Guidance Information Extraction From Broadcast Speech In Traffic Domain

Key information extraction from speech is an important field. It is a difficultythat language understanding in spoken language in a specific domain. Besides, routedecision of UGV (Unmanned Ground Vehicles) has become a hot spot. This systemrealize the function that guide route decision according to auditory information.Under robust Speech Recognition System, our group build a languageunderstanding system, which including following modules:1. In sentence boundary detection module, it extracts word class and its contextas features, and constructs the sequence tagging model. After comparing theperformance between Maximum Entropy and Conditional Random Field (CRF), itselects CRF. The F-measure can be up to81.69%after noise reduction.2. Structured information extraction module. It compares the advantages anddisadvantages of full parsing and partial parsing, and compares the performance ofapproach based on Context Free Grammar (CFG) and that based on shallow parsing.The F-measure of CFG can be96.0%. The F-measure of shallow parsing can be93.6%. Considering the robust of system, it uses shallow parsing.3. Vehicle routing decision interface converting module. The module has trafficnetwork data from OpenStreetMap. The degree of congestion are defined5levels.Based on these data, it propose an approach of converting the structured informationinto route decision interface.