Elamur Research⌕
☰
home / papers / 2026 / AlignAR: Generative Sentence Alignment for Arabic–English Parallel Corpora of Legal and Literary Texts

AlignAR: Generative Sentence Alignment for Arabic–English Parallel Corpora of Legal and Literary Texts

Paper · 2026 · Baorong Huang and Ali Asiri

Summary

The paper tackles sentence alignment for Arabic–English, a pair with few gold-standard parallel resources. Existing benchmarks are dominated by one-to-one sentence correspondences, which hides the weaknesses of aligners. The authors build a new gold corpus with two halves. The "Easy" half has five Saudi laws with official English translations. The "Hard" half has five classical and literary texts whose translations often split or merge sentences.

They also propose AlignAR, a zero-shot generative aligner. Source and target lines are numbered, and an LLM is asked to pick out the translation of each source line among the target lines, returning JSON. Only the line numbers are then kept as the alignment ("ladder").

On the legal texts every system scores above 0.90 F1, so that subset cannot tell methods apart. On the literary texts the classic aligners degrade sharply, while Gemini-2.5-flash with AlignAR reaches 0.855 strict F1, against 0.761 for the best baseline. The work shows that benchmark difficulty matters and that LLMs can handle structurally divergent translations.

Key points

  • Corpus sources:
    • Easy subset: laws such as the Companies Law and Anti-Bribery Law, obtained via Saudi Arabia's National Center for Archives and Records (NCAR).
    • Hard subset: texts from the Rasaifa platform, including Ḥayy ibn Yaqẓān and stories from Hassan Ghazala's A Textbook of Literary Translation.
  • Difficulty gap between subsets:
    • Share of 1-to-1 alignments is about 78–87% in the Easy documents but only about 29–46% in the Hard ones.
    • The Arabic/English sentence ratio drops to roughly 0.45 in the Hard documents.
    • Tokens were counted with Stanza.
    • The prose gives hard-subset sentence totals that do not match the per-document table.
  • Systems compared:
    • Baselines: BertAlign, BleuAlign (Bleualign) and VecAlign.
    • AlignAR was run with two LLMs: GPT-5.1-mini and Gemini-2.5-flash.
  • Evaluation: strict precision, recall and F1 (an alignment counts only if it exactly matches the reference), micro-averaged over documents.
  • Results on the Hard subset (overall F1):
    • Gemini: 0.855; GPT: 0.767; VecAlign: 0.761; BertAlign: 0.726; BleuAlign: 0.477.
    • BleuAlign falls from 0.942 on Easy to 0.477 on Hard.
    • Gemini beats GPT by more than 12 points on Hard 5.
    • The reported Gemini precision and recall (both 0.831) are inconsistent with its F1.

Interesting ideas

  • I1. Number every input line, have the LLM echo the text together with its index, then discard everything except the indices; this keeps output quality high while sidestepping fuzzy string matching against slightly altered LLM copies.
  • I2. Reframe alignment as asking the model to find which target lines translate each source line. This plays to what LLMs do well instead of asking for an abstract mapping.
  • I3. Pair a saturated "easy" set with a deliberately "hard" set built by reducing trivial one-to-one cases, so that real differences between methods become visible.
  • I4. Use strict exact-match scoring instead of lax overlap credit, because downstream users need precise mappings and partial overlaps are effectively errors.
  • I5. Simple corpus statistics, such as the sentence-count ratio and the share of 1-1 links, serve as cheap predictors of how hard a document will be to align.

For Mova

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

  • I1 (learn): LLM teacher tiers could return annotations keyed by token IDs while also echoing the tokens, so silver labels attach exactly to our tokenization even when the teacher misspells words.
  • I2 (back): to check text → graph → text or meta-world translation, an aligner of this kind could map each original sentence to the regenerated sentences that express it, exposing dropped or merged content.
  • I3 (learn): evaluation suites for parsers and readers should include subsets filtered to remove trivially easy items (e.g. short, canonical sentences), so that gains are not masked by ceiling effects.
  • I4 (frames): score extracted who-did-what-to-whom frames by exact role-filler match rather than partial span overlap, since a half-right role misleads the world model.
  • I5 (tree): analogous cheap statistics, such as the share of non-projective arcs or average sentence length, could flag which treebank sections or converted dialects will stress the parsers before training.

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

🎓Umm al-Qura UniversityMakkah al Mukarramah · Saudi Arabia21 works · 274 citations

Also on the byline: Huaihua University.

 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 3Research subjects and languages the work is about.English · 2,263 worksSemitic languages · 2,908 worksLegal AI · 134 works

Cite

Cite as Baorong Huang and Ali Asiri. 2026. AlignAR: Generative Sentence Alignment for Arabic–English Parallel Corpora of Legal and Literary Texts. The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/2c9daup5j4k7

@inproceedings{huang2026alignar,
  title = {{AlignAR: Generative Sentence Alignment for Arabic–English Parallel Corpora of Legal and Literary Texts}},
  author = {Baorong Huang and Ali Asiri},
  booktitle = {The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks},
  year = {2026},
  doi = {10.63317/2c9daup5j4k7},
  url = {https://doi.org/10.63317/2c9daup5j4k7}
}

References

16 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Maha Jarallah Althobaiti. 2022. A simple yet robust algorithm for automatic extraction of parallel sentences: A case study on arabic-english wikipedia articles. IEEE Access, 10:401–420. ↗
  2. Xiaohe Bo, Zeyu Zhang, Quanyu Dai, Xueyang Feng, Lei Wang, Rui Li, Xu Chen, and Ji-Rong Wen. 2024. Reflective multi-agent collaboration based on large language models. In Advances in Neural Information Processing Systems, volume 37, pages 138595–138631. Curran Associates, Inc. find
  3. Peter F. Brown, Jennifer C. Lai, and Robert L. Mercer. 1991. Aligning sentences in parallel corpora. Proceedings of the 29th Annual Meeting of the Association for Computational Linguistics, pages 169–176.
  4. William A. Gale and Kenneth W. Church. 1993. A program for aligning sentences in bilingual corpora. Computational Linguistics, 19(1):75–102. ↗
  5. Philipp Koehn, Vishrav Chaudhary, Ahmed El-Kishky, Naman Goyal, Peng-Jen Chen, and Francisco Guzmán. 2020. Findings of the wmt 2020 shared task on parallel corpus filtering and alignment. In Proceedings of the Fifth Conference on Machine Translation, pages 726–742.
  6. Philipp Koehn, Huda Khayrallah, Kenneth Heafield, and Mikel L Forcada. 2018. Findings of the wmt 2018 shared task on parallel corpus filtering. In EMNLP 2018 Third Conference on Machine Translation (WMT18), pages 726–739. Association for Computational Linguistics.
  7. Lei Liu and Min Zhu. 2023. Bertalign: Improved word embedding-based sentence alignment for chinese–english parallel corpora of literary texts. Digital Scholarship in the Humanities, 38(2):621–634. ↗
  8. Francesco Maria Molfese, Andrei Stefan Bejgu, Simone Tedeschi, Simone Conia, and Roberto Navigli. 2024. Crocoalign: A cross-lingual, context-aware and fully-neural sentence alignment system for long texts. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2209–2220. find
  9. Yiyue Qian, Shinan Zhang, Yun Zhou, Haibo Ding, Diego Socolinsky, and Yi Zhang. 2026. Collabeval: Enhancing llm-as-a-judge via multi-agent collaboration. ↗
  10. Rico Sennrich and Martin Volk. 2010. Mt-based sentence alignment for ocr-generated parallel texts. In Proceedings of the Ninth Conference of the Association for Machine Translation in the Americas (AMTA 2010), pages 1–10, Denver, USA. find
  11. Wei Shang and Xin Li. 2024. Construction and alignment features of english corpus based on particle swarm optimization algorithm. In 2024 International Conference on Distributed Systems, Computer Networks and Cybersecurity (ICDSCNC), pages 1–5. IEEE. find
  12. Steve Sloto, Brian Thompson, Huda Khayrallah, Tobias Domhan, Thamme Gowda, and Philipp Koehn. 2023. Findings of the wmt 2023 shared task on parallel data curation. In Proceedings of the Eighth Conference on Machine Translation, pages 95–102. find
  13. Steinþór Steingrímsson, Hrafn Loftsson, and Andy Way. 2023. SentAlign: Accurate and Scalable Sentence Alignment. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Singapore, Singapore. Association for Computational Linguistics. find
  14. Brian Thompson and Philipp Koehn. 2019. Vecalign: Improved sentence alignment in linear time and space. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP-IJCNLP), pages 1342–1348, Hong Kong, China. Association for Computational Linguistics.
  15. Xiaotian Wang, Takehito Utsuro, and Masaaki Nagata. 2024. Document alignment based on overlapping fixed-length segments. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop), pages 51–61. find
  16. Wu Zhang. 2022. Improve sentence alignment by divide-and-conquer. arXiv preprint arXiv:2201.06907. 10. Language Resource References N/A 10.1. Prompt Template and Sample Responses The following prompt template is used in the alignment process. You are an expert in translation. Align two documents and output bilingual pairs. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/huang-2026-alignar-generative-sentence-alignment-arabic/