--- title: "TTLab at Daleel 2026 — Verma 2026" description: "The paper describes STAR-Ar, the TTLab entry for Task 2 of Daleel 2026, the first shared task on Arabic argument mining." authors: ["Bhuvanesh Verma", "Ali Abusaleh", "Alexander Mehler"] year: 2026 venue: "arXiv:2609.39385" source: "https://arxiv.org/abs/2609.39385" doi: "10.48550/arxiv.2609.39385" arxiv: "2609.39385" licence: "CC BY 4.0 (arXiv)" licence_url: "http://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/verma-2026-ttlab-daleel-2026-star-ar/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/verma-2026-ttlab-daleel-2026-star-ar/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/10/) / TTLab at Daleel 2026: STAR-Ar, Sequence Tagging for Argument Recognition in Arabic # TTLab at Daleel 2026: STAR-Ar, Sequence Tagging for Argument Recognition in Arabic **Authors:** [Bhuvanesh Verma](https://md.research.elamur.ai/free/en/authors/bhuvanesh-verma/), [Ali Abusaleh](https://md.research.elamur.ai/free/en/authors/ali-abusaleh/), [Alexander Mehler](https://md.research.elamur.ai/free/en/authors/alexander-mehler/) · **Year:** 2026 · **Venue:** arXiv:2609.39385 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://arxiv.org/abs/2609.39385) · [PDF](https://arxiv.org/pdf/2609.39385) · [DOI 10.48550/arxiv.2609.39385](https://doi.org/10.48550/arxiv.2609.39385) · [arXiv 2609.39385](https://arxiv.org/abs/2609.39385) · [OpenAlex](https://openalex.org/W7215046023) **Citations:** 0 citations (openalex-snapshot · 2026-10-14) ## Summary The paper describes STAR-Ar, the TTLab entry for Task 2 of Daleel 2026, the first shared task on Arabic argument mining. In Task 2 a system must find the argumentative discourse units (ADUs) within paragraphs drawn from debates and editorials and give each one a class (Assumption, Common ground, Testimony, Statistics, Anecdote, Other). The authors treat span detection and labelling as a single BIO tagging problem. Their model stacks a MARBERTv2 encoder, a BiLSTM and a CRF. The CRF transition matrix is fixed so that ill-formed BIO sequences cannot be produced. Training combines the CRF objective with a class-weighted auxiliary token loss, uses rarest-label stratified 5-fold cross-validation and oversamples minority classes. At inference the five folds are merged by averaging their emissions before decoding. The system scores 72.69 F1 on dev and 73.74 on test, which places it 3rd of 14 teams. The work shows that a simple, well-constrained tagger is competitive on a small, imbalanced, low-resource task. It also shows that frequent classes absorb the spans of rarer ones. ## Key points - The data has 612 training, 217 dev and 213 test paragraphs, with 2,975 training spans. Coverage is dense: 84.6% of training paragraphs have 80–100% of their text inside labelled spans. - The architecture is MARBERTv2 → 2-layer BiLSTM (256 per direction) → linear emissions → CRF. Invalid transitions (O→I, a sequence starting with I, I-tags crossing labels) receive a score of −10000. MARBERTv2 beat the AraBERT variants across several seeds. - The loss is L_CRF + 0.3·L_CE. The CE term uses inverse-square-root class weights, and the O weight is capped at 0.3. Learning rates differ by layer: 2e-5 for the encoder and 1e-3 for the top layers. - Training on debates alone comes close to training on both domains. Editorial-only models are clearly weaker (60.80 vs 72.69 dev F1 when both training sets are used), which the authors attribute to the smaller editorial set. - Assumption has a recall of 88.49% but swallows other classes: 74.22% of Common ground characters are tagged as AS, leaving CO recall at 12.26%. Statistics is the rarest class yet reaches 64.77% recall thanks to its numeric surface cues. ## Interesting ideas - Hard-coding structural constraints into the CRF transition matrix turns well-formed output into a guarantee rather than a tendency the model must learn — For Mova (tokens): BIO-style chunk and multiword-expression tagging in the en tagger could use prohibitive transition scores so the decoder never produces broken spans. - The fold ensemble averages raw per-token emission scores and runs one Viterbi pass, instead of voting over already-decoded sequences, so the merged output stays structurally consistent — For Mova (learn): when several witness parsers agree on silver data, combining their arc or label scores before decoding may give cleaner consensus trees than voting on finished outputs. - An auxiliary token-level loss with damped (inverse-sqrt) class weights and a capped weight for the dominant "outside" tag counters imbalance without distorting the structured objective — For Mova (tree): gparse could reweight rare dependency relations in a similar auxiliary loss while keeping the main tree objective unweighted. - Character-overlap confusion analysis with an explicit NONE column separates "wrong class" errors from "missed span" errors and reveals which class absorbs the others — For Mova (learn): evaluations of frame or entity spans could report overlap-based confusion with a NONE column, exposing majority-label absorption that a single F1 hides. - Cross-validation folds are stratified by each document's rarest label, so scarce classes show up in every fold — For Mova (answer): splitting school tests or fable QA sets by their rarest question type would keep rare categories present in every evaluation split. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @misc{verma2026ttlab, title = {{TTLab at Daleel 2026: STAR-Ar, Sequence Tagging for Argument Recognition in Arabic}}, author = {Verma, Bhuvanesh and Abusaleh, Ali and Mehler, Alexander}, year = {2026}, doi = {10.48550/arxiv.2609.39385}, eprint = {2609.39385}, archivePrefix = {arXiv}, url = {https://doi.org/10.48550/arxiv.2609.39385} } ``` - [← previous TTLab at AlexandriaX-2026: A Fine-Tuned Surface Tagger for Arabic Machine-Translation Error-Span Detection and Classification](https://md.research.elamur.ai/free/en/papers/abusaleh-2026-ttlab-alexandriax-2026-fine-tuned/) - [next → TTLab at StanceEval-2026: A Cloze-Style Prompting Approach for Arabic-Language Stance Detection (CLASP-Ar)](https://md.research.elamur.ai/free/en/papers/verma-2026-ttlab-stanceeval-2026-cloze-style/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/10/) ## References 11 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Arne Binder, Leonhard Hennig, and Bhuvanesh Verma. 2022. Full-text argumentation mining on scientific publications. In Proceedings of the First Workshop on Information Extraction from Scientific Publications, pages 54–66, Online. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Arne+Binder%2C+Leonhard+Hennig%2C+and+Bhuvanesh+Verma%2E+2022%2E+Full%2Dtext+argumentation+mining+on+scientific+publications%2E+In+Proceedings+of+the+First+Workshop+on+Information+Extraction+from+Scientific+Publi) 2. Steffen Eger, Johannes Daxenberger, and Iryna Gurevych. 2017. Neural end-to-end learning for computational argumentation mining. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11–22. [find](https://scholar.google.com/scholar?q=Steffen+Eger%2C+Johannes+Daxenberger%2C+and+Iryna+Gurevych%2E+2017%2E+Neural+end%2Dto%2Dend+learning+for+computational+argumentation+mining%2E+In+Proceedings+of+the+55th+Annual+Meeting+of+the+Association+for+Comput) [our card](https://md.research.elamur.ai/free/en/papers/eger-2017-neural-end-end-learning-computational/) 3. Theodosis Goudas, Christos Louizos, Georgios Petasis, and Vangelis Karkaletsis. 2014. Argument extraction from news, blogs, and social media. In Hellenic conference on artificial intelligence, pages 287–299. Springer. [find](https://scholar.google.com/scholar?q=Theodosis+Goudas%2C+Christos+Louizos%2C+Georgios+Petasis%2C+and+Vangelis+Karkaletsis%2E+2014%2E+Argument+extraction+from+news%2C+blogs%2C+and+social+media%2E+In+Hellenic+conference+on+artificial+intelligence%2C+pages+2) 4. John Lawrence and Chris Reed. 2019. Argument mining: A survey. Computational linguistics, 45(4):765– 818. [find](https://scholar.google.com/scholar?q=John+Lawrence+and+Chris+Reed%2E+2019%2E+Argument+mining%3A+A+survey%2E+Computational+linguistics%2C+45%284%29%3A765%E2%80%93+818%2E) [our card](https://md.research.elamur.ai/free/en/papers/lawrence-2019-argument-mining-survey/) 5. Sara Nabhani, Nahla Bassyouni, Ali Al-Zawqari, Mohammad Khader, and Khalid Al-Khatib. 2026. [find](https://scholar.google.com/scholar?q=Sara+Nabhani%2C+Nahla+Bassyouni%2C+Ali+Al%2DZawqari%2C+Mohammad+Khader%2C+and+Khalid+Al%2DKhatib%2E+2026%2E) 6. Daleel 2026 shared task on arabic argumentative discourse mining. Budapest, Hungary. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Daleel+2026+shared+task+on+arabic+argumentative+discourse+mining%2E+Budapest%2C+Hungary%2E+Association+for+Computational+Linguistics%2E) 7. Andreas Peldszus and Manfred Stede. 2013. From argument diagrams to argumentation mining in texts: A survey. International Journal of Cognitive Informatics and Natural Intelligence (IJCINI), 7(1):1–31. [find](https://scholar.google.com/scholar?q=Andreas+Peldszus+and+Manfred+Stede%2E+2013%2E+From+argument+diagrams+to+argumentation+mining+in+texts%3A+A+survey%2E+International+Journal+of+Cognitive+Informatics+and+Natural+Intelligence+%28IJCINI%29%2C+7%281%29%3A1%E2%80%9331) 8. Christian Stab and Iryna Gurevych. 2017. Parsing argumentation structures in persuasive essays. Computational Linguistics, 43(3):619–659. [find](https://scholar.google.com/scholar?q=Christian+Stab+and+Iryna+Gurevych%2E+2017%2E+Parsing+argumentation+structures+in+persuasive+essays%2E+Computational+Linguistics%2C+43%283%29%3A619%E2%80%93659%2E) [our card](https://md.research.elamur.ai/free/en/papers/stab-2017-parsing-argumentation-structures-persuasive-essays/) 9. Simone Teufel. 1999. Argumentative zoning: Information extraction from scientific text. Ph.D. thesis, University of Edinburgh Edinburgh, Scotland. [find](https://scholar.google.com/scholar?q=Simone+Teufel%2E+1999%2E+Argumentative+zoning%3A+Information+extraction+from+scientific+text%2E+Ph%2ED%2E+thesis%2C+University+of+Edinburgh+Edinburgh%2C+Scotland%2E) 10. Stephen E. Toulmin. 1958. The Uses of Argument. Cambridge University Press, Cambridge, England. [find](https://scholar.google.com/scholar?q=Stephen+E%2E+Toulmin%2E+1958%2E+The+Uses+of+Argument%2E+Cambridge+University+Press%2C+Cambridge%2C+England%2E) 11. Douglas Walton, Christopher Reed, and Fabrizio Macagno. 2008. Argumentation Schemes. Cambridge University Press, Cambridge, England. [find](https://scholar.google.com/scholar?q=Douglas+Walton%2C+Christopher+Reed%2C+and+Fabrizio+Macagno%2E+2008%2E+Argumentation+Schemes%2E+Cambridge+University+Press%2C+Cambridge%2C+England%2E)