Large vocabulary speech recognition using concept networks

Summary form only given. A new algorithm for large-vocabulary speech recognition using a connectionist-model approach is described. The technique is to select reasonable word sequences using semantic information about a conceptual relationship among words. The conceptual relationship is realized by the knowledge representation method called concept networks, which represents knowledge by two-relation links, i.e. hierarchical is-a relations and instance relations. Theoretical evaluations were carried out using concept networks that defined the relationships among words in 3000-word vocabularies picked up from technical news articles. The authors report on formulation results of word relations and evaluation results of the relationship between the number of input words and the number of word sequences selected.<>