TTLab at Daleel 2026: STAR-Ar, Sequence Tagging for Argument Recognition in Arabic

Bhuvanesh Verma; Ali Abusaleh; Alexander Mehler

cs.CLSep 30, 2026

TTLab at Daleel 2026: STAR-Ar, Sequence Tagging for Argument Recognition in Arabic

Authors: Bhuvanesh Verma, Ali Abusaleh, Alexander Mehler

Organizations: Text Technology Lab (TTLab), Goethe University Frankfurt

Abstract

Argument Mining (AM) is a critical NLP task that remains significantly under-resourced in Arabic. This paper presents $\testtt{STAR-Ar}$ , a BERT-BiLSTM-CRF architecture for argument discourse detection and classification, as our system for Daleel 2026, the inaugural Arabic argument mining shared task. The task requires the identification and classification of argumentative discourse units (ADUs) in debate and editorial texts.We jointly model these two objectives as a token-level sequence labeling task using a BERT-BiLSTM-CRF architecture that combines contextual transformer embeddings with structural transition constraints to support accurate span detection. $\testtt{STAR-Ar}$ achieves an F1-score of 72.69 on validation and 73.7 on test data. Our domain-specific analysis shows that models trained exclusively on editorials underperform those trained on debates, a disparity we primarily attribute to the smaller size of the editorial dataset. The code for $\testtt{STAR-Ar}$ is available at ${[\faGithub~TTLab at Daleel 2026](https://github.com/ENTAILab/daleel_2026_Arabic-Argumentative-Discourse-Mining)}$

Figures & tables

Panel of Figure 1 — Figure 1: Overview of STAR-Ar , a BERT–BiLSTM–CRF sequence tagger for Task 2 (ADU span detection and classification) of Daleel 2026.

Split	Genre	Paragraphs	AS	OT	AN	TE	CO	ST
Train	Overall	612	462 (75.5%)	222 (36.3%)	182 (29.7%)	160 (26.1%)	36 (5.9%)	28 (4.6%)
	Debate	357	287 (80.4%)	202 (56.6%)	69 (19.3%)	100 (28.0%)	22 (6.2%)	6 (1.7%)
	Editorial	255	175 (68.6%)	20 (7.8%)	113 (44.3%)	60 (23.5%)	14 (5.5%)	22 (8.6%)
Dev	Overall	217	157 (72.4%)	63 (29.0%)	68 (31.3%)	76 (35.0%)	12 (5.5%)	9 (4.1%)
	Debate	129	106 (82.2%)	52 (40.3%)	33 (25.6%)	43 (33.3%)	7 (5.4%)	1 (0.8%)
	Editorial	88	51 (58.0%)	11 (12.5%)	35 (39.8%)	33 (37.5%)	5 (5.7%)	8 (9.1%)

Table 1: Paragraph-level label coverage in the Daleel 2026 dataset. Each value reports the number and percentage of paragraphs containing at least one span of the corresponding label. Percentages are computed with respect to the number of paragraphs in each row and do not sum to 100% because a paragraph may contain multiple labels.

Split	Evaluation	Editorial	Debate	Both
Dev	Editorial	60.80	51.30	66.52
	Debate	58.28	75.29	75.04
	Both	60.15	69.25	72.69
Test	Editorial	62.35	54.52	63.56
	Debate	60.07	76.17	77.84
	Both	61.25	70.68	73.74

Table 2: F1-score (%) for evaluation on the Daleel development and test sets. Columns denote the training domain, while rows denote the evaluation split and domain.

Figure 2: Cross-domain span-level confusion matrix. Rows are gold classes; cells show mean character-wise overlap with predictions. The diagonal represents span-level recall; ’NONE’ captures false negatives.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Encoder Model	Task 2 Overall $F_{1}$ (%)
AraBERTv02-Twitter-b	$63.75\pm 0.43$
AraBERTv02-Twitter-l	$65.82\pm 2.17$
MARBERTv2	$67.53\pm 1.53$

Appendix

Table 3: Ablation Study: Impact of different pre-trained text encoders on overall performance (models trained on the combined dataset).

Figure 3: Paragraph-level dev-set heatmap (cross-domain trained). Rows list gold labels with paragraph support; the diagonal gives recall. Off-diagonal cells show spurious predictions on missed gold paragraphs.

Explore similar work

From Argument Components to Graphs: A Multi-Agent Debate with Confidence Gating for Argument Relations
Jun 14, 2026Jakub Bąba, Jarosław A. ChudziakMulti-Agent Debate Debate
Faculty of Electronics and Information Technology, Warsaw University of Technology, Poland
Extracting Arguments, Not Just Classifying Them: Instruction-Tuned LLMs for Generative Component Detection
Sep 21, 2026Sofiane Elguendouze, Erwan Hain, Elena Cabrio +1Argumentation Frameworks Sequence Modeling
UniCA, I3S, MARIANNE · Universit´e Cˆote d’Azur, CNRS, INRIA, I3S Sophia-Antipolis, France · UniCA, MARIANNE +2
TTLab at StanceEval-2026: A Cloze-Style Prompting Approach for Arabic-Language Stance Detection (CLASP-Ar)
Sep 24, 2026Bhuvanesh Verma, Ali Abusaleh, Alexander MehlerCross-Topic Arabic Stance Detection Natural Language Processing
Text Technology Lab (TTLab), Goethe University Frankfurt

Jun 14, 2026, cs.CLJ/K move · Enter open · S save

From Argument Components to Graphs: A Multi-Agent Debate with Confidence Gating for Argument Relations

Jakub Bąba, Jarosław A. Chudziak

Faculty of Electronics and Information Technology, Warsaw University of Technology, Poland

Multi-Agent Debate Debate Large Language Model Decisions Bert-Based Models Gating Single-Agent

Open paper AskPDF arXiv

Large Language Models (LLMs) are increasingly assessed and utilized in the field of Argument Mining (AM), thanks to their strong general reasoning capabilities. However, standard training-free models often miss sophisticated details, specifically in contexts where two parts of the text have to be analyzed together. Furthermore, self-correction mechanisms tend to reinforce initial hallucinations in reasoning. Overcoming these limitations typically requires expensive, domain-specific supervised fine-tuning. Recent work has shown that a multi-agent paradigm can address such weaknesses for the component classification task through dialectical refinement with a Proponent-Opponent-Judge architecture, setting a promising direction for training-free approaches in the field. In this paper, we extend and evaluate this framework on the Argument Relation Identification and Classification (ARIC) task, reformulating it as a debate over component pairs. Besides that, we introduce a confidence gating mechanism that enables debating only on the uncertain cases and accepting the initial prediction when confidence is high. On the UKP Argument Annotated Essays v2 corpus, we demonstrate that the selective debate achieves the highest Macro F1 among all training-free methods, while debate over all samples degrades performance below that of one of the baselines. All generative approaches also outperform fine-tuned RoBERTa models on Macro F1, suggesting that the under-representation of the Attack class was more damaging to supervised fine-tuning than to inference-only models. Additionally, our framework produces human-readable debate transcripts, offering interpretability absent from both single-agent and supervised classifiers.

Abstract

Explore similar work

From Argument Components to Graphs: A Multi-Agent Debate with Confidence Gating for Argument Relations

Extracting Arguments, Not Just Classifying Them: Instruction-Tuned LLMs for Generative Component Detection

TTLab at StanceEval-2026: A Cloze-Style Prompting Approach for Arabic-Language Stance Detection (CLASP-Ar)