CN112765315A - Intelligent classification system and method for legal scenes - Google Patents
Intelligent classification system and method for legal scenes Download PDFInfo
- Publication number
- CN112765315A CN112765315A CN202110061753.3A CN202110061753A CN112765315A CN 112765315 A CN112765315 A CN 112765315A CN 202110061753 A CN202110061753 A CN 202110061753A CN 112765315 A CN112765315 A CN 112765315A
- Authority
- CN
- China
- Prior art keywords
- class
- sample
- module
- samples
- classification
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 46
- 239000013598 vector Substances 0.000 claims abstract description 98
- 238000012549 training Methods 0.000 claims abstract description 19
- 230000000694 effects Effects 0.000 claims description 11
- 238000012360 testing method Methods 0.000 claims description 8
- 230000006870 function Effects 0.000 claims description 6
- 230000003247 decreasing effect Effects 0.000 claims description 4
- 239000000284 extract Substances 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 230000003044 adaptive effect Effects 0.000 claims description 2
- 238000013135 deep learning Methods 0.000 abstract description 2
- 238000012986 modification Methods 0.000 abstract 1
- 230000004048 modification Effects 0.000 abstract 1
- 238000010586 diagram Methods 0.000 description 3
- 230000007547 defect Effects 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 230000001960 triggered effect Effects 0.000 description 1
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
- 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
- G06F16/3346—Query execution using probabilistic model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/216—Parsing using statistical methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/18—Legal services
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Business, Economics & Management (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Tourism & Hospitality (AREA)
- Probability & Statistics with Applications (AREA)
- Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Human Resources & Organizations (AREA)
- Human Computer Interaction (AREA)
- Marketing (AREA)
- Economics (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Technology Law (AREA)
- General Business, Economics & Management (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a system and a method for intelligently classifying legal scenes, wherein the system comprises three modules of classification, self-learning and self-adaption. Firstly, inputting a sample to be classified into a classification module fusing a mixed attention prototype network and a word vector similarity method, and predicting the class of the sample; meanwhile, a self-learning module is designed, and the module can set confidence coefficient for the prediction result and add a high-confidence prediction sample into a training set, so that a prediction library is expanded and the performance of a model is improved; in addition, the model also has the self-adaptive capability, and can automatically adapt to changes such as increase, reduction or modification of categories. Compared with the traditional deep learning, the technical method can realize high-efficiency legal scene classification under the condition of only a small amount of initial training samples.
Description
Technical Field
The invention relates to the technical field of legal natural language processing, in particular to a legal scene intelligent classification system and method based on fusion of a mixed attention prototype network and word vector similarity.
Background
The existing text classification models based on deep learning have a plurality of models and achieve good classification effect, but in the field of scarce samples, the models are difficult to play a role. Especially, in an intelligent question-answering system in the legal field, accurate classification of legal scenes is a necessary premise for realizing intelligent question-answering, but user consultation problems are often serious in spoken language, the labeling cost is high, and the effect of the traditional supervised learning method is poor. Aiming at the problem, a method for fusing prototype network classification and word vector similarity based on mixed attention is designed, and the problem of difficulty in classification of legal scenes under sparse labeling is solved. The model has self-learning and self-adaptive capabilities, wherein the self-learning performance is that the model performance can be improved by automatically expanding the corpus; the self-adaptation is represented by automatically adapting to the conditions of increasing, decreasing or modifying the categories and the like on the premise of not retraining the model.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a legal scene intelligent classification system and a legal scene intelligent classification method, which solve the defects in the prior art.
In order to realize the purpose, the technical scheme adopted by the invention is as follows:
a legal scene intelligent classification system comprising: the system comprises a classification module, a self-learning module and a self-adaption module;
the classification module comprises a prototype network module based on mixed attention and a word vector similarity module and is used for classifying user consultation problems. The prototype network module based on mixed attention is used for calculating the distance between the user consultation problem and each class prototype so as to judge the class to which the user consultation problem belongs; the word vector similarity module judges the category of the word vector by calculating the similarity between the word vector of the user consultation problem and various label vectors, so that the mixed attention prototype network is assisted, and the classification effect is improved;
a self-learning module: the method is used for automatically expanding the category data with insufficient corpus in the training set, so that the accuracy of the system in classification prediction is improved;
an adaptive module: when adding, subtracting or modifying categories to be classified, changes in categories can be automatically accommodated.
The invention also discloses a classification method of the intelligent classification system for legal scenes, which comprises the following steps:
step 1, firstly, inputting user consultation questions into a prototype network method based on mixed attention to obtain score vectors P on each class1Then, inputting the user consultation question into a word vector similarity method to obtain score vectors P on each class2Finally, by attention mechanism P1And P2And weighting and summing to obtain a final score vector P, thereby outputting the category with the highest score. Wherein the score vector P ═ α P1+βP2And alpha and beta are attention coefficients.
And 2, firstly, judging whether the training data of the final type obtained by model prediction is sufficient or not, when the data is insufficient, calculating the confidence coefficient of the prediction result of the sample, if the confidence coefficient exceeds a threshold value set by a system, storing the sample into a temporary corpus, and when the data volume of the temporary corpus reaches 200, expanding the temporary corpus into a training set to retrain the model, thereby improving the accuracy of the model.
And 3, when the categories to be classified are increased, decreased or modified, finely adjusting the model, testing the model by using the data in the test set, and triggering the model to be retrained if the accuracy on the test set is lower than a threshold value set by the system. In addition, when the categories are added, the number of samples corresponding to each category should not be less than 20.
Further, the prototype network method based on mixed attention comprises the following working steps:
1. the method adopts an N-way K-shot mode of small sample learning to train, firstly, N categories are extracted for training set data input into the module, and each category comprises K samples to form a support setAnd S, extracting Q samples from each class in the rest samples of the N classes to form a query set. Then inputting the samples in the support set into the coding layer, converting the natural language text into a vector form E which can be identified by a computer, and simultaneously inputting the samples q in the query set into the coding layer to convert the samples q into a feature vector xq. Wherein, the supporting set
J-th sample, w, representing the i-th classtDenotes the t-th word in the sample, n denotes the maximum length of the sample,
a feature vector representing the jth sample of the ith class;
2. because the contribution degrees of the features of the feature vectors in different dimensions to the categories of the feature vectors are different, for various samples in the support set of the input feature level attention module, the features of the various samples in each dimension are extracted, and the feature vector Z which can better represent the characteristics of the categories is obtained (Z is equal to the feature vector Z)1，z2，…zN). Wherein z isiRepresenting a weight vector of the ith class in a feature dimension;
3. in the classification of small samples, since training data input into the model is very small, noise contained in the samples has a great influence on the performance of the model. Therefore, each sample in the support set is weighted by the example-level attention module, so that the sample with higher similarity to the sample in the query set is given higher weight to obtain a weighted feature vector E', and noise data brought by the noise data is reducedInfluence. Wherein,
a weighted feature vector representing the jth sample of the ith class in the support set;
4. the weighted feature vector of each sample can be obtained through the example level attention module, so that the class prototype c of each class can be obtained through summing the weighted feature vectors of the samples of each classi；
5. The weight vectors of various samples on different dimensions can be obtained through the feature level attention module, and therefore the weight vectors z of various samples can be obtained according to various weight vectorsiGenerating a distance function, and calculating the distance d between the query set sample q and various prototypes according to the distance functioniFinally according to the distance diThe class to which the sample q belongs can be derived. Wherein the distance function is represented as di＝zi(ci-xq)2；
6. For the user consultation problem input into the module, the distance d between the module and various prototypes can be calculated according to the distance functioniAccording to the distance diObtaining the probability that the user's consultation problem belongs to various categories
Wherein,
representing the probability that the user consultation question belongs to the i-th class in the mixed attention prototype-based network method.
Further, the working steps of the word vector similarity method are as follows:
1. firstly, the class name of each class in a training set is taken as a keyword, and the keyword is input into word2vec to obtain a word vector K of each keywordiThen, the samples in each class are classifiedConverting the words into corresponding word vectors, and adding and averaging the sample word vectors in each class as a mark vector V of each classi；
2. Respectively matching the user consultation problems with various keyword vectors KiAnd a flag vector ViCarrying out similarity calculation to obtain a corresponding similarity score M1And M2Then to M1And M2Weighted summation is carried out to obtain the probability that the problem belongs to various types
Wherein,
indicating the probability that the user consultation question belongs to the ith class in the word vector similarity method.
Further, the corpus expansion in step 2 specifically includes: after the sample prediction is carried out, the category of the user consultation problem is obtained, whether the corpus is sufficient or not is judged, the confidence coefficient of the problem needs to be judged under the condition that the corpus is insufficient, and therefore a confidence coefficient selection strategy based on data uncertainty is adopted to judge whether the user consultation problem can be used for expanding the corpus or not.
Further, the confidence coefficient is calculated according to the formula
Wherein true _ num represents the number of words in the question appearing in the prediction class sample, and all _ num represents the number of words in the question; if the words in the problem exist in the samples of the class, the result corresponding to the problem is considered to be reliable and can be used for expanding the corpus, if part of the words do not exist in the samples of the class and the confidence coefficient smaller than the threshold set by the system is obtained, the problem cannot be used for expanding the corpus, the class to which the problem belongs needs to be judged by a lawyer and then put into the temporary corpus, and when the data volume in the temporary corpus reaches 200, the model is retrained, so that the model achieves a better classification effect.
Compared with the prior art, the invention has the advantages that:
the method not only can well realize the classification of legal scenes under the condition of scarce samples, but also has the self-learning and self-adaptive capabilities, and improves the flexibility of the model.
Drawings
FIG. 1 is a general framework diagram of an intelligent classification method for legal scenarios according to an embodiment of the present invention;
FIG. 2 is a block diagram of a hybrid attention based prototype network module according to an embodiment of the present invention;
FIG. 3 is a block diagram of a legal scene classification module based on word vector similarity according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings by way of examples.
As shown in fig. 1, the intelligent classification system for legal scenes includes: the system comprises a classification module, a self-learning module and a self-adaption module;
for the classification module, firstly, the user consultation questions are input into a prototype network method based on mixed attention to obtain score vectors P on each class1Then, inputting the user consultation question into a word vector similarity method to obtain score vectors P on each class2Finally, by attention mechanism P1And P2And weighting and summing to obtain a final score vector P, thereby outputting the category with the highest score.
For the self-learning module, the invention designs a mechanism for screening input corpora through confidence coefficient to expand the corpus of the model. Firstly, judging whether the training data of the final type obtained by model prediction needs to be expanded, if so, calculating the confidence coefficient of the sample prediction result, if the confidence coefficient exceeds a threshold value set by a system, storing the sample into a temporary corpus, and when the data volume of the temporary corpus reaches 200, expanding the sample into a training set, and automatically retraining the model, thereby improving the effect of the model.
For the self-adaptive module, when the categories to be classified are increased, decreased or modified, the model needs to be finely adjusted, the model is tested by using data in the test set, and if the accuracy on the test set is lower than a threshold value set by a system, the model is triggered to be retrained. In addition, when the categories are added, the number of samples corresponding to each category should not be less than 20.
As shown in fig. 2, the prototype network module based on mixed attention mainly includes the following points:
1. the method adopts an N-way K-shot mode of small sample learning to train, firstly extracts N types of training set data input into the module, wherein each type comprises K samples to form a support set S, and extracts Q samples from each type of the rest samples of the N types to form a query set. Then inputting the samples in the support set into the coding layer, converting the natural language text into a vector form E which can be identified by a computer, and simultaneously inputting the samples q in the query set into the coding layer to convert the samples q into a feature vector xq. Wherein, the supporting set
J-th sample, w, representing the i-th classtDenotes the t-th word in the sample, n denotes the maximum length of the sample,
representing the feature vector of the jth sample of the ith class.
2. Because the feature vectors have different contribution degrees of features on different dimensions to the categories to which the feature vectors belong, the attention module support set for the input feature levelThe characteristics of each sample in each dimension are extracted to obtain a characteristic vector Z (Z is the characteristic of each class of samples) which can better represent the class characteristics of the samples1，z2，…zN). Wherein z isiRepresenting the weight vector of class i in the feature dimension.
3. In the classification of small samples, since training data input into the model is very small, noise contained in the samples has a great influence on the performance of the model. Therefore, each sample in the support set is weighted by the example-level attention module, so that the sample with higher similarity to the sample in the query set is given higher weight to obtain a weighted feature vector E', and the influence caused by noise data is reduced. Wherein,
representing the weighted feature vector of the jth sample of the ith class.
4. The weighted feature vector of each sample can be obtained through the example level attention module, so that the class prototype c of each class can be obtained through adding the weighted feature vectors of the samples of each classi。
5. The weight vectors of various samples on different dimensions can be obtained through the feature level attention module, and therefore the weight vectors z of various samples can be obtained according to various weight vectorsiGenerating a distance function, and calculating the distance d between the query set sample q and various prototypes according to the distance functioniFinally according to the distance diAnd obtaining the belonged category of the sample q. Wherein the distance function is represented as di＝zi(ci-xq)2。
6. For the user consultation problem input into the module, the distance d between the module and various prototypes can be calculated according to the distance functioniAccording to the distance diObtaining the probability that the user's consultation problem belongs to various categories
Wherein,
representing the probability that the user consultation question belongs to the i-th class in the mixed attention prototype-based network method.
As shown in fig. 3, the legal scene classification module based on word vector similarity includes the following points:
1. firstly, the class name of each class in a training set is taken as a keyword, and the keyword is input into word2vec to obtain a word vector K of each keywordi. Then, the samples in each class are divided into words and converted into corresponding word vectors, and the sample word vectors in each class are added to calculate the average value to be used as a mark vector V of each classi。
2. Respectively matching the user consultation problems with various keyword vectors KiAnd a flag vector ViCarrying out similarity calculation to obtain corresponding similarity M1And M2Then to M1And M2Weighted summation is carried out to obtain the probability that the problem belongs to various types
Wherein,
indicating the probability that the user consultation question belongs to the ith class in the word vector similarity method.
The effect of the word vector similarity method is far inferior to that of the prototype network classification method, but the situation complementary to the prototype network classification method can be formed by adding the word vector similarity method, and when the prediction effect of the prototype network method is poor, the prediction result of the prototype network method can be supplemented by fusing the word vector similarity method.
Because the classified samples in the middle of the corpus are scarce, the model performance is poor, and the corpus is expanded by using a self-learning method to improve the effect of the model. After the sample prediction is carried out to obtain the category of the user consultation problem, whether the corpus is sufficient is judged, the confidence coefficient of the problem prediction result needs to be judged under the condition that the corpus is insufficient, and therefore a confidence coefficient selection strategy based on data uncertainty is adopted to judge whether the user consultation problem can be used for expanding the corpus.
The confidence of the problem is calculated by the formula
Where true _ num represents the number of words in the problem that appear in the prediction class sample and all _ num represents the number of words in the problem. If the words in the problem exist in the samples of the class, the result corresponding to the problem is considered to be reliable and can be used for expanding the corpus, if part of the words do not exist in the samples of the class and the confidence coefficient smaller than the threshold set by the system is obtained, the problem cannot be used for expanding the corpus, the class of the problem needs to be judged by a lawyer and then put into the temporary corpus, and when the data volume in the temporary corpus reaches 200, the model is retrained, so that the model achieves a better classification effect.
It will be appreciated by those of ordinary skill in the art that the examples described herein are intended to assist the reader in understanding the manner in which the invention is practiced, and it is to be understood that the scope of the invention is not limited to such specifically recited statements and examples. Those skilled in the art can make various other specific changes and combinations based on the teachings of the present invention without departing from the spirit of the invention, and these changes and combinations are within the scope of the invention.
Claims (6)
1. An intelligent legal scene classification system, comprising: the system comprises a classification module, a self-learning module and a self-adaption module;
the classification module comprises a prototype network module based on mixed attention and a word vector similarity module and is used for classifying user consultation problems. The prototype network module based on mixed attention is used for calculating the distance between the user consultation problem and each class prototype so as to judge the class to which the user consultation problem belongs; the word vector similarity module judges the category of the word vector by calculating the similarity between the word vector of the user consultation problem and various label word vectors, so that the mixed attention prototype network is assisted, and the classification effect is improved;
a self-learning module: the method is used for automatically expanding the category data with insufficient corpus in the training set, so that the accuracy of the system in classification prediction is improved;
an adaptive module: when adding, subtracting or modifying categories to be classified, changes in categories can be automatically accommodated.
2. The classification method of the intelligent classification system of legal scenes according to claim 1, characterized by comprising the following steps:
step 1, firstly, inputting user consultation questions into a prototype network method based on mixed attention to obtain score vectors P on each class1Then, inputting the user consultation question into a word vector similarity method to obtain score vectors P on each class2By attention mechanism pair P1And P2And weighting and summing to obtain a final score vector P, thereby outputting the category with the highest score. Wherein the score vector P ═ α P1+βP2And alpha and beta are attention coefficients.
Step 2, firstly, judging whether the training data of the final type obtained by model prediction is sufficient or not, when the data is insufficient, calculating the confidence coefficient of the sample prediction result, if the confidence coefficient exceeds a threshold value set by a system, storing the sample into a temporary corpus, and when the data volume of the temporary corpus reaches 200, expanding the temporary corpus into a training set to retrain the model again, thereby improving the accuracy of the model;
and 3, when the categories to be classified are increased, decreased or modified, finely adjusting the model, testing the model by using the data in the test set, and triggering the model to be retrained if the accuracy on the test set is lower than a threshold value set by the system. In addition, when the categories are added, the number of samples corresponding to each category should not be less than 20.
3. The classification method according to claim 2, characterized in that: the prototype network method based on mixed attention comprises the following working steps:
(1) the method adopts an N-way K-shot mode of small sample learning to train, firstly extracts N types of training set data input into the module, wherein each type comprises K samples to form a support set S, and extracts Q samples from each type of the rest samples of the N types to form a query set. Then inputting the samples in the support set into the coding layer, converting the natural language text into a vector form E which can be identified by a computer, and simultaneously inputting the samples q in the query set into the coding layer to convert the samples q into a feature vector xq. Wherein, the supporting set
J-th sample, w, representing the i-th classtDenotes the t-th word in the sample, n denotes the maximum length of the sample,
a feature vector representing the jth sample of the ith class;
(2) because the feature vectors have different contribution degrees on the features of different dimensions to the categories of the feature vectors, for various samples in the support set of the input feature level attention module, the features of the various samples on each dimension are extracted to obtain the features capable of better representing the categoriesIs (Z) is the feature vector Z ═ Z1,z2,…zN) (ii) a Wherein z isiRepresenting a weight vector of the ith class in a feature dimension;
(3) in the classification of small samples, because the training data input into the model are few, the noise contained in the sample can have great influence on the performance of the model; therefore, each sample in the support set is weighted by the example level attention module, so that the sample with higher similarity to the sample in the query set is endowed with higher weight to obtain a weighted feature vector E', and the influence brought by noise data is reduced; wherein,
a weighted feature vector representing the jth sample of the ith class in the support set;
(4) the weighted feature vector of each sample can be obtained through the example level attention module, so that the class prototype c of each class can be obtained through adding the weighted feature vectors of the samples of each classi；
(5) The weight vectors of various samples on different dimensions can be obtained through the feature level attention module, and therefore the weight vectors z of various samples can be obtained according to various weight vectorsiGenerating a distance function, and calculating the distance d between the query set sample q and various prototypes according to the distance functioniFinally according to the distance diThe class to which the sample q belongs can be derived. Wherein the distance function is represented as di＝zi(ci-xq)2；
(6) For the user consultation problem input into the module, the distance d between the module and various prototypes can be calculated according to the distance functioniAccording to the distance diObtaining the probability that the user's consultation problem belongs to various categories
Wherein,
representing the probability that the user consultation question belongs to the i-th class in the mixed attention prototype-based network method.
4. The classification method according to claim 2, characterized in that: the working steps of the word vector similarity method are as follows:
(1) firstly, the class name of each class in a training set is taken as a keyword, and the keyword is input into word2vec to obtain a word vector K of each keywordiThen, the samples in each class are divided into words and converted into corresponding word vectors, and the sample word vectors in each class are added and averaged to be used as a mark vector V of each classi；
(2) Respectively matching the user consultation problems with various keyword vectors KiAnd a flag vector ViCarrying out similarity calculation to obtain a corresponding similarity score M1And M2Then to M1And M2Weighted summation is carried out to obtain the probability that the problem belongs to various types
Wherein,
indicating the probability that sample q belongs to class i in the word vector similarity method.
5. The classification method according to claim 2, characterized in that: the corpus expansion in step 2 specifically comprises: after the sample prediction is carried out, the category of the user consultation problem is obtained, whether the corpus is sufficient or not is judged, the confidence coefficient of the problem needs to be judged under the condition that the corpus is insufficient, and therefore a confidence coefficient selection strategy based on data uncertainty is adopted to judge whether the user consultation problem can be used for expanding the corpus or not.
6. The classification method according to claim 5, characterized in that: the confidence coefficient is calculated by the formula
Wherein true _ num represents the number of words in the question appearing in the prediction class sample, and all _ num represents the number of words in the question; if the words in the problem exist in the samples of the class, the result corresponding to the problem is considered to be reliable and can be used for expanding the corpus, if part of the words do not exist in the samples of the class and the confidence coefficient smaller than the threshold set by the system is obtained, the problem cannot be used for expanding the corpus, the class to which the problem belongs needs to be judged by a lawyer and then put into the temporary corpus, and when the data volume in the temporary corpus reaches 200, the model is retrained, so that the model achieves a better classification effect.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110061753.3A CN112765315B (en) | 2021-01-18 | 2021-01-18 | Intelligent classification system and method for legal scenes |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110061753.3A CN112765315B (en) | 2021-01-18 | 2021-01-18 | Intelligent classification system and method for legal scenes |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112765315A true CN112765315A (en) | 2021-05-07 |
| CN112765315B CN112765315B (en) | 2022-09-30 |
Family
ID=75702802
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110061753.3A Active CN112765315B (en) | 2021-01-18 | 2021-01-18 | Intelligent classification system and method for legal scenes |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112765315B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113220853A (en) * | 2021-05-12 | 2021-08-06 | 燕山大学 | Automatic generation method and system for legal questions |
| CN113591469A (en) * | 2021-06-15 | 2021-11-02 | 杭州费尔斯通科技有限公司 | Text enhancement method and system based on word interpretation |
| CN114612708A (en) * | 2022-02-23 | 2022-06-10 | 广州市玄武无线科技股份有限公司 | Commodity identification method and device, terminal equipment and computer readable medium |
| CN115033689A (en) * | 2022-05-27 | 2022-09-09 | 重庆邮电大学 | Original network Euclidean distance calculation method based on small sample text classification |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180204111A1 (en) * | 2013-02-28 | 2018-07-19 | Z Advanced Computing, Inc. | System and Method for Extremely Efficient Image and Pattern Recognition and Artificial Intelligence Platform |
| CN111767400A (en) * | 2020-06-30 | 2020-10-13 | 平安国际智慧城市科技股份有限公司 | Training method and device of text classification model, computer equipment and storage medium |
| CN111985581A (en) * | 2020-09-09 | 2020-11-24 | 福州大学 | Sample-level attention network-based few-sample learning method |
| CN112015902A (en) * | 2020-09-14 | 2020-12-01 | 中国人民解放军国防科技大学 | Least-order text classification method under metric-based meta-learning framework |
-
2021
- 2021-01-18 CN CN202110061753.3A patent/CN112765315B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180204111A1 (en) * | 2013-02-28 | 2018-07-19 | Z Advanced Computing, Inc. | System and Method for Extremely Efficient Image and Pattern Recognition and Artificial Intelligence Platform |
| CN111767400A (en) * | 2020-06-30 | 2020-10-13 | 平安国际智慧城市科技股份有限公司 | Training method and device of text classification model, computer equipment and storage medium |
| CN111985581A (en) * | 2020-09-09 | 2020-11-24 | 福州大学 | Sample-level attention network-based few-sample learning method |
| CN112015902A (en) * | 2020-09-14 | 2020-12-01 | 中国人民解放军国防科技大学 | Least-order text classification method under metric-based meta-learning framework |
Non-Patent Citations (1)
| Title |
|---|
| 陈潇等: "基于网络表示的半监督问答文本情感分类方法", 《郑州大学学报(理学版)》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113220853A (en) * | 2021-05-12 | 2021-08-06 | 燕山大学 | Automatic generation method and system for legal questions |
| CN113591469A (en) * | 2021-06-15 | 2021-11-02 | 杭州费尔斯通科技有限公司 | Text enhancement method and system based on word interpretation |
| CN114612708A (en) * | 2022-02-23 | 2022-06-10 | 广州市玄武无线科技股份有限公司 | Commodity identification method and device, terminal equipment and computer readable medium |
| CN114612708B (en) * | 2022-02-23 | 2022-12-09 | 广州市玄武无线科技股份有限公司 | Commodity identification method and device, terminal equipment and computer readable medium |
| CN115033689A (en) * | 2022-05-27 | 2022-09-09 | 重庆邮电大学 | Original network Euclidean distance calculation method based on small sample text classification |
| CN115033689B (en) * | 2022-05-27 | 2023-04-18 | 重庆邮电大学 | Original network Euclidean distance calculation method based on small sample text classification |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112765315B (en) | 2022-09-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112765315B (en) | Intelligent classification system and method for legal scenes | |
| CN108875807B (en) | Image description method based on multiple attention and multiple scales | |
| CN108984745B (en) | Neural network text classification method fusing multiple knowledge maps | |
| CN108595632B (en) | Hybrid neural network text classification method fusing abstract and main body characteristics | |
| CN111931513B (en) | Text intention recognition method and device | |
| US20220147836A1 (en) | Method and device for text-enhanced knowledge graph joint representation learning | |
| CN110826337A (en) | Short text semantic training model obtaining method and similarity matching algorithm | |
| CN106845411B (en) | Video description generation method based on deep learning and probability map model | |
| CN110619051B (en) | Question sentence classification method, device, electronic equipment and storage medium | |
| CN113626589B (en) | Multi-label text classification method based on mixed attention mechanism | |
| CN111506732B (en) | Text multi-level label classification method | |
| CN111368142B (en) | Video intensive event description method based on generation countermeasure network | |
| CN112685504B (en) | Production process-oriented distributed migration chart learning method | |
| CN111506728B (en) | Hierarchical structure text automatic classification method based on HD-MSCNN | |
| CN110874411A (en) | Cross-domain emotion classification system based on attention mechanism fusion | |
| CN111177402A (en) | Evaluation method and device based on word segmentation processing, computer equipment and storage medium | |
| CN111144566A (en) | Neural network weight parameter training method, characteristic classification method and corresponding device | |
| CN113673535A (en) | Image description generation method of multi-modal feature fusion network | |
| CN115064154A (en) | Method and device for generating mixed language voice recognition model | |
| JPH0934863A (en) | Information integral processing method by neural network | |
| CN112989843B (en) | Intention recognition method, device, computing equipment and storage medium | |
| CN118227790A (en) | Text classification method, system, equipment and medium based on multi-label association | |
| CN115577111A (en) | Text classification method based on self-attention mechanism | |
| CN115170888A (en) | Electronic component zero sample identification model and method based on visual information and semantic attributes | |
| CN115168576A (en) | Method, model and medium for analyzing aspect emotion |
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 | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20221215 Address after: No. J24, Floor 17, No. 1, Zhongguancun Street, Haidian District, Beijing 100084 Patentee after: Shengming Jizhi (Beijing) Technology Co.,Ltd. Address before: 066004 No. 438, Hebei Avenue, Qinhuangdao, Hebei Patentee before: Yanshan University |