Research And Application Of Sub-slot Based Task-oriented Dialog System

In the traditional task-oriented dialog system,many advances have been achieved under an assumption that each slot value is informed or updated as a whole in a single turn by default.But in real-world dialogs,some slot values are often in a much more complicated manner.A slot value will be provided segment by segment over multiple-turn interaction in a dialog.Take name collection as an example,user often splits the name into characters and informs each character with its paraphrase.The final completed name will be acquired through multiple rounds of disambiguation and confirmation.In particular,there are many challenges such as speech recognition,Chinese character recognition which further increase the difficulty of slot filling.This kind of task that requires multiple-turn interaction to collect one slot value is called sub-slot based task-oriented dialog.In order to address the research vacancy on this issue.This paper elaborates the definition and application scenarios of sub-slot based taskoriented dialog,and constructs a sub-slot based task-oriented dialog dataset(Sub-slot Dialog dataset,SSD).On this basis,the research on this task is carried out.Specifically,the main contributions of this paper are as follows:1.This paper proposes a novel sub-slot based dialog task and elaborates its definition and application scenarios.Meanwhile,A Sub-slot Dialog dataset is built,covering four common scenarios including phone numbers,ID numbers,Chinese names and license plate numbers collection.And the quality of the dataset is analyzed and evaluated.2.This paper proposes a sub-slot based task-oriented dialog system framework.The framework introduces a knowledge prediction module on basis of the traditional dialog system which make the model can understand the user utterance with extra knowledgebase.On the base of the framework,two knowledge-enhanced sub-slot based task-oriented dialog system models,Chinese Knowledge Match based Model and Chinese Glyphbased Pretrained Model,are implemented.The experiments have proved the validity and superiority of the model.3.This paper designs and implements a sub-slot based dialog system.The system has the ability of collecting user information such as phone number and name through voice,and can accurately understand the user’s complex description and give a reasonable reply.Finally,the functional and performance tests of the system are completed.