Researchers have developed a multilingual question-answering system for financial exams, named DS@GT, which utilizes a retrieval-augmented pipeline built on LangGraph. The system identifies query language and retrieves relevant information from a large knowledge base using BGE-M3 embeddings and FAISS indexing. It then scores potential answers by analyzing next-token log-probabilities rather than generating free-form text, employing a strategy called Retrieval-Augmented Direct Scoring (RADS). For languages with fewer resources, it fuses retrieval indices using weighted Reciprocal Rank Fusion. The system routes queries to different models based on language, selecting Qwen3-14B for Arabic, Chinese, and Hindi; Qwen2.5-14B for English; and Llama-3.1-8B for Greek, after discovering significant performance disparities across languages. Notably, chain-of-thought prompting was found to degrade accuracy in Greek, and enabling a default thinking mode in Qwen3 negatively impacted Arabic performance. AI
IMPACT This research demonstrates advanced techniques for multilingual NLP in specialized domains, potentially improving access to financial knowledge in various languages.
RANK_REASON The cluster describes a research paper detailing a novel system for multilingual financial question answering. [lever_c_demoted from research: ic=1 ai=1.0]
Read on arXiv cs.IR (Information Retrieval) →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
