📄 Translation from the Information Bottleneck Perspective: an Efficiency Analysis of Spatial Prepositions in Bitexts

Summary

Studies using the Information Bottleneck (IB) framework have shown that languages trade informativity against simplicity near-optimally. So far, however, they have relied on naming data elicited with non-linguistic stimuli such as colour chips, images or videos. This paper extends IB analysis to linguistic input by treating translation as an IB problem. A word together with its sentence context in the source text is the stimulus, and the translator's choice of target-language word is its compressed encoding. The authors test this on spatial prepositions in a French novel and its English, German and Serbian translations, using a corpus released by LIFAT et al., 2024. Source prepositions were matched to their target equivalents to estimate Complexity. For Accuracy, a pile-sorting pilot with 35 participants supplied pairwise similarity judgements. A five-dimensional low-rank projection model reproduces these judgements with Spearman ρ = 0.78. Attested translations lie closer to the optimal IB frontier than random or counterfactual systems. The authors treat this as preliminary evidence of efficiency pressure on human translators. It also suggests that bitexts could stand in for costly controlled naming experiments.

Key points

• Model chain. The model chains context U → meaning M → target word W. Two occurrences count as one meaning only when their full sentential context is identical. A uniform prior p(m) is assumed, Complexity is I(M;W), and Accuracy is I(W;U).
• Alignment. Prepositions were aligned with an LLM using a few-shot prompt (reproduced in Appendix A). The prompt returns "None" when the target has no spatial term or renders the meaning only through a verb or noun.
• Similarity data. Similarity was collected through pile sorting of preposition-in-context cards (N = 35). It was modelled with a low-rank projection (D = 5), reaching Spearman ρ = 0.78.
• Embedding space. An appendix shows a t-SNE view of 30 K-means centroids built from final-layer xlm-roberta-large representations, alongside the most frequent French prepositions.
• Main result. Real translations into all three target languages are more efficient, meaning closer to the IB frontier, than counterfactual and random alternatives.