Scoring derivations with feature functions in cdec

From cdec Decoder

Jump to: navigation, search

cdec has been designed to support translation models parameterized with millions of weights that determine how much value real-valued feature functions have. cdec distinguishes two kinds of feature functions: those that are associated only with a rule, independent of its context, which are called rule features whose value are determined in the phrase table or grammar, and feature functions that are potentially sensitive to contextual information, which are called rescoring features and whose value is determined by a process of intersection with a FST whose edges encode the context-sensitive feature information in its structure.

Personal tools