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

Bhuvanesh Verma; Ali Abusaleh; Alexander Mehler

cs.CLSep 24, 2026

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

Authors: Bhuvanesh Verma, Ali Abusaleh, Alexander Mehler

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

Abstract

Arabic-language stance detection remains challenging, and previous shared-task systems have largely relied on multitask learning and ensembles. While these systems achieve state-of-the-art performance, their applicability and transferability are limited by the additional complexity introduced by multitask learning.To reduce this complexity, we introduce $\texttt{CLASP-Ar}$ , which reformulates the task as cloze-style masked language modeling. In this approach, the target, predicted sentiment, and text are combined into a single prompt whose $\texttt{[MASK]}$ prediction is restricted to a verbalizer-constrained label vocabulary.

Figures & tables

Dataset	Property	Value
ArabicStance-X	Usage	Pre-training
	Samples	11,500
	Targets	17
	Favor	4,980
	Against/None	4,452 / 2,068
Mawqif-XT	Usage	Main task

Table 1: Statistics of the datasets used in this work. ArabicStance-X Alkhathlan et al. (2025) is used only for intermediate stance pre-training, while Mawqif-XT Albalawi et al. (2026b) is used for model development and evaluation.

Track	Favg2	Favg3	Overall Accuracy
Track 1	0.7136	0.5313	0.6761
Track 2	0.7414	0.6564	0.7096

Table 2: Performance of CLASP-Ar on the final evaluation data of Track 1 (seen targets) and Track 2 (unseen targets).

Without Pre-training
Method	$F_{\text{avg2}}$	$F_{\text{avg3}}$	Accuracy
Baseline	85.75	71.91	83.20
Base Setup	85.95	73.62	83.94
+ Extended None	86.33	70.40	83.94
+ Class Weights	85.40	74.23	83.36
With Pre-training

Table 3: Impact of adding None instances and class weights to CLASP-Ar . Full experiment results with multiple seeds can be found in Appendix B

Ablation Setting	$F_{\text{avg2}}$	$F_{\text{avg3}}$	Accuracy
Baseline (None)	84.90	70.83	82.61
+ Sentiment	86.28	72.36	84.38
+ Sarcasm	84.91	71.44	82.98
+ Both	85.85	72.50	83.74

Table 4: Ablation analysis on the development set reporting

F_{\text{avg2}}

F_{\text{avg3}}

, and Overall Accuracy across 4 settings.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Encoder	Plain	Pretrain	Joint
AraBERT large-twitter †	84.13	85.24	84.42
MARBERTv2	82.65	82.35	82.49
CAMeLBERT-mix	79.50	80.18	81.15

Appendix

Table 5: Impact of transfer learning strategies on development

F_{\mathrm{avg2}}

. Pretrain denotes intermediate fine-tuning on ArabicStance-X followed by training on Mawqif-XT . Joint denotes training on the union of Mawqif-XT and ArabicStance-X . † Original shared-task baseline using the large model.

Method	$F_{\text{favor}}$	$F_{\text{against}}$	$F_{\text{none}}$	$F_{\text{avg2}}$	$F_{\text{avg3}}$	Accuracy
Baseline	89.77	81.72	44.25	85.75	71.91	83.20
CLASP-Ar
Without Pre-training
Base Setup	89.99 ± 0.84	81.91 ± 1.03	48.96 ± 4.60	85.95 ± 0.82	73.62 ± 1.78	83.94 ± 0.87
+ Extended None	89.94 ± 0.41	82.73 ± 0.22	38.55 ± 10.81	86.33 ± 0.30	70.40 ± 3.67	83.94 ± 0.42
+ Class Weights	89.66 ± 1.29	81.14 ± 0.67	51.88 ± 2.13	85.40 ± 0.86	74.23 ± 0.85	83.36 ± 1.33

Appendix

Table 6: Overall stance detection performance comparison. Our implementation is presented as mean and standard deviation over 5 runs using different seeds. All values are reported as percentages (%). Baseline uses same encoder as for the CLASP-Ar which is AraBERT-large-twitter

Method	Covid Vaccine		Digital Trans.		Women Emp.
Method	$F_{\text{avg2}}$	$F_{\text{avg3}}$	$F_{\text{avg2}}$	$F_{\text{avg3}}$	$F_{\text{avg2}}$	$F_{\text{avg3}}$
CLASP-Ar
Without Pre-training
Base Setup	80.78 ± 1.03	64.52 ± 2.83	86.30 ± 2.89	80.99 ± 2.42	88.87 ± 1.19	70.48 ± 4.44
+ Extended None	82.52 ± 1.00	64.63 ± 2.31	84.82 ± 2.10	74.17 ± 7.77	88.15 ± 0.55	67.50 ± 4.67
+ Class Weights	78.86 ± 1.71	64.38 ± 1.65	87.08 ± 0.92	81.89 ± 1.53	88.82 ± 0.99	74.89 ± 1.47

Appendix

Table 7: Target-specific stance detection performance comparison. Results are shown as mean and standard deviation over 5 runs. All values are reported as percentages (%).

Ablation	$F_{\text{favor}}$	$F_{\text{against}}$	$F_{\text{none}}$	$F_{\text{avg2}}$	$F_{\text{avg3}}$	Accuracy
None	89.41 $\pm$ 0.30	80.39 $\pm$ 2.47	42.67 $\pm$ 1.21	84.90 $\pm$ 1.38	70.83 $\pm$ 0.61	82.61 $\pm$ 1.04
Sentiment	90.91 $\pm$ 0.29	81.66 $\pm$ 0.69	44.51 $\pm$ 4.77	86.28 $\pm$ 0.20	72.36 $\pm$ 1.73	84.38 $\pm$ 0.49
Sarcasm	89.87 $\pm$ 0.78	79.95 $\pm$ 1.48	44.50 $\pm$ 4.23	84.91 $\pm$ 0.94	71.44 $\pm$ 1.57	82.98 $\pm$ 0.92
Both	90.28 $\pm$ 1.20	81.42 $\pm$ 1.39	45.80 $\pm$ 1.03	85.85 $\pm$ 1.28	72.50 $\pm$ 1.10	83.74 $\pm$ 1.35

Appendix

Table 8: Comprehensive evaluation showing label-specific

F_{1}

scores (

F_{\text{favor}}

F_{\text{against}}

F_{\text{none}}

), macro-averaged scores (

F_{\text{avg2}}

F_{\text{avg3}}

), and overall Accuracy across multi-seed experiments on the development set.

Explore similar work

PAST-TIDE: Prototype-Anchored Statement Tuning with Topic-Invariant Normalization for Stance Detection
Jul 6, 2026Md. Shakhoyat Rahman Shujon, MD Jahid Hasan Jim, Md. Milon Islam +2Cross-Topic Arabic Stance Detection Natural Language Processing
Department of Computer Science and Engineering, Khulna University of Engineering & Technology · Department of Electrical and Computer Engineering, University of Waterloo · Department of Machine Learning, Mohamed bin Zayed University of Artificial Intelligence
Mawqif-XT: An Arabic Benchmark Dataset for Cross-Target Stance Detection
Aug 10, 2026Rasha Albalawi, Nuha Albadi, Hamzah Luqman +4Cross-Topic Arabic Stance Detection Arabic
University of Tabuk, Saudi Arabia · SDAIA-KFUPM JRC for AI, KFUPM, Saudi Arabia · Information and Computer Science Department, KFUPM, Saudi Arabia
StanceNakba Shared Task: Actor and Topic-Aware Stance Detection in Public Discourse
Jun 10, 2026Kholoud K. Aldous, Md Rafiul Biswas, Mabrouka Bessghaier +3Cross-Topic Arabic Stance Detection Natural Language Processing
Northwestern University in Qatar · Hamad Bin Khalifa University

Jul 6, 2026, cs.CLJ/K move · Enter open · S save

PAST-TIDE: Prototype-Anchored Statement Tuning with Topic-Invariant Normalization for Stance Detection

Md. Shakhoyat Rahman Shujon, MD Jahid Hasan Jim, Md. Milon Islam, Md Rezwanul Haque+1

Department of Computer Science and Engineering, Khulna University of Engineering & Technology · Department of Electrical and Computer Engineering, University of Waterloo · Department of Machine Learning, Mohamed bin Zayed University of Artificial Intelligence

Cross-Topic Arabic Stance Detection Natural Language Processing Motion-Language Alignment Language Modeling Tuning Normalization

Open paper AskPDF arXiv

We introduce PAST-TIDE, our stance detection system addressing both subtasks of the StanceNakba Shared Task at NakbaNLP@LREC-COLING 2026. The main idea is statement tuning. We redefine stance as cloze-style masked language modeling (MLM), letting a verbalizer map label words to stance categories through the pre-trained MLM head rather than appending a randomly initialized classification head. We complement this with prototypical contrastive learning, which uses learnable class prototypes for batch-size independent contrastive training, and topic-conditional layer normalization for cross-topic Arabic stance detection. PAST-TIDE achieves macro-F1 scores of 0.75 for Subtask A and 0.74 for Subtask B on the official leaderboard, indicating that minimal architectural additions to a pre-trained model can remain competitive in low-resource settings.

Abstract

Explore similar work

PAST-TIDE: Prototype-Anchored Statement Tuning with Topic-Invariant Normalization for Stance Detection

Mawqif-XT: An Arabic Benchmark Dataset for Cross-Target Stance Detection

StanceNakba Shared Task: Actor and Topic-Aware Stance Detection in Public Discourse