We present an overview of Task 2 from CheckThat! at CLEF 2025, which focuses on claim normalization. The tasks asks systems to transform informal and often noisy social media posts into clear, concise, and verifiable statements known as normalized claims, which capture the core factual assertion of a post, which makes them much easier to verify and fact-check. The task is especially relevant in multilingual and low-resource contexts, where the diversity of languages and limited labelled data pose serious challenges. Task 2 was conducted in two distinct settings: (i) monolingual, where systems were trained and tested on the same language, and (ii) zero-shot, where models had to normalize claims in a new target language without any in-language training data. The monolingual track covered thirteen languages, including English, German, French, Spanish, Portugese, Hindi, Marathi, Punjabi, Tamil, Arabic, Thai, Indonesian, and Polish. While the zero-shot setting introduced seven more languages, such as Dutch, Romanian, Bengali, Telugu, Korean, Greek, and Czech. This structure allowed us to evaluate both language-specific performance and cross-lingual generalization. In total, 18 teams participated in Task 2, submitting 1,226 valid runs across the two settings. The submissions were evaluated using the METEOR score. Many teams leveraged transformer-based models, multilingual embeddings, and retrieval-augmented strategies. In this paper, we outline the task setup, give details about the datasets, and provide a detailed summary of the diverse approaches adopted by the participating teams.
Social media have revolutionized global communication, removing geographical barriers and allowing
global knowledge exchange. However, it has also become a breeding ground for misinformation,
spreading false claims quickly across languages and cultures [
Social media posts are often written in vague, informal language, frequently mixing opinions, using
rhetorical questions, and incomplete thoughts. This makes it dificult to extract clear, check-worthy
claims, defined as factual statements that can be verified or disproven [
The task is a precursor to fact-checking, distilling the essence of the claim and removing any unnecessary information, thereby increasing the eficiency and reliability of the fact-checking process.
Despite eforts to combat misinformation across a variety of languages [
The CheckThat! Lab aims to accelerate the development of tools and datasets that enable diferent
phases of the fact-checking pipeline. Since its beginning, the lab has organized several shared tasks that
represent real-world issues in misinformation detection and verification, with a focus on multilingual,
cross-domain, and practical applications. The 2025 edition of the lab included four tasks in monolingual,
multilingual, and cross-lingual settings, covering over 20 languages across these tasks [
Task Description. In this year’s CheckThat! Lab, Task 2 addressed the growing need to extract verified claims from the informal language found on social media. Unlike standard fact-checking pipelines that rely on well-formed input, our task aimed to rewrite user-generated content—often imprecise, opinionated, or fragmented—into clear, concise, and factual statements, the way that human fact-checkers formulate the claims they are checking.
The task is especially timely and relevant in multilingual and low-resource settings. To simulate realistic fact-checking scenarios, Task 2 was conducted in two settings: • Monolingual: In the monolingual setting, training, development datasets are provided for the language used for testing. The model is trained, validated, and tested on the same language, allowing it to learn language-specific structures and patterns. The languages included in this setup are English, German, French, Spanish, Portugese, Hindi, Marathi, Punjabi, Tamil, Arabic, Thai, Indonesian, and Polish. • Zero-shot: The zero-shot setting provides only the test data for the target language, without any corresponding training or development data for it. The participants may train their models using data from other languages or conduct zero-shot experiments with large language models (LLMs), evaluating the performance in the target language without prior exposure. This setup tests the model’s ability to generalize to unseen languages. The languages in this setting are Dutch, Romanian, Bengali, Telugu, Korean, Greek, and Czech.
In the following sections, we give details about our dataset, a detailed overview of the participating systems, and discussion of the approaches.
Fact-checking is critical for combating the spread of false claims. As fully automating manual
factchecking is very time-consuming, researchers have worked on specific subtasks that can help human
fact-checkers. This encompasses a spectrum of tasks, including claim detection [
The proliferation of false claims on social media platforms has led to the development of specialized
systems tailored for handling informal texts from these platforms [
Most existing methods aimed at combating misinformation have primarily focused on English
[
Despite the growing interest in fact-checking across multiple languages, the task of claim
normalization has been largely unexplored beyond English [
Inspired by the principle of dataset recycling Koch et al. [35], we identified and reused four datasets, which we repurposed for the task of claim normalization. This reduces annotation efort as well as subjective annotation biases. Below, we describe each dataset in detail:
(a) CLAN [
(b) MultiClaim [28]: It contains multilingual fact-checking pairs obtained from 142 fact-checking sites, making it the largest dataset of fact-checks released to date, encompassing 39 languages. Each fact-checking article is represented in the dataset by its claim, title, publication date, and URL. However, the entire text of the articles has not been published. In addition, the dataset includes relevant social media posts with text, OCR of attached images (if any), publication date, social media platform, and fact-checker rating for each post. We used this dataset to collect claims from fact-checking websites and corresponding social media posts for our study. This allowed us to extract 21k in post-claim pairs. It is worth noting that we only use monolingual pairs from this dataset in our work.
(c) X-Claim [
To ensure the data quality of the final compiled corpus, we randomly selected 50 examples from each language and asked native speakers to verify the post and the corresponding normalized claims. For languages where we could not find native speakers, we used the Google Translate API to translate them into English and cross-checked the quality of the examples. Table 1 shows a few examples from our mCLAN dataset in diferent languages. We consolidated all examples from these datasets and performed a combined analysis. Table 1 shows a few examples from mCLAN dataset in diferent languages.
Through data compilation, we obtained a total of 28,012 instances in twenty languages from all datasets. To maintain uniformity, we used the train/dev/test splits from the original datasets. For languages with a small number of instances, e.g., around 100, we only kept the test sets with no training data. Table 2 gives details about the final dataset and the train/dev/test splits.
To better apprehend the distribution of languages, we analysed the dataset linguistically. With its diverse vocabulary and flexible syntax, English is the primary language used on several social media platforms [38]. Thus, our dataset is also primarily composed of English examples. While German and Dutch are less dominant, they still benefit from a shared Latin script and similar grammatical structure. The Indic languages in the dataset encompass Hindi, Marathi, Punjabi, and Bengali. Hindi uses the Devanagari script. Marathi also uses the Devanagari script, albeit with some diferences in the characters. The Gurmukhi script is used for Punjabi, and Bengali is written using the Bengali script. Due to their diverse scripts and extensive use of diacritics, Indic languages pose unique computational challenges. The dataset also includes two languages from the Dravidian languages: Tamil and Telugu. Both are important representatives of the Dravidian language family, with scripts derived from the ancient Brahmic script.
We received submissions from 18 teams, totalling 1,226 valid runs across all the languages; 12 of these teams submitted their working notes. Table 3 lists all teams and their ranking for each language.
Baseline and Evaluation Metric. We used mT5-large as our baseline. For the monolingual setting, we fine-tuned the model using language-specific training data, translating the instruction “Identify the central claim in the given post: < >” into the language of the test claim. This allowed the model to operate directly in the target language. We used METEOR as an evaluation measure.
Table 6 presents the results for the monolingual setup, while Table 4 reports the scores for the zero-shot setup. Most of the teams outperformed the baseline, while dfkinit2b [39], DS@GT [40], TIFIN [41], and AKCIT-FN [42] consistently ranked among the top-performers across most languages. Team dfkinit2b [39] was ranked first in 6 out of 13 languages in the monolingual setting. In the zero-shot setting, they were first across all seven unseen languages.
Most teams used sequence-to-sequence generation strategies for claim normalization, typically relying on transformer-based models. The most prevalent approach involved fine-tuning pretrained models such as BART, T5, mBART, and LLaMA on monolingual data.
Team dfkinit2b [39] participated in both settings, testing zero- and few-shot prompting with models such as Gemma-3, Qwen-3, Qwen-2.5, Llama-3.3, and Mistral. They explored various prompts and used cosine similarity to select demonstrations for few-shot learning. They also included adapter fine-tuning, data pre-processing with language checks and emoji removal, and data augmentation via translation. For the final submission, they ensembled top-performing model outputs by computing embedding centroids with multilingual SentenceTransformers and selecting claims closest to these centroids.
Team DS@GT [40] embedded the unnormalized claims from the pooled train and development
datasets, as well as from the test set, using state-of-the-art embeddings for each language. For testing,
a GPT-4o mini model was prompted following the approach discussed in [
Team TIFIN [41] fine-tuned Qwen-14B using LoRA with 4-bit precision for eficiency. They preprocessed data by filtering meaningful post-claim pairs, removing duplicates, and creating a unified multilingual dataset. Instruction-based fine-tuning incorporated Chain-of-Thought prompting with 5W1H questions to guide claim extraction. During inference, context resolution replaced partial posts with complete ones, and few-shot prompting with similar examples improved claim structure. This approach aimed to boost claim extraction accuracy and multilingual performance.
Team AKCIT-FN [42] adopted a dual-strategy approach tailored to data availability. For the 13 supervised languages, they fine-tuned various language-specific and multilingual Small Language Models (SLMs) such as PTT5, AraT5, and Varta T5. For the seven zero-shot languages, they used prompting with Large Language Models (LLMs) such as the GPT series, Gemini, and Qwen 2.5. Their methodology also included a data cleaning algorithm to remove repetitive content and trailing None placeholders, as well as cross-split deduplication. Few-shot prompting experiments for monolingual settings involved selecting examples randomly, based on dificulty (METEOR score), or using HDBSCAN cluster prototypes for semantic diversity.
Team Factiverse and IAI [43] focused on the monolingual setting, comparing four main approaches: zero-shot prompting, fine-tuning, Fixed In-Context Learning (FICL), and Adaptive In-Context Learning (AICL). For the ICL methods, they used a ChromaDB vector store with all-MiniLM-L6-v2 embeddings to retrieve semantically similar examples from the training data based on cosine distance. While FICL used a fixed number of top-K examples, the team’s novel AICL approach dynamically selected examples by applying a cosine distance threshold, eliminating the need to pre-determine the number of shots. They also explored data augmentation via machine translation for low-resource languages.
The MMA team [44] focused on the monolingual setting, exploring several model architectures and training strategies. Their approaches included fine-tuning a unified multilingual umt5 model on all languages, as well as training separate umt5 models for each language. They also tested zero-shot prompting with Qwen2.5 models and employed a parameter-eficient fine-tuning (PEFT) method using LoRA, which involved a two-stage process of first extracting key points and then generating a claim from those points. For Arabic, they conducted specific experiments by fine-tuning ara-t5 and augmenting the training data with scraped post-claim pairs from the Google Fact Check Tools API.
The UmuTeam [49] used a generative approach based on the Flan-T5-Base model for the Claim Extraction and Normalization task. Their strategy varied based on the data setting: for the monolingual scenarios, they fine-tuned a separate instance of Flan-T5-Base for each language, using only that language’s specific training data to allow the models to specialize. For the zero-shot languages, they ifne-tuned a single Flan-T5-Base model on the concatenated training data from all other languages, aiming to leverage cross-lingual transfer for generalization.
The UNH team [45] only experimented with the English language. Their fine-tuning experiments included fully fine-tuning a Flan-T5 Large model, using LoRA for a Flan-T5 Base model, and fine-tuning a DeepSeek-Llama-8b model. Their prompting strategies involved few-shot prompting with keywordbased example selection, iterative self-refinement to improve claim quality, and a Max Multi-Prompt method that simulated choosing the best output from several targeted prompts.
Saivineetha [50] focused on Hindi and Telugu. For Hindi, which was in the monolingual setting, they performed Parameter-Eficient Fine-Tuning (PEFT) using QLoRA with 4-bit quantization on the Gemma 2 9B instruct model. The model was instruction fine-tuned on the provided Hindi dataset of posts and normalized claims. For Telugu, which was in the zero-shot setting, they used zero-shot prompting with the Gemma 3 12B instruct model, using a prompt template designed to convert unstructured Telugu posts into normalized claims.
The JU_NLP@M&S team [48] framed the claim normalization task as a monolingual sequenceto-sequence generation problem, centered on fine-tuning a BART-Large transformer model. Their methodology included a preprocessing module for tokenization using byte-level BPE, padding inputs to a fixed length, and truncating where necessary. Model training was conducted for 5 epochs using Hugging Face’s Seq2SeqTrainer, employing mixed-precision (FP16) to optimize memory usage and a learning rate of 3e-5. For inference, they used beam search with four beams to enhance the quality of the generated claims.
Team Investigators [46] focused on the claim normalization task by fine-tuning several models, including LLaMA-3.2, BART, and T5, with a particular focus on the flan-t5-base model for the final submission. Their methodology was primarily monolingual, with extensive experiments on the English and Spanish datasets. Before training, they implemented a pre-processing pipeline to filter out records that were not in the target language. For the zero-shot setting, they experimented with cross-lingual transfer by training a model on the Spanish dataset and then evaluating it on the Korean test data.
Team OpenFact [47] experimented with several decoder-only LLMs, including LLaMA 3.1, DeepSeekR1, and GPT-4.1-mini. Their methodology had three steps: (1) generating up to three initial claim candidates, (2) iteratively refining each candidate using a self-reflection technique where the model provides feedback on its output, and (3) using an LLM as a judge to select the best among the refined candidates. They also performed supervised fine-tuning on the GPT-4.1-mini model using the cleaned training data. dfkinit2b [39] § § DS@GT [40] § § TIFIN [41] § § AKCIT-FN [42] § § Factiverse and IAI [43] § MMA [44] § UNH [45] § Investigators [46] § § § OpenFact [47] § § § JU_NLP@M&S [48] § Saivineetha [50] § § UmuTeam [49] § §
The participating teams in CheckThat! 2025 Task 2 tried several strategies for multilingual claim normalization. These approaches can be analyzed through four major dimensions: model architecture, ifne-tuning vs. in-context learning paradigms, data handling, and performance across monolingual and zero-shot settings. An overview of the approaches is given in Table 5.
The primary area of divergence among the teams was their selection of model architecture. Some teams handled the task as a typical sequence-to-sequence problem, using encoder-decoder models that excel at summarization. For instance, the JU_NLP@M&S team fine-tuned BART-Large for monolingual textto-text generation. UmuTeam, Investigators, and MMA explored variants of T5, including multilingual models such as Flan-T5 and UMT5. In contrast, other teams used decoder-only large language models to improve their in-context learning and reasoning abilities. OpenFact evaluated models such as LLaMA 3.1, DeepSeek-R1, and GPT-4.1-mini. Similarly, TIFIN and dfkinit2b chose Qwen for their multilingual performance and eficiency in fine-tuning. This distinction highlights the trade-of between the recognised strengths of encoder-decoder in generation tasks and the growing potential of decoderonly models for flexible reasoning.
Fine-tuning was a common choice among the participating teams. Several teams employed parametereficient fine-tuning (PEFT) approaches, such as LoRA or QLoRA. For example, Saivineetha fine-tuned Gemma 2 for Hindi, while TIFIN and dfkinit2b applied LoRA to Qwen models. OpenFact’s supervised ifne-tuning of GPT-4.1-mini was reported to be their most efective configuration. In contrast, teams using decoder-only models emphasized in-context learning (ICL). DS@GT used retrieval-based ICL, pulling top-3 similar examples from the training set as dynamic prompts. dfkinit2b also adopted semantic similarity-driven selection for few-shot prompts. Factiverse used Adaptive In-Context Learning (AICL), which dynamically modifies the number of in-context examples depending on similarity thresholds. TIFIN implemented a 5W1H prompting strategy, structuring claim-related information into six categories (Who, What, Where, When, Why, and How) to guide model reasoning. OpenFact and UNH also used self-refinement, where an LLM iteratively critiques and improves its outputs.
Due to the noisy nature of social media data, data preprocessing becomes crucial. OpenFact used GPT-4.1-mini to filter out training instances with some mismatches with the ground truth. To augment the training data, MMA scraped additional Arabic samples using Google’s Fact Check Tools API, while Investigators used the Gemini API to generate synthetic examples. DS@GT created a retrieval-first pipeline that reused existing normalizations when similar posts were found. dfkinit2b employed an ensemble method to generate claims based on five diferent approaches. The output closest to the centroid of all created embeddings was then chosen. This strategy worked well in both monolingual and zero-shot settings.
In the monolingual setting, where training data for 13 languages was available, the participating teams either trained language-specific models or used the data to retrieve information for ICL. Saivineetha, for example, trained a dedicated Hindi model, while DS@GT and Factiverse retrieved similar examples to construct prompts dynamically. In the zero-shot setting, the teams had to rely on cross-lingual generalization. UmuTeam and MMA developed multilingual models based on merged monolingual data and applied them to zero-shot languages. Other teams, such as TIFIN and DS@GT, used English-centric prompts and relied on the inherent multilingual capacity of the LLMs to handle the target languages. Among the most efective zero-shot strategies were dfkinit2b’s ensemble approach and OpenFact’s ifne-tuned GPT-4.1-mini, both of which performed consistently well across languages without labeled data.
We presented a detailed overview of Task 2 from the CheckThat! Lab at CLEF 2025. It focused on claim normalization, the task of transforming informal and noisy social media content into clear, concise, and verifiable statements. In total, 18 teams participated in the task. Most of the participants used Transformer-based models, with a clear trend towards leveraging large language models from the T5, Qwen, and Llama families. Common and efective strategies included parameter-eficient ifne-tuning, retrieval-augmented in-context learning, and sophisticated data preprocessing. The dual setting for monolingual and zero-shot evaluation provided a valuable framework for assessing both language-specific adaptation and cross-lingual generalization.
In this study, we employed mT5-large as the baseline system. All experiments were carried out under controlled conditions. To help with spell check suggestions, OpenAI GPT-4o was accessed through a plugin on Overleaf. The authors thoroughly evaluated and edited all of the tool’s suggestions. No generative AI tools were used to generate the content of the main manuscript. The authors take full responsibility for the final content of the publication.
Social Media, PNAS nexus 3 (2024) pgae217. [27] A. Gupta, V. Srikumar, X-Fact: A New Benchmark Dataset for Multilingual Fact Checking, in: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), Association for Computational Linguistics, Online, 2021, pp. 675–682. [28] M. Pikuliak, I. Srba, R. Moro, T. Hromadka, T. Smoleň, M. Melišek, I. Vykopal, J. Simko, J. Podroužek, M. Bielikova, Multilingual Previously Fact-Checked Claim Retrieval, in: H. Bouamor, J. Pino, K. Bali (Eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Singapore, 2023, pp. 16477–16500. [29] Y.-C. Chang, C. Kruengkrai, J. Yamagishi, XFEVER: Exploring Fact Verification across Languages, in: Proceedings of the 35th Conference on Computational Linguistics and Speech Processing (ROCLING 2023), 2023, pp. 1–11. [30] P. Nakov, A. Barrón-Cedeno, T. Elsayed, R. Suwaileh, L. Màrquez, W. Zaghouani, P. Atanasova, S. Kyuchukov, G. Da San Martino, Overview of the CLEF-2018 CheckThat! Lab on Automatic Identification and Verification of Political Claims, in: Experimental IR Meets Multilinguality, Multimodality, and Interaction: 9th International Conference of the CLEF Association, CLEF 2018, Avignon, France, September 10-14, 2018, Proceedings 9, Springer, 2018, pp. 372–387. [31] S. Shaar, A. Nikolov, N. Babulkov, F. Alam, A. Barrón-Cedeno, T. Elsayed, M. Hasanain, R. Suwaileh, F. Haouari, G. Da San Martino, et al., Overview of CheckThat! 2020 English: Automatic Identification and Verification of Claims in Social Media., CLEF (Working Notes) 2696 (2020). [32] P. Nakov, G. Da San Martino, T. Elsayed, A. Barrón-Cedeño, R. Míguez, S. Shaar, F. Alam, F. Haouari, M. Hasanain, W. Mansour, et al., Overview of the CLEF-2021 CheckThat! Lab on Detecting Check-Worthy Claims, Previously Fact-Checked Claims, and Fake News, in: Proceedings of the 12th International Conference of the CLEF Association: Information Access Evaluation Meets Multiliguality, Multimodality, and Visualization, CLEF ’2021, Bucharest, Romania (online), 2021, pp. 264–291. [33] A. Barrón-Cedeño, F. Alam, A. Galassi, G. Da San Martino, P. Nakov, T. Elsayed, D. Azizov, T. Caselli, G. S. Cheema, F. Haouari, et al., Overview of the CLEF–2023 CheckThat! Lab on Checkworthiness, Subjectivity, Political Bias, Factuality, and Authority of News Articles and their Source, in: International conference of the cross-language evaluation forum for European languages, Springer, 2023, pp. 251–275. [34] A. Barrón-Cedeño, F. Alam, J. M. Struß, P. Nakov, T. Chakraborty, T. Elsayed, P. Przybyła, T. Caselli, G. Da San Martino, F. Haouari, et al., Overview of the CLEF-2024 CheckThat! Lab: CheckWorthiness, Subjectivity, Persuasion, Roles, Authorities, and Adversarial Robustness, in: International Conference of the Cross-Language Evaluation Forum for European Languages, Springer, 2024, pp. 28–52. [35] B. Koch, E. Denton, A. Hanna, J. G. Foster, Reduced, Reused and Recycled: The Life of a Dataset in Machine Learning Research, in: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. [36] Z.-Y. Dou, G. Neubig, Word Alignment by Fine-tuning Embeddings on Parallel Corpora, in: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, 2021, pp. 2112–2128. [37] V. Bhatnagar, D. Kanojia, K. Chebrolu, Harnessing Abstractive Summarization for Fact-Checked Claim Detection, in: Proceedings of the 29th International Conference on Computational Linguistics, 2022, pp. 2934–2945. [38] A. Petrosyan, Most used languages online by share of websites 2024, 2024. Accessed: 01 June 2024. [39] T. Anikina, I. Vykopal, S. Kula, R. K. Chikkala, N. Skachkova, J. Yang, V. Solopova, V. Schmitt, S. Ostermann, dfkinit2b at CheckThat! 2025: Leveraging LLMs and Ensemble of Methods for Multilingual Claim Normalization, in: [51], 2025. [40] A. Pramov, J. Ma, B. Patel, DS@GT at CheckThat! 2025: A Simple Retrieval-First, LLM-Backed
Framework for Claim Normalization, in: [51], 2025. [41] M. Sharma, A. Suneesh, M. Jain, P. K. Rajpoot, P. Devadiga, B. Hazarika, A. Shrivastava, K. Gurumurthy, A. B. Suresh, A. U. Baliga, TIFIN at CheckThat! 2025: Reasoning-Guided Claim Normalization for Noisy Multilingual Social Media Posts, in: [51], 2025. [42] F. L. N. Almada, K. D. P. Mariano, M. A. Dutra, V. E. d. S. Monteiro, J. R. S. Gomes, A. R. Galvão Filho, A. d. S. Soares, Akcit-FN at CheckThat!2025: Switching Fine-Tuned SLMs and LLM Prompting for Multilingual Claim Normalization, in: [51], 2025. [43] P. Amatya, V. Setty, Factiverse and IAI at CheckThat! 2025: Adaptive ICL for Claim Extraction, in: [51], 2025. [44] M. Saeed, M. Yasser, M. Torki, N. Elmakky, MMA at CheckThat! 2025: Multilingual Claim
Normalization of Social-Media Posts, in: [51], 2025. [45] J. Wilder, N. Kadapala, Y. Xu, M. Alsaadi, M. Rogers, P. Agrawal, A. Hassick, L. Dietz, UNH at
Check That! 2025: Fine-tuning Vs Prompting, in: [51], 2025. [46] S. M. A. Hashmi, S. Aamir, M. Anas, T. Usmani, F. Alvi, A. Samad, Investigators at CheckThat! 2025: Using LLMs to Improve Fact-Checking, in: [51], 2025. [47] M. Sawiński, K. Węcel, E. Księżniak, OpenFact at CheckThat! 2025: Application of self-reflecting and reasoning LLMs for fact-checking claim normalization, in: [51], 2025. [48] M. Mondal, S. Saha, D. Saha, D. Das, JU_NLP@M&S at CheckThat! 2025: Automated Claim Extraction and Normalization for Misinformation Detection in Social Media Content, in: [51], 2025. [49] T. B. Beltrán, R. Pan, J. A. García Díaz, R. Valencia García, UmuTeam at CheckThat! 2025:
Language-specific versus multilingual models for Fact-Checking, in: [51], 2025. [50] S. V. Baddepudi Venkata Naga Sri, Saivineetha at CheckThat! 2025: Exploring Fine-Tuning and
Zero-Shot Approaches for Claim Normalization, in: [51], 2025. [51] G. Faggioli, N. Ferro, P. Rosso, D. Spina (Eds.), Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CLEF 2025, Madrid, Spain, 2025.
Thai DS@GT 0.5859 AKCIT-FN 0.3179 dfkinit2b 0.2999 Baseline 0.2015 Factiverse and IAI 0.0965 OpenFact 0.0872 aryasuneesh 0.0464 UmuTeam 0.0147
Arabic dfkinit2b 0.5037 DS@GT 0.5035 MMA 0.4584 OpenFact 0.4175 TIFIN 0.3705 AKCIT-FN 0.3277 Factiverse and IAI 0.2457 Baseline 0.2186 UmuTeam 0.0003