TurCOMET · EMNLP 2026

+0.113system-level Kendall τvs. xCOMET-XL: 0.636 → 0.749

+0.024segment-level Kendall τ0.267 → 0.291, p = 0.022

330Ktrainable parameters<0.01% of frozen xCOMET-XL

1,322documents, 7 domainsTurkDoc-MT benchmark, EN↔TR

2,200expert MQM annotationsby 4 professional translators

Abstract

We introduce TurCOMET, the first Turkish-specific augmentation of COMET for evaluating translations between Turkish and English. By extending xCOMET-XL with Turkish morphological features and training it on additional Turkish-English data via an additive bottleneck layer, TurCOMET resolves neural evaluation failures caused by sub-word over-segmentation. To validate our metric, we introduce TurkDoc-MT, a meticulously curated Turkish-English document-level aligned dataset of 1,322 instances spanning seven domains. Meta-evaluation demonstrates that TurCOMET significantly outperforms the off-the-shelf baseline, yielding absolute Kendall’s τ correlation gains of +0.113 (system-level) and +0.024 (segment-level), with the largest gains in the TR→EN direction. All models, datasets, code, and annotations are publicly released for the benefit of Turkish natural language processing and benchmarking any translation system.

Paper & poster

Read the full camera-ready paper or the EMNLP 2026 poster right here. Click any page to open the PDF at that page.

How TurCOMET works

Turkish is agglutinative: a single root can carry a long chain of derivational and inflectional suffixes, which the XLM-R sub-word tokenizer fragments aggressively. This over-segmentation depresses surface overlap and dilutes the pooled sentence representation, so TurCOMET feeds the metric explicit Turkish-aware signals.

TurCOMET architecture: frozen xCOMET-XL backbone with a trainable bottleneck that fuses five Turkish lexical features into the pooled representation — **Figure 1.** The frozen xCOMET-XL backbone (3.5B parameters) is kept intact. The only addition is a violet bottleneck that fuses a five-dimensional Turkish feature vector into the pooled embedding: *h′ = h + W₂ tanh(W₁[h; f] + b₁) + b₂*, width 64.

Five Turkish-aware features

Character overlap that survives suffix chains, lemma-level edit rate, a reference-free semantic anchor, an omission detector and a direction flag.

chrF++lemma-TERLaBSE coslength-zdirection

Never worse at step zero

W₂ and b₂ are zero-initialized, so training starts from scores identical to xCOMET-XL. Only ~330K parameters (<0.01% of the network) are trained.

One model, both directions

Fine-tuned on 35,357 WMT EN↔TR direct-assessment ratings (2017–18) with AdamW and MSE loss on a single A100. Held-out segment τ: 0.368 → 0.394.

Results

An expert MQM audit of 550 translations (4 professional translators, 11 LLMs × 2 directions) shows that TurCOMET agrees best with human judgment.

Correlation with expert MQM judgments

Segment level n = 550, system level n = 22. Higher is better.

	Metric	Pearson	Spearman	Kendall
Segment	BLEU	0.385	0.335	0.256
	xCOMET-XL	0.419	0.350	0.267
	TurCOMET	0.436	0.383	0.291
System	BLEU	0.828	0.852	0.628
	xCOMET-XL	0.950	0.802	0.636
	TurCOMET	0.976	0.913	0.749

Where the gain comes from. Segment-level Δτ = +0.0247 overall (95% CI [+0.004, +0.049], p = 0.022), carried almost entirely by TR→EN (Δτ = +0.049, p < 0.001). In EN→TR the two neural metrics are statistically indistinguishable.

LLM leaderboard on TurkDoc-MT

Averaged over both directions, sorted by the human MQM judges (negated z-score, higher = better).

Model	Humans	BLEU	xCOMET-XL	TurCOMET
gemma-4-31B-it	+0.32	31.3	0.814	0.714
translategemma-27b-it	+0.26	21.4	0.805	0.709
translategemma-12b-it	+0.22	19.0	0.785	0.695
gpt-oss-120b	+0.21	22.5	0.782	0.688
gemma-3-12b-it	+0.21	25.5	0.766	0.675
gemma-3-27b-it	+0.17	28.7	0.782	0.686
gpt-oss-20b	−0.19	22.4	0.749	0.665
Qwen3.5-9B	−0.20	22.6	0.734	0.657
translategemma-4b-it	−0.30	10.1	0.732	0.651
Llama-3.1-8B-Instruct	−0.42	16.2	0.668	0.607
gemma-3-4b-it	−0.43	14.6	0.665	0.606

TurkDoc-MT benchmark

1,322 English–Turkish document pairs kept at their native document boundaries, preserving discourse-level context, across seven domains from curated Wikipedia prose to multi-thousand-word TED talks.

Domain	Source	Docs	Mean words	90th pct.	99th pct.
Wikipedia articles	FLORES-200	200	72.8	106.1	124.2
Technical & scientific	TWiST	200	119.9	205.4	363.3
Educational dialogue	XQuAD	200	134.1	190.4	330.5
Journalistic prose	NTREX-128	122	342.9	543.6	689.1
Legal rulings	AYM	200	107.2	193.1	426.3
Web crawl	DocHPLT	200	225.1	348.1	1,260.0
Spoken language	TED2020	200	2,646.3	3,285.4	4,495.6
Total		1,322

Released resources

Everything needed to use TurCOMET, reproduce the training, and re-run the meta-evaluation.

Model TurCOMET/XCOMET-XL-TR

The TurCOMET checkpoint (bf16, ~7 GB). Scores (source, translation, reference) triplets in both directions.

Dataset TurCOMET/TurkDoc-MT

The benchmark: 1,322 document-level EN↔TR pairs across seven domains.

Dataset TurCOMET/TurkDoc-MT-MQM-Annotations

2,200 expert MQM annotations (error spans, categories, severities, z-scores) from four professional translators.

Dataset TurCOMET/TurkDoc-MT_with_model-translations

TurkDoc-MT with the EN→TR and TR→EN translations produced by every evaluated LLM.

Dataset TurCOMET/wmt-da-en-tr_tr-en

Training data: 35,357 WMT EN↔TR DA ratings with precomputed features and the exact 90/5/5 splits.

Citation

If you use TurCOMET or TurkDoc-MT, please cite:

@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}
}