Intelligent personal assistant software, such as Apple’s Siri and Samsung’s S-Voice, is being widely used these days. One of the core modules of this kind of software is the spoken language understanding (SLU) module used to predict the user’s intention for determining the system actions. The SLU module usually consists of several connected recognition components on a pipeline framework, whereas the proposed SLU module is developed by a novel technique that can simultaneously recognize four recognition components, namely named entity, speech-act, target, and operation using conditional random fields. In the experiments, the proposed simultaneous recognition technique achieved a relative improvement as high as approximately 2.2% and a faster speed of approximately 15% compared to a pipeline framework. A significance test showed that this improvement was statistically significant because the p-value was smaller than 0.01.
JOINTLY PREDICTING DIALOG ACT AND NAMED ENTITY FOR SPOKEN LANGUAGE UNDERSTANDING
