Multilingual Fact-Checking Pipelines

• Multilingual fact-checking pipelines are automated systems that verify claims across languages by aggregating, normalizing, and aligning textual and multimodal evidence.
• They employ advanced neural architectures, cross-lingual retrieval techniques, and evidence categorization to generate interpretable justifications and consistent verdicts.
• These pipelines ensure robust misinformation verification by integrating human evaluations with automated assessments and standardized metadata.

Multilingual Fact-Checking Pipelines

A multilingual fact-checking pipeline refers to an automated or semi-automated system designed to verify the veracity of claims across multiple languages and modalities (text, images, videos) using structured methodologies. Such pipelines have become indispensable for combating global-scale misinformation, especially as online content, social media narratives, and disinformation campaigns transcend linguistic and national boundaries. Recent advances involve not only textual claim verification but also the extraction and alignment of multimodal evidence, verdict normalization across diverse rating scales, and the generation of interpretable, language-aware justifications. These pipelines leverage neural architectures (transformers, LLMs), knowledge-rich retrieval, evidence aggregation, and rigorous evaluation protocols to ensure factual fidelity under heterogeneous, multilingual conditions (Hüsünbeyi et al., 12 Jan 2026).

1. Pipeline Architectures and Key Modules

A state-of-the-art multilingual fact-checking pipeline is architecturally modular and typically decomposes into the following key stages, each with language-agnostic or language-aware variants:

1. Claim Aggregation and Input Filtering
   • Claims are collected from sources such as Google Fact-Check Search API and ClaimReview JSON feeds, stratified by publisher and language tag.
   • Preprocessing includes metadata extraction (dates, publisher, language), deduplication, content-type classification, and language verification (Hüsünbeyi et al., 12 Jan 2026).
2. Evidence Retrieval and Alignment
   • Multilingual retrieval is achieved using dense bi-encoders or dual-encoder architectures trained on claim-passage pairs across languages (Rastogi, 5 Aug 2025, Huang et al., 2022).
   • Techniques include cross-lingual retrieval (CONCRETE, X-ICT) where encoders are optimized to retrieve evidence from collections in any language, sometimes leveraging translation pivots or multilingual embedding spaces (Huang et al., 2022).
   • Visual media (images, video frames) are processed for alignment—keyframes are extracted using clustering on perceptual hash vectors, and captions/alt-text are associated with claims (Hüsünbeyi et al., 12 Jan 2026).
3. Verdict Normalization and Metadata Standardization
   • Ratings from heterogeneous fact-checking organizations are mapped to a unified verdict space {True, False, Partially-True, Other}, using pretrained lookup tables and fallback rules for ambiguous labels (Hüsünbeyi et al., 12 Jan 2026).
   • Metadata is enriched and standardized (e.g., ISO-8601 dates, publisher IDs, language codes).
4. Evidence Categorization and Extraction
   • Evidence is extracted from source articles into six structured categories: Expert Testimony, Quantitative Data, Official Records, Media Records, Multimedia Evidence, Eyewitness Accounts. Multilingual LLMs (e.g., Gemini-2.5-Pro) are prompted to output fine-grained structured evidence blocks with stable locators in the source content (Hüsünbeyi et al., 12 Jan 2026).
5. Justification Generation
   • Using only extracted evidence and the normalized verdict, the pipeline prompts an LLM to generate a multi-evidence, category-referenced justification, optionally referencing multimodal alignment (e.g., image captions, video frame timestamps) (Hüsünbeyi et al., 12 Jan 2026).
6. Evaluation and Human Verification
   • Outputs are evaluated by both native-speaker human annotators and LLM-based judges (e.g. GPT-4o) following detailed penalty-aware rubrics on criteria of correctness, coherence, and completeness.
   • Quantitative metrics and correlations between human and automated assessments are systematically reported (Hüsünbeyi et al., 12 Jan 2026).

2. Multilingual and Multimodal Handling

Robust multilingual handling is achieved by:

• Querying APIs and configuring scrapers with strict language filters, and by delivering prompts, metadata classification, and result evaluation in the target language (e.g., French, German) (Hüsünbeyi et al., 12 Jan 2026).
• Using language-specific LLMs for tasks such as content-type classification, evidence extraction, and justification generation (Gemini, Qwen, Llama3).
• Structuring evaluation to stratify by language, ensuring rubric consistency across native language annotators.
• For multimodal sources, extracting, clustering, and indexing visual content, and aligning it with textual claims and verdicts (e.g., storing representative video keyframes with timestamped references for seamless integration in justifications) (Hüsünbeyi et al., 12 Jan 2026).

3. Key Algorithms and Functions

Verdict Normalization

A lookup table associating publisher-specific labels to a standard verdict space is used. Where not available, deterministic fallbacks (e.g., substring matching for “true”/“false” in multilingual senses) are applied:

def normalize_verdict(r_raw): if r_raw in verdict_lookup: return verdict_lookup[r_raw] else: if contains(r_raw, ["true", "vrai", "wahr"]): return "True" if contains(r_raw, ["false", "faux", "falsch"]): return "False" if contains(r_raw, ["partiel", "teilweise"]): return "Partially-True" return "Other"

(Hüsünbeyi et al., 12 Jan 2026)

Evidence Extraction

A structured prompt is used for multilingual LLMs, producing six-category evidence blocks grounded strictly in the target article:

function extract_evidence(claim_text, article_text): response = call_LLM(model="Gemini-2.5-Pro", prompt=PROMPT) return parse_JSON(response)

Categories: E = {Expert, Quantitative, Official, MediaRecord, Multimedia, Eyewitness} (Hüsünbeyi et al., 12 Jan 2026).

Justification Generation

A multi-evidence, chain-of-thought prompt conditions on claim, normalized verdict, and categorized evidence, enforcing category references and verdict alignment:

function generate_justification(claim_text, normalized_verdict, evidence): return call_LLM(model="Gemini-2.5-Pro", prompt=PROMPT_JUST)

(Hüsünbeyi et al., 12 Jan 2026)

4. Structured Dataset Schema and Metadata

Every dataset record generated by such a pipeline contains:

Field	Type / Possible Values	Description
claim_id	UUID	Unique identifier
publisher	String	Original publisher name
language	"fr" / "de"	Language code
claim_date	ISO-8601	Date claim made
review_date	ISO-8601	Date of fact-check
claim_text	UTF-8 string	Text of the claim
content_type	{Text, Image, Video, Statistic}	Classified by LLM
review_url	URL	Original article link
original_rating	String	Publisher’s raw label
normalized_verdict	{True, False, Partially-True, Other}	Standardized verdict
evidence	List[6 blocks]	Each with items: {text_snippet, source_locator}
visual_media	Dict: images & videos	Media with {url, caption, locator}, keyframes

(Hüsünbeyi et al., 12 Jan 2026)

5. Evaluation Methodologies and Quantitative Findings

Evaluation leverages both LLM-based and human assessments:

• Criteria: Correctness (accuracy of verdict-evidence linkage), Coherence (logical, readable justification), Completeness (coverage of all relevant evidence).
• Metrics: Scores are reported on a 0–100 scale using GPT-4o and human annotators in the target language.
  • For Gemini-2.5-Pro on evidence extraction: French correctness ≈71.4, completeness ≈69.2; German correctness ≈76.5, completeness ≈73.7.
  • Justification generation: French correctness ≈94.7; German correctness ≈87.6; multimodal justification (text+visual): French correctness ≈97.1 (Hüsünbeyi et al., 12 Jan 2026).
• Human-AI Consistency: Human ratings are strongly correlated with G-Eval outputs; model ranking: Gemini > Qwen > Llama3.

6. Significance, Limitations, and Future Directions

This multilingual, multimodal pipeline design:

• Integrates structured evidence collection under realistic, journalistic standards.
• Normalizes heterogeneous verdict vocabularies, aligns multimodal evidence, and structures metadata for downstream interpretability.
• Provides an up-to-date basis and reproducible framework for evidence-grounded, cross-lingual misinformation verification.
• Enables the fine-grained comparison of fact-checking practices by organization or region.

Limitations include language/model-specific gaps in evidence extraction and justification quality, particularly in low-resource settings. The research identifies the need for further extensions to more languages, broader claim modalities, and tighter integration of multimodal retrieval mechanisms (Hüsünbeyi et al., 12 Jan 2026).

---

References

• "Multilingual, Multimodal Pipeline for Creating Authentic and Structured Fact-Checked Claim Dataset" (Hüsünbeyi et al., 12 Jan 2026)