Lexical Structure for Dialogue Act Recognition

摘要: This paper deals with automatic dialogue acts (DAs) recognition in Czech. Dialogue are sentence-level labels that represent different states of a dialogue, such as questions, hesitations, ... In our application, multimodal reservation system, four considered: statements, orders, yes/no questions and other questions. The main contribution this work is to propose compare several approaches recognize based on three types information: lexical information, prosody word positions. These tested Czech Railways corpus contains human-human dialogues, which transcribed both manually an speech recognizer for comparison. experimental results confirm every type feature (lexical, prosodic positions) bring relevant somewhat complementary information. proposed methods take into account positions especially interesting, they global information about the structure sentence, at opposite traditional n-gram models only capture local cues. When sequences estimated from recognizer, resulting decrease accuracy all very small (about 3 %), confirms capability perform well real applications.