This work extends algorithmic models of noisy-channel inference to model intercomprehension within a Bayesian framework, addressing how speakers of a related language (L1) achieve partial intelligibility of an unfamiliar language (L2).
- The model uses an L1-only language model for scoring latent hypotheses about translations of observed L2 utterances.
- A general-purpose noise model infers mappings between L2 and L1 words based on form-based similarity or symbolic rules.
- Human behavioral experiments elicited inferences for Dutch, Italian, and Ukrainian from speakers of English, Spanish, and Russian.
- The full model shows closer alignment to human intercomprehension performance than ablations and compares favorably to zero-shot prompting of larger models.
These results provide a cognitively plausible computational model of intercomprehension, highlighting the flexible inferences made by comprehenders under wide uncertainty in real-world cross-language scenarios.