An Industry Economic Activity Classification Method and System Based on a Semantic Understanding Model
Through the industry economic activity classification method based on semantic understanding model, the BERT model and softmax classifier are used to solve the problem of classification error of the enterprise business scope, and efficient and accurate industry classification is achieved.
Patent Information
- Application Number
- CN202111422110.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-26
- Publication Date
- 2025-07-25
- Estimated Expiration
- 2041-11-26
AI Technical Summary
The existing technology has errors in the national economic industry classification of the business scope of the enterprise, resulting in incorrect assessment of the industry development status, and the method based on word matching is inefficient.
The industry economic activity classification method based on semantic understanding model is adopted, and precise classification is achieved from a semantic perspective by obtaining data sets, cleaning, division, keyword extraction and BERT model classification, combined with softmax classifier and dictionary matching.
It improves the accuracy of industry classification, reduces the consumption of time and human resources, and ensures the compatibility of enterprise information and industry classification tables.
Smart Images
Figure CN114116979B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of industry classification, and particularly to a method and system for classifying industrial economic activities based on a semantic understanding model. Background Art
[0002] With the rapid economic development, various enterprises have emerged like mushrooms after rain. The business scope of enterprises involves hundreds or thousands of commodities. Classifying the company's commodities into the corresponding economic activities in the national economic industry classification table will help determine the company's subsequent development direction. However, classifying commodities into the national economic industry classification is a major problem.
[0003] The existing solutions generally rely on manual classification or machine classification. In machine classification, words and phrases are generally matched with the industry classification table, and the classification results are determined according to the similarity. This makes the matching results often superficial. In real life, the names of economic activities are diverse. Different words can represent the same semantics, and the same word can represent two different things in different contexts. This is a major difficulty in the classification task.
[0004] In view of the problems existing in the prior art, the present invention proposes a method and system for classifying industrial economic activities based on a semantic understanding model. Summary of the Invention
[0005] The purpose of the present invention is to provide a method and system for classifying industrial economic activities based on a semantic understanding model to address the deficiencies of the prior art. From the perspective of semantic understanding, specific products are matched to the most suitable industries one by one, and then the middle category, large category, and category are inferred from the information of the small category.
[0006] To achieve the above purpose, the present invention adopts the following technical solutions:
[0007] A method for classifying industrial economic activities based on a semantic understanding model, comprising:
[0008] S1. Obtain a data set corresponding to each industrial economic activity and the main products;
[0009] S2. Remove invalid data from the data set and classify the data set according to the national industry classification table;
[0010] S3. Input the industrial economic activities in the classified data set, extract keywords in the industrial economic activities, and use a similarity matching algorithm to select a candidate industry classification word set and a pseudo-candidate industry classification word set from the national economic industry classification table and the data set;
[0011] S4. Input the input economic activities, the candidate industry classification term set, and the pseudo-candidate industry classification term set into a preset BERT model, and output the classification result through a softmax classifier;
[0012] S5. Establish a dictionary mainly based on the national economic industry classification table and the content of the training set, match the output classification result with the content in the dictionary, and output the final result.
[0013] Further, the specific steps of step S2 are as follows:
[0014] S21. Clean the data in the obtained data set and delete the invalid information in the data;
[0015] S22. According to the information in the national economic industry classification table, divide the cleaned data into small categories in the national economic industry classification table as the output;
[0016] S23. Divide the data set into a training set, a test set, and a validation set.
[0017] Further, the specific steps of step S3 are as follows:
[0018] S31. Use the TextRank algorithm to extract keywords from the industrial economic activities in the data set;
[0019] S32. Use the Jaccard similarity algorithm for the extracted keywords to select the candidate industry classification term set and the pseudo-candidate industry classification term set.
[0020] Further, the keyword extraction using the TextRank algorithm in step S31 is expressed as:
[0021]
[0022] Among them, WS(V i ) represents the weight of phrase i; W ji represents the similarity between two sentences j and sentence i; WS(V j ) represents the weight of phrase j obtained in the previous iteration; d represents the damping coefficient; W jk represents the similarity between sentence j and sentence k; V k represents the node coefficient of sentence k.
[0023] Further, the specific steps of step S4 are as follows:
[0024] S41. Combine the input economic activities with the candidate industry classification terms, separate them using the sentence pair classification task SEP, and classify them through the single text classification task CLS to obtain the input information, and the input information is used as the input of the next fully connected layer;
[0025] S42. Input the obtained input information into the input layer of a preset BERT model for processing, and classify the processed information using a softmax classifier to obtain an output result.
[0026] Further, the output result in step S42 is expressed as:
[0027]
[0028] where y represents the softmax value of the output result; e i represents the exponent of the i-th element; ∑ j e j represents the sum of the exponents of all elements. That is to say, the output result is the ratio of the exponent of this element to the sum of the exponents of all elements.
[0029] Further, step S5 is specifically as follows:
[0030] S51. Establish a dictionary mainly based on the data in the national economic industry classification table, expand the number of sub-categories using the data in the training set, and associate the data in the training set with the sub-categories;
[0031] S52. Put the output classification result into the dictionary for matching and output the final result.
[0032] Further, the output of the final result in step S52 is expressed as:
[0033] result = match(dic, y)
[0034] where result represents the output final result; dic represents the dictionary.
[0035] Correspondingly, a classification system for industrial economic activities based on a semantic understanding model is also provided, including:
[0036] An acquisition module for acquiring a data set corresponding to each industrial economic activity and its main products;
[0037] A removal module for removing invalid data from the data set and classifying the data set according to the national industry classification table;
[0038] An extraction module for inputting the industrial economic activities in the classified data set, extracting keywords in the industrial economic activities, and using a similarity matching algorithm to select a candidate industry classification word set and a pseudo-candidate industry classification word set from the national economic industry classification table and the data set;
[0039] A classification module for inputting the input economic activities, the candidate industry classification word set, and the pseudo-candidate industry classification word set into a preset BERT model and outputting a classification result through a softmax classifier;
[0040] An output module, which is used to establish a dictionary mainly based on the national economic industry classification table and the content of the training set, match the output classification results with the content in the dictionary, and output the final results.
[0041] Furthermore, the extraction module includes:
[0042] A cleaning module, which is used to clean the data in the obtained data set and delete the invalid information in the data;
[0043] A first division module, which is used to divide the cleaned data into small categories in the national economic industry classification table according to the information in the national economic industry classification table as the output;
[0044] A second division module, which is used to divide the data set into a training set, a test set and a validation set.
[0045] Compared with the prior art, the present invention has the following beneficial effects:
[0046] 1. It solves the problem of incorrect evaluation of the industry development status caused by incorrect industry classification, and avoids the consumption of time, manpower, etc.
[0047] 2. Understand the economic activity information of the industry from the semantic perspective, so that the enterprise information classification results are more in line with the industry classification table. Description of the Drawings
[0048] Figure 1 It is a flowchart of a method for classifying industry economic activities based on a semantic understanding model provided in the first embodiment;
[0049] Figure 2 It is a schematic diagram of the data processing flow provided in the first embodiment;
[0050] Figure 3 It is a BERT model diagram provided in the first embodiment. Specific Embodiments
[0051] The following uses specific specific examples to illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.
[0052] The purpose of the present invention is to provide a method and system for classifying industry economic activities based on a semantic understanding model in view of the defects of the prior art.
[0053] Example 1
[0054] This embodiment provides a classification method for industrial economic activities based on a semantic understanding model, as Figure 1 shown, including:
[0055] S1. Obtain a dataset corresponding to each industrial economic activity and its main product;
[0056] S2. Remove invalid data from the dataset and classify the dataset according to the national industrial classification table;
[0057] S3. Input the industrial economic activities in the classified dataset, extract keywords in the industrial economic activities, and use a similarity matching algorithm to select a candidate industrial classification term set and a pseudo-candidate industrial classification term set from the national economic industry classification table and the dataset;
[0058] S4. Input the input economic activities, the candidate industrial classification term set, and the pseudo-candidate industrial classification term set into a preset BERT model, and output a classification result through a softmax classifier;
[0059] S5. Establish a dictionary based on the national economic industry classification table and the content of the training set, match the output classification result with the content in the dictionary, and output the final result.
[0060] This embodiment aims at the difficulty in matching current social economic activity information with the national economic industry classification table. When matching according to the literal information of words and phrases in industry information, it often leads to a situation where one piece of information is matched to multiple industries. In response to such a situation, the existing classification method is further optimized, and a classification method for industrial economic activities based on a semantic understanding model is proposed, which deeply mines text information, considers classification from the semantic aspect of words, and matches each economic activity information to the correct classification.
[0061] In step S1, obtain a dataset corresponding to each industrial economic activity and its main product.
[0062] The dataset can crawl the required data from Flush, Shanghai Stock Exchange, and the official websites of listed companies, such as information representing industrial economic activities like property leasing, urea production, and washing machine wholesale, and store this information in a database. Use sql commands to delete the collected duplicate data to form the data basis for the system's preliminary training.
[0063] In step S2, remove invalid data from the dataset and classify the dataset according to the national industrial classification table; specifically:
[0064] S21. Clean the data in the obtained dataset and delete the invalid information in the data;
[0065] For the data in the database, by adding filtering conditions in the WHERE and using the SELECT command to find abnormal data in the database, such as data that will interfere with the experimental results, like pictures and punctuation marks, and then using the DELETE command in combination with the data found by SELECT to delete the abnormal data in the database.
[0066] S22. According to the information in the national economic industry classification table, divide the cleaned data into small categories in the national economic industry classification table as the output.
[0067] For the data that has been processed, it is necessary to manually specify the output of the data set in the early stage, which takes a lot of time. In this process, we specify the output as small categories in the industry classification table, classify the data set, and further filter the data when manually specifying the output.
[0068] S23. Divide the data set into a training set, a test set, and a validation set.
[0069] Use the K-fold cross-validation method to divide the data set into K parts. The so-called K-fold means dividing the data set K times so that all data appears in both the training set and the test set. In this example, we choose K = 5. Then divide it according to the ratio of training set:test set:validation set = 6:2:2. Use the training set to train our model, use the test set to verify the performance of the model, and use the validation set to examine the generalization performance of the test model, whether it is overfitting or underfitting.
[0070] In step S3, input the industrial economic activities in the classified data set, extract the keywords in the industrial economic activities, and use the similarity matching algorithm to select the candidate industry classification word set and the pseudo-candidate industry classification word set from the national economic industry classification table and the data set. Specifically:
[0071] S31. Use the TextRank algorithm to extract keywords from the industrial economic activities in the data set.
[0072] In natural language processing, the selection of keywords is very important. Good keywords help us quickly understand the semantics of the text. Here we use the TextRank algorithm to extract keywords. This algorithm is improved from Google's PageRank algorithm. It can extract the keywords of the text from the given text, expressed as:
[0073]
[0074] Among them, WS(V i ) represents the weight of phrase i, and the summation on the right represents the contribution degree of each adjacent phrase to the short text; W jirepresents the similarity between two sentences j and i; WS(V j ) represents the weight of phrase j obtained in the previous iteration; d represents the damping factor, generally 0.85; W jk represents the similarity between sentence j and sentence k; V k represents the node coefficient of sentence k.
[0075] S32. Use the Jaccard similarity algorithm for the extracted keywords to select the candidate industry classification word set and the pseudo-candidate industry classification word set.
[0076] In this step, a candidate word set is constructed for the input words, which will reduce the matching time and improve efficiency. Based on the literal information matching of words, semantic information is combined for allocation. The steps are as follows: First, use the Jaccard similarity algorithm to match the input information, traverse the subcategories in the national economic industry classification table. For example, if the industrial information is iron ore concentrate smelting, use the Jaccard algorithm to search for candidate words from the subcategories literally, and candidate words such as iron ore mining, iron smelting, ferroalloy smelting, etc. can be found. Sort them according to the similarity size and select the top 30 candidate words to form the candidate industry classification words; To make full use of the hidden information in the training set, also use the Jaccard similarity algorithm to match the input information with the training set information. For example, if the industrial information is sika deer breeding and the keyword is deer, but no words related to deer can be found in the subcategories. However, in the training set, there is information such as red deer breeding, sika deer breeding, white-lipped deer breeding, etc. Sort them according to the similarity size and select the top 30 matching words to form the pseudo-candidate industry classification words. This step takes into account that it is sometimes difficult to directly match the input information with the subcategories in the classification table. For example, it is difficult to directly match sika deer breeding to other livestock breeding. But if there is an intermediate word as a jump, such as male sika deer breeding, which corresponds to other livestock breeding in the training set, then sika deer breeding can be indirectly matched to other livestock breeding. This measure will improve the matching accuracy and efficiency.
[0077] Such as Figure 2 shown in the schematic diagram of the data processing flow.
[0078] In step S4, the input economic activity, the candidate industry classification word set, and the pseudo-candidate industry classification word set are input into a preset BERT model, and the classification result is output through a softmax classifier. Specifically:
[0079] S41. Combine the input economic activity with the candidate industry classification words, separate them using the sentence pair classification task SEP, and classify them through the single text classification task CLS to obtain the input information, and the input information is used as the input of the next fully connected layer;
[0080] SEP and CLS are encodings in BERT:
[0081] [CLS] Single Text Classification Task: For the text classification task, the BERT model inserts a [CLS] symbol at the beginning of the text, and uses the output vector corresponding to this symbol as the semantic representation of the entire text for text classification. It can be understood that: compared with other existing words in the text, this symbol without obvious semantic information will more "fairly" integrate the semantic information of each word in the text. (CLS is not the representation of the first sentence, but a classification flag. Through self-attention, it can integrate the information related to classification of two sentences and then be used in pre-training or subsequent task classification).
[0082] [SEP] Sentence Pair Classification Task: The actual application scenarios of this task include: question answering (judging whether a question matches an answer), sentence matching (whether two sentences express the same meaning), etc. For this task, in addition to adding a [CLS] symbol and using the corresponding output as the semantic representation of the text, the BERT model also uses a [SEP] symbol to separate the two input sentences and attaches two different text vectors to the two sentences respectively for distinction.
[0083] In this embodiment, the information after [CLS] is put into the input information and used as the input of the fully connected layer, and [SEP] is used to separate the input information from the candidate set information; the input of the input layer of the BERT model is constructed.
[0084] S42. Input the obtained input information into the input layer of the preset BERT model for processing, and classify the processed information using a softmax classifier to obtain the output result.
[0085] Using the BERT pre-trained model as the model of this embodiment, because this model has the function of semantic recognition. For example, in daily life, the character "non" is common. For instance, the economic activity name is "manufacture of engines not for motor vehicles". Literally, it is very similar to the minor category "manufacture of engines for motor vehicles" in the industry classification table. However, from the semantic perspective, this economic activity is completely different from this classification. And the BERT model can exactly distinguish from the semantics. Take the input of the BERT model as the input of the softmax classifier for classification, and denote the output result as y. The calculation formula is:
[0086]
[0087] where y represents the softmax value of the output result; e i represents the exponential of the i-th element; ∑ j e j represents the sum of the exponentials of all elements. That is to say, the output result is the ratio of the exponential of this element to the sum of the exponentials of all elements.
[0088] As shown Figure 3 in the figure is the BERT model diagram.
[0089] In step S5, a dictionary is established based on the national economic industry classification table and the content of the training set, and the output classification result is matched with the content in the dictionary to output the final result. Specifically:
[0090] S51. Establish a dictionary mainly based on the data in the national economic industry classification table, use the data in the training set to expand the number of sub-categories, and associate the data in the training set with the sub-categories;
[0091] Establish a dictionary and fill it with the content in the national economic industry classification table as the content in the dictionary, as shown in Table 1:
[0092]
[0093] Table 1 Display of dictionary content
[0094] Use the valid information in the training set to supplement the sub-categories. Only supplement the sub-category information, and do not supplement the class, major category, and middle category, which is for the convenience of the second-round matching. Some information is shown in Table 2.
[0095] Subclass Aircraft manufacturing Onshore oil extraction Wheat cultivation Sika deer breeding Pigeon breeding Vanadium ore mining and dressing
[0096] Table 2 Training set + sub-category information
[0097] S52. Put the output classification result into the dictionary for matching and output the final result.
[0098] Determine the next operation according to the final output of the model. If the output result is a sub-category, reverse-lookup the middle category, major category, and class through the dictionary. For example, if the matching result is "wheat planting", according to the dictionary content, it can be deduced that its corresponding category is "grain planting", the major category is "agriculture", and the class is "agriculture, forestry, animal husbandry, and fishery"; if the output result is the content of the training set, then perform the second-round matching. For example, if the matching result is "sika deer breeding", according to the classification result in the training set, it is "other livestock breeding", and look up the corresponding sub-category according to the dictionary content, and reverse-lookup the middle category, major category, and class. The calculation formula is:
[0099]
[0100] Among them, result represents the final output result; dic represents the dictionary; if the matching result y' is a candidate industry classification word, the final output result is y'; if the matching result is a pseudo-candidate industry classification word y', then perform the second-round matching, and use the result y' of the first-round matching as the input of the second-round matching to obtain the final output.
[0101] Compared with the existing technology, this embodiment has the following advantages:
[0102] 1. It solves the problem of incorrect evaluation of the industry development status caused by incorrect industry classification, and avoids the consumption of time, manpower, etc.
[0103] 2. Understanding the economic activity information of the industry from the semantic perspective makes the enterprise information classification result more consistent with the industry classification table.
[0104] Embodiment Two
[0105] This embodiment provides an industry economic activity classification system based on a semantic understanding model, including:
[0106] An acquisition module, configured to acquire a data set corresponding to each industry economic activity and the main products.
[0107] A removal module, configured to remove invalid data from the data set and classify the data set according to the national industry classification table.
[0108] An extraction module, configured to input the industrial economic activities in the classified data set, extract keywords in the industrial economic activities, and select a candidate industry classification word set and a pseudo-candidate industry classification word set from the national economic industry classification table and the data set using a similarity matching algorithm.
[0109] A classification module, configured to input the input economic activities, the candidate industry classification word set, and the pseudo-candidate industry classification word set into a preset BERT model, and output a classification result through a softmax classifier.
[0110] An output module, configured to establish a dictionary mainly based on the national economic industry classification table and the content of the training set, match the output classification result with the content in the dictionary, and output the final result.
[0111] Furthermore, the removal module includes:
[0112] A cleaning module, configured to clean the data in the acquired data set and delete invalid information in the data.
[0113] A first partitioning module, configured to partition the cleaned data into small categories in the national economic industry classification table as the output according to the information in the national economic industry classification table.
[0114] A second partitioning module, configured to partition the data set into a training set, a test set, and a validation set.
[0115] It should be noted that the industry economic activity classification system based on a semantic understanding model provided in this embodiment is similar to Embodiment One, and will not be elaborated here.
[0116] Compared with the prior art, the present embodiment has the following beneficial effects:
[0117] 1. It solves the problem of incorrect evaluation of the industry development status caused by incorrect industry classification, and avoids the consumption of time, manpower, etc.
[0118] 2. Understanding the economic activity information of the industry from a semantic perspective makes the enterprise information classification result more consistent with the industry classification table.
[0119] Note that the above is only the preferred embodiment of the present invention and the technical principles applied. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein. Various obvious changes, re-adjustments, and substitutions can be made by those skilled in the art without departing from the protection scope of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments. Without departing from the concept of the present invention, more other equivalent embodiments can be included, and the scope of the present invention is determined by the scope of the appended claims.
Claims
1. A classification method for industrial economic activities based on a semantic understanding model, characterized in that, Including: S1. Obtain the dataset corresponding to the economic activities and main products in each industry; S2. Remove the invalid data in the dataset and classify the dataset according to the national industry classification table; S3. Input the industrial economic activities in the classified dataset, extract the keywords in the industrial economic activities, and use the similarity matching algorithm to select the candidate industry classification word set and the pseudo-candidate industry classification word set from the national economic industry classification table and the dataset; S31. Use the TextRank algorithm to extract keywords from the industrial economic activities in the dataset; S32. Use the Jaccard similarity algorithm for the extracted keywords to select the candidate industry classification word set and the pseudo-candidate industry classification word set; Use the Jaccard similarity algorithm to match the extracted keywords, traverse the sub-categories in the national economic industry classification table, first search from the sub-categories literally, sort according to the similarity size, and select the top 30 candidate words to form the candidate industry classification word set; Use the Jaccard similarity algorithm to match the extracted keywords with the training dataset, sort according to the similarity size, and select the top 30 matching words to form the pseudo-candidate industry classification word set; S4. Input the input economic activities, candidate industry classification word set, and pseudo-candidate industry classification word set into the preset BERT model, and output the classification result through the softmax classifier; S41. Put the input economic activities and candidate classification word set after the single text classification task CLS, and then separate the economic activities and candidate classification word set using the sentence pair classification task SEP, so as to construct the input information of the input layer of the BERT model; among them, the candidate classification word set includes the candidate industry classification word set and the pseudo-candidate industry classification word set; S42. Input the input information into the input layer of the preset BERT model for processing, and classify the processed information using the softmax classifier to obtain the output result; S5. Establish a dictionary based on the national economic industry classification table and the content of the training set, match the output classification result with the content in the dictionary, and output the final result.
2. The classification method of industrial economic activities based on a semantic understanding model according to claim 1, characterized in that The specific steps of S2 are as follows: S21. Clean the data in the obtained dataset and delete the invalid information in the data; S22. According to the information in the national economic industry classification table, divide the cleaned data into the sub-categories in the national economic industry classification table as the output; S23. Divide the dataset into a training set, a test set, and a validation set.
3. The industrial economic activity classification method based on a semantic understanding model according to claim 2, wherein, The use of the TextRank algorithm for keyword extraction in step S31 is expressed as: Among them, WS(V i ) represents the weight of phrase i; W ji represents the similarity between two sentences j and sentence i; WS(V j ) represents the weight of phrase j obtained in the previous iteration; d represents the damping coefficient; W jk represents the similarity between sentence j and sentence k; V k represents the node coefficient of sentence k.
4. A classification method for industrial economic activities based on a semantic understanding model according to claim 3, characterized in that The output result in step S42 is expressed as: Among them, y represents the softmax value of the output result; e i represents the exponent of the i-th element; ∑ j e j represents the sum of the exponents of all elements. That is to say, the output result is the ratio of the exponent of this element to the sum of the exponents of all elements.
5. A method for classifying industrial economic activities based on a semantic understanding model according to claim 2, characterized in that The specific steps of S5 are as follows: S51. Establish a dictionary based on the data in the national economic industry classification table, use the data in the training set to expand the number of sub-categories, and associate the data in the training set with the sub-categories; S52. Put the output classification result into the dictionary for matching and output the final result.
6. The industrial economic activity classification method based on a semantic understanding model according to claim 5, characterized in that The output of the final result in step S52 is expressed as: result = match(dic, y) Among them, result represents the final output result; dic represents the dot.
7. An industrial economic activity classification system based on a semantic understanding model, characterized in that, Including: An acquisition module for acquiring datasets corresponding to economic activities in various industries and main products; A removal module for removing invalid data from the dataset and classifying the dataset according to the national industry classification table; An extraction module for inputting industrial economic activities in the classified dataset, extracting keywords in the industrial economic activities, and using a similarity matching algorithm to select a candidate industry classification word set and a pseudo-candidate industry classification word set from the national economic industry classification table and the dataset; Using the TextRank algorithm to extract keywords from the industrial economic activities in the dataset; Using the Jaccard similarity algorithm for the extracted keywords to select a candidate industry classification word set and a pseudo-candidate industry classification word set; Using the Jaccard similarity algorithm to match the extracted keywords, traversing the subclasses in the national economic industry classification table, first searching from the subclasses literally, sorting according to the similarity size, and selecting the top 30 candidate words to form a candidate industry classification word set; Using the Jaccard similarity algorithm to match the extracted keywords with the training dataset, sorting according to the similarity size, and selecting the top 30 matching words to form a pseudo-candidate industry classification word set; A classification module for inputting the input economic activities, candidate industry classification word set, and pseudo-candidate industry classification word set into a preset BERT model and outputting a classification result through a softmax classifier; After the single text classification task CLS, put in the input economic activities and candidate classification word set, and then separate the economic activities and candidate classification word set using the sentence pair classification task SEP, thereby constructing the input information of the input layer of the BERT model; among them, the candidate classification word set includes the candidate industry classification word set and the pseudo-candidate industry classification word set; Input the input information into the input layer of the preset BERT model for processing, and classify the processed information using a softmax classifier to obtain an output result; An output module for establishing a dictionary mainly based on the national economic industry classification table and the content of the training set, matching the output classification result with the content in the dictionary, and outputting the final result.
8. An industrial economic activity classification system based on a semantic understanding model according to claim 7, characterized in that, The removal module includes: A cleaning module for cleaning the data in the acquired dataset and deleting the invalid information in the data; A first partitioning module for partitioning the cleaned data into subclasses in the national economic industry classification table as output according to the information in the national economic industry classification table; A second partitioning module for partitioning the dataset into a training set, a test set, and a validation set.
Citation Information
Patent Citations
Taxpayer industry classification method based on noise label learning
CN112765358A
Knowledge graph based national economy industry classification method and system
CN113377957A