| Slot Filling,aimed at correctly tagging each slot value with slot label in the utterance,is a key task in task-oriented dialogue system.Current methods for this task could achieve good performance only when giving sufficient data.However,different domains usually have different slots.It is impossible to label ample data for every domain.So,a new concept"cross-domain" was put forward to solve this problem.Models for crossdomain slot filling could be trained in source domain and fine-tuned in target domain.Traditionally,target domain faces more data-sparseness problem than source domain.Based on existing research,this paper’s work mainly includes the following two aspects:First,This paper proposes a model for Cross-Domain Slot Filling Base Multi Granularity Label.The model has four parts:Context Encoder,Slot Label Encoder,Sequence Labeling based on Coarse Labels,Sequence Labeling based on Fine Labels.Context Encoder is used to encode the utterance to get word representation vector;Slot Label Encoder is used to encode slot labels’ descriptions to get label representation vectors;Sequence Labeling based on Coarse Labels gets coarse slot labels through clustering or human induction,and uses these labels for sequence labeling task.Sequence Labeling based on Fine Labels computes the similarity between word and label representation vectors and does sequence labeling task with fine slot labels.In two public dataset SNIPS and MultiWOZ,this model has achieved stable improvement in F1,comparing with other baselines.Secondly,this paper has applied this technology in a dialogue system to design a "Sustainably Enhanced Cross-domain Dialogue System".The system assembles cross-domain slot filling model and could offer multi domain dialogue service like booking restaurants,ordering taxis,booking hotels,etc.Meanwhile,the system supports auxiliary annotation and model training for customer service staffs.The system could sustainably enhance its performance through this way.Following the software engineering development process,the system requirements for two roles are analyzed first,then the system design and implementation are performed,and finally the system is tested.The result tells that this system has perfect functions and can better meet system requirements. |
