---
title: "Balancing Fluency and Adherence — Abela 2026"
description: "The paper asks how to make machine translation use required domain terminology when the target language has little data and rich morphology, and…"
authors: ["Kurt Abela", "Marc Tanti", "Claudia Borg"]
year: 2026
venue: "Proceedings for the Ninth Workshop on Technologies for Machine Translation of Low Resource Languages (LoResMT 2026)"
source: "https://aclanthology.org/2026.loresmt-1.6/"
doi: "10.18653/v1/2026.loresmt-1.6"
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/abela-2026-balancing-fluency-adherence-hybrid-fallback/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/abela-2026-balancing-fluency-adherence-hybrid-fallback/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/4/) / Balancing Fluency and Adherence: Hybrid Fallback Term Injection in Low-Resource Terminology Translation

# Balancing Fluency and Adherence: Hybrid Fallback Term Injection in Low-Resource Terminology Translation

**Authors:** [Kurt Abela](https://md.research.elamur.ai/free/en/authors/kurt-abela/), [Marc Tanti](https://md.research.elamur.ai/free/en/authors/marc-tanti/), [Claudia Borg](https://md.research.elamur.ai/free/en/authors/claudia-borg/) · **Year:** 2026 · **Venue:** Proceedings for the Ninth Workshop on Technologies for Machine Translation of Low Resource Languages (LoResMT 2026) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.loresmt-1.6/) · [PDF](https://aclanthology.org/2026.loresmt-1.6.pdf) · [DOI 10.18653/v1/2026.loresmt-1.6](https://doi.org/10.18653/v1/2026.loresmt-1.6) · [OpenAlex](https://openalex.org/W7140134521)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
The paper asks how to make machine translation use required domain terminology when the target language has little data and rich morphology, and dictionaries are available but in-domain parallel text is scarce. The authors work on English-Maltese and English-Slovak in the fisheries domain. Terms come from IATE. Parallel material comes from European Parliament PECH committee documents and EUR-Lex. For Maltese, extra in-domain data was mined from the DGT Translation Memory with a BERTu-based topic classifier.

They compare two approaches. The static one trains on "acontextual drill" augmentation built from dictionary entries. The dynamic one applies Constrained Beam Search (CBS) at decoding time. CBS inserts nearly every term but sharply lowers BLEU. Static augmentation keeps output fluent and helps Maltese, but it does not work for Slovak.

Their proposal, Hybrid Fallback Term Injection, decodes first with the augmented model. It re-decodes with CBS only when a required term is missing. This recovers up to about 90% of missing terms while avoiding most of the quality loss. The result matters for any data-poor setting that needs faithful terminology.

## Key points
- **Data:**
  - The Maltese baseline was trained on the 2.9M-pair corpus of Williams et al. (2023), which is mostly legal, parliamentary and health text.
  - The target side was re-tokenized with the MLRS/BERTu vocabulary.
- **Domain mining:**
  - BERTu was fine-tuned on MultiEURLEX to predict EUROVOC top-level domains, reaching macro F1 77.4 over 21 domains.
  - The classifier was then run over DGT-TM to pull out fisheries documents.
- **Static augmentation results:**
  - For Maltese, the drill raised TIR from 5% to 55%.
  - Unseen-term adherence rose from 4% to 19%.
  - For Slovak it failed, which the authors link to morphological mismatch and a smaller fine-tuning set.
- **CBS versus Hybrid, from one results table:**
  - CBS reached 100% TIR but BLEU fell from 60.95 to 55.80. Across setups the drop was up to about 6 BLEU.
  - Hybrid scored 57.15 BLEU with 83.85% TIR.
  - Hybrid restricted to noun constraints scored 63.93 BLEU and 92.50% TIR, with 100% TIR on unseen terms.
- **Limitations:**
  - Hybrid is slower than CBS (4.42 vs 5.47 sent./sec), because the fallback fires on roughly half the sentences.
  - Term checking uses strict surface-form matching, which counts valid inflected forms as misses.

## Interesting ideas
- **Verify, then fall back.** Generate with the fluent but less reliable method, check the output against an explicit requirement, and call the forcing method only on failures; this keeps the cost of hard constraints confined to the cases that need them. — For Mova (back): Generation from the meaning graph could render text freely first, then check that every required entity and term appears. Only failed sentences would be regenerated with forced lexical choices.
- **Constrain nouns only.** Restricting forced terms to nouns gave both higher term adherence and better fluency than constraining everything, so the choice of which part of speech to force is a powerful dial. — For Mova (back): When the text → graph → text compressor must preserve wording, it could pin entity nouns and let verbs and function words inflect naturally.
- **Context blindness as an error type.** CBS forced the fish-processing sense of "header" and turned "Draft" into a catch of fish, showing that a dictionary match without a sense check corrupts meaning. — For Mova (frames): Before a frame or VerbNet role is filled, the lexical sense should be checked against the document domain. This would stop domain gazetteers from overriding the reading the context supports.
- **Report unseen-term adherence separately.** Measuring adherence on terms never seen in training, alongside overall TIR, shows whether a method truly generalizes or only memorizes. — For Mova (learn): Evaluations of self-induced repair rules and silver data could hold out lexical items never seen in training. This works like a negative control against memorization.
- **Mine in-domain data with a classifier.** Running a topic classifier trained on labelled legal text over a large generic memory yields a focused fine-tuning set cheaply. — For Mova (learn): A genre classifier could pick fable- or textbook-like sentences out of large corpora. Those sentences would then be prioritized for silver parsing and teacher labelling.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) · [Machine translation](https://md.research.elamur.ai/free/en/shelves/machine-translation/)

Topics: [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) · [Low-resource languages](https://md.research.elamur.ai/free/en/topics/low-resource-languages/)

## Cite

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{abela2026balancing,
  title = {{Balancing Fluency and Adherence: Hybrid Fallback Term Injection in Low-Resource Terminology Translation}},
  author = {Abela, Kurt and Tanti, Marc and Borg, Claudia},
  booktitle = {Proceedings for the Ninth Workshop on Technologies for Machine Translation of Low Resource Languages (LoResMT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.loresmt-1.6},
  url = {https://doi.org/10.18653/v1/2026.loresmt-1.6}
}
```

- [← previous AYN: A Tiny Yet Competitive Indian Legal Language Model Pretrained from Scratch](https://md.research.elamur.ai/free/en/papers/niyogi-2026-ayn-tiny-yet-competitive-indian/)
- [next → Bangla Key2Text: Text Generation from Keywords for a Low Resource Language](https://md.research.elamur.ai/free/en/papers/talukder-2026-bangla-key2text-text-generation-keywords/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/4/)

## References

22 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Eva Hasler, Adrià de Gispert, Gonzalo Iglesias, and Bill Byrne. 2018. Neural machine translation decoding with terminology constraints. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics, pages 506–512. [find](https://scholar.google.com/scholar?q=Eva+Hasler%2C+Adri%C3%A0+de+Gispert%2C+Gonzalo+Iglesias%2C+and+Bill+Byrne%2E+2018%2E+Neural+machine+translation+decoding+with+terminology+constraints%2E+In+Proceedings+of+the+2018+Conference+of+the+North+American+Chap) [our card](https://md.research.elamur.ai/free/en/papers/hasler-2018-neural-machine-translation-decoding-terminology/)
2. Mikko Aulamo, Sami Virpioja, and Jörg Tiedemann. 2020. Opusfilter: A flexible tool for filtering and combining parallel corpora. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 4254–4261. [find](https://scholar.google.com/scholar?q=Mikko+Aulamo%2C+Sami+Virpioja%2C+and+J%C3%B6rg+Tiedemann%2E+2020%2E+Opusfilter%3A+A+flexible+tool+for+filtering+and+combining+parallel+corpora%2E+In+Proceedings+of+the+12th+Language+Resources+and+Evaluation+Conference)
3. Lucia Benkova, Dasa Munkova, L’ubomír Benko, and Michal Munk. 2021. Evaluation of english–slovak neural and statistical machine translation. Applied Sciences, 11(7):2948. [find](https://scholar.google.com/scholar?q=Lucia+Benkova%2C+Dasa+Munkova%2C+L%E2%80%99ubom%C3%ADr+Benko%2C+and+Michal+Munk%2E+2021%2E+Evaluation+of+english%E2%80%93slovak+neural+and+statistical+machine+translation%2E+Applied+Sciences%2C+11%287%29%3A2948%2E)
4. Chris Hokamp and Qun Liu. 2017. Lexically constrained decoding for sequence generation using grid beam search. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1535–1546. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Chris+Hokamp+and+Qun+Liu%2E+2017%2E+Lexically+constrained+decoding+for+sequence+generation+using+grid+beam+search%2E+In+Proceedings+of+the+55th+Annual+Meeting+of+the+Association+for+Computational+Linguistic) [our card](https://md.research.elamur.ai/free/en/papers/hokamp-2017-lexically-constrained-decoding-sequence-generation/)
5. Toms Bergmanis and Marcis Pinnis. 2021. Context-independent terminology translation with neural ma-Sejoon Kim, Mingi Sung, Jeonghwan Lee, Hyunkuk Lim, and Jorge Gimenez Perez. 2024. Efficient terminology integration for LLM-based translation in specialized domains. In Proceedings of the Ninth Conference on Machine Translation, pages 636–642, Miami, Florida, USA. Association for Computational Linguistics. Annual Meeting of the Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Toms+Bergmanis+and+Marcis+Pinnis%2E+2021%2E+Context%2Dindependent+terminology+translation+with+neural+ma%2DSejoon+Kim%2C+Mingi+Sung%2C+Jeonghwan+Lee%2C+Hyunkuk+Lim%2C+and+Jorge+Gimenez+Perez%2E+2024%2E+Efficient+terminol)
6. Kai Song, Yue Zhang, Heng Yu, Weihua Luo, Kun Wang, and Min Zhang. 2019. Code-switching for enhancing nmt with pre-specified translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics, pages 449–459. [find](https://scholar.google.com/scholar?q=Kai+Song%2C+Yue+Zhang%2C+Heng+Yu%2C+Weihua+Luo%2C+Kun+Wang%2C+and+Min+Zhang%2E+2019%2E+Code%2Dswitching+for+enhancing+nmt+with+pre%2Dspecified+translation%2E+In+Proceedings+of+the+2019+Conference+of+the+North+American+Ch)
7. Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, pages 28–39. [↗](https://doi.org/10.18653/v1/w17-3204)
8. Ralf Steinberger, Mohamed Ebrahim, Alexandros Poulis, Carrasco-Benitez Manuel, Schlüter Patrick, Przybyszewski Marek, and Gilbro Signe. 2012. DGT-TM: A freely available translation memory in 22 languages. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), pages 454–459. European Language Resources Association (ELRA). [find](https://scholar.google.com/scholar?q=Ralf+Steinberger%2C+Mohamed+Ebrahim%2C+Alexandros+Poulis%2C+Carrasco%2DBenitez+Manuel%2C+Schl%C3%BCter+Patrick%2C+Przybyszewski+Marek%2C+and+Gilbro+Signe%2E+2012%2E+DGT%2DTM%3A+A+freely+available+translation+memory+in+22+langua) [our card](https://md.research.elamur.ai/free/en/papers/steinberger-2012-dgt-tm-freely-available-translation/)
9. Kurt Micallef, Albert Gatt, Marc Tanti, Lonneke van der Plas, and Claudia Borg. 2022. Pre-training data quality and quantity for a low-resource language: New corpus and BERT models for Maltese. In Proceedings of the Third Workshop on Deep Learning for Low-Resource Natural Language Processing, pages 90–101, Hybrid. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Kurt+Micallef%2C+Albert+Gatt%2C+Marc+Tanti%2C+Lonneke+van+der+Plas%2C+and+Claudia+Borg%2E+2022%2E+Pre%2Dtraining+data+quality+and+quantity+for+a+low%2Dresource+language%3A+New+corpus+and+BERT+models+for+Maltese%2E+In+Pro)
10. Jörg Tiedemann and Santhosh Thottingal. 2020. OPUS-MT – building open translation services for the world. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation, pages 479–480, Lisboa, Portugal. European Association for Machine Translation. [find](https://scholar.google.com/scholar?q=J%C3%B6rg+Tiedemann+and+Santhosh+Thottingal%2E+2020%2E+OPUS%2DMT+%E2%80%93+building+open+translation+services+for+the+world%2E+In+Proceedings+of+the+22nd+Annual+Conference+of+the+European+Association+for+Machine+Translati)
11. Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Li, Ashish Ghouri, Michael Dauphin, Michael Auli, and David Grangier. 2019. fairseq: A fast, extensible toolkit for sequence modeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), pages 48–53. [find](https://scholar.google.com/scholar?q=Myle+Ott%2C+Sergey+Edunov%2C+Alexei+Baevski%2C+Angela+Fan%2C+Sam+Li%2C+Ashish+Ghouri%2C+Michael+Dauphin%2C+Michael+Auli%2C+and+David+Grangier%2E+2019%2E+fairseq%3A+A+fast%2C+extensible+toolkit+for+sequence+modeling%2E+In+Proce) [our card](https://md.research.elamur.ai/free/en/papers/ott-2019-fairseq-fast-extensible-toolkit-sequence/)
12. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information processing systems, pages 5998–6008. [find](https://scholar.google.com/scholar?q=Ashish+Vaswani%2C+Noam+Shazeer%2C+Niki+Parmar%2C+Jakob+Uszkoreit%2C+Llion+Jones%2C+Aidan+N+Gomez%2C+%C5%81ukasz+Kaiser%2C+and+Illia+Polosukhin%2E+2017%2E+Attention+is+all+you+need%2E+In+Advances+in+neural+information+processi)
13. Matúš Pikuliak, Štefan Grivalský, Martin Konôpka, Miroslav Blšták, Martin Tamajka, Viktor Bachratý, Marián Šimko, Pavol Balážik, Michal Trnka, and Filip Uhlárik. 2021. Slovakbert: Slovak masked language model. Preprint, arXiv:2109.15254. [↗](https://arxiv.org/abs/2109.15254)
14. Aiden Williams, Kurt Abela, Rishu Kumar, Martin Bär, Hannah Billinghurst, Kurt Micallef, Ahnaf Mozib Samin, Andrea DeMarco, Lonneke van der Plas, and Claudia Borg. 2023. UM-DFKI Maltese speech translation. In Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 433–441, Toronto, Canada (in-person and online). Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Aiden+Williams%2C+Kurt+Abela%2C+Rishu+Kumar%2C+Martin+B%C3%A4r%2C+Hannah+Billinghurst%2C+Kurt+Micallef%2C+Ahnaf+Mozib+Samin%2C+Andrea+DeMarco%2C+Lonneke+van+der+Plas%2C+and+Claudia+Borg%2E+2023%2E+UM%2DDFKI+Maltese+speech+transla) [our card](https://md.research.elamur.ai/free/en/papers/williams-2023-um-dfki-maltese-speech-translation/)
15. Matt Post and David Vilar. 2018. Fast lexically constrained decoding with dynamic beam allocation for neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1314–1324. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Matt+Post+and+David+Vilar%2E+2018%2E+Fast+lexically+constrained+decoding+with+dynamic+beam+allocation+for+neural+machine+translation%2E+In+Proceedings+of+the+2018+Conference+of+the+North+American+Chapter+of) [our card](https://md.research.elamur.ai/free/en/papers/post-2018-fast-lexically-constrained-decoding-dynamic/)
16. Weijia Xu and Marine Carpuat. 2021. EDITOR: An edit-based transformer with repositioning for neural machine translation with soft lexical constraints. Transactions of the Association for Computational Linguistics, 9:311–328. [find](https://scholar.google.com/scholar?q=Weijia+Xu+and+Marine+Carpuat%2E+2021%2E+EDITOR%3A+An+edit%2Dbased+transformer+with+repositioning+for+neural+machine+translation+with+soft+lexical+constraints%2E+Transactions+of+the+Association+for+Computational)
17. Publications Office of the European Union. 2023. EuroVoc: the eu’s multilingual thesaurus. [find](https://scholar.google.com/scholar?q=Publications+Office+of+the+European+Union%2E+2023%2E+EuroVoc%3A+the+eu%E2%80%99s+multilingual+thesaurus%2E)
18. Vikas Raunak, Arul Menezes, and Marcin Junczys-Dowmunt. 2021. The curious case of hallucinations in neural machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Vikas+Raunak%2C+Arul+Menezes%2C+and+Marcin+Junczys%2DDowmunt%2E+2021%2E+The+curious+case+of+hallucinations+in+neural+machine+translation%2E+In+Proceedings+of+the+2021+Conference+of+the+North+American+Chapter+of+t) [our card](https://md.research.elamur.ai/free/en/papers/raunak-2021-curious-case-hallucinations-neural-machine/)
19. Biao Zhang, Philip Williams, Ivan Titov, and Rico Sennrich. 2020a. Improving massively multilingual neural machine translation and zero-shot translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1628– 1639. [find](https://scholar.google.com/scholar?q=Biao+Zhang%2C+Philip+Williams%2C+Ivan+Titov%2C+and+Rico+Sennrich%2E+2020a%2E+Improving+massively+multilingual+neural+machine+translation+and+zero%2Dshot+translation%2E+In+Proceedings+of+the+58th+Annual+Meeting+of+t) [our card](https://md.research.elamur.ai/free/en/papers/zhang-2020-improving-massively-multilingual-neural-machine/)
20. Kirill Semenov, Xu Huang, Vilém Zouhar, Nathaniel Berger, Dawei Zhu, Arturo Oncevay, and Pinzhen Chen. 2025. Findings of the WMT25 terminology translation task: Terminology is useful especially for good MTs. In Proceedings of the Tenth Conference on Machine Translation, pages 554–576, Suzhou, China. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Kirill+Semenov%2C+Xu+Huang%2C+Vil%C3%A9m+Zouhar%2C+Nathaniel+Berger%2C+Dawei+Zhu%2C+Arturo+Oncevay%2C+and+Pinzhen+Chen%2E+2025%2E+Findings+of+the+WMT25+terminology+translation+task%3A+Terminology+is+useful+especially+for+go)
21. Boliang Zhang, Ajay Nagesh, and Kevin Knight. 2020b. Parallel corpus filtering via pre-trained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8545–8554, Online. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Boliang+Zhang%2C+Ajay+Nagesh%2C+and+Kevin+Knight%2E+2020b%2E+Parallel+corpus+filtering+via+pre%2Dtrained+language+models%2E+In+Proceedings+of+the+58th+Annual+Meeting+of+the+Association+for+Computational+Linguisti)
22. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th [find](https://scholar.google.com/scholar?q=Rico+Sennrich%2C+Barry+Haddow%2C+and+Alexandra+Birch%2E+2016%2E+Improving+neural+machine+translation+models+with+monolingual+data%2E+In+Proceedings+of+the+54th)