End-to-end hierarchical decoding task-based dialogue system

Abstract

The invention discloses an end-to-end hierarchical decoding task-type dialogue system, including: a sentence encoder module, which is used to convert the natural language sentence u input by the user in the first round l The output is encoded as a fixed-length vector through a bidirectional recurrent neural network. The dialogue history encoder module is used to encode the dialogue history into s through a unidirectional recurrent neural network. l : The user intent introduction module is used to input the dialogue history and the embedding of the current user sentence, and the user’s intent representation is obtained through sampling. The database retrieval module is used to input the slot extracted from the natural language sentence input by the user, search in the database, and return a variable d l , indicating whether there is data that meets the requirements in the database; the decoding output module is used to input the dialogue history s l , The embedding database of the user sentence returns the variable d l and user intent hidden variable z l , output natural language sentences as the system's reply to user input:

Description

technical field [0001] The invention relates to an end-to-end hierarchical decoding task type dialogue system. Background technique [0002] The current task-based dialogue systems are mainly based on finite state machines and statistical dialogue models. There are two main models: [0003] The first is a finite state machine model based on slot-filling. This model is based on slot-filling and manually designed dialog states, and uses the slot filled state as the representation of dialog states, and manually designs dialog states and state transition rules based on business logic. This method relies on the accuracy of slot extraction, and the dialogue process relies on manual design. It needs to design complex rules to support complex dialogue strategies and reply templates, which is difficult to apply to different fields. [0004] The second is a dialogue system based on a statistical model, which was proposed by Steveyoung and includes four main modules: natural language...

AI Technical Summary

Problems solved by technology

Due to the structure of the pipeline, an error generated by a certain module will greatly affect subsequent modules, resulting in the propagation and amplification of errors

In addition, the natural language understanding and dialogue state tracking modules require a large amount of manually labeled data to train the corresponding neural network classifier, which is difficult to use in cold start or when the amount of data is small

Images