TIFIN India at SemEval-2025: Harnessing Translation to Overcome Multilingual IR Challenges in Fact-Checked Claim Retrieval

2025-04-23

AI Summary

To address the challenge of retrieving verified factual claims in both monolingual and cross-lingual settings for misinformation governance, this paper proposes a two-stage retrieval framework: initial retrieval via a fine-tuned semantic embedding model, followed by LLM-driven fine-grained re-ranking. We present the first empirical validation of LLM-based translation for cross-lingual fact-checking retrieval, and introduce LLM-assisted cross-lingual semantic alignment alongside lightweight inference optimization techniques—enabling efficient, reproducible deployment on consumer-grade GPUs. Experiments show that our method achieves Success@10 of 0.938 on monolingual test sets and 0.810 on cross-lingual ones, significantly outperforming existing baselines. The approach offers a practical, resource-efficient solution for multilingual fact-checking in low-resource scenarios.

Technology Category

• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Machine Learning: Large Multimodal Models (LMMs)
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

We address the challenge of retrieving previously fact-checked claims in monolingual and crosslingual settings - a critical task given the global prevalence of disinformation. Our approach follows a two-stage strategy: a reliable baseline retrieval system using a fine-tuned embedding model and an LLM-based reranker. Our key contribution is demonstrating how LLM-based translation can overcome the hurdles of multilingual information retrieval. Additionally, we focus on ensuring that the bulk of the pipeline can be replicated on a consumer GPU. Our final integrated system achieved a success@10 score of 0.938 and 0.81025 on the monolingual and crosslingual test sets, respectively.

Problem

• Retrieving fact-checked claims in monolingual and crosslingual settings
• Overcoming multilingual information retrieval hurdles using LLM-based translation
• Ensuring pipeline replicability on consumer GPUs for efficiency

Innovation

• Two-stage retrieval with fine-tuned embedding model
• LLM-based translation for multilingual IR
• Consumer GPU-compatible pipeline design