CN112000793A - Man-machine interaction oriented dialogue target planning method - Google Patents

Man-machine interaction oriented dialogue target planning method Download PDF

Info

Publication number
CN112000793A
CN112000793A CN202010888963.5A CN202010888963A CN112000793A CN 112000793 A CN112000793 A CN 112000793A CN 202010888963 A CN202010888963 A CN 202010888963A CN 112000793 A CN112000793 A CN 112000793A
Authority
CN
China
Prior art keywords
target
vector
enhancement
network module
graph
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010888963.5A
Other languages
Chinese (zh)
Other versions
CN112000793B (en
Inventor
张柏林
涂志莹
初佃辉
申义
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Institute of Technology
Original Assignee
Harbin Institute of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Harbin Institute of Technology filed Critical Harbin Institute of Technology
Priority to CN202010888963.5A priority Critical patent/CN112000793B/en
Publication of CN112000793A publication Critical patent/CN112000793A/en
Application granted granted Critical
Publication of CN112000793B publication Critical patent/CN112000793B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/332—Query formulation
    • G06F16/3329—Natural language query formulation or dialogue systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/34—Browsing; Visualisation therefor
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

The invention discloses a human-computer interaction oriented dialogue target planning method, which comprises the following steps: step S1, collecting and labeling a large-scale corpus; step S2, calculating a target transfer vector matrix; step S3, constructing a target planning joint learning model; and step S4, implementing the target planning scheme. The invention divides the dialogue target planning problem into a target prediction subtask and a target completion subtask, and discloses endogenous potential association between the two subtasks, thereby clarifying the essential reason that the two subtasks can carry out joint learning. The invention provides a series of methods for constructing and vectorizing a conversation object transition diagram. The invention provides a novel multi-task joint learning neural network model, which models the inherent relevance between two subtasks by designing a target prediction and a target completion two sub-network and improves the performance of each network feature representation by adopting a circulation enhancement mechanism so as to improve the effect of a target planning model.

Description

Man-machine interaction oriented dialogue target planning method
Technical Field
The invention belongs to the technical field of computer services, relates to a human-computer interaction-oriented dialogue target planning method, and particularly relates to a multi-task joint learning model-based dialogue target planning method.
Background
The man-machine interaction technology is characterized in that a certain dialogue language is used between a person and a computer, and a bridge between the person and a machine is built for the information exchange process between the person and the computer for completing a determined task in a certain interaction mode. Human-computer interaction technology, especially human-computer dialogue technology for machine understanding and using natural language to realize human-computer communication, is an important challenge for artificial intelligence. With the recent rise of deep learning, the field of man-machine conversation has been advanced greatly. However, the current man-machine dialog system is still in a starting stage: often only passively respond to the user's questions and not actively direct the topic targets in the conversation.
The traditional man-machine conversation system takes the dialogue corpus of one question and one answer as a data set, trains an end-to-end sequence generation model, can only return a proper output as a response according to the input of a user, cannot actively ask a question to the user, and cannot naturally guide the whole dialogue to be carried out.
But in real world applications such as chatting, task-type conversations, recommended conversations, and even question and answer, man-machine interaction conversations are more of a form of multi-turn conversations. Therefore, for many practical applications, it is crucial how the system actively and naturally conducts the guidance of the dialog in multiple rounds of man-machine dialog. For example, the system enhances the user experience by actively asking questions in a question-and-answer or task-type conversation, or the system may introduce recommendations for a given item as a commercial in a chat.
In a real situation, a human-computer conversation is shown in fig. 1, and chat contents of a user and a machine are switched among different conversation targets. The first dialog object [ small talk ] is actively initiated by the machine, the dialog object is directed to [ disease diagnosis ] and [ recommended treatment scheme ] according to the user's reply machine, and finally the dialog is terminated by the user, the dialog object is [ goodbye ]. The complete dialog target sequence in this example is [ small speech, disease diagnosis, recommended treatment, see again ], unknown to the user. The machine judges which conversation target the current conversation target should be classified into according to the output text of the user, and simultaneously judges whether the current conversation target is finished, and if so, the next conversation target is predicted. When the user inputs the text as 'next toothache test, try one's test. When declaring that the current session goal recommended a treatment regimen is complete. When predicting a dialog target, it is necessary to know whether the current target is completed or not to find a more accurate prediction direction.
The dialogue objective planning can thus be split into two subtasks with a certain relevance:
(1) target prediction: judging a current conversation target according to the user text, wherein the current conversation target can be abstracted into a multi-classification task;
(2) the target is completed: and judging whether the conversation target is finished according to the user text and the current conversation target, wherein the conversation target can be abstracted into two classification tasks.
Most machine learning tasks today are single task learning. For complex problems, the problem is decomposed into simple and independent sub-problems to be solved separately, and then results are combined to obtain the result of the original complex problem. However, this approach has two disadvantages, namely, on one hand, the mutual relevance among the sub-problems is ignored, and the sub-problems can be actually linked together through a sharing factor or a sharing representation. On the other hand, the independent modeling subproblem brings significant increase of model parameters, can not exert the end-to-end advantage of the deep network, and brings inconvenience to actual services.
Therefore, a plurality of subtasks can be subjected to joint learning, and multitask refers to optimization of a plurality of targets instead of a single target. The multi-task learning can realize information migration and sharing among different tasks, and information learned by the other task can be fully utilized among all subtasks to improve the learning effect of the subtasks. The traditional joint learning implementation mode is that a plurality of tasks share the same parameter, or loss functions of two tasks are reasonably assembled into one loss function, and parameter training is carried out on the assembled loss function. This traditional approach makes the different subtasks unbiased, makes no use of the information contributions of the different subtasks to distinguish, and does not explicitly model the potential dependencies between different tasks.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a human-computer interaction-oriented dialogue target planning method. The invention designs a conditional probability-based generation scheme of a directed weighted target transfer graph, and calculates vector representation of each target node in the graph based on a Deepwalk algorithm.
The purpose of the invention is realized by the following technical scheme:
a human-computer interaction-oriented dialogue target planning method comprises the following steps:
step S1, collecting and labeling a large-scale corpus:
(1) collecting and integrating a common multi-round dialogue corpus, and cleaning and sorting texts of an original corpus;
(2) manually marking the dialogue targets related to the input text of the user in each pair of dialogs and whether the dialogue targets are finished or not for the processed corpus, and sorting the dialogue targets related to the whole section of dialogue formed by multiple turns of dialogs into a target sequence;
step S2, calculating a target transfer vector matrix:
integrating the target sequence constructed for the whole corpus in the step S1, converting the target sequence into a target transition graph, calculating the weight of edges in the transition graph by using conditional probability, and finally performing vectorization representation on each target node in the graph by using a Deepwalk algorithm, wherein the steps are as follows:
(1) merging all target transfer sequences appearing in the corpus, and converting the merged target transfer sequences into a transfer graph G ═ V, E, wherein V is a node set, and E is an edge set;
(2) calculated at a known target node viSubject to condition (v) of the target nodejConditional probability of occurrence P (v)j|vi) Taking the weighted value as the weight on the edge e to form a directed weighted target transfer graph G;
(3) converting the nodes in the graph into low-dimensional dense entity vectors by utilizing a graph representation learning algorithm Deepwalk, and fully utilizing the information of random walk sequences in the graph structure;
step S3, constructing a target planning joint learning model:
the model mainly comprises five modules: attention network module, target completion network module, target prediction network module, circulation enhancement module and classification module, wherein: the attention network module is used for calculating the contribution degree of the network structure characteristics of each target node in the target transfer graph to the target to which the user input text belongs; the target completion network module and the target prediction network module are respectively used for enhancing and enriching vector representation of corresponding subtasks; the circulation enhancing module designs a circulation mechanism for updating the vector representation of each enhanced subtask; the classification module is used for completing the specific classification problem corresponding to each subtask;
the specific construction steps are as follows:
(1) inputting the target transfer vector matrix obtained in the step S2 and the text semantic feature vector subjected to ERNIE fine tuning into an attention network module, and outputting a target prediction vector;
(2) inputting the text semantic feature vector and the target prediction vector obtained in the step (1) into a target completion network module to improve the performance of feature representation, and obtaining a target completion enhancement vector as output;
(3) inputting the target prediction vector and the target completion enhancement vector respectively obtained in the step (1) and the step (2) into a target prediction network module to share information between the two vectors, and obtaining a target prediction enhancement vector as output;
(4) setting cycle times, continuously updating the input in the step (2) and the step (3), and taking the enhanced vector as the new input in the step (2) and the step (3) so that a cycle mechanism is arranged between the target prediction network module and the target completion network module;
(5) inputting the target prediction vector and the target completion vector after cyclic enhancement into a classification module, and obtaining classification results by respectively passing the two vectors through two different softmax functions;
step S4, implementing a target planning scheme:
and (4) training the target planning joint learning model constructed in the step (S3) based on the real corpus obtained in the step (S1), selecting the model with the best effect through parameter tuning, deploying the model on line, and exposing the model into a Web service interface for a service user to call.
Compared with the prior art, the invention has the following advantages:
1. the invention divides the dialogue target planning problem into a target prediction subtask and a target completion subtask, and discloses endogenous potential association between the two subtasks, thereby clarifying the essential reason that the two subtasks can carry out joint learning.
2. The invention provides a series of methods for constructing and vectorizing a conversation object transition diagram, which comprises the following steps:
(1) a method of constructing a conversation target transition graph;
(2) a method for calculating the edge weight based on the conditional probability on the target transition graph;
(3) and obtaining vector embedded representation of the target transition graph node based on the Deepwalk algorithm.
3. The invention provides a novel multi-task joint learning neural network model based on a conversation target transfer diagram, and the novel multi-task joint learning neural network model is characterized in that the intrinsic relevance between two subtasks is modeled by designing a target prediction and a target completion two sub-networks, and the performance of each network feature representation is improved by adopting a circulation enhancement mechanism, so that the effect of a target planning model is improved.
Drawings
FIG. 1 is an exemplary diagram of a session target switching process in normal human-computer interaction;
FIG. 2 is a flow chart of a human-computer interaction-oriented dialog goal planning method of the present invention;
FIG. 3 is a flowchart of an algorithm for calculating a target transfer vector matrix;
FIG. 4 is a block diagram of an attention network module;
fig. 5 is a structure diagram of a target planning joint learning model network.
Detailed Description
The technical solution of the present invention is further described below with reference to the accompanying drawings, but not limited thereto, and any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention shall be covered by the protection scope of the present invention.
The invention provides a human-computer interaction-oriented dialog target planning method, which comprises the following steps of:
and step S1, collecting and labeling a large-scale corpus.
The method mainly comprises the steps of collecting a disclosed common multi-round man-machine conversation data set, requiring that each section of conversation relates to a plurality of conversation targets, having multi-round man-machine conversation contents under each conversation target, and having obvious sequential relation among the conversation targets, so as to better simulate rich and variable scenes of the conversation targets in practical application. The data set is manually marked, merged and arranged into a language material library after operations such as word segmentation, word stop removal, special character removal, word list statistics and the like. The complete dialog segment contains a plurality of dialog contents of multiple rounds under different targets, each dialog comprises input texts of a user and response texts of a machine, and the label contents are the dialog targets related to the input texts of the user in each dialog and whether the targets are completed. In order to label the above text, a large amount of data needs to be observed and summarized, and 21 common dialogue targets are found. After labeling, the dialog targets related to the whole dialog formed by multiple rounds of dialogs are arranged into a target sequence.
Step S2, calculating a target transfer vector matrix
In this step, a target transfer vector matrix needs to be calculated for the data collected and labeled in step S1, and a specific flowchart is shown in fig. 3. Firstly, constructing a conversation target transition graph, calculating conditional probability to obtain weights on edges in the transition graph, and finally converting all nodes in the graph into a low-dimensional dense entity vector matrix by using a Deepwalk algorithm. Each row of the matrix is a vector representation of a node, the number of rows is the number of target nodes, and the number of columns is the dimension of the node vector.
A whole dialog between a user and a machine involves multiple dialog targets, which can be regarded as one target transition sequence, and all transition sequences appearing in the corpus are merged and can be converted into a transition graph G ═ V, E. Where V is the set of nodes and E is the set of edges. Each dialog target is a node, and the edge e ═ vi,vj) E represents the target viTo vjHas a transfer relationship. Calculated at a known target viUnder the condition of vjConditional probability of occurrence P (v)j|vi) The weighted values are used as the weights on the edge e to form a directed weighted target transition graph G.
Deepwalk is a graph representation learning algorithm, fully utilizes information of random walk sequences in a network structure, and converts nodes in a graph into low-dimensional dense entity vectors. Deepwalk firstly generates a large number of central nodes v by sampling on the networkiRandom walk sequence with window size k (v)i-w,...,vi-1,vi+1,...,vi+w) Then carrying out probability modeling on the node pairs in each local window in the random walk sequence, and calculating the node pairs consisting of the central nodes viAnd generating the probability of nodes on two sides, maximizing the likelihood probability of the random walk sequence, and finally using the random gradient descent learning parameters.
And step S3, constructing a target planning joint learning model.
The target transfer graph obtained in step S2 represents the compliance relationship between different chat targets in the conversation process, and the trained node vector reflects the abstract network structure characteristics of the transfer graph. ERNIE is an unsupervised word vector pre-training model, and low-dimensional dense word vectors in a semantic space are obtained by uniformly modeling lexical structures, grammatical structures and semantic information in training data, so that the general semantic representation capability is greatly enhanced. The ERNIE-based pre-training model excavates semantic features in the text, and the semantic features and the network structural features jointly express the target features of the current conversation. And simultaneously, constructing a target planning joint learning model by using the network structure characteristics and the semantic characteristics. The model mainly comprises:
(1) attention network module: designing an attention network for calculating the contribution degree of each target node network structure feature to the target of the text, inputting the target transfer vector matrix obtained in the step S2 and the text semantic feature vector after ERNIE fine adjustment, and outputting the target prediction vector apThe concrete structure is shown in FIG. 4;
(2) a target completion network module: directly using text semantic feature vector as target completion vector acA obtained from attention network modulepTogether, this goal is entered to complete the network module. Calculating an enhancement factor fc＝tanh(Wc·ac+Vp·ap+bc) Wherein W iscAnd VcAre respectively acAnd apWeight matrix of bcIs a bias vector for the perturbation. Will f iscAnd acThe product is made to obtain a target completion enhancement vector hc＝fc·acAs an output;
(3) a target prediction network module: input is h obtained by the target completion network modulecAnd a obtained by the attention network modulep. Calculating an enhancement factor fr＝sigmoid(Wr·hc+Vr·ap+br) And an attenuation factor fd＝sigmoid(Wd·hc+Vd·ap+bd) Wherein W isrAnd VrAre respectively hcAnd apEnhanced weight matrix of, WdAnd VdAre respectively hcAnd apAttenuation weight matrix of brAnd bdIs a bias vector for the perturbation. After attenuation and enhancement in proper proportion, calculating to obtain a target prediction enhancement vector hp＝tanh(fr·hc+fd·ap)·ap；
(4) A circulation enhancing module: setting cycle times to continuously complete the input a of the target in the network modulepIs replaced by hp+apWherein h ispIs the output of the target prediction network moduleAnd a circulation mechanism is arranged between the target completion network module and the target prediction network module.
(5) Classification module
Soft max (W. concat (a) was usedc,hc) Calculate the probability of each target, where concat is the fusion operation of the vectors, and the one with the highest probability is the target of the current dialog. While using soft max (W. concat (a)p,hp) The probability of whether each target is completed is calculated, and if the probability is more than 0.5, the current target of the task is already finished and the task needs to be transferred to the next target.
And constructing a target planning joint learning model based on the modules to carry out user target planning, wherein the specific model structure is shown in FIG. 5.
And step S4, implementing the target planning scheme.
The step is mainly to train the target planning joint learning model constructed in the step S3 based on the real corpus obtained in the step S1. And selecting the model with the best effect through parameter tuning, deploying the model on line, and exposing the model into a Web service interface for a service user to call.
Example (b):
as shown in fig. 1, when the machine is expressed as "toothache is a common symptom of oral cavity affection, various causes may cause the symptom to appear. "when the goal is" small "in the conversation between the machine and the user is predicted at the same time using the goal planning joint learning model, and the goal is completed and the next goal should be done. What is what when the user's expression is "what that? ", the goal is to recommend a treatment plan and not complete in the machine to user's conversation using the goal planning co-learning model.

Claims (8)

1. A human-computer interaction-oriented dialog target planning method is characterized by comprising the following steps:
step S1, collecting and labeling a large-scale corpus:
(1) collecting and integrating a common multi-round dialogue corpus, and cleaning and sorting texts of an original corpus;
(2) manually marking the dialogue targets related to the input text of the user in each pair of dialogs and whether the dialogue targets are finished or not for the processed corpus, and sorting the dialogue targets related to the whole section of dialogue formed by multiple turns of dialogs into a target sequence;
step S2, calculating a target transfer vector matrix:
integrating the target sequence constructed for the whole corpus in the step S1, converting the target sequence into a target transition graph, calculating the weight of edges in the transition graph by using conditional probability, and finally performing vectorization representation on each target node in the graph by using a Deepwalk algorithm;
step S3, constructing a target planning joint learning model:
the target planning joint learning model mainly comprises five modules: attention network module, target completion network module, target prediction network module, circulation enhancement module and classification module, wherein: the attention network module is used for calculating the contribution degree of the network structure characteristics of each target node in the target transfer graph to the target to which the user input text belongs; the target completion network module and the target prediction network module are respectively used for enhancing and enriching vector representation of corresponding subtasks; the cyclic enhancement module is used for updating the vector representation of each sub task after enhancement; the classification module is used for completing the specific classification problem corresponding to each subtask;
step S4, implementing a target planning scheme:
and (4) training the target planning joint learning model constructed in the step (S3) based on the real corpus obtained in the step (S1), selecting the model with the best effect through parameter tuning, deploying the model on line, and exposing the model into a Web service interface for a service user to call.
2. The human-computer interaction-oriented dialog goal planning method according to claim 1, characterized in that the specific steps of step S2 are as follows:
(1) merging all target transfer sequences appearing in the corpus, and converting the merged target transfer sequences into a transfer graph G ═ V, E, wherein V is a node set, and E is an edge set;
(2) calculated at a known target node viSubject to condition (v) of the target nodejConditional probability of occurrence P (v)j|vi) Taking the weighted value as the weight on the edge e to form a directed weighted target transfer graph G;
(3) and (3) converting the nodes in the graph into low-dimensional dense entity vectors by utilizing a graph representation learning algorithm Deepwalk, and fully utilizing the information of the random walk sequence in the graph structure.
3. The human-computer interaction-oriented dialog goal planning method according to claim 1, characterized in that the specific construction steps of step S3 are as follows:
(1) inputting the target transfer vector matrix obtained in the step S2 and the text semantic feature vector subjected to ERNIE fine tuning into an attention network module, and outputting a target prediction vector;
(2) inputting the text semantic feature vector and the target prediction vector obtained in the step (1) into a target completion network module to improve the performance of feature representation, and obtaining a target completion enhancement vector as output;
(3) inputting the target prediction vector and the target completion enhancement vector respectively obtained in the step (1) and the step (2) into a target prediction network module to share information between the two vectors, and obtaining a target prediction enhancement vector as output;
(4) setting cycle times, continuously updating the input in the step (2) and the step (3), and taking the enhanced vector as the new input in the step (2) and the step (3) so that a cycle mechanism is arranged between the target prediction network module and the target completion network module;
(5) and inputting the target prediction vector and the target completion vector after cyclic enhancement into a classification module, and obtaining a classification result by respectively passing the two vectors through two different softmax functions.
4. The human-computer interaction-oriented dialog goal planning method of claim 3, characterized in that the goal achievement enhancement vector hc＝fc·acWherein: a iscFor the target completion vector, fcIs an enhancement factor.
5. Human-computer interaction-oriented dialog goal planning method according to claim 4, characterized in that fc＝tanh(Wc·ac+Vp·ap+bc) Wherein: wcAnd VcAre respectively acAnd apWeight matrix of apAs target prediction vector, bcIs a bias vector for the perturbation.
6. The human-computer interaction-oriented dialog goal planning method of claim 3, characterized in that the goal prediction enhancement vector hp＝tanh(fr·hc+fd·ap)·apWherein: a ispIs the target prediction vector, hcCompleting the enhancement vector for the target, fdAs attenuation factor, frIs an enhancement factor.
7. Human-computer interaction-oriented dialog goal planning method according to claim 6, characterized in that the enhancement factor fr＝sigmoid(Wr·hc+Vr·ap+br) Attenuation factor fd＝sigmoid(Wd·hc+Vd·ap+bd) Wherein: wrAnd VrCompleting the enhancement vector h for the targets respectivelycAnd a target prediction vector apEnhanced weight matrix of, WdAnd VdCompleting the enhancement vector h for the targets respectivelycAnd a target prediction vector apAttenuation weight matrix of brAnd bdIs a bias vector for the perturbation.
8. Human-computer interaction-oriented dialog goal planning method according to claim 1 or 3, characterized in that the classification module uses softmax (W concat (a)c,hc) Calculate a probability for each target, wherein: a iscFor the target completion vector, hcCompleting the vector enhancement for the target, concat is the fusion operation of the vectors, and the one with the maximum probability isA target of a current conversation; while using softmax (W. concat (a)p,hp) Calculating the probability of whether each target is completed, if the probability is more than 0.5, the current target of the task is already finished, and the task needs to be transferred to the next target, wherein: a ispIs the target prediction vector, hpAn enhancement vector is predicted for the target.
CN202010888963.5A 2020-08-28 2020-08-28 Man-machine interaction oriented dialogue target planning method Active CN112000793B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010888963.5A CN112000793B (en) 2020-08-28 2020-08-28 Man-machine interaction oriented dialogue target planning method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010888963.5A CN112000793B (en) 2020-08-28 2020-08-28 Man-machine interaction oriented dialogue target planning method

Publications (2)

Publication Number Publication Date
CN112000793A true CN112000793A (en) 2020-11-27
CN112000793B CN112000793B (en) 2022-08-09

Family

ID=73466027

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010888963.5A Active CN112000793B (en) 2020-08-28 2020-08-28 Man-machine interaction oriented dialogue target planning method

Country Status (1)

Country Link
CN (1) CN112000793B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114579892A (en) * 2022-05-06 2022-06-03 南京航空航天大学 User remote access position prediction method based on cross-city interest point matching
CN115186147A (en) * 2022-05-31 2022-10-14 华院计算技术(上海)股份有限公司 Method and device for generating conversation content, storage medium and terminal

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170337036A1 (en) * 2015-03-12 2017-11-23 Kabushiki Kaisha Toshiba Dialogue support apparatus, method and terminal
CN108282587A (en) * 2018-01-19 2018-07-13 重庆邮电大学 Mobile customer service dialogue management method under being oriented to strategy based on status tracking
CN109165284A (en) * 2018-08-22 2019-01-08 重庆邮电大学 A kind of financial field human-computer dialogue intension recognizing method based on big data
CN111241828A (en) * 2020-01-10 2020-06-05 平安科技（深圳）有限公司 Intelligent emotion recognition method and device and computer readable storage medium

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170337036A1 (en) * 2015-03-12 2017-11-23 Kabushiki Kaisha Toshiba Dialogue support apparatus, method and terminal
CN108282587A (en) * 2018-01-19 2018-07-13 重庆邮电大学 Mobile customer service dialogue management method under being oriented to strategy based on status tracking
CN109165284A (en) * 2018-08-22 2019-01-08 重庆邮电大学 A kind of financial field human-computer dialogue intension recognizing method based on big data
CN111241828A (en) * 2020-01-10 2020-06-05 平安科技（深圳）有限公司 Intelligent emotion recognition method and device and computer readable storage medium

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114579892A (en) * 2022-05-06 2022-06-03 南京航空航天大学 User remote access position prediction method based on cross-city interest point matching
CN115186147A (en) * 2022-05-31 2022-10-14 华院计算技术(上海)股份有限公司 Method and device for generating conversation content, storage medium and terminal
CN115186147B (en) * 2022-05-31 2023-08-25 华院计算技术(上海)股份有限公司 Dialogue content generation method and device, storage medium and terminal

Also Published As

Publication number Publication date
CN112000793B (en) 2022-08-09

Similar Documents

Publication Publication Date Title
CN110609891B (en) Visual dialog generation method based on context awareness graph neural network
Hasanbeig et al. Deepsynth: Automata synthesis for automatic task segmentation in deep reinforcement learning
CN109299237B (en) Cyclic network man-machine conversation method based on actor critic reinforcement learning algorithm
CN106445147B (en) The behavior management method and device of conversational system based on artificial intelligence
CN110032630A (en) Talk about art recommendation apparatus, method and model training equipment
Hinoshita et al. Emergence of hierarchical structure mirroring linguistic composition in a recurrent neural network
CN112000793B (en) Man-machine interaction oriented dialogue target planning method
CN107451230A (en) A kind of answering method and question answering system
Delgado et al. A multiobjective genetic algorithm for obtaining the optimal size of a recurrent neural network for grammatical inference
CN112116137A (en) Student class dropping prediction method based on mixed deep neural network
CN114327483A (en) Graph tensor neural network model establishing method and source code semantic identification method
Bryndin Communicative-associative development of smart artificial intelligence by criteria with the help of ensembles of diversified agents
CN110297894B (en) Intelligent dialogue generating method based on auxiliary network
CN108364066A (en) Artificial neural network chip and its application process based on N-GRAM and WFST models
CN115408502A (en) Cognitive learning in a synchronous conference to generate scripts that simulate live agent actions
CN112560440B (en) Syntax dependency method for aspect-level emotion analysis based on deep learning
CN113779988A (en) Method for extracting process knowledge events in communication field
Carcassi et al. The emergence of monotone quantifiers via iterated learning
Xu et al. CNN-based skip-gram method for improving classification accuracy of chinese text
CN110046709A (en) A kind of multi-task learning model based on two-way LSTM
Hasanbeig et al. Symbolic task inference in deep reinforcement learning
Vargas et al. Artificial immune systems as complex adaptive systems
KR20210152979A (en) Learning contents recommendation system based on artificial intelligence learning and operation method thereof
Chetoui et al. Course recommendation model based on Knowledge Graph Embedding
Garciarena et al. Automatic Structural Search for Multi-task Learning VALPs

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant