publications

2026

1. arXiv 2026

   

   Last Translation Benchmark

   Vilém Zouhar, Niyati Bafna, Mukund Choudhary, and 4 more authors

   2026

   Abs arXiv Bib Code Demo

   Machine translation benchmarks are often either too easy or too artificial, and automatic metrics grow less reliable as models improve. The Last Translation Benchmark collects human-authored, peer-reviewed examples (text, images, audio, video) that break state-of-the-art MT systems, each with handcrafted verification rules for objective evaluation. It is a live dataset open to contributions.

   @misc{zouhar-2026-ltb,
     title = {Last Translation Benchmark},
     author = {Zouhar, Vil{\'e}m and Bafna, Niyati and Choudhary, Mukund and Z{\"u}fle, Maike and Rajaee, Sara and Chen, Pinzhen and others},
     journal = {arXiv preprint arXiv:2609.04173},
     year = {2026},
   }

2. ICML-W 2026

   

   Who Flips? Self- and Cross-Model Counterarguments Reveal Answer Instability in LLMs

   Nafiseh Nikeghbal, Amir Hossein Kargaran, Shaghayegh Kolli, and 1 more author

   In ICML 2026 Workshops: Agents in the Wild (AIWILD); Trustworthy AI for Good (AI4GOOD), 2026

   Abs arXiv Bib Code

   Accuracy benchmarks check whether LLMs reach the right answer, not whether they keep it. We challenge correct answers with a coherent argument for a wrong option and measure how often models flip. Across seven frontier models and 57 MMLU subjects, flip rates range from 17.5% to 97.3%; self-attribution raises them further, and our cross-model MaxFlip benchmark amplifies flips by up to 23.6pp.

   @inproceedings{nikeghbal-2026-whoflips,
     title = {Who Flips? Self- and Cross-Model Counterarguments Reveal Answer Instability in {LLM}s},
     author = {Nikeghbal, Nafiseh and Kargaran, Amir Hossein and Kolli, Shaghayegh and Diesner, Jana},
     booktitle = {ICML 2026 Workshops: Agents in the Wild (AIWILD); Trustworthy AI for Good (AI4GOOD)},
     year = {2026},
   }

3. ICML-W 2026

   

   GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts

   Amir Hossein Kargaran, Nafiseh Nikeghbal, Jana Diesner, and 2 more authors

   In ICML 2026 Workshops: Culture x AI: Evaluating AI as a Cultural Technology; GlobalSouthML (Affinity Event), 2026

   Abs arXiv Bib Code Demo

   OCR evaluation has focused on a handful of high- and mid-resource scripts. GlotOCR Bench tests OCR across 100+ Unicode scripts with clean and degraded rendered images. Most vision-language models do well on fewer than ten scripts, and even the strongest fail beyond thirty; performance tracks pretraining coverage, and on unfamiliar scripts models output noise or hallucinate characters from similar scripts.

   @inproceedings{kargaran-2026-glotocr,
     title = {{GlotOCR} Bench: {OCR} Models Still Struggle Beyond a Handful of {U}nicode Scripts},
     author = {Kargaran, Amir Hossein and Nikeghbal, Nafiseh and Diesner, Jana and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {ICML 2026 Workshops: Culture x AI: Evaluating AI as a Cultural Technology; GlobalSouthML (Affinity Event)},
     year = {2026},
   }

4. HF 2026

   

   FineTranslations

   Guilherme Penedo, Hynek Kydlı́ček, Amir Hossein Kargaran, and 1 more author

   2026

   Abs Bib Code Demo

   FineTranslations is a parallel dataset of over one trillion tokens between English and 500+ languages, built by translating FineWeb2 web documents into English with Gemma 3 27B. It is meant to improve machine translation, especially from English into lower-resource languages, and its English side also works well for English pretraining.

   @misc{penedo-2026-finetranslations,
     title = {{FineTranslations}},
     author = {Penedo, Guilherme and Kydl{\'\i}{\v{c}}ek, Hynek and Kargaran, Amir Hossein and von Werra, Leandro},
     journal = {Hugging Face repository},
     year = {2026},
     publisher = {Hugging Face},
   }

5. WWW 2026

   

   GlotWeb: Web Indexing for Minority Languages

   Abdullah Al Sefat, Amir Hossein Kargaran, François Yvon, and 1 more author

   In Proceedings of the ACM Web Conference 2026, 2026

   Abs DOI Bib ACM Code Demo

   GlotWeb is a web indexing system for minority languages. It finds web content in 400+ languages, verifies the language with GlotLID and filters it for quality, yielding 169K+ verified links, 47% of them in languages missing from major multilingual datasets. The pipeline and an interactive demo are open.

   @inproceedings{sefat-2026-glotweb,
     title = {{GlotWeb}: Web Indexing for Minority Languages},
     author = {Sefat, Abdullah Al and Kargaran, Amir Hossein and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {Proceedings of the ACM Web Conference 2026},
     year = {2026},
     pages = {8469--8472},
     publisher = {Association for Computing Machinery},
     doi = {10.1145/3774904.3792887},
   }

2025

1. arXiv 2025

   

   Insights from the ICLR Peer Review and Rebuttal Process

   Amir Hossein Kargaran, Nafiseh Nikeghbal, Jing Yang, and 1 more author

   2025

   Abs arXiv Bib Code

   A large-scale analysis of the ICLR 2024 and 2025 peer review, before and after rebuttal. Combining score statistics with LLM-based analysis of reviews and rebuttals, we find that initial scores and co-reviewers’ ratings best predict score changes, and that rebuttals matter most for borderline papers. We derive practical advice for authors and for designing fairer review processes.

   @misc{kargaran-2025-iclr,
     title = {Insights from the {ICLR} Peer Review and Rebuttal Process},
     author = {Kargaran, Amir Hossein and Nikeghbal, Nafiseh and Yang, Jing and Ousidhoum, Nedjma},
     journal = {arXiv preprint arXiv:2511.15462},
     year = {2025},
   }

2. EMNLP 2025

   

   CoBia: Constructed Conversations Can Trigger Otherwise Concealed Societal Biases in LLMs

   Nafiseh Nikeghbal, Amir Hossein Kargaran, and Jana Diesner

   In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 2025

   Abs DOI arXiv Bib ACL Code

   LLMs pass standard safety checks yet can still reveal harmful biases in conversation. CoBia builds conversations in which the model appears to have made a biased claim about a social group, then tests whether it recovers and rejects biased follow-ups. Across 11 LLMs and six socio-demographic categories, constructed conversations reliably surface and amplify hidden biases.

   @inproceedings{nikeghbal-2025-cobia,
     title = {{C}o{B}ia: Constructed Conversations Can Trigger Otherwise Concealed Societal Biases in {LLM}s},
     author = {Nikeghbal, Nafiseh and Kargaran, Amir Hossein and Diesner, Jana},
     booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
     year = {2025},
     publisher = {Association for Computational Linguistics},
     address = {Suzhou, China},
     doi = {10.18653/v1/2025.emnlp-main.84},
   }

3. EMNLP 2025

   

   On Relation-Specific Neurons in Large Language Models

   Yihong Liu, Runsheng Chen, Lea Hirlimann, and 6 more authors

   In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 2025

   Abs DOI arXiv Bib ACL Code

   Do some neurons in LLMs encode a relation itself, independent of entities? Studying Llama-2 with a statistics-based method, we find relation-specific neurons and show three properties: cumulativity (many neurons jointly encode a relation), versatility (neurons are shared across relations and even languages), and interference (deactivating one relation’s neurons can improve recall for others).

   @inproceedings{liu-2025-relation,
     title = {On Relation-Specific Neurons in Large Language Models},
     author = {Liu, Yihong and Chen, Runsheng and Hirlimann, Lea and Hakimi, Ahmad Dawar and Wang, Mingyang and Kargaran, Amir Hossein and Rothe, Sascha and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
     year = {2025},
     publisher = {Association for Computational Linguistics},
     address = {Suzhou, China},
     doi = {10.18653/v1/2025.emnlp-main.52},
   }

4. EMNLP 2025

   

   Tracing Multilingual Factual Knowledge Acquisition in Pretraining

   Yihong Liu, Mingyang Wang, Amir Hossein Kargaran, and 5 more authors

   In Findings of the Association for Computational Linguistics: EMNLP 2025, 2025

   Abs DOI arXiv Bib ACL Code

   We trace how multilingual factual recall and crosslingual consistency develop during pretraining, using OLMo-7B. Both improve over time, driven mainly by how often a fact appears in the corpus. Some rare non-English facts are still recalled thanks to crosslingual transfer from English, an effect that emerges early in pretraining.

   @inproceedings{liu-2025-tracing,
     title = {Tracing Multilingual Factual Knowledge Acquisition in Pretraining},
     author = {Liu, Yihong and Wang, Mingyang and Kargaran, Amir Hossein and K{\"o}rner, Felicia and Nie, Ercong and Plank, Barbara and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2025},
     year = {2025},
     publisher = {Association for Computational Linguistics},
     address = {Suzhou, China},
     doi = {10.18653/v1/2025.findings-emnlp.113},
   }

5. COLM 2025

   

   FineWeb2: One Pipeline to Scale Them All – Adapting Pre-Training Data Processing to Every Language

   Guilherme Penedo, Hynek Kydlı́ček, Vinko Sabolčec, and 7 more authors

   In Second Conference on Language Modeling, 2025

   Abs arXiv Bib Code Demo

   Training strong multilingual LLMs is hard because filtering and deduplication pipelines are difficult to tailor to many languages. We introduce a FineWeb-based curation pipeline that adapts automatically to any language, validate it with extensive ablations, and add a principled rebalancing method. Scaled to 1,000+ languages, it produces FineWeb2, a 20TB (5B-document) multilingual dataset.

   @inproceedings{penedo-2025-fineweb2,
     title = {{FineWeb2}: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language},
     author = {Penedo, Guilherme and Kydl{\'\i}{\v{c}}ek, Hynek and Sabol{\v{c}}ec, Vinko and Messmer, Bettina and Foroutan, Negar and Kargaran, Amir Hossein and Raffel, Colin and Jaggi, Martin and Von Werra, Leandro and Wolf, Thomas},
     booktitle = {Second Conference on Language Modeling},
     year = {2025},
   }

6. ACL 2025

   

   How Programming Concepts and Neurons Are Shared in Code Language Models

   Amir Hossein Kargaran, Yihong Liu, François Yvon, and 1 more author

   In Findings of the Association for Computational Linguistics: ACL 2025, 2025

   Abs arXiv Bib ACL Code

   How do LLMs represent multiple programming languages (PLs) relative to English? Decoding intermediate layers during translation across 21 PL pairs shows a concept space closer to English. Language-specific neurons concentrate in the bottom layers, PL-exclusive neurons in the top layers, and PLs aligned with many others have no identifiable specific neurons.

   @inproceedings{kargaran-2025-programming,
     title = {How Programming Concepts and Neurons Are Shared in Code Language Models},
     author = {Kargaran, Amir Hossein and Liu, Yihong and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {Findings of the Association for Computational Linguistics: ACL 2025},
     year = {2025},
     publisher = {Association for Computational Linguistics},
     address = {Vienna, Austria},
   }

7. ACL 2025

   

   MEXA: Multilingual Evaluation of English-Centric LLMs via Cross-Lingual Alignment

   Amir Hossein Kargaran, Ali Modarressi, Nafiseh Nikeghbal, and 3 more authors

   In Findings of the Association for Computational Linguistics: ACL 2025, 2025

   Abs arXiv Bib ACL Code Demo

   MEXA estimates the multilingual ability of English-centric LLMs without task benchmarks: it measures how well each language aligns with English in the model’s intermediate layers using parallel sentences. MEXA’s scores correlate strongly with downstream tasks (average Pearson 0.90 across nine models), making it a reliable, low-cost way to evaluate many languages.

   @inproceedings{kargaran-2025-mexa,
     title = {{MEXA}: Multilingual Evaluation of {E}nglish-Centric {LLM}s via Cross-Lingual Alignment},
     author = {Kargaran, Amir Hossein and Modarressi, Ali and Nikeghbal, Nafiseh and Diesner, Jana and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {Findings of the Association for Computational Linguistics: ACL 2025},
     year = {2025},
     publisher = {Association for Computational Linguistics},
     address = {Vienna, Austria},
   }

8. SIGTYP 2025

   

   Construction-Based Reduction of Translationese for Low-Resource Languages: A Pilot Study on Bavarian

   Peiqin Lin, Marion Thaler, Daniela Goschala, and 4 more authors

   In Proceedings of the 7th Workshop on Research in Computational Linguistic Typology and Multilingual NLP, 2025

   Abs Bib ACL Code

   LLMs translating into low-resource languages tend to produce translationese. We train on synthetic data in which a construction unique to the target language is made more frequent. For Bavarian, the model learns the construction and native speakers judge its translations as more natural.

   @inproceedings{lin-2025-construction,
     title = {Construction-Based Reduction of Translationese for Low-Resource Languages: A Pilot Study on {B}avarian},
     author = {Lin, Peiqin and Thaler, Marion and Goschala, Daniela and Kargaran, Amir Hossein and Liu, Yihong and Martins, Andr{\'e} F. T. and Sch{\"u}tze, Hinrich},
     booktitle = {Proceedings of the 7th Workshop on Research in Computational Linguistic Typology and Multilingual NLP},
     year = {2025},
     publisher = {Association for Computational Linguistics},
     address = {Vienna, Austria},
   }

9. COLING 2025

   

   How Transliterations Improve Crosslingual Alignment

   Yihong Liu, Mingyang Wang, Amir Hossein Kargaran, and 6 more authors

   In Proceedings of the 31st International Conference on Computational Linguistics, 2025

   Abs arXiv Bib ACL Code

   Post-aligning multilingual models on original and transliterated text improves crosslingual alignment, but why? Training models for Polish-Ukrainian and Hindi-Urdu, we find that transliterations alone raise overall similarity and that contrastive transliteration objectives help distinguish matched from random pairs. Better alignment, however, does not always improve downstream performance.

   @inproceedings{liu-2025-transliterations,
     title = {How Transliterations Improve Crosslingual Alignment},
     author = {Liu, Yihong and Wang, Mingyang and Kargaran, Amir Hossein and Imani, Ayyoob and Xhelili, Orgest and Ye, Haotian and Ma, Chunlan and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {Proceedings of the 31st International Conference on Computational Linguistics},
     year = {2025},
     publisher = {Association for Computational Linguistics},
     address = {Abu Dhabi, UAE},
   }

2024

1. NeurIPS 2024

   

   GlotCC: An Open Broad-Coverage CommonCrawl Corpus and Pipeline for Minority Languages

   Amir Hossein Kargaran, François Yvon, and Hinrich Schütze

   In Advances in Neural Information Processing Systems (Datasets and Benchmarks Track), 2024

   Abs arXiv Bib Code Demo

   No existing corpus covers many minority languages, comes from an open reproducible pipeline and is rigorously cleaned. GlotCC is a clean, document-level, 2TB general-domain corpus from CommonCrawl covering 1,000+ languages, released with its full pipeline, language identification model and filters.

   @inproceedings{kargaran-2024-glotcc,
     title = {{GlotCC}: An Open Broad-Coverage {C}ommon{C}rawl Corpus and Pipeline for Minority Languages},
     author = {Kargaran, Amir Hossein and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {Advances in Neural Information Processing Systems (Datasets and Benchmarks Track)},
     year = {2024},
   }

2. ACL 2024

   

   MaskLID: Code-Switching Language Identification through Iterative Masking

   Amir Hossein Kargaran, François Yvon, and Hinrich Schütze

   In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 2024

   Abs arXiv Bib ACL PDF Code Demo

   Sentence-level language identifiers usually return only the dominant language of a code-switched sentence. MaskLID is a training-free method that masks the features of the dominant language so the identifier can detect the second language in the next round, using the LID model itself. We apply it to GlotLID and OpenLID.

   @inproceedings{kargaran-2024-masklid,
     title = {{MaskLID}: Code-Switching Language Identification through Iterative Masking},
     author = {Kargaran, Amir Hossein and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)},
     year = {2024},
     publisher = {Association for Computational Linguistics},
     address = {Bangkok, Thailand},
   }

3. LREC 2024

   

   GlotScript: A Resource and Tool for Low Resource Writing System Identification

   Amir Hossein Kargaran, François Yvon, and Hinrich Schütze

   In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 2024

   Abs Oral Presentation arXiv Bib ACL Code Slides

   GlotScript is an open resource and tool for writing system identification. GlotScript-R lists the attested scripts of 7,000+ languages, and GlotScript-T identifies all 161 Unicode 15.0 scripts in text (ISO 15924). We show its use for cleaning multilingual corpora and for analyzing how well language model tokenizers cover low-resource scripts.

   @inproceedings{kargaran-2024-glotscript,
     title = {{GlotScript}: A Resource and Tool for Low Resource Writing System Identification},
     author = {Kargaran, Amir Hossein and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
     year = {2024},
     address = {Torino, Italia},
   }

4. MSR 2024

   

   GIRT-Model: Automated Generation of Issue Report Templates

   Nafiseh Nikeghbal, Amir Hossein Kargaran, and Abbas Heydarnoori

   In 21st IEEE/ACM International Conference on Mining Software Repositories (MSR), Apr 2024

   Abs DOI arXiv Bib ACM Code

   Issue report templates (IRTs) help manage GitHub issues but are rarely adopted. GIRT-Model is a T5-based assistant, instruction-tuned on our GIRT-Instruct dataset, that generates IRTs from a developer’s instructions. It outperforms general models in automatic and human evaluation, and user-study participants found it useful.

   @inproceedings{nikeghbal-2024-girtmodel,
     title = {{GIRT}-Model: Automated Generation of Issue Report Templates},
     author = {Nikeghbal, Nafiseh and Kargaran, Amir Hossein and Heydarnoori, Abbas},
     booktitle = {21st IEEE/ACM International Conference on Mining Software Repositories (MSR)},
     year = {2024},
     month = apr,
     publisher = {IEEE/ACM},
     address = {Lisbon, Portugal},
     doi = {10.1145/3643991.3644906},
   }

2023

1. EMNLP 2023

   

   GlotLID: Language Identification for Low-Resource Languages

   Amir Hossein Kargaran, Ayyoob Imani, François Yvon, and 1 more author

   In Findings of the Association for Computational Linguistics: EMNLP 2023, 2023

   Abs arXiv Bib ACL PDF Code Poster Demo

   No language identifier combined wide low-resource coverage, reliability and efficiency. GlotLID-M identifies 1,665 languages and outperforms CLD3, FT176, OpenLID and NLLB when balancing F1 and false positive rate. We also analyze the specific challenges of low-resource LID, such as wrong corpus metadata, leakage from high-resource languages and closely related languages.

   @inproceedings{kargaran-2023-glotlid,
     title = {{GlotLID}: Language Identification for Low-Resource Languages},
     author = {Kargaran, Amir Hossein and Imani, Ayyoob and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2023},
     year = {2023},
     publisher = {Association for Computational Linguistics},
     address = {Singapore},
   }

2. ACL 2023

   

   Glot500: Scaling Multilingual Corpora and Language Models to 500 Languages

   Ayyoob ImaniGooghari, Peiqin Lin, Amir Hossein Kargaran, and 8 more authors

   In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2023

   Abs Area Chair Award Oral Presentation arXiv Bib ACL PDF Code

   Instead of making LLMs better for  100 languages, we scale them horizontally. We build Glot500-c, a corpus for 511 mostly low-resource languages, and continue pretraining to create Glot500-m, which improves substantially over XLM-R on five tasks. Representation quality depends on a mix of factors, including corpus size, script, related languages and model capacity.

   @inproceedings{imanigooghari-2023-glot,
     title = {{Glot500}: Scaling Multilingual Corpora and Language Models to 500 Languages},
     author = {ImaniGooghari, Ayyoob and Lin, Peiqin and Kargaran, Amir Hossein and Severini, Silvia and Sabet, Masoud Jalili and Kassner, Nora and Ma, Chunlan and Schmid, Helmut and Martins, Andr{\'e} and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
     booktitle = {Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
     year = {2023},
     publisher = {Association for Computational Linguistics},
     address = {Toronto, Canada},
   }

3. MSR 2023

   

   GIRT-Data: Sampling GitHub Issue Report Templates

   Nafiseh Nikeghbal, Amir Hossein Kargaran, Abbas Heydarnoori, and 1 more author

   In IEEE/ACM 20th International Conference on Mining Software Repositories (MSR), 2023

   Abs arXiv Bib IEEE PDF Code

   GitHub introduced issue report templates (IRTs) in 2016, yet only about 5% of repositories use them, and research on them is limited. GIRT-Data is the first and largest dataset of IRTs in YAML and Markdown, covering 1,084,300 repositories (50,032 with IRTs), released with an open-source crawler.

   @inproceedings{nikeghbal-2023-girtdata,
     title = {{GIRT}-Data: Sampling {G}it{H}ub Issue Report Templates},
     author = {Nikeghbal, Nafiseh and Kargaran, Amir Hossein and Heydarnoori, Abbas and Sch{\"u}tze, Hinrich},
     booktitle = {IEEE/ACM 20th International Conference on Mining Software Repositories (MSR)},
     year = {2023},
   }

4. arXiv 2023

   

   MenuCraft: Interactive Menu System Design with Large Language Models

   Amir Hossein Kargaran, Nafiseh Nikeghbal, Abbas Heydarnoori, and 1 more author

   2023

   Abs arXiv Bib PDF Demo

   Designing menu systems involves many options and human factors, such as the semantic relations between commands. MenuCraft is an AI-assisted menu designer in which the designer and a dialogue system built on large language models design menus together, making design and customization easier through conversation.

   @misc{kargaran-2023-menucraft,
     title = {{MenuCraft}: Interactive Menu System Design with Large Language Models},
     author = {Kargaran, Amir Hossein and Nikeghbal, Nafiseh and Heydarnoori, Abbas and Sch{\"u}tze, Hinrich},
     journal = {arXiv preprint arXiv:2303.04496},
     year = {2023},
   }

2022

1. AACL 2022

   

   Hengam: An Adversarially Trained Transformer for Persian Temporal Tagging

   Sajad Mirzababaei, Amir Hossein Kargaran, Hinrich Schütze, and 1 more author

   In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2022

   Abs Bib ACL PDF Video Code

   Hengam is an adversarially trained transformer for Persian temporal tagging. We build a rule-based tagger to weakly label a large, diverse Persian corpus, train a transformer that generalizes beyond the rules, and release HengamGold, the first high-quality gold standard for the task. Hengam reaches a type F1 of 95.42 and handles ambiguity and misspellings.

   @inproceedings{kargaran-2022-hengam,
     title = {{H}engam: An Adversarially Trained Transformer for {P}ersian Temporal Tagging},
     author = {Mirzababaei, Sajad and Kargaran, Amir Hossein and Sch{\"u}tze, Hinrich and Asgari, Ehsaneddin},
     booktitle = {Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)},
     year = {2022},
     pages = {1013--1024},
     publisher = {Association for Computational Linguistics},
     address = {Online only},
   }

2021

1. WebSci 2021

   

   Wide-AdGraph: Detecting Ad Trackers with a Wide Dependency Chain Graph

   Amir Hossein Kargaran, Mohammad Sadegh Akhondzadeh, Mohammad Reza Heidarpour, and 3 more authors

   In 13th ACM Web Science Conference, 2021

   Abs Best Student Paper DOI arXiv Bib ACM PDF_V2 PDF_V1 Video Code

   Ad and tracker blocking relies on crowd-sourced filter lists that are hard to maintain. We build a large graph connecting resource requests across many websites and train a classifier on its features to detect ads and trackers. On the Alexa top-10K sites it is accurate and robust to evasion, and it finds services missing from existing lists.

   @inproceedings{kargaran-2021-wideadgraph,
     title = {{Wide-AdGraph}: Detecting Ad Trackers with a Wide Dependency Chain Graph},
     author = {Kargaran, Amir Hossein and Akhondzadeh, Mohammad Sadegh and Heidarpour, Mohammad Reza and Manshaei, Mohammad Hossein and Salamatian, Kave and Nejad Sattary, Masoud},
     booktitle = {13th ACM Web Science Conference},
     year = {2021},
     pages = {253--261},
     publisher = {Association for Computing Machinery},
     address = {Virtual Event, United Kingdom},
     doi = {10.1145/3447535.3462549},
   }

2. ADCHEM 2021

   

   Analytical Derivation and Comparison of Alarm Similarity Measures

   Amir Hossein Kargaran, Amir Neshastegaran, Iman Izadi, and 1 more author

   IFAC-PapersOnLine, 2021

   Abs DOI arXiv Bib PDF_V2 PDF_V1 Video Code Website

   Alarms in industrial plants are correlated, so one fault can trigger many alarms and overwhelm operators. We analytically derive and compare the similarity measures used to find correlated alarms, and validate the results with Monte Carlo simulation.

   @article{kargaran-2021-alarmsimilarity,
     title = {Analytical Derivation and Comparison of Alarm Similarity Measures},
     author = {Kargaran, Amir Hossein and Neshastegaran, Amir and Izadi, Iman and Yazdian, Ehsan},
     journal = {IFAC-PapersOnLine},
     year = {2021},
     volume = {54},
     number = {3},
     pages = {360--365},
     publisher = {Elsevier},
     doi = {10.1016/j.ifacol.2021.08.268},
   }