Speech Act Identification Using Semantic Dependency Graphs with Probabilistic Context-Free Grammars

Abstract

We propose an approach for identifying the speech acts of speakers’ utterances in conversational spoken dialogue that involves using semantic dependency graphs with probabilistic context-free grammars (PCFGs). The semantic dependency graph based on the HowNet knowledge base is adopted to model the relationships between words in an utterance parsed by PCFG. Dependency relationships between words within the utterance are extracted by decomposing the semantic dependency graph according to predefined events. The corresponding values of semantic slots are subsequently extracted from the speaker's utterances according to the corresponding identified speech act. The experimental results obtained when using the proposed approach indicated that the accuracy rates of speech act detection and task completion were 95.6% and 77.4% for human-generated transcription (REF) and speech-to-text recognition output (STT), respectively, and the average numbers of turns of each dialogue were 8.3 and 11.8 for REF and STT, respectively. Compared with Bayes classifier, partial pattern tree, and Bayesian-network-based approaches, we obtained 14.1%, 9.2%, and 3% improvements in the accuracy of speech act identification, respectively.