--- title: "Nullpointer at ArAIEval Shared Task: Arabic Propagandist" description: "The paper addresses ArAIEval 2024 Task 1: finding which propaganda techniques appear in Arabic… 4 references, cited by 1 paper here." authors: ["Abrar Abir", "Kemal Oflazer"] year: 2024 venue: "Proceedings of the Second Arabic Natural Language Processing Conference" source: "https://aclanthology.org/2024.arabicnlp-1.49/" doi: "10.18653/v1/2024.arabicnlp-1.49" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/abir-2024-nullpointer-araieval-shared-task-arabic/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/abir-2024-nullpointer-araieval-shared-task-arabic/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2024](https://md.research.elamur.ai/free/en/papers/2024/page/13/) / Nullpointer at ArAIEval Shared Task: Arabic Propagandist Technique Detection with Token-to-Word Mapping in Sequence Tagging # Nullpointer at ArAIEval Shared Task: Arabic Propagandist Technique Detection with Token-to-Word Mapping in Sequence Tagging **Authors:** [Abrar Abir](https://md.research.elamur.ai/free/en/authors/abrar-abir/) and [Kemal Oflazer](https://md.research.elamur.ai/free/en/authors/kemal-oflazer/) · **Year:** 2024 · **Venue:** [Proceedings of the Second Arabic Natural Language Processing Conference](https://md.research.elamur.ai/free/en/venues/arabicnlp/) ([ArabicNLP 2024](https://md.research.elamur.ai/free/en/venues/arabicnlp/2024/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2024.arabicnlp-1.49/) · [PDF](https://aclanthology.org/2024.arabicnlp-1.49.pdf) · [DOI 10.18653/v1/2024.arabicnlp-1.49](https://doi.org/10.18653/v1/2024.arabicnlp-1.49) · [OpenAlex](https://openalex.org/W4402671357) **Citations:** 0 citations (OpenAlex, 2026-10-14) · [1](https://md.research.elamur.ai/free/en/papers/abir-2024-nullpointer-araieval-shared-task-arabic/cited-by/) in our library, of the 149,439 papers whose bibliographies we read ## Summary The paper addresses ArAIEval 2024 Task 1: finding which propaganda techniques appear in Arabic tweets and news paragraphs, and the exact character spans where each one occurs. The authors first clean the data. They blank out invisible Unicode characters so that character offsets stay valid. They recover spans by substring search when the given offsets are wrong. They align "@USER" placeholders between texts and annotations, and they fix a few leftover cases by hand. The model is AraBERT v2. Each subword's contextual vector is concatenated with the [CLS] vector, and a one-hot genre flag can be added. A linear classifier sits on top. The study's focus is how to move between subword tokens and words. The authors compare four options: token-level labels throughout, majority vote over a word's tokens, the first token's label, and max-pooled word embeddings. Taking the first subword's label, with genre included, worked best (micro-F1 26.68 after the deadline; 25.41 and 4th place in the official ranking). The result shows that the token-to-word decision matters a lot in sequence tagging for morphologically rich languages. ## Key points - Data is from ArAIEval 2024 Task 1 (JSONL, UTF-8): - train: 995 tweets and 6002 paragraphs - dev: 249 tweets and 672 paragraphs - test: 260 tweets and 786 paragraphs - Test micro-F1 with and without genre: | Approach | With genre | Without genre | |---|---|---| | Token-to-token | 24.34 | 22.62 | | Majority-label | 20.73 | 16.57 | | First-label | 26.68 | 24.37 | | Word-to-word (max-pool) | 12.94 | 13.22 | - The genre feature helped every approach except word-to-word. It also slowed convergence from about 3 epochs to about 10. - Concatenating the [CLS] vector beat both adding it to the token vector and leaving it out. - A classifier with no hidden layer (about 37k weights over 1536-dim input) outperformed variants with 1–2 hidden layers of 32–128 units. The final model was trained on train+dev with settings chosen by grid search. ## Interesting ideas - When subwords must yield one decision per word, reading only the first subword's prediction beat both voting and pooled word vectors, likely because it matches how BERT-style tagging is usually trained. - Unwanted invisible characters were replaced with spaces instead of deleted, so every downstream character offset stays correct without any remapping. - The authors treated annotation offsets as unreliable: they discarded them and relocated each span by searching for its annotated text, fixing the rest by hand. - A tiny one-hot genre flag added to every token vector gave a clear gain. This suggests cheap metadata about text type can steer a tagger. - Concatenating a sentence-level vector with each token vector, plus a deliberately minimal linear head, outperformed more elaborate options. Simple global-plus-local inputs can beat added depth. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (tree): gparse can take arc and label scores from each word's first subword, and an ablation can compare this against pooling or voting. - **I2** (tokens): the en tokenizer can normalize odd Unicode the same way, keeping character spans valid for explanations and for tracing nodes back to the source text. - **I3** (learn): treebank and silver-data import can check every span against the raw string and re-anchor mismatches by search, logging failures rather than training on silently misaligned labels. - **I4** (tree): a register flag (fable, school test, word problem) could be fed to gparse or the perceptron parser, so attachments adapt to each text type. - **I5** (frames): a role or speech-act classifier in prag could pair each predicate's vector with a whole-clause summary vector, keeping a linear, inspectable head that stays explainable. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* ## Institutions - [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/) — University, Pittsburgh, United States · 539 works · 179,533 citations - [Carnegie Mellon University Qatar](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university-qatar/) — University, Doha, Qatar · 47 works · 2,773 citations 📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🤝 Social Sciences](https://md.research.elamur.ai/free/en/shelves/social-sciences/) · 1,772 works · [🤝 Sociology and Political Science](https://md.research.elamur.ai/free/en/shelves/social-sciences/sociology-and-political-science/) · 577 works 🧭 Topics (1): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works ## Cite Cite as: Abrar Abir and Kemal Oflazer. 2024. Nullpointer at ArAIEval Shared Task: Arabic Propagandist Technique Detection with Token-to-Word Mapping in Sequence Tagging. Proceedings of the Second Arabic Natural Language Processing Conference. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{abir2024nullpointer, title = {{Nullpointer at ArAIEval Shared Task: Arabic Propagandist Technique Detection with Token-to-Word Mapping in Sequence Tagging}}, author = {Abrar Abir and Kemal Oflazer}, booktitle = {Proceedings of the Second Arabic Natural Language Processing Conference}, year = {2024}, doi = {10.18653/v1/2024.arabicnlp-1.49}, url = {https://doi.org/10.18653/v1/2024.arabicnlp-1.49} } ``` - [← previous Noun phrases in early Germanic languages](https://md.research.elamur.ai/free/en/papers/bech-2024-noun-phrases-early-germanic/) - [next → Pathways of Knowledge Transmission: From the Greek to the Arabic Tradition](https://md.research.elamur.ai/free/en/papers/olivieri-2024-pathways-knowledge-transmission-greek-arabic/) - [↑ back to papers of 2024](https://md.research.elamur.ai/free/en/papers/2024/page/13/) ## References 4 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Different sequence tagging approaches The Firoj Alam, Abul Hasnat, Fatema Ahmed, Md Arid word-level-first sequence tagging technique Hasan, and Maram Hasanain. 2024. ArMeme: Pro-emerged as the most effective method. This result pagandistic content in Arabic memes. arXiv preprint is consistent with the nature of the Arabic lan-arXiv:2406.03916. guage, where the essence or primary meaning of a Firoj Alam, Hamdy Mubarak, Wajdi Zaghouani, Gio-word is often captured by the first token, especially vanni Da San Martino, and Preslav Nakov. 2022. for verbs. Conversely, the word-level-majority Overview of the WANLP 2022 shared task on propa-sequence tagging technique performed poorly ganda detection in Arabic. In Proceedings of the The Seventh Arabic Natural Language Processing Work-compared to other methods. Despite aggregating shop (WANLP), pages 108–118, Abu Dhabi, United token-level predictions at the word level, this Arab Emirates (Hybrid). Association for Computa-approach did not yield significant improvements tional Linguistics. in accuracy. This suggests that the majority label Maram Hasanain, Fatema Ahmed, and Firoj Alam. assignment strategy may dilute the relevance of 2023a. Large language models for propaganda span the primary propagandistic technique within each annotation. arXiv preprint arXiv:2311.09812. word, particularly in Arabic. 492 [↗](https://arxiv.org/abs/2406.03916) 2. Maram Hasanain, Fatema Ahmed, and Firoj Alam. 2024a. Can GPT-4 identify propaganda? annotation and detection of propaganda spans in news articles. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources And Evaluation, LREC-COLING 2024, Torino, Italy. [find](https://scholar.google.com/scholar?q=Maram+Hasanain%2C+Fatema+Ahmed%2C+and+Firoj+Alam%2E+2024a%2E+Can+GPT%2D4+identify+propaganda%3F+annotation+and+detection+of+propaganda+spans+in+news+articles%2E+In+Proceedings+of+the+2024+Joint+International+Confer) 3. Maram Hasanain, Firoj Alam, Hamdy Mubarak, Samir Abdaljalil, Wajdi Zaghouani, Preslav Nakov, Giovanni Da San Martino, and Abed Alhakim Freihat. 2023b. [ArAIEval Shared Task: Persuasion techniques and disinformation detection in Arabic text](https://md.research.elamur.ai/free/en/papers/hasanain-2023-araieval-shared-task-persuasion-techniques/). In Proceedings of the First Arabic Natural Language Processing Conference (ArabicNLP 2023), Singapore. Association for Computational Linguistics. Maram Hasanain, Md. Arid Hasan, Fatema Ahmed, Reem Suwaileh, Md. 4. Rafiul Biswas, Wajdi Zaghouani, and Firoj Alam. 2024b. [ArAIEval Shared Task: Propagandistic techniques detection in unimodal and multimodal Arabic content](https://md.research.elamur.ai/free/en/papers/hasanain-2024-araieval-shared-task-propagandistic-techniques/). In Proceedings of the Second Arabic Natural Language Processing Conference (ArabicNLP 2024), Bangkok. Association for Computational Linguistics. *Updated 31 October 2026*