應用遷移式學習於跨領域意圖偵測與槽值辨識

Smartphones are what modern people carry with all the time. In many cases,users could not operate their mobile phone to reply emails, send messages,check itineraries, play music, etc. because they are driving, cooking, or jogging.Therefore, the demand of virtual assistant through spoken language is growing.For example, the advent of Actions on Google and Alexa Skills enabledevelopers to more easily connect their services through a conversationalinterface. In this project, we will collect dialog of the following domains: includinge-mail, calendar, messages, events, playing (finding) songs, etc. to establish across-domain dialogue system that supports daily life.The two-year plan is as follows: the first year plan mainly collect the crossdomain dialogue of the above-mentioned services through Human-to-Human(H2H) method. We will trains the NLU (natural language understanding) modulefor intent detection and slot filling, and integrates dialogue state tracking, dialoguepolicy decision, and natural language generation modules. In the second year,we will use adopt Machine-to-Machine (M2M) method to speed up the dialogcollection process and use transfer learning in cross-domain intent detection andslot value identification, hoping to reduce the training cost of the model andexpand the generalization ability of the model in cross-domain dialogue systems.