Efficient communication requires balancing informativity and simplicity when encoding meanings. The Information Bottleneck (IB) framework captures this trade-off formally, predicting that natural language systems cluster near an optimal accuracy-complexity frontier. While supported in visual domains such as colour and motion, linguistic stimuli such as words in sentential context remain unexplored. We address this gap by framing translation as an IB optimisation problem, treating source sentences as stimuli and target sentences as compressed meanings. This allows IB analyses to be performed directly on bitexts rather than controlled naming experiments. We applied this to spatial prepositions across English, German and Serbian translations of a French novel. To estimate informativity, we conducted a pile-sorting pilot-study (N=35) and obtained similarity judgements of pairs of prepositions. We trained a low-rank projection model (D=5) that predicts these judgements (Spearman correlation: 0.78). Attested translations of prepositions lie closer to the IB optimal frontier than counterfactual alternatives, offering preliminary evidence that human translators exhibit communicative efficiency pressure in the spatial domain. More broadly, this work suggests that translation can serve as a window into the cognitive efficiency pressures shaping cross-linguistic semantic systems.
Figures & tables
Figure 1: Excerpt of ( LIFAT et al., 2024 ) . Spatial prepositions in the French source sentences are emphasised. id refers to the original identifier as released in the corpus.
Figure 2: French to English excerpt of computed alignments of spatial prepositions from LIFAT et al. (2024) . Each row is a one-hot encoding of the source preposition over the target lexicon. The source preposition is emphasised in bold.
Figure 3: Empirical similarity data from the pile-sorting task.
Model
Spearman ρ
Cosine Similarity (Baseline)
0.079
Ridge Regression
0.647±0.056
Low-rank Projection
D=1
0.263±0.140
D=5
0.780±0.062
D=10
0.762±0.053
Table 1: Spearman correlation between model-predicted pairwise similarities and human pile-sorting judgements. Models all take pairs of contextual embeddings as input. Learned models performance is computed from 6-fold nested cross validation.
Figure 4: How much do attested and random systems deviate from optimality?
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure C-1 : Distribution of the number of piles across participants (N=35).
Figure D-1 : t-SNE visualisation of the 30 K-means centroids (computed from stacked final-layer representations of xlm-roberta-large ) and embeddings of the most frequent French prepositions
Translated texts exhibit systematic differences from comparable texts originally written in the target language. Explaining this phenomenon, commonly known as translationese, remains an open challenge. Translationese has been attributed to production tendencies (e.g. interference, simplification), socio-cultural variables, and language-pair effects, yet a unified explanatory account is lacking. We investigate the hypothesis that translationese is a response to the cognitive load inherent in the translation task. We test whether observable translationese can be predicted from quantifiable measures of translation task difficulty. Translationese is measured as a segment-level probability of being a translation produced by an automatic classifier (translatedness score). Translation task difficulty includes source-text and cross-lingual transfer components. They are captured by information-theoretic metrics based on LLM surprisal and by established syntactic and semantic alternatives. We use a bidirectional English-German corpus comprising written and spoken subcorpora. Results indicate that translationese can only be explained in part by translation task difficulty, especially in the English-to-German direction. For most experiments, cross-lingual transfer difficulty contributes more than source-text complexity. Information-theoretic indicators match or outperform traditional features in written mode, but offer no advantage in spoken mode. Source-text syntactic complexity and translation-solution entropy emerged as the strongest predictors of translationese across language pairs and modes.
Maria Kunilovskaya
Department of Language Science and Technology, Saarland University
Modern translation workflows demand more than semantic equivalence. Users routinely require models to preserve JSON or HTML schemas, honor curated glossaries, disambiguate with provided context, and match prescribed registers, often several at once. Conventional metrics such as BLEU and xCOMET capture semantic fidelity but provide little signal on constraint adherence, while general instruction following benchmarks ignore the cross-lingual nature of translation. We introduce \bench, a benchmark for multilingual translation instruction following covering seven languages, with 4,506 single-constraint and 2,838 multi-constraint items spanning six constraint dimensions and five compositional patterns with instructions issued in all seven languages. Constraints are split into a gating subset verified by deterministic checkers and a continuous subset scored by a rubric-based LLM judge, combined under a multiplicative rule that resists reward hacking. Evaluating 15 models reveals systematic gaps that prior protocols miss: Instruction following scales with size more sharply than translation quality, glossary and structured-format constraints dominate the difficulty gradient, and general instruction following rankings correlate only weakly with translation behavior. Our benchmark are available at https://github.com/Tencent-Hunyuan/Hy-MT2/tree/main/IFMTBench.
A standard pipeline for symbolic reasoning over natural-language problems translates them into first-order logic and invokes a theorem prover. The translation step is the bottleneck: swap "every" for "some" and every inference that follows is corrupted. Yet today's metrics often score more broken translations higher than less broken ones, because BLEU, BERTScore, and Smatch++ reward surface overlap that the worst errors happen to preserve. We introduce SIV, which derives two kinds of probes from the target formula and uses a theorem prover to verify the candidate translation against each. Positive probes are statements the candidate must entail, which detect translations that drop content; contrastive probes are statements the candidate must not entail, which detect translations that assert more than the original. On a controlled pool of perturbed FOLIO translations, the severity of the error accounts for 80% of SIV's score variance, compared with at most 17% for any prior metric. Across six error classes on a disjoint pool, SIV scores the reference above the perturbed candidate in over 99% of pairs. Because each probe is labeled with what it tests, the failure pattern also supplies a labeled error trace, recovering the perturbation class at macro-F1 0.638, nearly double the score-only baseline. On 434 expert-audited real LLM translations, SIV attains the top AUC, uniquely detects and grades expert-labeled major errors, and abstains, rather than mis-scoring, on out-of-vocabulary translations.