We Hebben Een Serieus Translatie: Modeling Intercomprehension as Probabilistic Inference

How do humans achieve partial, zero-shot comprehension of unfamiliar but related languages (intercomprehension)? The authors investigate whether this ability can be modeled as a rational, probabilistic inference process where a listener uses their native language (L1) knowledge to interpret noisy input from a related language (L2).

The researchers propose a Bayesian framework that treats intercomprehension as a noisy-channel decipherment problem. The model consists of two primary components: a generative model and an inference algorithm. The generative model uses a monolingual language model (L1) to provide linguistic priors and a noise model that accounts for word-level transformations, including both surface-form edit distances and learned symbolic substitution rules.

To perform inference, the authors use Sequential Monte Carlo (SMC) with a tempering schedule. This allows the model to gradually shift from relying on naive, surface-level similarities (low-beta) to incorporating more complex linguistic priors (high-beta). The model maintains a set of weighted particles representing different translation hypotheses, which are refined through MCMC rejuvenation moves that propose new word edits or rule library updates.

The model was evaluated against human behavioral data from three language pairs: Dutch-English, Italian-Spanish, and Ukrainian-Russian. The results show that the full Bayesian model, which integrates both form-based likelihoods and linguistic priors, aligns more closely with human performance than ablations that rely solely on surface-level edit distance or those that lack the ability to learn symbolic rules. Furthermore, the model compares favorably to much larger, zero-shot prompted Large Language Models, suggesting that the proposed cognitive architecture effectively captures the flexible, uncertainty-driven nature of human cross-language inference.