Token-Level Interpolation For Class-Based Language Models

摘要

Optimization language model is that by iteration, arthroplasty method is used for interpolation language model (LM) from multiple component LMs, according to the purpose of interpolation weights optimization for application in domain. Component LMs may include class-based LMs, and interpolation can context-specific or independence. By iterative process, component LMs can be interpolated and be used to express material as alternative representation or parsing token. Posterior probability can determine that the component for parsing and being used for determining the LM of new (or update) interpolation weights, in this way combination or interposer LMs advanced optimize field. Component LMs can be merged, and according to optimization weight, become single combined type LM, the application of deployment.