Elamur Research⌕
☰
home / papers / 2026 / Linear Semantic Segmentation for Low-Resource Spoken Dialects

Linear Semantic Segmentation for Low-Resource Spoken Dialects

Paper · 2026 · Kirill Chirkunov, Younes Samih, Abed Alhakim Freihat et al.

Summary

Topic segmentation tools are usually built for tidy written prose. Paragraphs, punctuation and regular syntax give them boundary hints. Dialectal Arabic transcripts lack most of these hints. They have inconsistent spelling, Arabic–English code-switching and weakly signalled topic shifts. The authors build DialSeg-Ar, a benchmark of more than 1000 samples with gold linear segmentation. It covers MSA news (OPUS), dialogue from Gulf dialect novels (Rewayat), Moroccan broadcast transcripts (MGB-5), Gulf/Iraqi/Levantine telephone speech (LDC) and Gulf–English podcasts (Mixat). Every source is first recast as a sequence of speaker-tagged utterances. gpt-oss-120b then drafts boundaries, and native speakers annotate and validate them. The authors compare classical lexical methods, embedding methods, supervised segmenters and Arabic LLMs. They also finetune a Gemma3-4B segmenter, partly on deliberately corrupted inputs. Methods that do well on MSA news drop on dialectal speech, and prompting or few-shot LLMs barely beat classical baselines. The finetuned model gets the best Pk and WindowDiff overall and stays stable when inputs are machine-translated. The main lesson is that supervised adaptation to the language variety, not model scale, drives gains.

Key points

  • The benchmark has five sources in three domain groups: clean text (OPUS News Commentary, Rewayat), noisy text (MGB-5, LDC CallHome/CallFriend-style telephone corpora) and code-switching (Mixat, 2 shows × 14 episodes).
  • LDC material is released only as file/block identifiers plus boundary metadata, so users must license the transcripts themselves. Mixat-derived data carries a non-commercial research restriction.
  • Baselines include:
    • classical: TextTiling, AraTextTiling, C99, ArabC99;
    • embedding and supervised: TeT+CLS variants, SaT;
    • LLMs: ALLaM-7B-It, Fanar-1-9B-It, Nile-Chat-12B.
  • On the original-language macro average, the finetuned Gemma3-4B scores F1 0.52, Pk 0.43 and WD 0.43. These are the best Pk/WD in the table, while the top F1 (0.54) belongs to a TeT+CLS variant. Scores stay similar on EN MT input (Pk 0.42, WD 0.42) and EN→MSA MT input (Pk 0.43, WD 0.44).
  • In the ablation, Pk/WD move from 0.49/0.55 without finetuning, to 0.46/0.47 with plain finetuning, to 0.43/0.43 when corruption is added during finetuning. Stated limitations are that the study uses text only (no prosody or ASR errors) and measures no downstream effect.

Interesting ideas

  • I1. All genres, from news and novels to calls and podcasts, are mapped into one format: a flat list of utterances with speaker IDs. A single segmentation protocol and metric then apply everywhere.
  • I2. An LLM proposes the initial boundaries, and native annotators then correct and validate them. Gold data becomes cheaper without giving up human verification.
  • I3. Finetuning on corrupted inputs clearly improved boundary quality over clean-only finetuning. Exposure to noise appears to make the model rely less on surface cues.
  • I4. Robustness is measured by running the same task on the original text and on its machine translations. A stable score suggests the model tracks meaning rather than surface form.
  • I5. Licensed corpora are distributed as standoff references (document IDs plus boundary metadata) instead of raw text. This keeps a benchmark shareable without breaching licenses.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (world): A shared utterance/speaker representation would let the world builder split dialogues and narrative prose into scenes the same way, as a step before tracking states and event order.
  • I2 (learn): Mova's silver pipeline could use an LLM-teacher draft followed by targeted human checks for discourse-level labels, which witness parsers cannot agree on as easily as on trees.
  • I3 (tree): gparse could be trained with deliberately perturbed copies of treebank sentences (spelling noise, missing punctuation) to harden it against informal English text.
  • I4 (back): Segment or frame boundaries found before and after Mova's text → graph → text round trip or meta-world translation could be compared as a meaning-preservation check.
  • I5 (tree): Converted UD treebanks or silver annotations built on restricted sources could be shipped as offsets and labels only, with users rebuilding the text from their own copies.

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

🏢IBMArmonk · United States139 works · 70,691 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Semitic languages · 2,908 worksLow-resource languages · 267 works

Cite

Cite as Kirill Chirkunov, Younes Samih, Abed Alhakim Freihat and Hanan Aldarmaki. 2026. Linear Semantic Segmentation for Low-Resource Spoken Dialects. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1740

@inproceedings{chirkunov2026linear,
  title = {{Linear Semantic Segmentation for Low-Resource Spoken Dialects}},
  author = {Kirill Chirkunov, Younes Samih, Abed Alhakim Freihat and Hanan Aldarmaki},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1740},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1740}
}

References

47 references, 19 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Khadige Abboud, Olga Golovneva, and Christopher DiPersio. 2022. Cross-lingual transfer for low-resource Arabic language understanding. In Proceedings of the Seventh Arabic Natural Language Processing Workshop (WANLP), pages 225–237, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.
  2. Muhammad Abdul-Mageed, AbdelRahim Elmadany, and El Moatez Billah Nagoudi. 2021. ARBERT & MARBERT: Deep bidirectional transformers for Arabic. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 7088–7105, Online. Association for Computational Linguistics. ↗
  3. Kathrein Abu Kwaik, Motaz Saad, Stergios Chatzikyriakidis, and Simon Dobnik. 2018. Shami: A corpus of Levantine Arabic dialects. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).
  4. Maryam Khalifa Al Ali and Hanan Aldarmaki. 2024. Mixat: A data set of bilingual Emirati-English speech. In Proceedings of the 3rd Annual Meeting of the Special Interest Group on Under-resourced Languages @ LREC-COLING 2024, pages 222–226, Torino, Italia. ELRA and ICCL. ↗
  5. Ahmed Ali, Suwon Shon, Younes Samih, Hamdy Mubarak, Ahmed Abdelali, James Glass, Steve Renals, and Khalid Choukri. 2019. The MGB-5 Challenge: Recognition and dialect identification of dialectal Arabic speech. In 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 1026–1033. find
  6. Malik H. Altakrori and 1 others. 2025. Dialectal Arabic MMLU: Benchmarking Dialectal Capabilities in Arabic and Multilingual Language Models. Preprint, arXiv:2510.27543. ArXiv:2510.27543v1 [cs.CL], submitted 31 Oct 2025. ↗
  7. Fakhraddin Alwajih, Abdellah El Mekki, Samar Mohamed Magdy, AbdelRahim A. Elmadany, Omer Nacar, El Moatez Billah Nagoudi, Reem Abdel-Salam, Hanin Atwany, Youssef Nafea, and 1 others. 2025. Palm: A culturally inclusive and linguistically diverse dataset for Arabic LLMs. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 32871–32894, Vienna, Austria. Association for Computational Linguistics.
  8. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. Arabert: Transformer-based model for arabic language understanding. In LREC 2020 Workshop Language Resources and Evaluation Conference 11–16 May 2020, page 9. ↗
  9. Appen Pty Ltd. 2006a. Gulf arabic conversational telephone speech, transcripts. https://catalog.ldc.upenn.edu/LDC2006T15. LDC Catalog No. LDC2006T15. ↗
  10. Appen Pty Ltd. 2006b. Iraqi arabic conversational telephone speech, transcripts. https://catalog.ldc.upenn.edu/LDC2006T16. LDC Catalog No. LDC2006T16. ↗
  11. Appen Pty Ltd. 2007. Levantine arabic conversational telephone speech, transcripts. https://catalog.ldc.upenn.edu/LDC2007T01. LDC Catalog No. LDC2007T01. ↗
  12. ArabicSpeech. 2025. MGB-5 (hugging face dataset). https://huggingface.co/datasets/ ArabicSpeech/MGB-5. Accessed: 2025-12-12. ↗
  13. M Saiful Bari and 1 others. 2025. ALLam: Large language models for arabic and english. In The Thirteenth International Conference on Learning Representations.
  14. Gagan Bhatia, El Moatez Billah Nagoudi, Abdellah El Mekki, Fakhraddin Alwajih, and Muhammad Abdul-Mageed. 2025. Swan and ArabicMTEB: Dialect-aware, Arabic-centric, cross-lingual, and cross-cultural embedding models and benchmarks. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 4654–4670, Albuquerque, New Mexico. Association for Computational Linguistics.
  15. Houda Bouamor, Nizar Habash, Mohammad Salameh, Wajdi Zaghouani, Owen Rambow, Dana Abdulrahim, Ossama Obeid, Salam Khalifa, Fadhl Eryani, Alexander Erdmann, and Kemal Oflazer. 2018. The MADAR Arabic dialect corpus and lexicon. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). find
  16. Anja Habacha Chaibi, Marwa Naili, and Samia Sam-moud. 2014. Topic segmentation for textual document written in arabic language. Procedia Computer Science, 35:437–446. Knowledge-Based and Intelligent Information & Engineering Systems 18th Annual Conference, KES-2014 Gdynia, Poland, September 2014 Proceedings. find
  17. Freddy Y. Y. Choi. 2000. Advances in domain independent linear text segmentation. In 1st Meeting of the North American Chapter of the Association for Computational Linguistics. ↗
  18. Kareem Darwish, Nizar Habash, Mourad Abbas, Hend Al-Khalifa, Huseein T. Al-Natsheh, Samhaa R. El-Beltagy, Houda Bouamor, Karim Bouzoubaa, Violetta Cavalli-Sforza, Wassim El-Hajj, Mustafa Jarrar, and Hamdy Mubarak. 2021. A panoramic survey of natural language processing in the arab world. Preprint, arXiv:2011.12631. ↗
  19. Jonathan Dunn. 2023. Variation and instability in dialect-based embedding spaces. In Tenth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial 2023), pages 67–77, Dubrovnik, Croatia. Association for Computational Linguistics. ↗
  20. Abdellah El Mekki, Houdaifa Atou, Omer Nacar, Shady Shehata, and Muhammad Abdul-Mageed. 2025. NileChat: Towards linguistically diverse and culturally aware LLMs for local communities. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 10978–11002, Suzhou, China. Association for Computational Linguistics. ↗
  21. Yaxin Fan, Feng Jiang, Peifeng Li, and Haizhou Li. 2024. Uncovering the potential of ChatGPT for discourse analysis in dialogue: An empirical study. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 16998–17010, Torino, Italia. ELRA and ICCL. Fanar-Team, Ummar Abbas, Mohammad Shahmeer Ahmad, Firoj Alam, Enes Altinisik, Ehsannedin Asgari, Yazan Boshmaf, Sabri Boughorbel, Sanjay Chawla, Shammur Chowdhury, Fahim Dalvi, Kareem Darwish, Nadir Durrani, Mohamed Elfeky, Ahmed Elmagarmid, Mohamed Eltabakh, Masoomali Fatehkia, Anastasios Fragkopoulos, Maram Hasanain, and 23 others. Fanar: An arabic-centric multimodal generative ai platform.
  22. Markus Frohmann, Igor Sterner, Ivan Vulić, Benjamin Minixhofer, and Markus Schedl. 2024. Segment any text: A universal approach for robust, efficient and adaptable sentence segmentation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 11908–11941, Miami, Florida, USA. Association for Computational Linguistics. find
  23. Iacopo Ghinassi, Lin Wang, Chris Newell, and Matthew Purver. 2024a. Recent trends in linear text segmentation: A survey. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3084–3095, Miami, Florida, USA. Association for Computational Linguistics. find
  24. Iacopo Ghinassi, Lin Wang, Chris Newell, and Matthew Purver. 2024b. Recent trends in linear text segmentation: A survey. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3084–3095, Miami, Florida, USA. Association for Computational Linguistics. find
  25. Reshmi Ghosh, Harjeet Singh Kajal, Sharanya Kamath, Dhuri Shrivastava, Samyadeep Basu, and Soundararajan Srinivasan. 2022. Topic segmentation in the wild: Towards segmentation of semi-structured unstructured chats. Preprint, arXiv:2211.14954. ↗
  26. Goran Glavaš and Swapna Somasundaran. 2020. Two-level transformer and auxiliary coherence modeling for improved text segmentation. Preprint, arXiv:2001.00891. ↗
  27. Guokan, Shang and others. 2025b. Atlas-Chat: Adapting Large Language Models for Low-Resource Moroccan Arabic Dialect. In Proceedings of the First Workshop on Language Models for Low-Resource Languages, pages 9–30, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
  28. Nizar Habash and David Palfreyman. 2022. ZAEBUC: An annotated Arabic-English bilingual writer corpus. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 79–88, Marseille, France. European Language Resources Association. find
  29. Marti A. Hearst. 1997. Text tiling: Segmenting text into multi-paragraph subtopic passages. Computational Linguistics, 23(1):33–64.
  30. Jais2Team. 2025. Jais 2: A family of Arabic-centric open large language models. Technical report, MBZUAI-IFM.
  31. Mustafa Jarrar, Nizar Habash, Faeq Alrimawi, Diyam Akra, and Nasser Zalmout. 2016. Curras: an annotated corpus for the palestinian arabic dialect. Language Resources and Evaluation, 51:745 – 775. ↗
  32. Michal Lukasik, Boris Dadachev, Kishore Papineni, and Gonçalo Simões. 2020. Text segmentation by cross segment attention. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4707–4716, Online. Association for Computational Linguistics. find
  33. Mohamed Maamouri, David Graff, and Christopher Cieri. 2006. Arabic broadcast news transcripts. Catalog No. LDC2006T20. ↗
  34. Omer Nacar and Anis Koubaa. 2024. Enhancing semantic similarity understanding in arabic nlp with nested embedding learning. Preprint, arXiv:2407.21139. ↗
  35. Marwa Naili, Anja Habacha Chaïbi, and Henda Hajjami Ben Ghézala. 2018. The contribution of stemming and semantics in arabic topic segmentation. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 17:1 – 25. ↗
  36. Lev Pevzner and Marti A. Hearst. 2002. A critique and improvement of an evaluation metric for text segmentation. Computational Linguistics, 28(1):19– 36. ↗
  37. Omar F. Zaidan and Chris Callison-Burch. 2014. Arabic dialect identification. Computational Linguistics, 40(1):171–202. ↗
  38. Ming Zhong, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. 2022. Dialoglm: Pre-trained model for long dialogue understanding and summarization. Preprint, arXiv:2109.02492. ↗
  39. Matthew Purver. 2011. Topic segmentation. SPOKEN LANGUAGE UNDERSTANDING, page 291. find
  40. Elnaz Rahmati, Alireza Salkhordeh Ziabari, and Morteza Dehghani. 2025. CoCo-CoLa: Evaluating and improving language adherence in multilingual LLMs. In Proceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025), pages 62–77, Suzhuo, China. Association for Computational Linguistics. find
  41. Fabian Retkowski and Alexander Waibel. 2024. From text segmentation to smart chaptering: A novel benchmark for structuring video transcriptions. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 406–419, St. Julian’s, Malta. Association for Computational Linguistics. ↗
  42. V. Schechter. 2025. Embeddinggemma: Powerful and lightweight text representations. arxiv. find
  43. Guokan Shang, Hadi Abdine, Ahmad Chamma, Amr Mohamed, Mohamed Anwar, Abdelaziz Bounhar, Omar El Herraoui, Preslav Nakov, Michalis Vazirgiannis, and Eric P. Xing. 2025a. Nile-chat: Egyptian language models for Arabic and Latin scripts. In Proceedings of The Third Arabic Natural Language Processing Conference, pages 306–322, Suzhou, China. Association for Computational Linguistics.
  44. Abderrahman Skiredj, Ferdaous Azhari, Houdaifa Atou, Nouamane Tazi, and Ismail Berrada. 2025. Gemmaroc: Unlocking darija proficiency in llms with minimal data. Preprint, arXiv:2505.17082. ArXiv:2505.17082v1 [cs.CL], submitted 20 May 2025. Accessed: 2025-12-14. ↗
  45. Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier García, Jason Wei, Xuezhi Wang, Hyung Won Chung, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Neil Houlsby, and Donald Metzler. 2022. Ul2: Unifying language learning paradigms. In International Conference on Learning Representations. ↗
  46. Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), pages 2214–2218, Istanbul, Turkey. European Language Resources Association (ELRA). find
  47. Yi Xu, Hai Zhao, and Zhuosheng Zhang. 2021. Topic-aware multi-turn dialogue modeling. Proceedings of the AAAI Conference on Artificial Intelligence, 35(16):14176–14184. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/chirkunov-2026-linear-semantic-segmentation-low-resource/