---
title: "Adaptive CAT-embedded MT for low-memory — Sabo 2026"
description: "Adapting MT in real time inside CAT tools usually means either updating model weights after each segment, which typical translator laptops cannot handle…"
authors: ["Marek Sabo"]
year: 2026
venue: "Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 2)"
source: "https://aclanthology.org/2026.eamt-2.14/"
licence: "CC BY-ND 4.0 (EAMT)"
licence_url: "https://creativecommons.org/licenses/by-nd/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/sabo-2026-adaptive-cat-embedded-mt-low/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/sabo-2026-adaptive-cat-embedded-mt-low/)

[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/) / Adaptive CAT-embedded MT for low-memory, low-compute end-user devices

# Adaptive CAT-embedded MT for low-memory, low-compute end-user devices

**Authors:** Marek Sabo · **Year:** 2026 · **Venue:** Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 2) · **Licence:** [CC BY-ND 4.0 (EAMT)](https://creativecommons.org/licenses/by-nd/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.eamt-2.14/) · [PDF](https://aclanthology.org/2026.eamt-2.14.pdf)

## Summary
Adapting MT in real time inside CAT tools usually means either updating model weights after each segment, which typical translator laptops cannot handle, or making translators maintain glossaries by hand. This paper presents ACATMT, a compact English–Swedish encoder-decoder system initialized from mT5-base with a smaller 50,000-token unigram vocabulary. It is quantized and exported to ONNX, so it runs locally on a CPU in under 1 GB of RAM. Training began with bilingual masked span prediction, followed by fine-tuning on translations with optional term pairs marked by special tokens. The training glossaries were built from Stanza-based noun-phrase extraction, with target terms aligned using LaBSE. The model also learned to extract glossaries itself. It runs this extraction before and after a post-edit, detects changed terminology, and stores it for later segments without any user action. Translation-memory matches are supplied through decoder prefilling. On 5,021 unseen medical-technology segments, providing terms raised ACATMT well above both its own unaided output and the Helsinki opus-mt-en-sv baseline. This suggests that private, adaptive MT is practical on modest hardware.

## Key points
- **Speed and hardware:** Generation reached 16, 21, 33 and 46 tokens/s across power modes, from maximum battery saving to plugged in. The test machine was an Intel Core i5-1135G7 laptop with 16 GB RAM and no GPU.
- **Training data:** Training glossaries came from parallel corpora via the Stanza pipeline plus custom rules, with LaBSE alignment. Data for the glossary-extraction task was produced by the DeepSeek API from a few thousand bilingual texts.
- **Scores (COMET / BLEU):**
  - ACATMT without terms: 0.8195 / 35.23, slightly below opus-mt-en-sv at 0.8368 / 35.68.
  - ACATMT with decoder prefilling of terms: 0.8733 / 47.57, the best result. Glossary mode gave 0.8719 / 45.19.
  - opus-mt-en-sv with term prefilling: 0.8436 / 37.29.
- **Evaluation caveats:** Evaluation was automatic only, with no human post-editors. The test terms were extracted from the same reference material by ACATMT's own extraction mode. The implicit post-edit adaptation loop (term capture and apply rates) has not yet been evaluated.
- **Interface and availability:** A browser interface streams output token by token and lets translators click a word to get alternatives and regenerate from that point. The system is proprietary and planned for integration into STAR Transit, with licensing and pricing still open.

*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/)

## Cite

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

```bibtex
@inproceedings{sabo2026adaptive,
  title = {{Adaptive CAT-embedded MT for low-memory, low-compute end-user devices}},
  author = {Sabo, Marek},
  booktitle = {Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 2)},
  year = {2026},
  url = {https://aclanthology.org/2026.eamt-2.14/}
}
```

- [← previous ACID: On the Perception of Online Classism](https://md.research.elamur.ai/free/en/papers/muti-2026-acid-perception-online-classism/)
- [next → Adaptive Emotion Management in Human-robot Dialogue using Online Group Relative Policy Optimization](https://md.research.elamur.ai/free/en/papers/manaseryan-2026-adaptive-emotion-management-human-robot/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/)