Stochastic Estimation of Transduced Language Models
An unbiased sampler makes prefix probabilities tractable when a transducer maps many source strings to one target prefix. It matters mainly for specialized text or biological sequence systems.
The method samples source prefixes **without replacement** and corrects each with its inverse inclusion probability. Recursive correction yields an **unbiased estimator** and estimates probability mass lost through threshold pruning.
Builders implementing transduced models can replace an unknown-error lower bound with an estimator whose compute and variance are measurable. The approach is most relevant when transformations admit exponentially many or infinite source strings.
The method samples source prefixes **without replacement** and corrects each with its inverse inclusion probability. Recursive correction yields an **unbiased estimator** and estimates probability mass lost through threshold pruning. Builders implementing transduced models can replace an unknown-error lower bound with an estimator whose compute and variance are measurable. The approach is most relevant when transformations admit exponentially many or infinite source strings. Evidence covers encyclopedic text and DNA. The DNA-to-amino-acid case ran **several orders of magnitude** faster than threshold-pruned beam summing, but the supplied material gives no exact runtime or broader production results.
This replaces threshold-pruned lower bounds with an unbiased, compute–variance-controllable estimate of transformed-model probability and the mass pruning discards. It narrows the practical case to transformations with too many source strings to enumerate; the strong DNA speed result supports that setting but does not establish general production performance.