MTRACE: Multilingual Retrieval-Augmented Generation for Temporally Diverse Text Corpora

Souhail Bakkali; Anthony Mudet

cs.DLDec 14, 2025

MTRACE: Multilingual Retrieval-Augmented Generation for Temporally Diverse Text Corpora

Authors: Souhail Bakkali, Anthony Mudet

Organizations: Univ Rennes, CNRS, IRISA - UMR 6074, Rennes, France · L3i-lab, La Rochelle Universit´e, France

Abstract

Large multilingual knowledge bases expose temporally diverse information, yet retrieval quality remains sensitive to lexical variation and cross-lingual terminology shifts. We develop and evaluate MTRACE (Multilingual Temporal Retrieval-Augmented Generation with evidence grounding), a pipeline designed to test whether query expansion and multi-query fusion mitigate vocabulary mismatch in temporally layered text corpora, on the French and English subsets of MIRACL. Our approach integrates: (i) semantic query expansion (SQE) and multi-query fusion via Reciprocal Rank Fusion (RRF), targeting retrieval stability under query variation; (ii) a generation prompt enforcing strict grounding in retrieved evidence and explicit abstention when evidence is insufficient; and (iii) a modular architecture enabling systematic component evaluation. Ablation studies on Named Entity Recognition (NER) and embedding model selection demonstrate the importance of syntactic coherence in entity extraction and of self-retrieval and efficiency measurements for retriever selection. Our end-to-end evaluation over 50 constructed queries shows faithful answers for well-supported queries, correct abstention on unanswerable questions, and no re-scored similarity gains from multi-query fusion over single-query dense retrieval. By scoping our claims to a clean, text-only baseline, we separate these effects from OCR-noise confounds; direct measurement of diachronic lexical drift is left to future work. Code and configurations are available at \url{https://anonymous.4open.science/r/MIRAGE-8EAA/

Figures & tables

Figure 1: Overview of MTRACE, our multilingual RAG pipeline. Ingestion (Phase 1): preprocessing, chunking, NER, and vector indexing. Query (Phase 2): the query is expanded via an LLM, the original and reformulated query rankings are fused with RRF, and a generator constrained to retrieved evidence answers the query, abstaining when evidence is insufficient.

Model	Extracted Entity Examples
wikineural	[(’Walter Porzig’, ’PER’), (’Mei’, ’PER’)]
bert-base-multilingual-cased	[(’Selon le lingu’, ’LABEL_1’), (’##iste allemand Walter Porzig’, ’LABEL_0’), ...]

Table 1: Qualitative comparison of entity extraction.

Model	Top-5	Drop	Time(s)	Dim.	Semantic	Efficiency
e5-small-v2	0.9073	0.105	125	384	Excellent	Good compromise
multilingual-e5-large	0.9134	0.097	360.6	1024	Excellent	Average
SFR-Embedding-Mistral	0.8123	0.2090	6143	4096	Good	Long + heavy
linq-embed-mistral	0.7410	0.2918	6184	4096	Fairly good	Long + heavy
MiniLM	0.7222	0.3087	72.91	384	Low	Very fast/lightweight

Table 2: Comprehensive evaluation of embedding models (self-retrieval protocol; multilingual-e5-large denotes intfloat/multilingual-e5-large-instruct ).

Model	Silhouette	DB ( $\downarrow$ )	CH ( $\uparrow$ )
multilingual-e5-large	0.014	5.66	10430.90
MiniLM	0.01	6.13	10286.34
SFR	0.0106	6.15	7032.89
e5-small-v2	-0.001	5.08	7627.89
linq-mistral	0.01	6.69	9065.89

Table 3: Clustering diagnostics of the vector space.

Model	@1 (D)	@5 (D)	$\Delta 1\rightarrow 2$ (D)	@1 (F)	@5 (F)	$\Delta 1\rightarrow 2$ (F)	Time (s)
multilingual-e5-large	0.8753	0.8645	0.0085	0.8653	0.8505	0.0102	5.0
e5-mistral-7b-instruct	0.6796	0.5729	0.0707	0.5928	0.4685	0.0776	46.3
SFR-Embedding-Mistral	0.7208	0.6863	0.0252	0.6987	0.6604	0.0263	49.2
Linq-Embed-Mistral	0.6080	0.5679	0.0247	0.5777	0.5344	0.0265	50.4

Table 4: Single-query dense retrieval (D) vs. multi-query fusion (F): mean cosine Sim@1/Sim@5 and top-1/top-2 gap when each condition’s top-

K

candidates are re-scored by the listed model, averaged over the 50 evaluation queries. Time (s) is total re-ranking time over the 50 queries under both conditions (model load excluded), measured on an NVIDIA T4 in fp16.

Question (Category)	Faithfulness	Answer Relevancy
Event-oriented Query: What were the primary reasons for the start of the American Civil War?	0.833	0.826
Generated Answer: Les principales raisons pour le déclenchement de la Guerre civile américaine sont identifiées comme étant la question de l’esclavage et les différences économiques entre le Nord industriel et le Sud agricole. Cependant, les causes de la sécession sont complexes et ont été controversées depuis le début de la guerre. Cet événement a commencé le 12 avril 1861, lorsque l’artillerie de la Caroline du Sud a ouvert le feu sur Fort Sumter.
Entity-focused Query: Qui est Antoine Meillet? (French)	0.500	1.000
Generated Answer: Antoine Meillet est un linguiste français, comme l’a souligné le linguiste français Antoine Meillet dans l’article "Arménien" (Source ID: 273).
Event-oriented Query: Quelles ont été les conséquences majeures de la Première Guerre mondiale en Europe? (French)	0.000	0.000
Generated Answer: Je ne peux pas répondre à cette question en me basant uniquement sur les informations fournies. Les extraits de journaux ne fournissent pas de détails sur les conséquences majeures de la Première Guerre mondiale en Europe.

Table 5: RAGAS evaluation results with generated answers (4 representative queries of 50; categories as in section 4.1 ).

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Query Variation Generation
Task: Reformulate the following question in { num_variations } different ways for the purpose of efficient search.
Constraints: • Preserve the original meaning • Be concise • One reformulation per line • Do not number the reformulations Input: { original_query }
Output: Reformulations:

Appendix

Table 6: Prompt template for query variation generation.

Answer Generation Prompt
Task: Answer the question using exclusively the provided "Segment Extracts".
Constraints: • Do not use outside knowledge or make assumptions. • If information is missing, state: "I cannot answer this question based solely on the provided information." • Verify that extracted details relate to the main event, not unrelated mentioned events. • Ensure relationships between entities are explicitly described before asserting them. • Do not refer to yourself as an AI model. • Note: A consequence is a result occurring after an event; a cause is not a consequence. Input: : { context_text }, Question: { query }
Output: Answer: {}

Appendix

Table 7: Structured prompt for answer generation.

Text QA Prompt
You must answer the following question using the provided text excerpts. Your task is to answer the question using EXCLUSIVELY the information contained in the excerpts provided below.

Input: Segment Excerpts: { context_text }
Input: Question: { query }

Constraints: • Make no assumptions; do not use any external knowledge. • If the excerpts don’t contain the necessary information to answer the question, you MUST explicitly state: "I cannot answer this question based solely on the provided information." • Answer in the same language as the question. • Carefully verify that each piece of information extracted pertains solely to the main event of the question, excluding those mentioned in the context, unless the causal link is explicit. • If you identify actors, ensure their relationships are explicitly described in the excerpts before asserting them. • Do not refer to yourself as an "AI model". • A consequence is a result or effect that occurs after an event. An event that triggers or causes another event is not a consequence of that event itself. Output: {}

Appendix

Table 8: Structured prompt for text QA.

Panel of Figure 2 — Figure 2: Comparative evaluation of NER models on multilingual text, assessing: (a) computational efficiency; (b) extraction capability; and (c) output quality (syntactic coherence).

Panel of Figure 3 — Figure 3: Evaluation of embedding models: (a) computational efficiency; (b) semantic retrieval performance; and (c) retrieval confidence.

Condition	e5-large	e5-mistral	SFR	Linq
Dense	0.8755 / 0.8647	0.6880 / 0.5780	0.7277 / 0.6941	0.6171 / 0.5785
Dense+exp	0.8657 / 0.8511	0.5949 / 0.4701	0.7056 / 0.6689	0.5882 / 0.5459
BM25	0.8679 / 0.8482	0.6802 / 0.5850	0.7168 / 0.6693	0.6059 / 0.5579
BM25+exp	0.8579 / 0.8400	0.6294 / 0.5250	0.6972 / 0.6538	0.5810 / 0.5378
Hybrid	0.8754 / 0.8635	0.6879 / 0.5870	0.7279 / 0.6911	0.6190 / 0.5777
Hybrid+exp	0.8676 / 0.8534	0.6303 / 0.5137	0.7117 / 0.6757	0.5982 / 0.5586

Appendix

Table 9: Six retrieval conditions under the protocol of table 4 : mean cosine Sim@1 / Sim@5 after re-scoring each condition’s top-10 candidates with the listed model, over the 50 queries. Dense and Dense+exp replicate the two conditions of table 4 within one session and one scorer precision (see provenance below). Hybrid fuses the dense and BM25 lists with RRF (

k=60

). Model abbreviations: e5-large = multilingual-e5-large , e5-mistral = e5-mistral-7b-instruct , SFR = SFR-Embedding-Mistral , Linq = Linq-Embed-Mistral .

Contrast	e5-large	e5-mistral	SFR	Linq
BM25 $-$ Dense	$-0.0076^{*}$ $[-0.0112,-0.0045]$	$-0.0078$ $[-0.0298,+0.0141]$	$-0.0109^{*}$ $[-0.0178,-0.0045]$	$-0.0112^{*}$ $[-0.0198,-0.0030]$
Hybrid $-$ Dense	$-0.0001$ $[-0.0002,+0.0000]$	$-0.0001$ $[-0.0125,+0.0124]$	$+0.0001$ $[-0.0018,+0.0022]$	$+0.0019$ $[-0.0018,+0.0058]$
Hybrid $-$ BM25	$+0.0075^{*}$ $[+0.0044,+0.0111]$	$+0.0078$ $[-0.0094,+0.0250]$	$+0.0110^{*}$ $[+0.0054,+0.0172]$	$+0.0131^{*}$ $[+0.0063,+0.0206]$
Dense+exp $-$ Dense	$-0.0098^{*}$ $[-0.0143,-0.0058]$	$-0.0931^{*}$ $[-0.1289,-0.0599]$	$-0.0221^{*}$ $[-0.0325,-0.0130]$	$-0.0289^{*}$ $[-0.0423,-0.0166]$
BM25+exp $-$ BM25	$-0.0100^{*}$ $[-0.0167,-0.0040]$	$-0.0508^{*}$ $[-0.0893,-0.0162]$	$-0.0196^{*}$ $[-0.0345,-0.0061]$	$-0.0250^{*}$ $[-0.0417,-0.0097]$
Hybrid+exp $-$ Hybrid	$-0.0078^{*}$ $[-0.0122,-0.0039]$	$-0.0577^{*}$ $[-0.0917,-0.0260]$	$-0.0161^{*}$ $[-0.0251,-0.0083]$	$-0.0208^{*}$ $[-0.0312,-0.0114]$

Appendix

Table 10: Paired per-query differences in cosine Sim@1 under the single-session protocol above: each cell gives the mean difference and, beneath it, the bootstrap 95% CI (10 000 resamples, seed 42). * marks intervals excluding zero. Dense+exp

-

Dense reproduces the effect in table 4 ; the final two rows give Sim@5.

Explore similar work

Apr 28, 2026cs.CL

CORAL: Adaptive Retrieval Loop for Culturally-Aligned Multilingual RAG

Multilingual retrieval-augmented generation (mRAG) is often implemented within a fixed retrieval space, typically via query or document translation or multilingual embedding vector representations. However, this approach may be inadequate for culturally grounded queries, in which retrieval-condition misalignment may occur. Even strong retrievers and generators may struggle to produce culturally relevant answers when sourcing evidence from inappropriate linguistic or regional contexts. To this end, we introduce CORAL (COntext-aware Retrieval with Agentic Loop, an adaptive retrieval methodology for mRAG that enables iterative refinement of both the retrieval space (corpora) and the retrieval probe (query) based on the quality of the evidence. The overall process includes: (1) selecting corpora, (2) retrieving documents, (3) critiquing evidence for relevance and cultural alignment, and (4) checking sufficiency. If the retrieved documents are insufficient to answer the query correctly, the system (5) reselects corpora and rewrites the query. Across two cultural QA benchmarks, CORAL achieves up to a 3.58%p accuracy improvement on low-resource languages relative to the strongest baselines.

Nayeon Lee, Jiwoo Song, Byeongcheol Kang

Naver · Samsung Research

Question-Answering Benchmarks Coral

Ask PDF

Sep 23, 2026cs.CL

TEMPS: Temporal Sentence Embeddings for Temporal Information Retrieval

Modern information retrieval (IR) systems rarely represent time, yet many information needs depend on it: in clinical, journalistic, and legal search, when an event occurred can decide whether a document is relevant. Dense retrievers and Retrieval-Augmented Generation (RAG) pipelines match queries to documents well on topic but poorly on time, so they surface content that is on-topic yet temporally wrong. We introduce Temporal Textual Similarity (TTS), a task that measures how well two anchored texts align in time, independent of their topical similarity. We then present TEMPS (Temporal Embedding Model for Precise Search), a modular temporal branch that attaches to a frozen semantic retriever and trains on that signal. It resolves anchored temporal expressions to intervals and moment-matches each one to a Gaussian; the resulting ordering supervises an anchor-date-conditioned encoder, whose score we fuse with the semantic score at inference. Grounding supplies the supervision, so training uses no hand-labeled temporal data. The temporal score itself is the Gaussian-KL inclusion measure from distributional embeddings; what TEMPS adds is the grounding and the moment-matched supervision. On three temporal benchmarks, TEMPS improves MRR for every semantic backbone tested and, on TS- Retriever, lifts R@1 from 19.92 to 25.39 over the prior temporal state of the art.

Mourad Hassani, Julien Romero, Amel Bouzeghoub +1

SAMOVAR, Télécom SudParis, Institut Polytechnique de Paris, France · Aldebaran Care, France

Timestep Embeddings Text Embeddings Similarity and Metrics

Ask PDF

Apr 22, 2026cs.CL

All Languages Matter: Understanding and Mitigating Language Bias in Multilingual RAG

Multilingual Retrieval-Augmented Generation (mRAG) leverages cross-lingual evidence to ground Large Language Models (LLMs) in global knowledge. However, we show that current mRAG systems suffer from a language bias during reranking, systematically favoring English and the query's native language. By introducing an estimated oracle evidence analysis, we quantify a substantial performance gap between existing rerankers and the achievable upper bound. Further analysis reveals a critical distributional mismatch: while optimal predictions require evidence scattered across multiple languages, current systems systematically suppress such ``answer-critical'' documents, thereby limiting downstream generation performance. To bridge this gap, we propose \textit{\textbf{L}anguage-\textbf{A}gnostic \textbf{U}tility-driven \textbf{R}eranker \textbf{A}lignment (LAURA)}, which aligns multilingual evidence ranking with downstream generative utility. Experiments across diverse languages and generation models show that LAURA effectively mitigates language bias and consistently improves mRAG performance.

Dan Wang, Guozhao Mo, Yafei Shi +9

Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences · University of Chinese Academy of Sciences · MYbank, AntGroup

Cross-Lingual Consistency Multilingual Language Pairs Linguistics Reranking

Ask PDF