Skip to content

X-CLAIM

FreeProfessional

Operator: github.com/mbzuai-nlp/X-CLAIM | Type: Non-detector | Cost: Free

TL;DR

Academic 7,000-claim multilingual dataset and baseline models (mBERT / mDeBERTa / XLM-R) for claim-span identification across English, Hindi, Punjabi, Tamil, Telugu, and Bengali: the toolkit's available Tamil entry in 2B.2. Pan-Indian by training-data composition; for Sri Lankan Tamil content the tool is operationally usable but the training data does not cover Sri-Lanka-specific named entities or communal-memory politics, and that adaptation gap is the binding caveat.

What it does

X-CLAIM is the dataset and accompanying baseline models from MBZUAI's NLP group, published at EMNLP 2023 main conference. The dataset contains 7,000 real-world claims across six languages (English, Hindi, Punjabi, Tamil, Telugu, Bengali), annotated for claim-span identification (which spans of a sentence are factual claims that warrant fact-check attention). The baseline models (multilingual BERT, mDeBERTa, XLM-R) provide a starting point for downstream Tamil-language claim-extraction work.

The tool's structural place in the toolkit is the Tamil- language entry point in 2B.2. The cell needs Tamil coverage because Sri Lanka is a focus country and Tamil is one of its two main local languages. X-CLAIM is the most documented Tamil-capable academic baseline ; it is not a Sri-Lanka-specific tool. The claim-span identification task the dataset trains is exactly the structural decomposition work a fact-checker needs at the front end of a verification workflow on Tamil content.

When to use it

  • A Sri Lankan fact-check team or a regional partner needs Tamil-language claim-span extraction as a baseline starting point and accepts that local adaptation will be required for Sri-Lanka-specific content.
  • A research group is building Tamil-language fact-check NLP capacity and needs a documented EMNLP-published dataset and baseline models as the academic substrate for their work.
  • A pan-South-Asian project is processing Tamil content from multiple jurisdictions and X-CLAIM's pan-Indian-plus-Sri- Lanka-applicability framing fits the input distribution.
  • A trainer is teaching Tamil-language claim-extraction concepts in a workshop setting and needs an academic reference point with a published dataset.

Limitations

Limitations

  • Indian-language-centred. The 7,000-claim dataset is built from pan-Indian sources rather than Sri Lanka-specific Tamil discourse; the baseline models reflect that training data.
  • Sri Lanka-specific Tamil adaptation gap. LIRNEasia research records that most Tamil-specific resources useful for Sri Lankan monitoring come from Indian or pan-Tamil ecosystems rather than Sri Lanka-specific ones, including the training data X-CLAIM is built on. Pan-Indian Tamil training data does not cover Sri Lankan named entities, Sri Lankan political actors, or Sri Lankan communal- memory politics with the depth a Sri-Lanka-specific tool would. Sri Lankan Tamil workflows therefore need more human review, more source-list curation, and more care with named entities, dialectal nuance, and communal memory politics than the bare X-CLAIM output provides (Synthesis section C).
  • Academic baseline rather than productionised tool. Operational deployment requires Python NLP capacity and developer integration; X-CLAIM is not a press-button web tool for front-line fact-checkers.
  • No Sri Lankan deployment of X-CLAIM is documented in the source set; institutional adoption would be the operator's first documented use rather than building on a documented Sri Lankan precedent.
  • Languages outside the EMNLP six (Bahasa Indonesia, Malay, Filipino / Tagalog, Thai, Sinhala, Lao) are not covered.

Privacy and threat model

X-CLAIM is an academic dataset and baseline-model release. Operational use means downloading the dataset and code and running it on the operator's own infrastructure, typically Python with the Hugging Face transformers stack. There is no hosted service to which a user submits claims; the data flow is fully under the operator's control.

For Sri Lanka deployment the threat-model dimension that matters is the broader operating environment documented in the Sri Lanka country page. Sri Lanka's Online Safety Act 2024 establishes broad state powers; institutional partners running Tamil-language NLP work on politically sensitive content should pair the technical deployment with the same source-protection discipline that applies elsewhere in the toolkit (Sherloq for offline forensics at 1B.1, the Sri Lanka country page's surveillance- environment framing). For research-publication outputs that identify or characterise specific Tamil discourse patterns, standard institutional research-publication ethics apply — named-entity exposure in published research can have operational consequences in a tense communal-political environment.

Country and platform applicability

Decision 7 framing applies: X-CLAIM is an NLP-class tool, and content-language coverage is the operational selection criterion. Tamil is the documented operational language; the b2-pass-Tamil-pan-Tamil-not-SL-adapted flag is the explicit acknowledgement that this tool enters the shortlist with the pan-Indian-not-Sri-Lanka-adapted caveat.

  • Indonesia: not applicable. Bahasa Indonesia is not among X-CLAIM's six languages; Indonesian content routes through Yudistira and Alegre.
  • Laos: not applicable. Lao is not covered; the 2B.2 Lao gap is structural across all NLP-class entries in this cell. The Lao path remains Google Cloud Translation routing (2A.1 cross-link) plus partner-mediated review.
  • Malaysia: not applicable as a primary tool. Malay is not in the X-CLAIM language set; Malaysian Tamil content in principle could route to X-CLAIM, but the documented use is pan-Indian Tamil rather than Malaysian Tamil, so the same adaptation-gap framing applies. Malaysian content routes through Alegre and MaLLaM.
  • Philippines: not applicable; Filipino / Tagalog is not covered.
  • Sri Lanka: the available Tamil tool, with the pan-Indian adaptation gap as the binding caveat. For Sri Lankan Tamil content X-CLAIM is the documented option in the 2B.2 shortlist, but its training data is pan-Indian; Sri-Lanka-specific named entities, Sri Lankan political actors, and Sri Lankan communal-memory politics may not be captured. The pairing of X-CLAIM's pan-Indian framing with Dissect's Sinhala asymmetric-strong framing (Dissect (Watchdog Sri Lanka / LIRNEasia)) is the toolkit's honest closure of the Decision 7 Sinhala-Tamil parity question per LIRNEasia research: Sri Lanka's Sinhala stack is more locally coherent (LIRNEasia MisinformationCorpusSinhala, University of Moratuwa SinLlama, SLTK tokenizer) while Sri Lankan Tamil tooling leans on pan-Indian resources with the adaptation gap visible.
  • Thailand: not applicable; Thai is not covered.

Platform applicability: not platform-specific; X-CLAIM is a dataset and baseline-model release, deployable as Python NLP infrastructure. Its outputs feed whichever editorial workflow the operator runs at the front-end.

How to access

Free under the academic licence; the dataset and baseline models are distributed through standard academic NLP channels (MBZUAI NLP, Hugging Face, the EMNLP 2023 publication's documented release path). Operational use requires Python infrastructure and the Hugging Face transformers stack; institutional partners considering integration should clone the public release and follow the documented training and inference scripts. There is no hosted X-CLAIM service.

Cost (current as of 2026-05)

Free under the academic open licence. Operational cost is infrastructure (the operator's compute for inference, GPU if fine-tuning) and developer time on integration. For Sri-Lanka- specific adaptation, the cost is primarily the human work of curating Sri Lankan Tamil training data (the locally specific material the pan-Indian X-CLAIM corpus does not carry) and fine-tuning the baseline on that material.

Quickstart

(Quickstart for an institutional research or developer deployment; X-CLAIM is not a press-button tool.)

  1. Confirm the content is Tamil and that the Sri-Lanka-specific adaptation gap is acceptable for the use case, or that adaptation work is in scope.
  2. Clone the X-CLAIM dataset and baseline-model release through MBZUAI NLP / Hugging Face channels.
  3. Stand up a Python environment with the Hugging Face transformers stack; load the chosen baseline (mBERT, mDeBERTa, or XLM-R) and the dataset.
  4. For research-baseline use, run the published evaluation scripts on Tamil to reproduce the EMNLP results before trusting the tool on new content.
  5. For Sri-Lanka-specific adaptation, curate Sri Lankan Tamil training material (Hashtag Generation, Watchdog, Fact Crescendo SL, FactSeeker Tamil archives are the candidate sources) and fine-tune the baseline on the local data.
  6. Integrate the inference layer into the operator's editorial workflow with explicit human review on every output: a high-confidence claim-span flag is a starting point for verification, not evidence the claim is suspect.
  7. Pair the deployment with Sherloq at 1B.1 for any image / video material accompanying the Tamil-language text (the Sri Lankan surveillance-environment threat model documented in the Sri Lanka country page).

In the toolkit's workflow

Source-history pillar non-detector tool per Architectural Anchor 1. Sits in 2B.2 AI claim extraction as the Tamil entry alongside the Sri-Lanka-specific Dissect (Sinhala), the multilingual Alegre, the Bahasa-specific Yudistira, and the English baseline ClaimBuster. The asymmetric framing (Dissect Sinhala-strong via the LIRNEasia stack, X-CLAIM Tamil pan-Indian with adaptation gap) is the honest account of Sri Lanka's Sinhala / Tamil tooling asymmetry per LIRNEasia research rather than a single multilingual entry that would mask the gap.

Standard combinations:

  • With Dissect at 2B.2 – Sinhala primary; Tamil alternative. The two cards together close the Sri Lankan Sinhala / Tamil shortlist with the asymmetry visible.
  • With Meedan Alegre at 2B.2 – Alegre's indian_sbert layer carries the same pan-Indian-Tamil training inheritance as X-CLAIM; both cards carry the verbatim Sri-Lanka-Tamil adaptation caveat. Operationally, Alegre is the multilingual back-end of a tipline; X-CLAIM is the research-baseline dataset and model. They are complementary rather than substitutable.
  • With Sherloq at 1B.1 – for offline image / video forensics on Tamil-language content in the Sri Lankan surveillance-environment threat model.
  • With AI Disinfo Hub at 1B.5 – institutional knowledge layer including the EU DisinfoLab "King of slop" Sri Lanka coverage; a cross-reference for the broader research environment X-CLAIM-based work fits inside.
  • With Google Cloud Translation at 2A.1 – for routing Tamil-language outputs into English-language editorial layers and for partner-mediated cross-country review.

Decision-tree references: T7 tipline routing names X-CLAIM as the Philippines -ph tipline tool of record (VERA Files / X-CLAIM coalition); T5 escalation at T5.9 (text professional) routes Sri Lankan Tamil-language claim triage to X-CLAIM with the verbatim adaptation-gap caveat applied, and cross-cell escalation into 1B.5 source-reliability scoring and 2C.1 automated claim monitoring is the standard pipeline for institutional Tamil-language work.

This card carries no detector signal class: X-CLAIM produces a non-detector source-history signal (claim-span identification on the input text). Per Anchor 2 the claim-span output is one signal class supporting human verification; the verification work that follows is the next signal class in the editorial workflow.

Override notes

Override notes

  • Tamil real-gap Sri-Lanka-specific (tamil-real-gap-sri-lanka-specific): the verbatim caveat from LIRNEasia research, surfaced in the Limitations admonition and the Country and platform applicability section: pan-Indian Tamil training data does not cover Sri Lankan named entities, Sri Lankan political actors, or Sri Lankan communal-memory politics with the depth a Sri-Lanka- specific tool would. The card does not soften this; the pairing with Dissect is the toolkit's honest closure of the Sinhala / Tamil parity question.

  • Pan-Tamil pass with not-SL-adapted disclaimer (b2-pass-Tamil-pan-Tamil-not-SL-adapted): X-CLAIM enters the 2B.2 shortlist on Tamil-language coverage with the pan-Indian-not-Sri-Lanka-adapted disclaimer explicit in the card text. The disclaimer is part of the selection record, not a footnote.

Sources