Research On A Joint Algorithm For Multiple Intention Detection And Slot Filling In Task-Oriented Dialogue

With the increasing popular of task-oriented human-machine dialogue systems,multi-intent spoken language understanding has received unprecedented attention.Spoken language understanding consists of two main sub-tasks: intent detection(ID)and slot filling(SF)In realistic dialogue systems,the user’s utterance usually contains multiple intents and corresponding semantic information.The current optimal approach utilizes graph attention networks to model explicit correlation between intentions and slots to promote slot filling task.However,this approach does not consider the information about the edges between nodes and the correlation between intentions.In addition,existing methods decode slot filling with BIO sequence annotation,which not only cannot solve the problem of nested slots,but also the performance and inference rate are not high enough.To address the above issues,the following work and its contributions are performed in this work:(1)A model based on an iterative prediction mechanism and graph networks for joint multi-intent-slot extraction is proposed for inter-intent correlation and information about edges between nodes that can be applied to facilitate intent detection and slot filling.The model solves the problem of insufficient node information aggregation by constructing a weighted global-local undirected graph interaction network and proposing an edgeweighted graph attention network(EGAT),which enables each node to aggregate the information of surrounding nodes and edges.Moreover,an iterative prediction mechanism is proposed for the problem of inter-intent correlation and enhancing the robustness of the model,which treats the decoded intent as a soft label of the pre-decoded intent to achieve iterative prediction.Comparative experiments and ablation experiments are conducted on several publicly available(Chinese and English)datasets.The experimental results demonstrate that all metrics of both tasks are improved,indicating the generalization capability and effectiveness of the proposed method in this work.(2)To address the problems of slot nesting,slow inference rate,and poor performance of the model,a unified approach to solve nested and non-nested slots based on global pointers is proposed.Firstly,the model utilizes BERT as a parameter-sharing encoder for two tasks,which can deeply mine the text semantic representation;Then,the obtained semantic representation is fed into the global pointer network to obtain interword related information and use span-style pointers to decode slots,which can solve the problem of slot nesting;Finally,constructing a multidimensional type-slot label interaction network and combining slot-type information with the sentence-level semantic representation of BERT output for decoding intentions can further enhance the implicit association between intents and slots.Comparative experiments and ablation experiments are conducted on two publicly available multi-intent datasets.The experimental results demonstrate that the proposed model can effectively solve the slot nesting problem,improve the inference rate and performance,and provide a reference for future model design.