Man-machine dialogue model training method and man-machine dialogue method

A technology of human-machine dialogue and training methods, applied in the field of artificial intelligence, can solve problems such as high cost, time-consuming, labor-consuming, etc., to achieve the effect of improving performance and avoiding the accumulation of errors

Active Publication Date: 2021-06-29
SHANGHAI XIAOI ROBOT TECH CO LTD
View PDF4 Cites 1 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

End-to-end dialogue models often require a large amount of corpus for training, but the process of collecting corpus is time-consuming, labor-intensive, and costly; pipeline-type models usually have multiple separate modules for independent updates and optimizations, but It is necessary to follow the direction of the pipeline, and the results of the previous module must be obtained before the next module can be run, which can easily cause errors to accumulate and propagate between modules. In addition, when new training data is obtained, it must be completed after the training of the previous module. for subsequent training

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Man-machine dialogue model training method and man-machine dialogue method
  • Man-machine dialogue model training method and man-machine dialogue method
  • Man-machine dialogue model training method and man-machine dialogue method

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0040] In this embodiment, a training method of a man-machine dialogue model is given. Such as figure 1 shown, including the following steps:

[0041] Step S110: Obtain training samples;

[0042] Wherein, the training samples include input data and supervised label sequences corresponding to different decoding networks, and the input data includes the current round of user dialogue data U t Reply R with the last round of the system t-1 , the supervised label sequence includes at least the cumulative label sequence S of the dialogue state of the current round t_label And the current round system replies to the label sequence R t_label , where t represents the round number of the current round;

[0043] That is, the data set used for training must at least include user dialog data (that is, the source of the dialog) and supervisory labels, which are used to verify the training results of the decoding network.

[0044] Step S120: Encoding the input data;

[0045] For the c...

Embodiment 2

[0100] The training method of the man-machine dialogue model given in this embodiment is based on the basis of the first embodiment, and a hidden state connection is added between the encoder and each decoder of the model, instead of simply establishing a relationship with the output results of text symbols. By sharing the hidden state, the transfer of knowledge between networks is realized, and the initialization of each network is assisted, so that the connection between the networks is closer.

[0101] Optionally, in an implementation manner of this embodiment, in step S120 of Embodiment 1, when encoding the input data, in addition to converting the scalar data in the input data into a vector sequence, the encoder also Get the hidden state of the encoder It is used to assist each decoding network to perform this round of initialization.

[0102] In this implementation, the encoder hidden state The role of is not limited to each training treatment, It is used to initia...

Embodiment 3

[0112] In the third embodiment, a man-machine dialogue method is disclosed, such as image 3 shown, including the following steps:

[0113] Step S310: Obtain the current round of user dialogue data U t ;

[0114] Step S320: Send the current round of user dialog data U t Input utilizes in the man-machine dialogue model trained by any training method in embodiment one or embodiment two, obtain current round system reply data R t ;

[0115] Step S330: Reply the current round system with data R t sent to the user.

[0116] Taking the human-machine dialogue model including NLU network, DST network, DLP network and NLG network as an example, when performing multiple rounds of human-computer dialogue and communication, first, in step S310, obtain the dialogue data input by the user in this round, for example: I Look for an upscale restaurant on the south side of town.

[0117] received user data U 1 Afterwards, step S320 is performed to obtain the system reply of the current ...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

No PUM Login to view more

Abstract

The invention provides a man-machine dialogue model training method and a man-machine dialogue method. The man-machine dialogue model integrates sub-modules of a pipelined dialogue system into an overall end-to-end structural framework; the current round of dialogue data and the previous round of historical data of a user are obtained and coded into a vector sequence; and finally, through four sub-modules of natural language understanding, dialogue state tracking, dialogue strategy learning and natural language generation in sequence, a current round of reply of the system is obtained. The invention further provides a man-machine dialogue method which is suitable for performing man-machine interaction by utilizing the model. According to the invention, the structures of the sub-modules can be flexibly selected according to the types of the supervision labels contained in the training data when the man-machine dialogue model is trained, all the sub-modules can be optimized at the same time, the problem that errors are continuously accumulated and spread is avoided, the system reply accuracy can be improved, and the number of samples used during training is greatly reduced.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, in particular to a training method of a man-machine dialogue model and a man-machine dialogue method. Background technique [0002] A task-based dialogue system is a man-machine dialogue system that can provide information or services to users with a clear purpose under specific conditions. It can be applied in many fields such as air tickets, restaurant reservations, fees, and address inquiries. [0003] Traditional task-based dialogue system models can be divided into end-to-end and pipeline models, both of which have their own advantages, but also have obvious shortcomings. End-to-end dialogue models often require a large amount of corpus for training, but the process of collecting corpus is time-consuming, labor-intensive, and costly; pipeline-type models usually have multiple separate modules for independent updates and optimizations, but It is necessary to follow the direct...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Application Information

Patent Timeline
no application Login to view more
Patent Type & Authority Applications(China)
IPC IPC(8): G06F3/01G06N20/00
CPCG06F3/011G06N20/00
Inventor 霍沛沈大框陈成才
Owner SHANGHAI XIAOI ROBOT TECH CO LTD
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products