CN112905750A - Generation method and device of optimization model - Google Patents

Generation method and device of optimization model Download PDF

Info

Publication number
CN112905750A
CN112905750A CN202110289744.XA CN202110289744A CN112905750A CN 112905750 A CN112905750 A CN 112905750A CN 202110289744 A CN202110289744 A CN 202110289744A CN 112905750 A CN112905750 A CN 112905750A
Authority
CN
China
Prior art keywords
similarity
intent
model
score
text
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.)
Pending
Application number
CN202110289744.XA
Other languages
Chinese (zh)
Inventor
姜姗
刘升平
梁家恩
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.)
Unisound Intelligent Technology Co Ltd
Xiamen Yunzhixin Intelligent Technology Co Ltd
Original Assignee
Unisound Intelligent Technology Co Ltd
Xiamen Yunzhixin Intelligent Technology Co Ltd
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 Unisound Intelligent Technology Co Ltd, Xiamen Yunzhixin Intelligent Technology Co Ltd filed Critical Unisound Intelligent Technology Co Ltd
Priority to CN202110289744.XA priority Critical patent/CN112905750A/en
Publication of CN112905750A publication Critical patent/CN112905750A/en
Pending legal-status Critical Current

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
    • G06F16/3331—Query processing
    • 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
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent

Abstract

The embodiment of the invention provides a method and equipment for generating an optimization model, wherein the method comprises the following steps: acquiring training data; organizing the training data; splicing the user question type and the question type in the knowledge base into a whole through a first character, and respectively adding a second character and the first character into the head and the tail of the whole to be used as input; modeling using BERT based on the input to obtain a coded vector; selecting a vector corresponding to the second character as a feature vector; respectively inputting the feature vectors into a first full-connection layer for text similarity calculation and a second full-connection layer for intention classification; training the model based on the text similarity score and the intention classification score and a loss function of the model combining the first full-connected layer and the second full-connected layer; and selecting the model with the highest F1 value from the trained models as the best model. The scheme improves the performance of model text matching.

Description

Generation method and device of optimization model
Technical Field
The invention relates to the technical field of model training, in particular to a method and equipment for generating an optimization model.
Background
The text matching task is a core task in a search-type question-answering system, and returns the best answer corresponding to the user question by calculating the matching degree of the user question and the question in the knowledge base. Wherein, the BERT (a language model) model is a bidirectional language model based on large-scale corpus pre-training and can be migrated to a text matching task through fine tuning. At present, based on BERT model fine tuning is the mainstream method of text matching task.
However, the text matching model in the current retrieval type question-answering system directly takes the user question as input, and does not fully utilize the implied intention information of the user question, which results in insufficient accuracy.
Therefore, there is a need for a better method to solve the problems of the prior art.
Disclosure of Invention
The invention provides a method and equipment for generating an optimization model, which can solve the technical problem of accuracy in the prior art.
The technical scheme for solving the technical problems is as follows:
the embodiment of the invention provides a generation method of an optimization model, which comprises the following steps:
acquiring training data;
organizing the training data into user questions, questions in a knowledge base, intention category labels and similarity category labels of the user questions;
splicing the user question types and the question types in the knowledge base into a whole through first characters, and adding second characters and the first characters into the head and the tail of the whole respectively to be used as input;
modeling using BERT based on the input to obtain a coded vector;
selecting the vector corresponding to the second character as a feature vector;
respectively inputting the feature vectors into a first full-connection layer for text similarity calculation and a second full-connection layer for intention classification to obtain a text similarity score and an intention classification score;
training the model based on the text similarity score and the intent classification score and a loss function of the model that combines the first fully-connected layer and the second fully-connected layer;
and selecting the model with the highest F1 value from the trained models as the best model.
In a specific embodiment, the text similarity score is determined by the following formula:
ysimilarity＝Softmax(Fsimilarity(hi))；
wherein, ysimilarityScoring the text similarity; h isiIs a feature vector; fsimilarityIs the first fully connected layer for text similarity calculation.
In a particular embodiment, the intent classification score is determined by the following formula:
yintent＝Softmax(Fintent(hi))；
wherein, yintentClassifying a score for the intent; h isiIs a feature vector; fintentIs the second fully connected layer for the intended classification.
In a specific embodiment, the loss function is:
L＝αLsimilarity+(1-α)Lintent；
wherein L is the loss of the model combining the first fully-connected layer and the second fully-connected layer; l issimilarityMatching the loss of tasks for the text; l isintentFor the loss of the intent classification task, α (0 ≦ α ≦ 1) is a coefficient parameter for controlling LsimilarityAnd LintentThe ratio of the two losses.
In a specific embodiment, the loss of the text matching task is:
Figure BDA0002979385240000021
wherein N is the number of samples, i is the ith (i is more than or equal to 1 and less than or equal to N) sample in the training data, and yiIs similar to the sampleDegree category tag, y'iThe probability of a positive sample is predicted for the model.
In a specific embodiment, the loss of the intent classification task is:
Figure BDA0002979385240000031
wherein K is the total number of the sample intention categories, j is the jth (j is more than or equal to 1 and less than or equal to K) category in the intention categories, and ljFor sample intention class labels, pjThe intention class computed for the model is the probability of j.
In a specific embodiment, the network parameters of the model are updated by a back-propagation algorithm.
The embodiment of the present invention further provides a device for generating an optimization model, including:
the acquisition module is used for acquiring training data;
the organizing module is used for organizing the training data into user questions, questions in a knowledge base, intention category labels and similarity category labels of the user questions;
the splicing module is used for splicing the user question types and the question types in the knowledge base into a whole through first characters, and adding second characters and the first characters into the whole from head to tail respectively to be used as input;
a vector module for modeling using BERT based on the input to obtain a coded vector;
the selection module is used for selecting the vector corresponding to the second character as a feature vector;
the input module is used for respectively inputting the feature vectors into a first full-connection layer used for text similarity calculation and a second full-connection layer used for intention classification so as to obtain a text similarity score and an intention classification score;
a training module to train a model of a first fully-connected layer and a second fully-connected layer based on the text similarity score and the intent classification score and a loss function of the model;
and the processing module is used for selecting the model with the highest F1 value from the trained models as the best model.
In a specific embodiment, the text similarity score is determined by the following formula:
ysimilarity＝Softmax(Fsimilarity(hi))；
wherein, ysimilarityScoring the text similarity; h isiIs a feature vector; fsimilarityIs the first fully connected layer for text similarity calculation.
In a particular embodiment, the intent classification score is determined by the following formula:
yintent＝Softmax(Fintent(hi))；
wherein, yintentClassifying a score for the intent; h isiIs a feature vector; fintentIs the second fully connected layer for the intended classification.
The invention has the beneficial effects that:
the embodiment of the invention provides a method and equipment for generating an optimization model, wherein the method comprises the following steps: acquiring training data; organizing the training data into user questions, questions in a knowledge base, intention category labels and similarity category labels of the user questions; splicing the user question types and the question types in the knowledge base into a whole through first characters, and adding second characters and the first characters into the head and the tail of the whole respectively to be used as input; modeling using BERT based on the input to obtain a coded vector; selecting the vector corresponding to the second character as a feature vector; respectively inputting the feature vectors into a first full-connection layer for text similarity calculation and a second full-connection layer for intention classification to obtain a text similarity score and an intention classification score; training the model based on the text similarity score and the intent classification score and a loss function of the model that combines the first fully-connected layer and the second fully-connected layer; and selecting the model with the highest F1 value from the trained models as the best model. In the scheme, the intention classification and the text matching task are combined, so that the model learns the characteristics of different intentions, the comprehension of the model on the input problems of the user can be enhanced, the text matching task is assisted, and the text matching performance of the model is improved.
Drawings
Fig. 1 is a schematic flow chart of a method for generating an optimization model according to an embodiment of the present invention;
fig. 2 is a schematic diagram of a framework of a model in a method for generating an optimization model according to an embodiment of the present invention;
fig. 3 is a schematic structural diagram of an optimization model generation device according to an embodiment of the present invention;
fig. 4 is a frame structure diagram of a terminal according to an embodiment of the present invention.
Detailed Description
The principles and features of this invention are described below in conjunction with the following drawings, which are set forth by way of illustration only and are not intended to limit the scope of the invention.
Example 1
The embodiment 1 of the invention discloses a method for generating an optimization model, which comprises the following steps as shown in fig. 1:
step 101, acquiring training data;
step 102, organizing the training data into user questions, questions in a knowledge base, intention category labels and similarity category labels of the user questions;
in particular, the training data is organized into Sennce 1i,Sentence2i,li,yiForm (I) of the formula (I), wherein Sennce 1iIndicating a user question, Sennce 2iRepresenting a problem in a knowledge base,/iIntention category label, y, representing user questioniRepresenting similarity category labels.
103, splicing the user question type and the question type in the knowledge base into a whole through first characters, and adding second characters and the first characters into the head and the tail of the whole respectively to be used as input;
specifically, Sentence1i,Sentence2iBy [ SEP ]]Splicing characters and adding [ CLS ] at head and tail]And [ SEP ]]Characters are used as input.
104, modeling by using BERT based on the input to obtain a coded vector;
specifically, the input is modeled using BERT to obtain a coded vector.
Step 105, selecting the vector corresponding to the second character as a feature vector;
specific selection [ CLS]The vector corresponding to the character is used as a feature vector hi. Inputting the parameters into a full connection layer with two different parameters respectively, wherein the full connection layer F _ similarity is used for calculating the text similarity
Step 106, respectively inputting the feature vectors into a first full-connection layer for text similarity calculation and a second full-connection layer for intention classification to obtain a text similarity score and an intention classification score;
in a specific embodiment, the text similarity score is determined by the following formula:
ysimilarity＝Softmax(Fsimilarity(hi))；
wherein, ysimilarityScoring the text similarity; h isiIs a feature vector; fsimilarityIs the first fully connected layer for text similarity calculation.
Further, the intent classification score is determined by the following formula:
yintent＝Softmax(Fintent(hi))；
wherein, yintentClassifying a score for the intent; h isiIs a feature vector; fintentIs the second fully connected layer for the intended classification.
Further, the loss function is:
L＝αLsimilarity+(1-α)Lintent；
wherein L is the loss of the model combining the first fully-connected layer and the second fully-connected layer; l issimilarityFor matching tasks to textLoss; l isintentFor the loss of the intent classification task, α (0 ≦ α ≦ 1) is a coefficient parameter for controlling LsimilarityAnd LintentThe ratio of the two losses.
In a specific embodiment, the loss of the text matching task is:
Figure BDA0002979385240000061
wherein N is the number of samples, i is the ith (i is more than or equal to 1 and less than or equal to N) sample in the training data, and yiFor sample similarity class labels, y' i is the probability that the model predicts positive samples.
In a specific embodiment, the loss of the intent classification task is:
Figure BDA0002979385240000062
wherein K is the total number of the sample intention categories, j is the jth (j is more than or equal to 1 and less than or equal to K) category in the intention categories, and ljFor sample intention class labels, pjThe intention class computed for the model is the probability of j.
Step 107, training the model based on the text similarity score and the intention classification score and a loss function of the model combining the first full-connected layer and the second full-connected layer;
specifically, the model is a BERT joint learning model in which intent classifications are matched with text, which is constructed by means of intent classification tags of user questions, as shown in fig. 2.
Specifically, the network parameters of the model are updated by a back propagation algorithm. After each update, a trained model is obtained and step 108 is performed.
And 108, selecting the model with the highest F1 value from the trained models as the best model.
Specifically, F1 is also referred to as the F value, F-measure, or other variant).
The F1 value is defined on a per category basis and includes two broad concepts: precision (precision) and recall (recall). The accuracy rate refers to the proportion of individuals whose prediction results belong to a certain class, and actually belong to the class. Recall refers to the ratio of the number of individuals correctly predicted to be of a certain category to the total number of individuals of that category in the data set.
In the scheme, the intention classification and the text matching task are combined, so that the model learns the characteristics of different intentions, the comprehension of the model on the input problems of the user can be enhanced, the text matching task is assisted, and the text matching performance of the model is improved.
Example 2
Embodiment 2 of the present invention further discloses an optimization model generation device, as shown in fig. 3, including:
an obtaining module 201, configured to obtain training data;
an organizing module 202 for organizing the training data into user questions, questions in a knowledge base, intention category labels and similarity category labels of the user questions;
the splicing module 203 is used for splicing the user question type and the question type in the knowledge base into a whole through a first character, and adding a second character and the first character into the whole as input respectively at the head and the tail of the whole;
a vector module 204 for modeling using BERT based on the input to obtain a coded vector;
a selecting module 205, configured to select the vector corresponding to the second character as a feature vector;
an input module 206, configured to input the feature vectors into a first full-link layer for text similarity calculation and a second full-link layer for intent classification, respectively, so as to obtain a text similarity score and an intent classification score;
a training module 207 for training the model based on the text similarity score and the intent classification score and a loss function of the model in conjunction with the first fully-connected layer and the second fully-connected layer;
and the processing module 208 is used for selecting the model with the highest F1 value from the trained models as the best model.
In a specific embodiment, the text similarity score is determined by the following formula:
ysimilarity＝Softmax(Fsimilarity(hi))；
wherein, ysimilarityScoring the text similarity; h isiIs a feature vector; fsimilarityIs the first fully connected layer for text similarity calculation.
In a particular embodiment, the intent classification score is determined by the following formula:
yintent＝Softmax(Fintent(hi))；
wherein, yintentClassifying a score for the intent; h isiIs a feature vector; fintentIs the second fully connected layer for the intended classification.
In a specific embodiment, the loss function is:
L＝αLsimilarity+(1-α)Lintent；
wherein L is the loss of the model combining the first fully-connected layer and the second fully-connected layer; l issimilarityMatching the loss of tasks for the text; l isintentFor the loss of the intent classification task, α (0 ≦ α ≦ 1) is a coefficient parameter for controlling LsimilarityAnd LintentThe ratio of the two losses.
In a specific embodiment, the loss of the text matching task is:
Figure BDA0002979385240000081
wherein N is the number of samples, i is the ith (i is more than or equal to 1 and less than or equal to N) sample in the training data, and yiIs a sample similarity class label, y'iThe probability of a positive sample is predicted for the model.
In a specific embodiment, the loss of the intent classification task is:
Figure BDA0002979385240000082
wherein K is the total number of the sample intention categories, j is the jth (j is more than or equal to 1 and less than or equal to K) category in the intention categories, and ljFor sample intention class labels, pjThe intention class computed for the model is the probability of j.
In a specific embodiment, the network parameters of the model are updated by a back-propagation algorithm.
Example 3
Embodiment 3 of the present invention also discloses a terminal, as shown in fig. 4, including: a memory and a processor that executes the method of embodiment 1 when running an application in the memory.
The embodiment of the invention provides a method and equipment for generating an optimization model, wherein the method comprises the following steps: acquiring training data; organizing the training data into user questions, questions in a knowledge base, intention category labels and similarity category labels of the user questions; splicing the user question types and the question types in the knowledge base into a whole through first characters, and adding second characters and the first characters into the head and the tail of the whole respectively to be used as input; modeling using BERT based on the input to obtain a coded vector; selecting the vector corresponding to the second character as a feature vector; respectively inputting the feature vectors into a first full-connection layer for text similarity calculation and a second full-connection layer for intention classification to obtain a text similarity score and an intention classification score; training the model based on the text similarity score and the intent classification score and a loss function of the model that combines the first fully-connected layer and the second fully-connected layer; and selecting the model with the highest F1 value from the trained models as the best model. In the scheme, the intention classification and the text matching task are combined, so that the model learns the characteristics of different intentions, the comprehension of the model on the input problems of the user can be enhanced, the text matching task is assisted, and the text matching performance of the model is improved.
While the invention has been described with reference to specific embodiments, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (10)

1. A method for generating an optimization model, comprising:
acquiring training data;
organizing the training data into user questions, questions in a knowledge base, intention category labels and similarity category labels of the user questions;
splicing the user question types and the question types in the knowledge base into a whole through first characters, and adding second characters and the first characters into the head and the tail of the whole respectively to be used as input;
modeling using BERT based on the input to obtain a coded vector;
selecting the vector corresponding to the second character as a feature vector;
respectively inputting the feature vectors into a first full-connection layer for text similarity calculation and a second full-connection layer for intention classification to obtain a text similarity score and an intention classification score;
training the model based on the text similarity score and the intent classification score and a loss function of the model that combines the first fully-connected layer and the second fully-connected layer;
and selecting the model with the highest F1 value from the trained models as the best model.
2. The method of claim 1, wherein the text similarity score is determined by the formula:
ysimilarity＝Softmax(Fsimilarity(hi))；
wherein, ysimilarityScoring the text similarity; h isiIs a feature vector; fsimilarityIs the first fully connected layer for text similarity calculation.
3. The method of claim 1, wherein the intent classification score is determined by the formula:
yintent＝Softmax(Fintent(hi))；
wherein, yintentClassifying a score for the intent; h isiIs a feature vector; fintentIs the second fully connected layer for the intended classification.
4. The method of claim 1, wherein the loss function is:
L＝αLsimilarity+(1-α)Lintent；
wherein L is the loss of the model combining the first fully-connected layer and the second fully-connected layer; l issimilarityMatching the loss of tasks for the text; l isintentFor the loss of the intent classification task, α (0 ≦ α ≦ 1) is a coefficient parameter for controlling LsimilarityAnd LintentThe ratio of the two losses.
5. The method of claim 4, wherein the penalty for the text matching task is:
Figure FDA0002979385230000021
wherein N is the number of samples, i is the ith (i is more than or equal to 1 and less than or equal to N) sample in the training data, and yiIs a sample similarity class label, y'iThe probability of a positive sample is predicted for the model.
6. The method of claim 4, wherein the loss of the intent classification task is:
Figure FDA0002979385230000022
wherein K is the total number of the sample intention categories, j is the jth (j is more than or equal to 1 and less than or equal to K) category in the intention categories, and ljFor sample intention class labels, pjThe intention class computed for the model is the probability of j.
7. The method of claim 1, wherein the network parameters of the model are updated by a back-propagation algorithm.
8. An optimization model generation apparatus, comprising:
the acquisition module is used for acquiring training data;
the organizing module is used for organizing the training data into user questions, questions in a knowledge base, intention category labels and similarity category labels of the user questions;
the splicing module is used for splicing the user question types and the question types in the knowledge base into a whole through first characters, and adding second characters and the first characters into the whole from head to tail respectively to be used as input;
a vector module for modeling using BERT based on the input to obtain a coded vector;
the selection module is used for selecting the vector corresponding to the second character as a feature vector;
the input module is used for respectively inputting the feature vectors into a first full-connection layer used for text similarity calculation and a second full-connection layer used for intention classification so as to obtain a text similarity score and an intention classification score;
a training module to train a model of a first fully-connected layer and a second fully-connected layer based on the text similarity score and the intent classification score and a loss function of the model;
and the processing module is used for selecting the model with the highest F1 value from the trained models as the best model.
9. The apparatus of claim 8, wherein the text similarity score is determined by the formula:
ysimilarity＝Softmax(Fsimilarity(hi))；
wherein, ysimilarityScoring the text similarity; h isiIs a feature vector; fsimilarityIs the first fully connected layer for text similarity calculation.
10. The apparatus of claim 8, wherein the intent classification score is determined by the formula:
yintent＝Softmax(Fintent(hi))；
wherein, yintentClassifying a score for the intent; h isiIs a feature vector; fintentIs the second fully connected layer for the intended classification.
CN202110289744.XA 2021-03-16 2021-03-16 Generation method and device of optimization model Pending CN112905750A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110289744.XA CN112905750A (en) 2021-03-16 2021-03-16 Generation method and device of optimization model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110289744.XA CN112905750A (en) 2021-03-16 2021-03-16 Generation method and device of optimization model

Publications (1)

Publication Number Publication Date
CN112905750A true CN112905750A (en) 2021-06-04

Family

ID=76105396

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110289744.XA Pending CN112905750A (en) 2021-03-16 2021-03-16 Generation method and device of optimization model

Country Status (1)

Country Link
CN (1) CN112905750A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113515620A (en) * 2021-07-20 2021-10-19 云知声智能科技股份有限公司 Method and device for sorting technical standard documents of power equipment, electronic equipment and medium
CN114139497A (en) * 2021-12-13 2022-03-04 国家电网有限公司大数据中心 Text abstract extraction method based on BERTSUM model

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111538824A (en) * 2020-05-25 2020-08-14 武汉烽火普天信息技术有限公司 BERT-based intelligent question and answer implementation method and system
CN111651603A (en) * 2020-06-04 2020-09-11 上海电力大学 Power industry single-level text classification method and system based on LAV parameter fine adjustment
CN111860669A (en) * 2020-07-27 2020-10-30 平安科技（深圳）有限公司 Training method and device of OCR recognition model and computer equipment
WO2020248471A1 (en) * 2019-06-14 2020-12-17 华南理工大学 Aggregation cross-entropy loss function-based sequence recognition method
CN112214998A (en) * 2020-11-16 2021-01-12 中国平安财产保险股份有限公司 Method, device, equipment and storage medium for joint identification of intention and entity

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020248471A1 (en) * 2019-06-14 2020-12-17 华南理工大学 Aggregation cross-entropy loss function-based sequence recognition method
CN111538824A (en) * 2020-05-25 2020-08-14 武汉烽火普天信息技术有限公司 BERT-based intelligent question and answer implementation method and system
CN111651603A (en) * 2020-06-04 2020-09-11 上海电力大学 Power industry single-level text classification method and system based on LAV parameter fine adjustment
CN111860669A (en) * 2020-07-27 2020-10-30 平安科技（深圳）有限公司 Training method and device of OCR recognition model and computer equipment
CN112214998A (en) * 2020-11-16 2021-01-12 中国平安财产保险股份有限公司 Method, device, equipment and storage medium for joint identification of intention and entity

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113515620A (en) * 2021-07-20 2021-10-19 云知声智能科技股份有限公司 Method and device for sorting technical standard documents of power equipment, electronic equipment and medium
CN114139497A (en) * 2021-12-13 2022-03-04 国家电网有限公司大数据中心 Text abstract extraction method based on BERTSUM model

Similar Documents

Publication Publication Date Title
CN112528676B (en) Document-level event argument extraction method
CN111522839B (en) Deep learning-based natural language query method
CN110457675A (en) Prediction model training method, device, storage medium and computer equipment
CN109325547A (en) Non-motor vehicle image multi-tag classification method, system, equipment and storage medium
CN107729309A (en) A kind of method and device of the Chinese semantic analysis based on deep learning
CN113987209A (en) Natural language processing method and device based on knowledge-guided prefix fine tuning, computing equipment and storage medium
CN110866542B (en) Depth representation learning method based on feature controllable fusion
CN111666406B (en) Short text classification prediction method based on word and label combination of self-attention
CN111898374B (en) Text recognition method, device, storage medium and electronic equipment
CN112256866B (en) Text fine-grained emotion analysis algorithm based on deep learning
CN112015868A (en) Question-answering method based on knowledge graph completion
CN113392209A (en) Text clustering method based on artificial intelligence, related equipment and storage medium
CN112115242A (en) Intelligent customer service question-answering system based on naive Bayes classification algorithm
CN112905750A (en) Generation method and device of optimization model
CN114510570A (en) Intention classification method and device based on small sample corpus and computer equipment
CN112988970A (en) Text matching algorithm serving intelligent question-answering system
CN114091406A (en) Intelligent text labeling method and system for knowledge extraction
CN113641809A (en) XLNET-BiGRU-CRF-based intelligent question answering method
CN117056451A (en) New energy automobile complaint text aspect-viewpoint pair extraction method based on context enhancement
CN111767720A (en) Title generation method, computer and readable storage medium
CN111222533B (en) Deep learning visual question-answering method and system based on dependency tree
CN115906846A (en) Document-level named entity identification method based on double-graph hierarchical feature fusion
US20220253630A1 (en) Optimized policy-based active learning for content detection
CN112015760B (en) Automatic question-answering method and device based on candidate answer set reordering and storage medium
CN115223021A (en) Visual question-answering-based fruit tree full-growth period farm work decision-making method

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