-lingual generalization when training and benchmarks are English-centered
- Credibility of offline metrics (nDCG@10, MRR, Recall@k) versus real user satisfaction
- Latency, cost, and safety constraints in production systems
- Hallucination risks when retrieval feeds generation, and adversarial risks on open retrieval sources
- A Comprehensive Evaluation of Embedding Models and LLMs for IR and QA
- CL2CM: Improving cross-lingual cross-modal retrieval
- Cross-Lingual Cross-Modal Retrieval With Noise-Robust Fine-Tuning (IEEE)
- Evaluating Large Language Models for Cross-Lingual Retrieval (arXiv 2509.14749)
- Multimodal LLM Enhanced Cross-lingual Cross-modal Retrieval (MM 2024, DOI: 10.1145/3664647.3680886)
- The Cross-Lingual Cost: Retrieval Biases in RAG over Arabic-English Corpora (arXiv 2507.07543)
Typical Pipeline Described in the Post
1. Input and representation — encode queries, documents, and user context into dense or sparse representations, or structured prompts 2. Core modules — retriever, re-ranker, planner, memory, and tool interfaces, chained or combined per task 3. Learning strategies — supervised fine-tuning, contrastive learning, distillation, reinforcement learning (including process rewards), and synthetic data bootstrapping 4. Inference strategies — single-pass retrieval, iterative retrieval, parallel sub-queries, early stopping, and budget control
Open Problems Highlighted
Notes for Readers
This forum post is a structured overview built from the paper's abstract and public metadata. Specific experimental results, datasets, and metrics are not reproduced in detail here — consult the original PDF on arXiv before citing quantitative conclusions.