CN113495943B - Man-machine dialogue method based on knowledge tracking and transferring - Google Patents
Man-machine dialogue method based on knowledge tracking and transferring Download PDFInfo
- Publication number
- CN113495943B CN113495943B CN202010253520.9A CN202010253520A CN113495943B CN 113495943 B CN113495943 B CN 113495943B CN 202010253520 A CN202010253520 A CN 202010253520A CN 113495943 B CN113495943 B CN 113495943B
- Authority
- CN
- China
- Prior art keywords
- knowledge
- tracker
- shi
- posterior
- tracking
- 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.)
- Active
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
- G06F16/33—Querying
- G06F16/3331—Query processing
- G06F16/334—Query execution
- G06F16/3344—Query execution using natural language analysis
-
- 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
- G06N3/0463—Neocognitrons
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Human Computer Interaction (AREA)
- Machine Translation (AREA)
Abstract
The invention discloses a man-machine dialogue method based on knowledge tracking and transferring, which comprises the following steps: step one, constructing a model with knowledge tracking and transferring functions; the model adopts a coding-decoding framework based on deep learning, and comprises a coding layer, a knowledge tracking transfer layer and a decoding layer; training model parameters by using a priori-posterior dual learning mechanism according to the constructed model; and thirdly, after training, fixing all model parameters, and then carrying out actual dialogue application. The method disclosed by the invention further improves the suitability of knowledge selection, and helps the model generate replies with higher user experience; secondly, an unsupervised prior-posterior dual learning mechanism not only enhances interaction between knowledge tracking and transfer, ensures that prediction accuracy of the knowledge tracking and the transfer is improved simultaneously, but also remarkably reduces dependence of a model on manual annotation data.
Description
Technical Field
The invention belongs to the field of intelligent man-machine conversation, and particularly relates to a man-machine conversation method based on knowledge tracking and transferring.
Background
Man-machine conversations, i.e., humans can interact naturally with machines in the form of natural language (i.e., human language). The degree of intelligence of a human-machine dialog system can often be used to measure the extent of development of today's artificial intelligence technology, so building a sufficiently intelligent human-machine dialog model is a long-term goal of the artificial intelligence era. At present, related products of man-machine conversation are gradually applied to human real life, and great convenience is brought to human life.
The dialogue model has many challenges to be solved, and the information of the generated reply of the model is guaranteed. Currently, many research methods promote the "information content" of replies by introducing an external knowledge base (text fragments can be crawled from on-line resources such as the hundred degrees encyclopedia) containing a large number of knowledge text fragments, and this method is a knowledge-based dialogue method. Specifically, knowledge-based dialog methods are to solve two tasks: 1. knowledge selection, namely selecting a piece of knowledge to be chat from a knowledge base according to a dialogue environment; 2. and generating a reply, and generating a final reply according to the selected knowledge piece. Since the selected knowledge content directly determines the topic of the reply content, improper knowledge can directly lead to improper replies, so that the improvement of the knowledge selection suitability of the knowledge-based dialogue method is important.
However, the current mainstream methods still have two significant disadvantages in knowledge selection. First, in terms of model construction, almost all methods take only dialogue context (user's current input and previous dialogue history) to match knowledge in the knowledge base, without explicit modeling knowledge tracking and transfer. Knowledge tracking and transfer is to locate the already chatted knowledge in the dialogue history (knowledge tracking) and then to infer the next chatted knowledge together based on the chatted knowledge and the dialogue context (knowledge transfer). Knowledge tracking and transfer may additionally capture interactions and inferences relationships between already and about to chat knowledge, which may further promote suitability of knowledge selection as compared to using only a single context information.
Secondly, in the aspect of model training, the current mainstream method is data-driven, and knowledge selection and reply generation are all very dependent on large-scale manual annotation data for supervised learning, so that the acquisition cost of the data is very high. However, few studies are currently performed to explore the use of an unsupervised learning method to promote knowledge selection, thereby reducing reliance on annotation data.
Disclosure of Invention
In order to solve the technical problems, the invention provides a man-machine conversation method based on knowledge tracking and transferring, which further improves the suitability of knowledge selection and helps a model to generate a reply with higher user experience; secondly, an unsupervised prior-posterior dual learning mechanism not only enhances interaction between knowledge tracking and transfer, ensures that prediction accuracy of the knowledge tracking and the transfer is improved simultaneously, but also remarkably reduces dependence of a model on manual annotation data.
In order to achieve the above purpose, the technical scheme of the invention is as follows:
a man-machine conversation method based on knowledge tracking and transferring comprises the following steps:
step one, constructing a model with knowledge tracking and transferring functions;
the model adopts a coding-decoding framework based on deep learning, and comprises a coding layer, a knowledge tracking transfer layer and a decoding layer;
training model parameters by using a priori-posterior dual learning mechanism according to the constructed model;
and thirdly, after training, fixing all model parameters, and then carrying out actual dialogue application.
In the above scheme, the coding layer comprises a BERT encoder for encoding the knowledge base and the dialog context into hidden state representations respectively;
the knowledge tracking transfer layer comprises a priori knowledge tracker pri, a knowledge translator shi and a posterior knowledge tracker pos, wherein the priori knowledge tracker pri takes the hidden state representation of the dialogue context as input, predicts a priori knowledge tracking distribution on all text fragments in a knowledge base, and can sample a tracked knowledge (i.e. "chatty knowledge") and the hidden state representation thereof; the knowledge transferor shi takes the tracked knowledge and the hidden state representation of the dialogue context as input together, predicts a knowledge transfer distribution over all text segments in the knowledge base, from which a transferred knowledge (i.e. "knowledge to chat") and its hidden state representation can be sampled; the posterior knowledge tracker pos additionally takes as input a hidden state representation of the transferred knowledge, predicts a posterior knowledge tracking distribution over all text segments in the knowledge base, from which a tracked knowledge and its hidden state representation can be sampled;
the decoding layer includes a transform decoder that uses the transferred knowledge as input along with the hidden state representation of the dialog context to generate a final reply word by word.
In the above scheme, a dual closed loop is formed between the posterior knowledge tracker pos and the knowledge transferor shi, and the posterior knowledge tracker pos is only executed during model training, but not executed during model application.
In the above scheme, in the training process of the second step, the posterior knowledge tracker pos and the knowledge transferer shi are mutually dual tasks, so that the posterior knowledge tracker pos and the knowledge transferer shi are mutually guided and lifted in an unsupervised manner, iterated for a plurality of times until convergence is achieved, and the prior knowledge tracker pri benefits from dual interaction of the posterior knowledge tracker pos and the knowledge transferer shi at the same time; the specific training process is as follows:
step1: training the warming up, namely maximizing the probability of marking data in a training set by using maximum likelihood estimation, and ending the warming up training after the training is performed until parameters are converged;
step2: starting single-round iteration, firstly guiding and lifting a knowledge transferor shi through a posterior knowledge tracker pos;
step3: lifting the posterior knowledge tracker pos through knowledge transferer shi guidance;
step4: using KL divergence loss, enabling the prior knowledge tracking distribution to simulate and approximate the posterior knowledge tracking distribution, and ensuring that the prior knowledge tracker pri can acquire benefits from dual learning even if the prior knowledge tracker pri is not in a closed loop of the dual learning; so far, the single round iteration is ended;
step5: step2-Step4 are repeatedly executed to form a plurality of iterations until the model parameters further converge.
In the above scheme, step three, in the actual dialogue application, given the knowledge base and the dialogue context including the user input, the BERT encoder is firstly executed to obtain the hidden state representations of the knowledge base and the dialogue context, then the prior knowledge tracker pri and the knowledge transferer shi are sequentially executed to complete knowledge tracking and transferring, and finally the transferred knowledge representation pushed by the knowledge transferer shi is fed to the converter decoder to generate the final reply.
Through the technical scheme, the man-machine dialogue method based on knowledge tracking and transferring provided by the invention has the advantages that in the aspect of model construction, knowledge tracking and knowledge transferring are explicitly modeled; in the aspect of model training, an unsupervised prior-posterior dual learning mechanism is used for training model parameters, the learning mechanism regards knowledge tracking and transfer as dual tasks, and the knowledge tracking and transfer are automatically guided by each other in an unsupervised manner, so that the model parameters are lifted together, and a plurality of iterations are performed until convergence. Compared with the prior art, the method has the following beneficial effects:
1. explicit modeling knowledge tracking and knowledge transfer may additionally capture interactions and reasoning relationships between already and about to chat knowledge, and additional cues may further promote suitability of knowledge selection, as compared to using only a single context information, thereby helping the model to generate replies for higher user experiences.
2. The prior-posterior dual learning mechanism enables knowledge tracking and transferring to be guided automatically and mutually in an unsupervised mode, and the knowledge tracking and transferring are promoted jointly. The method not only enhances interaction between knowledge tracking and transfer, ensures that prediction accuracy of the knowledge tracking and the transfer is improved simultaneously, but also obviously reduces dependence of the model on manual annotation data.
3. The invention further distinguishes knowledge tracking into priori knowledge tracking and posterior knowledge tracking, and uses the priori-posterior dual learning mechanism to match and optimize, which successfully solves the problem of dual incompatibility existing in model training and application (the posterior knowledge tracking can not obtain transferred knowledge as input when the model is applied, thus only the priori knowledge tracking can be executed), and ensures that the priori knowledge tracking can acquire benefits from dual learning even if the prior knowledge tracking is not in a closed loop of dual learning.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below.
FIG. 1 is a schematic flow diagram of a human-machine conversation method based on knowledge tracking and transfer according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of the method of the present invention during model training;
FIG. 3 is a schematic diagram of an optimization knowledge translator in a single iteration of the method of the present invention during model training;
FIG. 4 is a schematic diagram of an optimization posterior knowledge tracker in a single iteration of the method of the present invention during model training;
FIG. 5 is a schematic diagram of a priori knowledge tracking distribution approaching posterior knowledge tracking distribution during model training in the method of the present invention;
FIG. 6 is a schematic representation of the method of the present invention when the model is applied.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.
The invention provides a man-machine dialogue method based on knowledge tracking and transferring, which specifically comprises the following steps:
step1: and constructing a model with knowledge tracking and transferring functions.
In the τ -th round of the dialogue, a knowledge base composed of |K| text segments is given(text segment K) i From |K i Phrase composition), given dialog context C τ =(X τ-1 ,Y τ-1 ,X τ ) (X is the user's input and Y is the model's reply, where interaction records for the user at τ -1 round and the model's input at τ are defined as contexts), the task of the model is to first select an appropriate text segment from the knowledge base KThen generate a text segment from the text segment τ Reply of word composition
At the coding layer, for dialog context C τ Using BERT base Encoder and averaging pooling operations to obtain its hidden state representation
Where d represents the hidden state dimension of 768 dimensions and p represents the average pooling operation. Knowledge base is built by the same methodCoded as hidden state representation +.>
At the knowledge tracking and transfer layer, the a priori knowledge tracker pri represents the hidden state of the dialog context of τ -1 roundAs input, a priori knowledge tracking distribution P (k|pri) on the knowledge base K is predicted:
wherein Mlp (·) =.w+b represents the multi-layer perceptron (Multilayer Perceptron), W and b are trainable parameters. [ ·; and (4) representing vector splicing operation, and T representing matrix transposition operation.
Knowledge transferor shi presents the hidden state of the dialog contextAnd hidden state representation of the tracked knowledgeAs input, a knowledge transfer distribution P (K|K) on a knowledge base K is predicted tra ,shi):
The posterior knowledge tracker pos represents the hidden state of the tau-1 round of dialog contextAnd the implicit state representation of the knowledge transferred +.>As input, a posterior knowledge tracking profile P (K|K) on the knowledge base K is predicted shi ,pos):
Generating Y word by word at the decoding layer using a transducer as a decoder τ . The decoder will talk context X τ And knowledge K transferred to shi Corresponding unadjusted pooled hidden state representationAnd->As input. During model training, K shi The transferred knowledge marked in the training set; k when the model is applied shi The knowledge transfer distribution P (k|k) predicted for knowledge transfer shi tra Shi) the most probable knowledge text segment.
Since the decoding process is a multi-time-step cyclic process, the decoder will be described in detail below to generate the word y at the t decoding time step τ,t Is a detailed procedure of (2). Given the hidden state vector of the t-th decoding time stepWherein emb (·) represents the word embedding representation of the fetch · h is expressed as τ,t Mapping to a predefined vocabulary v= { V 1 ,v 2 ,…,v |V| On }, a probability distribution P (y) τ,t ):
P(y τ,t )=softmax(Mlp(hτ ,t ))∈R |V| , (13)
At the time of model application, P (y τ,t ) Word v with maximum corresponding probability in distribution as generated word y of t-th time step τ,t . So far, the calculation of the t time step is finished, and the decoder is updated to obtain the decoder hidden state h tau of the t+1st time step ,t+1 A new cycle is started:
when the model is applied, after the decoding cycle is finished, the words output by each step sequentially form a complete final reply Y τ 。
Step2: model parameters are trained using a priori-posterior dual learning mechanism from the model obtained in step1 using knowledge-based dialogue datasets currently disclosed in industry and academia. Specifically, the training mechanism makes the posterior knowledge tracker pos and the knowledge translator shi be dual tasks, makes the posterior knowledge tracker pos and the knowledge translator shi guide and promote each other in an unsupervised manner, makes the posterior knowledge tracker pos and the knowledge translator shi iterate until convergence (in each iteration, the posterior knowledge tracker pos and the knowledge translator shi are alternately optimized), and ensures that the prior knowledge tracker pri can benefit from dual interactions. Figure 2 shows a schematic representation of the method of the invention during model training.
This step is achieved by the following procedure.
Step1: the goal of the warm-up training is to maximize the probability of labeling data in the training set using maximum likelihood estimation (MLE for short), where the following loss functions are defined:
L pri (θ)=-logP(K tra_label |pri), (15)
L pos (θ)=-logP(K tra_label |K shi_label ,pos), (16)
L shi (θ)=-logP(K shi_label |K tra_label ,shi), (17)
wherein θ is all trainable parameters in the model, and tra_label and shi_label correspond to the tracked knowledge and the transferred knowledge marked in the training set respectively. L (L) pri (θ) is a priori knowledge tracking loss, L pos (θ) is posterior knowledge tracking loss, L shi (θ) is knowledge transfer loss, L g And (θ) is the recovery generation loss.
Obtaining a final loss function L (theta):
L(θ)=L pri (θ)+L pos (θ)+L shi (θ)+L g (θ). (19)
all parameters of the model and the word embedding matrix are then updated with a back propagation algorithm commonly used for deep learning (abbreviated as BP algorithm) to reduce the loss. After the training is converged, the warm-up training phase is ended.
Step2: a single iteration starts. A training sample is sampled from the training set, and then a posterior knowledge tracker pos is used for guiding and lifting the knowledge transferor shi. As shown in FIG. 3, the tracked knowledge of the annotations is representedAs input to the knowledge transferor shi, the knowledge transferor shi generates a knowledge transfer profile P (k|k tra_label Shi). Sampling a knowledge from the distribution as transferred knowledge K shi_sample And express it +.>Feeding a posterior knowledge tracker pos to obtain annotated tracked knowledge K tra_label Is the inverse probability P (K) tra_label |K shi_sample Pos), consider the thrust back probability as a "reward":
E(R)=E[RlogP(K shi_sample |K tra_label ,shi)], (20)
R=log[P(K tra_label |K shi_sample ,pos)], (21)
wherein R is the output distribution of the secondary shiP(K|K tra_label Knowledge K of the transfer of samples in shi) shi_sample Is a "reward" of (a). E [. Cndot.]To achieve the desired result. Subsequently, the strategy gradient (policy gradient) method is used to maximize the desired E (R) of the reward and calculate the parameter θ 1 =[θ embedding ,θ encoder ,θ shi ](θ embedding For word embedding matrix, θ encoder θ, the parameter of the encoder shi Parameters for knowledge transferor shi):
Step3: similar to Step2, the posterior knowledge tracker pos will be lifted in a similar manner by the knowledge transferor shi, as shown in FIG. 4, to finally optimize the parameter θ 2 =[θ embedding ,θ encoder ,θ pos ](θ pos Parameters for a posterior knowledge tracker pos).
Step4: after the above-mentioned dual optimization of the knowledge transferer shi and the a priori knowledge tracker pos, the a priori knowledge tracker pos needs to be transferred from the revenue in the process to the a priori knowledge tracker pri. As shown in FIG. 5, the knowledge K to which the annotations in the training set are transferred shi_label Latent state representation of (1)As input to the already optimized posterior tracker pos, a posterior tracking profile P (k|k) shi_label Pos). Causing the distribution P (K|pri) to approach P (K|K) shi_label Pos), loss of divergence by KL: />
Because training is done in an unsupervised manner, inaccurate "rewards" are difficult to avoid. To mitigate the adverse effects of this phenomenon, KL divergence loss L is reduced kl (θ) and MLE loss [ L ] pos (θ),L shi (θ),L g (θ)]Linear summation for joint training:
L(θ)=L kl (θ)+γ[L pos (θ)+L shi (θ)+L g (θ)] (24)
where γ is a hyper-parameter that acts to control the proportion of MLE loss. The setting of γ in this method is 0.5. So far, a single iteration ends.
Step5: step2-Step4 are repeatedly executed to form a plurality of iterations until the model parameters further converge. Convergence is the end of the whole training process.
Step3: actual dialog applications.
After model training is completed, parameters of the model are all fixed. At this point, the model can be applied to the actual dialog scene.
As shown in fig. 6, a given knowledge base K is associated with a dialog context C containing user input τ The encoder is executed to acquire hidden state representations of the two, the prior knowledge tracker pri and the knowledge transferer shi are executed in sequence to complete knowledge tracking and transferring, and finally the transferred knowledge representation deduced by the knowledge transferer shi is fed to the decoder to generate a final reply Y τ 。
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (3)
1. A man-machine conversation method based on knowledge tracking and transferring is characterized by comprising the following steps:
step one, constructing a model with knowledge tracking and transferring functions;
the model adopts a coding-decoding framework based on deep learning, and comprises a coding layer, a knowledge tracking transfer layer and a decoding layer;
training model parameters by using a priori-posterior dual learning mechanism according to the constructed model;
step three, after training is completed, all model parameters are fixed, and then actual dialogue application is carried out;
the coding layer comprises a BERT coder which respectively codes the knowledge base and the dialogue context into a hidden state representation;
the knowledge tracking transfer layer comprises a priori knowledge tracker pri, a knowledge translator shi and a posterior knowledge tracker pos, wherein the priori knowledge tracker pri takes the hidden state representation of the dialogue context as input, predicts a priori knowledge tracking distribution on all text fragments in a knowledge base, and can sample a tracked knowledge and the hidden state representation thereof from the distribution; the knowledge transferor shi takes the tracked knowledge and the hidden state representation of the dialogue context as input together, predicts a knowledge transfer distribution over all text segments in the knowledge base, from which a transferred knowledge and its hidden state representation can be sampled; the posterior knowledge tracker pos additionally takes as input a hidden state representation of the transferred knowledge, predicts a posterior knowledge tracking distribution over all text segments in the knowledge base, from which a tracked knowledge and its hidden state representation can be sampled;
the decoding layer comprises a transform decoder which takes the transferred knowledge and the hidden state representation of the dialogue context as inputs together, and generates a final reply word by word;
in the training process of the second step, the posterior knowledge tracker pos and the knowledge transferer shi are mutually dual tasks, so that the posterior knowledge tracker pos and the knowledge transferer shi are mutually guided and lifted in an unsupervised mode, multiple iterations are performed until convergence is achieved, and the prior knowledge tracker pri benefits from dual interaction of the posterior knowledge tracker pos and the knowledge transferer shi at the same time; the specific training process is as follows:
step1: training the warming up, namely maximizing the probability of marking data in a training set by using maximum likelihood estimation, and ending the warming up training after the training is performed until parameters are converged;
step2: starting single-round iteration, firstly guiding and lifting a knowledge transferor shi through a posterior knowledge tracker pos;
step3: lifting the posterior knowledge tracker pos through knowledge transferer shi guidance;
step4: performing joint training on the linear sum of KL divergence loss and MLE loss, so that the prior knowledge tracking distribution is simulated and approximates to posterior knowledge tracking distribution; so far, the single round iteration is ended;
step5: step2-Step4 are repeatedly executed to form a plurality of iterations until the model parameters further converge.
2. A human-machine conversation method based on knowledge tracking and transferring as claimed in claim 1 wherein a dual loop is formed between a posterior knowledge tracker pos and a knowledge transferor shi, and the posterior knowledge tracker pos is only executed during model training and is not executed during model application.
3. The method according to claim 1, wherein, in the actual dialogue application, given the knowledge base and the dialogue context including the user input, the BERT encoder is firstly executed to obtain the hidden state representation of the knowledge base and the dialogue context, then the prior knowledge tracker pri and the knowledge transferor shi are sequentially executed to complete the knowledge tracking and transferring, and finally the transferred knowledge pushed by the knowledge transferor shi is fed to the transferor decoder to generate the final reply.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010253520.9A CN113495943B (en) | 2020-04-02 | 2020-04-02 | Man-machine dialogue method based on knowledge tracking and transferring |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010253520.9A CN113495943B (en) | 2020-04-02 | 2020-04-02 | Man-machine dialogue method based on knowledge tracking and transferring |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113495943A CN113495943A (en) | 2021-10-12 |
CN113495943B true CN113495943B (en) | 2023-07-14 |
Family
ID=77994291
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010253520.9A Active CN113495943B (en) | 2020-04-02 | 2020-04-02 | Man-machine dialogue method based on knowledge tracking and transferring |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113495943B (en) |
Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106599196A (en) * | 2016-12-14 | 2017-04-26 | 竹间智能科技(上海)有限公司 | Artificial intelligence conversation method and system |
WO2018133761A1 (en) * | 2017-01-17 | 2018-07-26 | 华为技术有限公司 | Method and device for man-machine dialogue |
WO2018157700A1 (en) * | 2017-03-02 | 2018-09-07 | 腾讯科技(深圳)有限公司 | Method and device for generating dialogue, and storage medium |
CN108763504A (en) * | 2018-05-30 | 2018-11-06 | 浙江大学 | It is a kind of that generation method and system are replied based on the dialogue for strengthening binary channels Sequence Learning |
CN109657041A (en) * | 2018-12-04 | 2019-04-19 | 南京理工大学 | The problem of based on deep learning automatic generation method |
CN109887484A (en) * | 2019-02-22 | 2019-06-14 | 平安科技(深圳)有限公司 | A kind of speech recognition based on paired-associate learning and phoneme synthesizing method and device |
CN110188182A (en) * | 2019-05-31 | 2019-08-30 | 中国科学院深圳先进技术研究院 | Model training method, dialogue generation method, device, equipment and medium |
CN110297895A (en) * | 2019-05-24 | 2019-10-01 | 山东大学 | A kind of dialogue method and system based on free text knowledge |
CN110297817A (en) * | 2019-06-25 | 2019-10-01 | 哈尔滨工业大学 | A method of the structure of knowledge is constructed based on personalized Bayes's knowledge tracing model |
CN110309170A (en) * | 2019-07-02 | 2019-10-08 | 北京大学 | A kind of Task takes turns the complicated intension recognizing method in dialogue more |
CN110807509A (en) * | 2018-08-06 | 2020-02-18 | 北京博智天下信息技术有限公司 | Depth knowledge tracking method based on Bayesian neural network |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107885756B (en) * | 2016-09-30 | 2020-05-08 | 华为技术有限公司 | Deep learning-based dialogue method, device and equipment |
-
2020
- 2020-04-02 CN CN202010253520.9A patent/CN113495943B/en active Active
Patent Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106599196A (en) * | 2016-12-14 | 2017-04-26 | 竹间智能科技(上海)有限公司 | Artificial intelligence conversation method and system |
WO2018133761A1 (en) * | 2017-01-17 | 2018-07-26 | 华为技术有限公司 | Method and device for man-machine dialogue |
WO2018157700A1 (en) * | 2017-03-02 | 2018-09-07 | 腾讯科技(深圳)有限公司 | Method and device for generating dialogue, and storage medium |
CN108763504A (en) * | 2018-05-30 | 2018-11-06 | 浙江大学 | It is a kind of that generation method and system are replied based on the dialogue for strengthening binary channels Sequence Learning |
CN110807509A (en) * | 2018-08-06 | 2020-02-18 | 北京博智天下信息技术有限公司 | Depth knowledge tracking method based on Bayesian neural network |
CN109657041A (en) * | 2018-12-04 | 2019-04-19 | 南京理工大学 | The problem of based on deep learning automatic generation method |
CN109887484A (en) * | 2019-02-22 | 2019-06-14 | 平安科技(深圳)有限公司 | A kind of speech recognition based on paired-associate learning and phoneme synthesizing method and device |
CN110297895A (en) * | 2019-05-24 | 2019-10-01 | 山东大学 | A kind of dialogue method and system based on free text knowledge |
CN110188182A (en) * | 2019-05-31 | 2019-08-30 | 中国科学院深圳先进技术研究院 | Model training method, dialogue generation method, device, equipment and medium |
CN110297817A (en) * | 2019-06-25 | 2019-10-01 | 哈尔滨工业大学 | A method of the structure of knowledge is constructed based on personalized Bayes's knowledge tracing model |
CN110309170A (en) * | 2019-07-02 | 2019-10-08 | 北京大学 | A kind of Task takes turns the complicated intension recognizing method in dialogue more |
Non-Patent Citations (3)
Title |
---|
基于深度学习的个性化对话内容生成方法;王豪;郭斌;郝少阳;张秋韵;於志文;;图学学报(第02期);全文 * |
基于统计学习的自然语言对话***的设计与实现;李伟通;皮德常;;微计算机应用(第07期);全文 * |
多知识点知识追踪模型与可视化研究;徐墨客;吴文峻;周萱;蒲彦均;;电化教育研究(第10期);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN113495943A (en) | 2021-10-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110188182B (en) | Model training method, dialogue generating method, device, equipment and medium | |
Liu et al. | Dialogue learning with human teaching and feedback in end-to-end trainable task-oriented dialogue systems | |
CN111666427B (en) | Entity relationship joint extraction method, device, equipment and medium | |
CN108920468B (en) | A kind of bilingual kind of inter-translation method of illiteracy Chinese based on intensified learning | |
CN111401084B (en) | Method and device for machine translation and computer readable storage medium | |
Liu et al. | Adversarial learning of task-oriented neural dialog models | |
CN112541060B (en) | End-to-end task type dialogue learning framework and method based on confrontation training | |
CN110162766B (en) | Word vector updating method and device | |
CN110929114A (en) | Tracking digital dialog states and generating responses using dynamic memory networks | |
US11475225B2 (en) | Method, system, electronic device and storage medium for clarification question generation | |
CN112380863A (en) | Sequence labeling method based on multi-head self-attention mechanism | |
CN111368545A (en) | Named entity identification method and device based on multi-task learning | |
CN114443828B (en) | Training method and device for universal dialogue model, electronic equipment and medium | |
CN114168707A (en) | Recommendation-oriented emotion type conversation method | |
CN115906815B (en) | Error correction method and device for modifying one or more types of error sentences | |
CN113971837A (en) | Knowledge-based multi-modal feature fusion dynamic graph neural sign language translation method | |
CN113887471A (en) | Video time sequence positioning method based on feature decoupling and cross comparison | |
CN114579699A (en) | Training method and device for pre-training language model | |
CN114528387A (en) | Deep learning conversation strategy model construction method and system based on conversation flow bootstrap | |
CN116932722A (en) | Cross-modal data fusion-based medical visual question-answering method and system | |
CN114757188A (en) | Standard medical text rewriting method based on generation of confrontation network | |
CN113495943B (en) | Man-machine dialogue method based on knowledge tracking and transferring | |
CN113297374A (en) | Text classification method based on BERT and word feature fusion | |
CN112183062A (en) | Spoken language understanding method based on alternate decoding, electronic equipment and storage medium | |
JP7520085B2 (en) | Text error correction and text error correction model generation method, device, equipment, and medium |
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 |