CN113887643B - New dialogue intention recognition method based on pseudo tag self-training and source domain retraining - Google Patents

New dialogue intention recognition method based on pseudo tag self-training and source domain retraining Download PDF

Info

Publication number
CN113887643B
CN113887643B CN202111187641.9A CN202111187641A CN113887643B CN 113887643 B CN113887643 B CN 113887643B CN 202111187641 A CN202111187641 A CN 202111187641A CN 113887643 B CN113887643 B CN 113887643B
Authority
CN
China
Prior art keywords
model
training
pseudo
self
retraining
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
Application number
CN202111187641.9A
Other languages
Chinese (zh)
Other versions
CN113887643A (en
Inventor
田锋
安文斌
郑庆华
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.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong University
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 Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN202111187641.9A priority Critical patent/CN113887643B/en
Publication of CN113887643A publication Critical patent/CN113887643A/en
Application granted granted Critical
Publication of CN113887643B publication Critical patent/CN113887643B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • 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
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • 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
    • 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

Abstract

The invention discloses a new dialogue intention recognition method based on pseudo tag self-training and source domain retraining, and belongs to the technical field of language processing. According to the new dialogue intention recognition method based on the pseudo tag self-training and the source domain retraining, the pseudo tag is generated for unlabeled data containing the new dialogue intention, and model parameters are iteratively updated by using the self-training method, so that the recognition accuracy is continuously improved; meanwhile, a retraining strategy is provided, so that knowledge can be better migrated between a source domain and a target domain, and the expression capacity of the model is improved; finally, the invention integrates the output of the three models for integrated learning, and improves the robustness of the models.

Description

New dialogue intention recognition method based on pseudo tag self-training and source domain retraining
Technical Field
The invention belongs to the technical field of language processing, and particularly relates to a new dialogue intention recognition method based on pseudo tag self-training and source domain retraining.
Background
The core module of the intelligent dialog system is user intention recognition. New dialog intention recognition aims at finding new generated dialog intention based on existing dialog intention, and only uses a small amount of marked known intention data to find and classify new intention from a large amount of unmarked data. Because the data containing the new intention is all unlabeled data, the existing dialog intention classification model cannot process the data, thereby causing user intention recognition errors and affecting subsequent responses of the intelligent dialog system.
In order to solve the above problems, two methods are mainly adopted in the current academia: 1. a self-adaptive clustering model based on contrast similarity is provided on the basis of a contrast learning method, such as Lin and the like, and high-quality data are selected for self-marking and training through the contrast learning method. 2. A new dialogue intention discovery model based on depth alignment clustering is proposed by a depth clustering-based method, such as Zhang, and pseudo labels are generated for unlabeled data through clustering and alignment operation and model training is carried out. The technical scheme has the following defects: first, the granularity of pseudo labels generated by the existing model for unlabeled data is coarse, and the model cannot be well trained to find out new dialogue intention. Second, the existing model only uses the labeling data to initialize the model, but fails to fully use the labeling data in the training process, making knowledge migration difficult. Thirdly, the existing model only utilizes a clustering model to generate final prediction, and integrated learning is not carried out by fusing other models, so that the robustness of the model is reduced.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide a new dialog intention recognition method based on pseudo tag self-training and source domain retraining.
In order to achieve the purpose, the invention is realized by adopting the following technical scheme:
a new dialog intention recognition method based on pseudo tag self-training and source domain retraining comprises the following steps:
1. model training
1) Feature extraction is performed on the input by using a model Net1 and a model Net2 to respectively obtain vector representations of the ith inputAnd->For->And->Weighted combination is performed to obtain the final representation of the i-th input +.>
The model Net1 and the model Net2 are BERT models with the same structure;
2) Final vector representation of input using cluster model KmeansClustering to obtain pseudo tags of the category to which each sample belongs>
Vector representation of input using Softmax classifierAnd->Respectively performing classification operation to obtain two other groups of pseudo tags ++>And->
For a pair ofAnd->Performing linear transformation to obtain predictive probability distribution vector of the model on the ith input>And
calculating predictive probability distribution vector by cross entropy loss by adopting self-training methodAnd pseudo tag->Is predictive of the difference of probability distribution vector +.>And pseudo tag->Is a predictive probability distribution vector>With pseudo tagsSumming the losses of all input samples to obtain a loss value +.>For loss valueWeighting combination is carried out to obtain final loss value of pseudo tag self-training +.>
3) Retraining model Net1 and model Net2 using labeled data, computing predictive probability distribution vectors via cross entropy loss functionsAnd->And (3) true label->The difference between them to obtain the loss value with labeling training
4) For a pair ofAnd->Weighting and combining to obtain final loss value +.>Continuously updating model parameters through back propagation, so that the difference between the prediction and the true value of the model is within a preset range;
5) And respectively inputting the dialogue text to be classified into a model Kmeans, a model Net1 and a model Net2 for label prediction, carrying out weighted combination on the obtained prediction labels, and finally classifying the combined result by using a Softmax classifier.
Further, the specific operation of feature extraction in step 1) is:
for the ith input text s i Feature extraction is performed by using a model Net1 and a model Net2 respectively to obtain vector representationAnd->
Where MeanPooling represents an average pooling of the final output of the BERT model,and->For hidden layer representation of the input, σ represents the ReLU activation function，W a ,b a Is a learnable parameter.
Further, the final representation is found in step 1)The operation of (1) is as follows:
for s i Is weighted and combined to obtain s i Final vector representation of (a)
Wherein lambda is f The super-parameters are preset values and are used for balancing the weights occupied by the two vector representations.
Further, the specific operation of generating the pseudo tag in the step 2) is as follows:
will s i Final vector representation of (a)Inputting into a cluster model Kmeans to obtain pseudo tags of corresponding categories
Wherein mu j For the vector representation of the center of the j-th cluster in the clustering process,in order to indicate the function, I.I. | 2 Is the Euclidean distance;
input vectors extracted from model Net1 and model Net2 using Softmax classifierAnd->Performing classification operation to obtain two other groups of pseudo tags belonging to each sample>And->
Input vector by a linear layer pairAnd->Performing linear transformation to obtain predictive probability distribution vector +.f of model Net1 and model Net2 for ith input>Is->
Wherein W is c ,b c ,W m ,b m Are all learnable parameters.
Further, the self-training of step 2) results in a loss valueThe specific operation of (a) is as follows:
by a self-training method, respectively calculating prediction probability distribution vectors by using a cross entropy loss functionAnd pseudo tag->Is predictive of the difference of probability distribution vector +.>And pseudo tag->Is a predictive probability distribution vector>And pseudo tag->Summing the losses of all input samples to obtain a loss value +.>
Wherein N is the number of samples, exp () is an exponential function;
for loss valueWeighting combination is carried out to obtain final loss value of pseudo tag self-training +.>
Wherein lambda is s1 And lambda (lambda) s2 All are super parameters, and are preset values.
Further, in step 3), a loss value with labeling training is obtainedThe specific operation of (a) is as follows:
predicting probability distribution vectors by cross entropy loss function calculation model Net1And (3) true label->The difference between them is summed up for all input samples to obtain a loss value +.>
Prediction probability distribution vector through cross entropy loss function calculation model Net2And (3) true label->The difference between them is summed up for all input samples to obtain a loss value +.>
Wherein M is the number of marked data;
loss by retraining model Net1 and model Net2 on labeled dataAnd->Weighting and combining to obtain the total loss value +.>
Wherein lambda is l1 Is a super parameter and is a preset value.
Further, in step 4), the specific operation of updating the model parameters using back propagation is:
self-training loss values for pseudo tagsAnd with a labeling training loss value->Weighting and combining to obtain the overall loss of the model>
Wherein lambda is t Is super-parameter, is a preset value and is used for balancing lossAnd->The weight of the vehicle is occupied;
obtaining the integral lossAnd then, continuously updating parameters of the model Net1 and the model Net2 by using a back propagation algorithm, so that the difference between the prediction and the true value of the model is within a preset range.
Further, the specific operation of predicting the input category when the model is used is as follows:
respectively inputting the dialogue text to be classified into a trained model Net1, a trained model Net2 and a clustering model Kmeans to respectively obtain a prediction label y 1 ，y 2 Y 3 The obtained three prediction labels are weighted and combined to obtain a final prediction label y vote ：
y vote ＝λ y1 y 1 +λ y2 y 2 +(1-λ y1 -λ y2 )y 3 (20)
Prediction tag y using Softmax classifier pair vote Classifying, namely taking the maximum value y of the classification as the final prediction category:
y＝max(Softmax(y vote )) (21)
compared with the prior art, the invention has the following beneficial effects:
according to the new dialogue intention recognition method based on the pseudo tag self-training and the source domain retraining, the pseudo tag is generated for unlabeled data containing the new dialogue intention, and model parameters are iteratively updated by using the self-training method, so that the recognition accuracy is continuously improved; meanwhile, a retraining strategy is provided, so that knowledge can be better migrated between a source domain and a target domain, and the expression capacity of the model is improved; finally, the invention integrates the output of the three models for integrated learning, and improves the robustness of the models.
Drawings
Fig. 1 is a diagram showing the overall network configuration in the present invention.
Detailed Description
In order that those skilled in the art will better understand the present invention, a technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in which it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present invention without making any inventive effort, shall fall within the scope of the present invention.
It should be noted that the terms "first," "second," and the like in the description and the claims of the present invention and the above figures are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate such that the embodiments of the invention described herein may be implemented in sequences other than those illustrated or otherwise described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.
In order to solve the problems in the prior art, the invention provides a new dialog intention recognition method based on self-training of a pseudo tag and retraining of a source domain, which generates the pseudo tag for unlabeled data containing new dialog intention, and iteratively updates model parameters by using the self-training method, so that the recognition accuracy is continuously improved; meanwhile, the invention provides a retraining strategy, so that knowledge can be better migrated between a source domain and a target domain, thereby improving the expression capacity of the model; finally, the invention integrates the output of the three models for integrated learning, and improves the robustness of the models.
The invention is described in further detail below with reference to the attached drawing figures:
referring to fig. 1, fig. 1 is a schematic diagram of a network model of the present invention, which includes a cluster model Kmeans and two BERT models Net1, net2. The three models can generate different pseudo tags, and the obtained pseudo tags are utilized for self-training, so that the performance of the models is continuously improved. And after training, combining the outputs of the three models to obtain a final predicted value.
1. Model training:
step 1: feature extraction of the input using two structurally identical pre-trained models BERT (model Net1 and model Net 2) to obtain vector representations of the ith input, respectivelyAnd->
Where MeanPooling represents an average pooling of the final output of the BERT model,and->For hidden layer representation of the input, σ represents the ReLU activation function, W a ,b a Is a learnable parameter.
Thereafter, for s i Is represented by two vectors of (a)Weighted combination is carried out to obtain s i Final vector representation of (a)
Wherein lambda is f The super-parameters are set in advance and used for balancing the weights occupied by the two vector representations.
Step 2: final vector representation of input using cluster model KmeansClustering to obtain pseudo tags of the category to which each sample belongs>
Will s i Final vector representation of (a)Inputting into a cluster model Kmeans to obtain pseudo tags of corresponding categories
Wherein mu j For the vector representation of the center of the j-th cluster in the clustering process,in order to indicate the function, I.I. | 2 Is the Euclidean distance.
Step 3: vector representation of input using Softmax classifierAnd->Respectively performing classification operation to obtain two other groups of pseudo tags ++>And->
Thereafter, by the pair ofAnd->Performing linear transformation to obtain predictive probability distribution vector of the model on the ith input>And->
Step 4: by a self-training method, respectively calculating prediction probability distribution vectors by using a cross entropy loss functionAnd pseudo tag->Is predictive of the difference of probability distribution vector +.>And pseudo tag->Is a difference of (1) and a predictive probability distribution vectorAnd pseudo tag->And summing all input samples to obtain a loss value +.>
Where N is the number of samples and exp () is an exponential function.
Then, for the loss valueWeighting and combining to obtain final loss value of pseudo tag self-training
Wherein lambda is s1 And lambda (lambda) s2 All are super parameters, and need to be set in advance for balancing the weights occupied by the three losses.
Step 5: retraining model Net1 and model Net2 using labeled data, calculating model predictive probability distribution vectors via cross entropy loss functionsAnd->And (3) true label->The difference between them, a loss value of +.>And->
Wherein M is the number of marked data.
Then, model Net1 and model Net2 are retrained with the loss obtained by the labeling dataAnd->Weighting and combining to obtain the total loss value +.>
Wherein lambda is l1 Is super-parameter and needs to be set in advance for balancing the lossAnd->The weight occupied.
Step 6: self-training penalty for pseudo tagsAnd->Weighted combination is carried out to obtain a final loss value
Wherein lambda is t Is super-parameter and needs to be set in advance for balancing the lossAnd->The weight occupied.
Step 7: obtaining the integral lossAnd then, continuously updating parameters of the model Net1 and the model Net2 by using a back propagation algorithm, so that the predicted value of the model gradually approaches to the true value.
Step 8: model training
The gradient is updated by using an Adam optimizer, the learning rate is set to 0.0001, the first-order momentum parameter of Adam is set to 0.1, the second-order momentum parameter is set to 0.999, the number of training iterations (epochs) of the data set is set to 100, the parameters of the pre-trained BERT model are fixed, and the super-parameters are selected according to the representation of the model on the model validation set.
2. Model use
Respectively inputting the dialogue text to be classified into a trained model Net1, a trained model Net2 and a clustering model Kmeans to respectively obtain a prediction label y 1 ，y 2 Y 3 The obtained three prediction labels are weighted and combined to obtain a final prediction label y vote ：
y vote ＝λ y1 y 1 +λ y2 y 2 +(1-λ y1 -λ y2 )y 3 (20)
Prediction tag y using Softmax classifier pair vote Classifying, namely taking the maximum value y of the classification as the final prediction category:
y＝max(Softmax(y vote )) (21)
to measure model performance, a comparative test was performed on three widely used public intent recognition datasets, training sets of datasets, test set divisions and the number and lexicon sizes of known intent categories as shown in Table 1. Table 2 shows the results of the comparison experiment, compared with fifteen common models, the index accuracy (Acc) and ARI values and NMI values, and it can be seen from the table that the model PTRN of the present invention has the best results on all indexes of all data sets, and has a larger improvement in performance compared with the current best method.
Table 1 statistics of data sets used to measure model performance
Table 2 shows the accuracy (Acc), ARI and NMI values of the comparative model over different data sets, wherein PTRN is the method of the present invention.
Table 2 performance of comparative models on different data sets
The above is only for illustrating the technical idea of the present invention, and the protection scope of the present invention is not limited by this, and any modification made on the basis of the technical scheme according to the technical idea of the present invention falls within the protection scope of the claims of the present invention.

Claims (8)

1. A new dialogue intention recognition method based on pseudo tag self-training and source domain retraining is characterized by comprising the following steps:
1) Feature extraction is performed on the input by using a model Net1 and a model Net2 to respectively obtain vector representations of the ith inputAnd->For->And->Weighted combination is performed to obtain the final representation of the i-th input +.>
The model Net1 and the model Net2 are BERT models with the same structure;
2) Final vector representation of input using cluster model KmeansClustering operation is carried out to obtain pseudo tags of the category to which each sample belongs>
Vector representation of input using Softmax classifierAnd->Respectively performing classification operation to obtain two other groups of pseudo tags belonging to each sample>And->
For a pair ofAnd->Performing linear transformation to obtain predictive probability distribution vector of the model on the ith input>And->
Calculating predictive probability distribution vector by cross entropy loss by adopting self-training methodAnd pseudo tag->Is predictive of the difference of probability distribution vector +.>And pseudo tag->Is a predictive probability distribution vector>And pseudo tag->Summing the losses of all input samples to obtain a loss value +.>For loss value->Weighting combination is carried out to obtain final loss value of pseudo tag self-training +.>
3) Retraining model Net1 and model Net2 using labeled data, computing predictive probability distribution vectors via cross entropy loss functionsAnd->And (3) true label->The difference between them, a loss value of +.>
4) For a pair ofAnd->Weighting and combining to obtain final loss value +.>Continuously updating model parameters through back propagation, so that the difference between the prediction and the true value of the model is within a preset range;
5) And respectively inputting the dialogue text to be classified into a model Kmeans, a model Net1 and a model Net2 for label prediction, carrying out weighted combination on the obtained prediction labels, and finally classifying the combined result by using a Softmax classifier.
2. The new dialog intention recognition method based on pseudo tag self-training and source domain retraining of claim 1, wherein the feature extraction in step 1) is specifically performed as follows:
for the ith input text s i Feature extraction is performed by using a model Net1 and a model Net2 respectively to obtain vector representationAnd->
Where MeanPooling represents an average pooling of the final output of the BERT model,and->For hidden layer representation of the input, σ represents the ReLU activation function, W a ,b a Is a learnable parameter.
3. The new dialog intention recognition method based on pseudo-tag self-training and source domain retraining of claim 2, wherein the final representation is found in step 1)The operation of (1) is as follows:
for s i Is weighted and combined to obtain s i Final vector representation of (a)
Wherein lambda is f The super-parameters are preset values and are used for balancing the weights occupied by the two vector representations.
4. The new dialog intention recognition method based on pseudo tag self-training and source domain retraining of claim 1, wherein the specific operations of pseudo tag generation of step 2) are:
will s i Final vector representation of (a)Inputting into a cluster model Kmeans to obtain pseudo tags of corresponding categories>
Wherein mu j For the vector representation of the center of the j-th cluster in the clustering process,in order to indicate the function, I.I. | 2 Is the Euclidean distance;
input vectors extracted from model Net1 and model Net2 using Softmax classifierAnd->Performing classification operation to obtain two other groups of pseudo tags belonging to each sample>And->
Input vector by a linear layer pairAnd->Performing linear transformation to obtain predictive probability distribution vector +.f of model Net1 and model Net2 for ith input>Is->
Wherein W is c ，b c ，W m ，b m Are all learnable parameters.
5. The new dialog intention recognition method based on pseudo-tag self-training and source domain retraining of claim 4, wherein the self-training of step 2) results in a loss valueThe specific operation of (a) is as follows:
by a self-training method, respectively calculating prediction probability distribution vectors by using a cross entropy loss functionWith pseudo tagsIs predictive of the difference of probability distribution vector +.>And pseudo tag->Is a predictive probability distribution vector>And pseudo tag->Summing the losses of all input samples to obtain a loss value +.>
Wherein N is the number of samples, exp () is an exponential function;
for loss valueWeighting combination is carried out to obtain final loss value of pseudo tag self-training +.>
Wherein lambda is s1 And lambda (lambda) s2 All are super parameters, and are preset values.
6. The new dialog intention recognition method based on pseudo-tag self-training and source domain retraining as claimed in claim 1, wherein in step 3) a trained penalty value is obtainedThe specific operation of (a) is as follows:
predicting probability distribution vectors by cross entropy loss function calculation model Net1And (3) true label->The difference between them is summed up for all input samples to obtain a loss value +.>
Prediction probability distribution vector through cross entropy loss function calculation model Net2And (3) true label->The difference between them is summed up for all input samples to obtain a loss value +.>
Wherein M is the number of marked data;
loss by retraining model Net1 and model Net2 on labeled dataAnd->Weighting and combining to obtain the total loss value +.>
Wherein lambda is I1 Is a super parameter and is a preset value.
7. The new dialog intention recognition method based on pseudo tag self-training and source domain retraining of claim 1, wherein in step 4), the specific operations of updating model parameters using back propagation are:
self-training loss values for pseudo tagsAnd with a labeling training loss value->Weighting and combining to obtain the overall loss of the model>
Wherein lambda is t Is super-parameter, is a preset value and is used for balancing lossAnd->The weight of the vehicle is occupied;
obtaining the integral lossAnd then, continuously updating parameters of the model Net1 and the model Net2 by using a back propagation algorithm, so that the difference between the prediction and the true value of the model is within a preset range.
8. The new dialog intention recognition method based on pseudo tag self-training and source domain retraining of claim 1, wherein the specific operation of predicting the input class when the model of step 5) is used is:
the dialogue text to be classified is respectively input into a trained model Net1, a model Net2 and a clustering model Kmeans,respectively obtaining the prediction labels y 1 ，y 2 Y 3 The obtained three prediction labels are weighted and combined to obtain a final prediction label y vote ：
y vote ＝λ y1 y 1 +λ y2 y 2 +(1-λ y1 -λ y2 )y 3 (20)
Prediction tag y using Softmax classifier pair vote Classifying, namely taking the maximum value y of the classification as the final prediction category:
y＝max(Softmax(y vote )) (21)。
CN202111187641.9A 2021-10-12 2021-10-12 New dialogue intention recognition method based on pseudo tag self-training and source domain retraining Active CN113887643B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111187641.9A CN113887643B (en) 2021-10-12 2021-10-12 New dialogue intention recognition method based on pseudo tag self-training and source domain retraining

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111187641.9A CN113887643B (en) 2021-10-12 2021-10-12 New dialogue intention recognition method based on pseudo tag self-training and source domain retraining

Publications (2)

Publication Number Publication Date
CN113887643A CN113887643A (en) 2022-01-04
CN113887643B true CN113887643B (en) 2023-07-18

Family

ID=79006290

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111187641.9A Active CN113887643B (en) 2021-10-12 2021-10-12 New dialogue intention recognition method based on pseudo tag self-training and source domain retraining

Country Status (1)

Country Link
CN (1) CN113887643B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114170484B (en) * 2022-02-11 2022-05-27 中科视语(北京)科技有限公司 Picture attribute prediction method and device, electronic equipment and storage medium
CN114637848A (en) * 2022-03-15 2022-06-17 美的集团(上海)有限公司 Semantic classification method and device
CN115168593B (en) * 2022-09-05 2022-11-29 深圳爱莫科技有限公司 Intelligent dialogue management method capable of self-learning and processing equipment
CN115512696A (en) * 2022-09-20 2022-12-23 中国第一汽车股份有限公司 Simulation training method and vehicle

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113378632A (en) * 2021-04-28 2021-09-10 南京大学 Unsupervised domain pedestrian re-identification algorithm based on pseudo label optimization

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10580176B2 (en) * 2018-06-28 2020-03-03 Microsoft Technology Licensing, Llc Visualization of user intent in virtual agent interaction
JP6695004B1 (en) * 2019-04-03 2020-05-20 ガンホー・オンライン・エンターテイメント株式会社 A terminal device that uses the detected intention of the user
CN110298391B (en) * 2019-06-12 2023-05-02 同济大学 Iterative incremental dialogue intention type recognition method based on small sample
CN110377911B (en) * 2019-07-23 2023-07-21 中国工商银行股份有限公司 Method and device for identifying intention under dialog framework
CN113297360B (en) * 2021-04-29 2022-05-27 天津汇智星源信息技术有限公司 Law question-answering method and device based on weak supervised learning and joint learning mechanism

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113378632A (en) * 2021-04-28 2021-09-10 南京大学 Unsupervised domain pedestrian re-identification algorithm based on pseudo label optimization

Also Published As

Publication number Publication date
CN113887643A (en) 2022-01-04

Similar Documents

Publication Publication Date Title
CN113887643B (en) New dialogue intention recognition method based on pseudo tag self-training and source domain retraining
Zhang et al. Discovering new intents with deep aligned clustering
CN106469560B (en) Voice emotion recognition method based on unsupervised domain adaptation
Peng et al. Accelerating minibatch stochastic gradient descent using typicality sampling
CN106709754A (en) Power user grouping method based on text mining
CN103514170B (en) A kind of file classification method and device of speech recognition
CN109165387A (en) A kind of Chinese comment sentiment analysis method based on GRU neural network
CN111626336A (en) Subway fault data classification method based on unbalanced data set
CN112529638B (en) Service demand dynamic prediction method and system based on user classification and deep learning
CN109583635A (en) A kind of short-term load forecasting modeling method towards operational reliability
CN104091038A (en) Method for weighting multiple example studying features based on master space classifying criterion
CN110795736B (en) Malicious android software detection method based on SVM decision tree
CN115577357A (en) Android malicious software detection method based on stacking integration technology
CN116050419B (en) Unsupervised identification method and system oriented to scientific literature knowledge entity
CN116380438A (en) Fault diagnosis method and device, electronic equipment and storage medium
CN115083511A (en) Peripheral gene regulation and control feature extraction method based on graph representation learning and attention
Tang et al. Chinese spam classification based on weighted distributed characteristic
CN115600595A (en) Entity relationship extraction method, system, equipment and readable storage medium
CN114357166A (en) Text classification method based on deep learning
CN113535928A (en) Service discovery method and system of long-term and short-term memory network based on attention mechanism
Sun et al. Analysis of English writing text features based on random forest and Logistic regression classification algorithm
Chun et al. Research on music classification based on MFCC and BP neural network
He et al. Label correlation mixture model: A supervised generative approach to multilabel spoken document categorization
CN111274359A (en) Query recommendation method and system based on improved VHRED and reinforcement learning
CN111310971B (en) Prospect analysis method, device and equipment for O2O business model

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