Morphemic and syntactic analyses in a Japanese dictation system

Abstract

A Japanese dictation system is being developed for task‐independent continuous speech recognition. To achieve task‐independency, syntactic and semantic constraints arc used in the morphemic and syntactic analysis subsystems. The structure of the Japanese sentence is effectively described by a two‐level grammar that consists of an intra‐Bunsetsu part and an inter‐Bunsetsu part. The Bunsetsu structure consists of one content word followed by several function words. Because most of the function words are short, they cannot be spotted individually with high accuracy. In the morphemic analysis subsystem, a syntax‐driven continuous DP matching algorithm spots Bunsetsu‐like units and generates a Bunsetsu lattice. In the spotting process, function words following a content word are predicted using the intra‐Bunsetsu grammar. In the syntactic analysis subsystem, a Bunsetsu lattice is parsed by an algorithm based on the CYK algorithm. Each Bunsetsu candidate has syntactic markers. The inter‐Bunsetsu grammar is expressed as the correspondence between a FEATURE marker in the modifier‐Bunsetsu and a SLOT‐FILLER marker in the head‐Bunsetsu. A recognition rate of 81.5% was obtained from simulator‐generated phoneme sequences with an 85% phoneme recognition score.