TurCOMET: A Morphologically-Aware Evaluation System for Document-Level Turkish-English Machine Translation

Atahan Uz · Hüseyin Emir Akdağ · Susan Üsküdarlı
Boğaziçi University · EMNLP 2026, Budapest, Hungary

📄 Paper (PDF) · 🖼️ Poster (PDF) · 🌐 Project page · 🤖 Model · 📚 TurkDoc-MT

TurCOMET is the first Turkish-specific augmentation of COMET. It adds a 330K-parameter bottleneck to a frozen xCOMET-XL that fuses five Turkish-aware lexical features (chrF++, lemma-TER via Zemberek, LaBSE cosine, length-z and a direction flag) into the pooled representation. This counters the sub-word over-segmentation that makes neural metrics miss errors in agglutinative Turkish. On TurkDoc-MT, our new 1,322-document, 7-domain EN↔TR benchmark, an expert MQM audit of 11 LLMs shows that TurCOMET agrees with professional translators better than both BLEU and xCOMET-XL.

📖 Read the paper & poster

Full paper · 21 pages	EMNLP 2026 poster · A0

Read online · PDF	View online · PDF

📊 Key results

Agreement with expert MQM judgments from 4 professional translators (550 rated translations, 22 model × direction systems):

Metric	Segment Kendall τ	System Kendall τ	System Pearson r
BLEU	0.256	0.628	0.828
xCOMET-XL	0.267	0.636	0.950
TurCOMET (ours)	0.291	0.749	0.976

The segment-level gain is significant overall (Δτ = +0.0247, p = 0.022) and comes almost entirely from TR→EN (Δτ = +0.049, p < 0.001). In EN→TR, TurCOMET and xCOMET-XL are statistically tied.

🤗 Released resources

	Repository	Contents
🤖 Model	`TurCOMET/XCOMET-XL-TR`	TurCOMET checkpoint (bf16, ~7 GB) for scoring EN↔TR translations
📚 Benchmark	`TurCOMET/TurkDoc-MT`	1,322 document-level EN↔TR pairs across 7 domains
✍️ Annotations	`TurCOMET/TurkDoc-MT-MQM-Annotations`	2,200 expert MQM annotations with error spans, categories and severities
🔁 Outputs	`TurCOMET/TurkDoc-MT_with_model-translations`	TurkDoc-MT translations from all 11 evaluated LLMs, both directions
🏋️ Training data	`TurCOMET/wmt-da-en-tr_tr-en`	35,357 WMT EN↔TR DA ratings with features and the exact 90/5/5 splits

📝 Citation

@inproceedings{uz-etal-2026-turcomet,
  title     = {{TurCOMET}: A Morphologically-Aware Evaluation System for
               Document-Level {T}urkish-{E}nglish Machine Translation},
  author    = {Uz, Atahan and Akda{\u{g}}, H{\"u}seyin Emir and
               {\"U}sk{\"u}darl{\i}, Susan},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods
               in Natural Language Processing},
  year      = {2026},
  address   = {Budapest, Hungary},
  publisher = {Association for Computational Linguistics}
}