Sani et al. (2025) introduced the first dedicated Hausa machine-generated text (MGT) detec- tor, an AfroXLMR-large classifier trained on 2,586 paired articles from seven Hausa news outlets and Gemini 2.0 Flash, reporting F1 of 0.9921 on a headline-grouped split. They iden- tified three limitations of their work: a single news domain, a small batch and short train- ing schedule, and a single generator (Gemini 2.0 Flash) for the machine-generated data. We extend their work along the third axis. First, we replicate their reported result across four random seeds, obtaining F1 of 0.992 ±0.005. Test recall on the AI class is 1.000 at every seed. Second, we analyse the detector’s false positives. All errors fall on Human articles, and two native-speaker reviewers confirm these are genuine, natural Hausa journalism. The clear- est correlate of misclassification is the absence of numerical content: the median digit frac- tion is 0.0% in misclassified articles vs. 0.6% in correctly classified ones. Third, we test the detector on six non-training generators span- ning four model families. Detection rates span from 0% (Aya-Expanse 8B) to nearly 100% (Gemma 4). All three Google-family models in our experiment (training-distribution Gemini 2.0 Flash and the two non-training Gemma 4 variants) are detected at rates above 98% with mean P(machine) above 0.98. No non-Google generator exceeds 21%. Three contamination checks fail to explain this away: generated articles overlap very little with their human sources, detection barely moves when we re- move every article that shares a 13-gram with its source, and the Google output that gets de- tected is at least as diverse as everyone else’s, so it is not memorised or templated text. What the detector has learned looks less like a general Hausa MGT signal and more like a Gemini- family stylistic fingerprint. We discuss what this means for deployment, including a 3.1% false-positive rate on real Hausa journalists.

This paper describes our submission toSemEval-2026 Task 9, Subtask 2 (Hausa). Thetask involves identifying specific categories ofpolarization (Political, Religious, Ethnic, etc.)in Hausa social media comments. The datasetpresented significant challenges, primarily extreme class imbalance and the low-resourcenature of the language. Our system uses a pretrained multilingual transformer (Afro-XLMRLarge) fine-tuned with Weighted Binary CrossEntropy loss and dynamic undersampling (1:3ratio) to mitigate the scarcity of polarized examples. On the official test set, our systemachieved an official Macro-F1 score of 0.2346and a Micro-F1 score of 0.2581. Our model isrecall-oriented (Micro-Recall: 0.6166), demonstrating strong capability in detecting polarization, though precision remains a challenge(0.1632). We achieved our best per-class performance in the Political domain (F1: 0.48).

This paper presents Team HausaNLP’s submission to SemEval-2026 Task 4 (Track A),which requires identifying the more narrativelysimilar of two candidate stories relative to ananchor. Narrative similarity is defined alongthree dimensions: abstract theme, course ofaction, and story outcomes. We conduct a systematic ablation comparing five approaches:a lexical TF-IDF baseline, two bi-encoderSBERT variants (all-MiniLM-L6-v2 andall-mpnet-base-v2), a paraphrase-focusedembedding model, and a cross-encoder reranker. On the 200-instance development set,all-mpnet-base-v2 achieves the best performance (61.5% accuracy, 61.48 macro-F1), outperforming both TF-IDF (54.5%) and the official SBERT baseline (55.0%). Surprisingly,the cross-encoder re-ranker (55.5%) does notimprove on the bi-encoders, which we attributeto the long-document nature of Wikipedia storysummaries exceeding the model’s effective context window. On the official test set, our primary SBERT MiniLM submission achieved61.50% accuracy (33rd of 44 teams). Our erroranalysis over 200 development instances identifies five systematic failure categories, distinctfrom the All Correct / Partial cases, including23 Lexical Trap cases, 23 Hard Cases, and 24Proposed-Recovery cases, thereby informingconcrete directions for future work.

This paper describes our participation in SemEval2026 Task 3: Dimensional Aspect-Based SentimentAnalysis (DimABSA) (Yu et al., 2026). We utilizeda pre-trained DeBERTa-V3 backbone to capturesemantic meaning through disentangled attention.While standard Mean Squared Error (MSE) loss establishes a performance floor, we propose a HybridMSE-CCCLoss to identify distributional relationships that simple regression missed. Our resultsdemonstrate a 54.6% reduction in validation losscompared to the baseline, significantly improvingdetection in high-intensity emotional bins by mitigating the “regression to the mean” phenomenon.

We describe HausaNLP’s submission toSemEval-2026 Task 7 Track 1 (short-answercultural question answering). Our system is atraining-free, prompt-based pipeline targetingnative Hausa (ha-NG). Two design decisionsdistinguish it from a generic zero-shot baseline.We use locale-conditional prompting: ha-NGquestions receive a system prompt instructingconcise standard Hausa output with explicitBoko-script characters (á, â, Î, ű). Second, weuse a two-model fallback pipeline: GPT-4o handles the primary pass, and Gemini 1.5 Flash retries any rows where the primary call returnedan error or empty output, separating modelknowledge failures from API-availability failures. On the official development leaderboard,our best run reached 36.4 accuracy. Error analysis shows that a non-trivial fraction of failures are placeholder strings caused by APIerrors rather than incorrect generations, andthat surface-level mismatches (verbosity, orthographic variation) account for many of the remaining errors. Code, prompts, and processingscripts are released for reproducibility.

This paper describes our system submission for the Nakba Archive Image Classification task, which requires predicting the presence of structural destruction in historical archival photographs. We framed this as a binary computer vision classification problem (destruction vs. not_destruction). Our system utilizes a pre-trained ResNet-50 convolutional neural network, adapted for binary output, combined with strategic prediction threshold tuning. Evaluated on the unseen final test set, our model achieved a macro F1-score of 0.450 and a balanced accuracy of 0.527, serving as an exploratory baseline that highlights the unique challenges of processing degraded historical imagery.

We describe our submission to DialectSentEval 2026 Subtask 1, a three-class sentiment task covering Moroccan, Egyptian, Lebanese, Jordanian, and Saudi Arabic. Our system fine-tunes MARBERTv2 in five stratified folds and averages their test probabilities. Using raw text without explicit dialect conditioning or external task-specific data, the official submission obtained 0.7939 macro-F1, 0.8152 accuracy, 0.7917 macro precision, and 0.7987 macro recall, ranking eighth among twelve systems in the supplied snapshot. Separately, a post-submission character 3-5-gram diagnostic falls from 0.9157 OOF macro-F1 with row-stratified folds to 0.7520 when repeated ID-label groups and normalized duplicates are kept together. This result establishes dataset-level split sensitivity; it does not quantify the effect on MARBERTv2. We document the submitted configuration, a group-aware diagnostic protocol, and the experiments still needed to isolate the effects of strict splitting and ensembling on the transformer system.

We describe Team Wasl’s systems for the KnowledgeGraphEval 2026 shared task at ArabicNLP 2026, which decomposes Arabic knowledge-graph construction into two evaluated stages: nested NER under domain shift and relation extraction between given mentions. For Subtask 1 (AdaptNER), we train a multi-head BIO tagger (one three-way O/B/I head per entity type over a shared pretrained Arabic encoder) on Wojood and combine three random seeds by span-level majority vote. For Subtask 2, forced-choice classification over the 40 WojoodRelations relation types, we fine-tune a typed-marker classifier with type-constraint decoding, ensembling five grouped folds by logit averaging. Our submissions scored 0.7483 micro-F1 on AdaptNER’s blind Konooz slice (ten unseen domains), 4th of the eight officially ranked teams and 11.8 points above the official baseline, and 0.9571 micro-F1 on relation extraction with the no-relation class masked at decoding, 2nd at the results release; re-scored by the organizers over all 41 classes, the same ensemble reaches 0.9328, 4th in the organizers’ final ranking. We analyze where the 18.6-point domain-shift tax on NER falls: chiefly on recall. We further show that type-constraint decoding and ensembling are partially redundant (+0.9 points on a single RE model, +0.1 on the ensemble) and that the surviving errors share an argument-type signature that no type constraint can separate.

Language identification (LID) is a fundamental step in curating multilingual corpora. However, LID models still perform poorly for many languages, especially on the noisy and heterogeneous web data often used to train multilingual language models. In this paper, we introduce CommonLID, a community-driven, human-annotated LID benchmark for the web domain, covering 109 languages. Many of the included languages have been previously under-served, making CommonLID a key resource for developing more representative high-quality text corpora. We show CommonLID’s value by using it, alongside five other common evaluation sets, to test eight popular LID models. We analyse our results to situate our contribution and to provide an overview of the state of the art. In particular, we highlight that existing evaluations overestimate LID accuracy for many languages in the web domain. We make CommonLID and the code used to create it available under an open, permissive license.