Zhenyun Deng

Postdoc @Cambridge_Uni | #NLProc automated fact checking, natural language understanding, and question answering | hiking, marathon

[Andreas Vlachos @vlachos_nlp]

Enjoying #EACL2026 and looking forward to https://t.co/3DLRwP0V2a tomorrow to present the results of our shared task on checking claims with images and text, the talks and posters, and the invited talks by @fredahshi @anne_lauscher Moath from @Fatabyyano_com and @nikaletras !

[Andreas Vlachos @vlachos_nlp]

Hi #FactChecking NLPers! As we are preparing to wrap up our image-text claims shared task, a reminder that @FEVERworkshop is welcoming submissions on all related topics with deadline on Dec19. https://t.co/Qtw4d155GL https://t.co/S9F7hROg9b

[Andreas Vlachos @vlachos_nlp]

Excited to announce the 9th @FEVERworkshop collocated with #EACL2026 in Morocco: https://t.co/4bGGntzZEZ with a new shared task on #factchecking image-text claims with evidence from the web using the AVerImaTeC dataset: https://t.co/5ugCkmUnji Deadline 2nd of December! And...

[Andreas Vlachos @vlachos_nlp]

Pleased to announce the next @FEVERworkshop at ACL2025! Regular workshop papers (ARR and direct submissions) due 15th of April! And new shared task focusing on reproducible and efficient verification of real world claims! Check https://t.co/S9F7hROg9b and get keen!

[Andreas Vlachos @vlachos_nlp]

@MichaelSejr @c_christodoulop @chenxi_jw @j6mes organising the @FEVERworkshop https://t.co/S9F7hRONYJ and presenting the shared task https://t.co/DN3tQVK6Kg (with @akhtarmubashara @Ramiyaly Rui Cao @Yulongchen1010 Oana Cocarascu @ZhenyunDeng @ZhijiangG @IamArpitMittal on Friday

[Michael Schlichtkrull @michael_sejr]

Hi #NLProc folks 👋 Got a paper on #factchecking or a related task with ARR reviews? Submit it to @FEVERworkshop! Our commitment deadline is upcoming on September 23rd AoE.

Colocated with #EMNLP2024 - come join us in Miami! 🏝️

Details at https://t.co/qiIAq5DHpF

[Workshop on Large Language Model Memorization @l2m2_workshop]

📣Hey #NLProc! We* are planning to organize a *ACL workshop on memorization in LLM. Goal: provide a central venue to discuss memorization from different angles (e.g., technical, legal, social, etc). We want to gauge community interest.

Would you consider attending?

(1/3)

[Andreas Vlachos @vlachos_nlp]

Time to report on the AVERITEC shared task @FEVERworkshop ! We had 21 teams making participating (https://t.co/y9XvCEbBdW), with the best results by the teams TUDA_MAI (https://t.co/PsKHvHuosL, 1st) HUMANE (https://t.co/amxhPFQqvV, 2nd) and CTU AIC (https://t.co/qEkTmV37Hy, 3rd)

[Andreas Vlachos @vlachos_nlp]

With the great team @akhtarmubashara @Ramiyaly Rui Cao @Yulongchen1010 @c_christodoulop Oana Cocarascu @ZhenyunDeng @ZhijiangG @IamArpitMittal @MichaelSejr @j6mes Chenxi Whithouse!

[Andreas Vlachos @vlachos_nlp]

The call for papers for the 7th @FEVERworkshop at #EMNLP is out: https://t.co/3DLRwP0V2a Deadline for ARR and non-ARR papers on August 15th! And the deadline for the new shared task on real-world #factchecking using the AVERITEC dataset is July 20th https://t.co/vxdfgCeuwk !

[Zhenyun Deng @ZhenyunDeng]

(9/9) When evaluated for evidence retrieval potential, the decontextualised claims obtained by enriching original sentences with the necessary context, are better than the original claim sentences, with an average 1.08 improvement in precision.

[Zhenyun Deng @ZhenyunDeng]

(7/9) Our method achieves a Precision@1 score of 47.8 on identifying central sentences, a 10% improvement over Claimbuster. Our method achieved a chrf score of 26.4 against gold decontextualised claims, outperforming all baselines.

[Zhenyun Deng @ZhenyunDeng]

(8/9) The results were verified further by a fact-checking professional, as the sentences returned by our method were deemed central to the document more often, and check-worthy more often than those extracted by Claimbuster.

[Zhenyun Deng @ZhenyunDeng]

(6/9) To evaluate our method, we derive a claim extraction dataset containing decontextualised claims from AVeriTeC (https://t.co/ba7Gzd1ss2), a recently proposed dataset containing claims from real-world fact-checking articles.

[Zhenyun Deng @ZhenyunDeng]

(4/9) Assuming that salient claims are derived from central sentences, i) we recast the document-level CE task into the extractive summarization task to extract central sentences and reduce redundancy;

[Zhenyun Deng @ZhenyunDeng]

(5/9) ii) we decontextualise central sentences to be understandable out of context by enriching them with the necessary context; iii) we introduce a QA-based framework to obtain the necessary context by resolving ambiguous information units in the extracted sentence.

[Zhenyun Deng @ZhenyunDeng]

(3/9) We propose a novel method for document-level claim extraction and decontexualisation for fact-checking, aiming to extract salient check-worthy claims from documents that can be understood outside the context of the document.

[Zhenyun Deng @ZhenyunDeng]

(2/9) In real-world scenarios, Claims often need to be extracted from documents containing multiple claims, not all of which are relevant to the central idea of the document, and verifying all claims manually or even automatically would be inefficient.

[Zhenyun Deng @ZhenyunDeng]

(1/9) Existing claim extraction methods mainly focus on detecting whether a sentence contains a claim or the boundaries of the claim within a sentence.

[Zhenyun Deng @ZhenyunDeng]

Happy to share that our paper "Document-level Claim Extraction and Decontextualisation for Fact-Checking" has been to ACL 2024. Joint work with my collaborators @vlachos_nlp and @michael_sejr.

arXiv: https://t.co/Zruko9epgO
github: https://t.co/7hpgBA7wxH