In this paper, we introduce a new datarepresentation format for language processing,the syntactic and semantic graphs(SSGs), and show its use for call classifi-cation in spoken dialog systems. For eachsentence or utterance, these graphs includelexical information (words), syntacticinformation (such as the part of speechtags of the words and the syntactic parse ofthe utterance), and semantic information(such as the named entities and semanticrole labels). In our experiments, weused written language as the training datawhile computing SSGs and tested on spokenlanguage. In spite of this mismatch,we have shown that this is a very promisingapproach for classifying complex examples,and by using SSGs it is possibleto reduce the call classification error rateby 4.74% relative.
展开▼