Function expression analysis device, feature weight learning device, function expression analysis method, feature weight learning method, program

摘要

PROBLEM TO BE SOLVED: To select a meaning label which is appropriate for a morphemic string even when there exists two or more function expressions having the same surface layer form.SOLUTION: Disclosed is a feature weight learning device and a function expression analysis device including a decoding part having lattice construction means and maximum likelihood path searching means. The lattice construction means is configured to create a phrase lattice in which a meaning label is applied to an input morphemic string. The phrase lattice includes a phrase in which a function expression dictionary and a morphemic string notation match, the phrase of a morpheme having parts of speech classified into function words, the phrase of the morpheme having parts of speech classified into predicates, and the phrase of the morpheme indicating neither predicate nor function expression. The maximum likelihood path searching means is configured to search the maximum likelihood path as the maximum likelihood path from the phrase lattice by using a parameter table in which weights corresponding to features are stored. The features include both a feature associating the morphemic string with the meaning label and a feature expressing a meaning label string.