Label marking model for deprotection business field of supervision official document

Abstract

The invention provides a deprotection business field label marking model for supervision official documents, and relates to the technical fields of data mining, text keyword extraction and natural language processing. The label marking model for the insurance elimination business field of the supervision official document specifically comprises the following steps of S1, keyword extraction, S101, model training, S102, model application, S103, keyword output, S2, keyword normalization, S201, similar keyword extraction, S202, similarity and similar word calculation, S203, key phrase and segmented word output, S3, text classification, S301, classifier establishment, S302, tradeoff of accuracy and recall ratio, and S303, classification result output. On the basis of a deprotection word library and a deprotection compliance judgment rule library, A < l > document pre-auditing with a model algorithm as a core is adopted, document analysis, key point stripping and standard extraction are achieved, the problems that materials are too many and cannot be checked completely, key points are too fine and cannot be checked, and content is too special and cannot be checked are solved, the breadth and depth of deprotection work are improved, timeliness and accuracy are improved, and the work efficiency is improved. The professional threshold is simplified.

Description

Technical Field

[0001] The present invention relates to the technical fields of data mining, text keyword extraction, and natural language processing, and specifically to a label marking model for the consumer protection business field of regulatory documents. Background Art

[0002] As the financial industry attaches more importance to consumer rights protection, various group financial companies need to pay timely attention to various official documents issued by the China Banking and Insurance Regulatory Commission and the People's Bank of China. They need to be equipped with professionals from multiple different business fields to pay attention to and interpret the documents, and send them to the company's corresponding business departments for response and processing.

[0003] Most of the existing models use OCR to automatically recognize the content of documents, and use a combination of Jieba word segmentation, LDA+TF-IDF, text similarity, and financial and consumer protection vocabulary to identify and mark the business areas targeted by regulatory documents. Based on the recognition results, the documents are automatically sent to the corresponding business departments within the company for response and processing. Overall, the use of these models requires a high level of professionalism from the user, and the timeliness and accuracy of overall use are poor. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a label marking model for the consumer protection business field of regulatory documents, which solves the problems of high professional threshold requirements, poor timeliness and accuracy when using the existing label marking model.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a label marking model for the consumer protection business field of regulatory documents specifically includes the following steps:

[0006] S1. Keyword extraction

[0007] S101. Model training;

[0008] S102. Model application;

[0009] S103. Keyword output;

[0010] S2. Keyword Normalization

[0011] S201. Similar keyword extraction;

[0012] S202. Calculate similarity and similar words;

[0013] S203. Output of key phrases and word segmentation;

[0014] S3. Text classification

[0015] S301. Classifier establishment;

[0016] S302. Accuracy and recall trade-off;

[0017] S303. Output classification results.

[0018] Preferably, the S1 specifically includes the following process:

[0019] S101. Model training

[0020] Obtain text files in the field of consumer protection, input the text files in the field of consumer protection into the first-order CRF model, train the CRF model using the ADF training method, and after training, convert the CRF model into a word segmentation model in the field of consumer protection for output;

[0021] S102. Model application

[0022] Input the custom dictionary into the word segmentation model in the consumer protection field, match the proprietary words in the custom dictionary with the word segmentation model in the consumer protection field through the dictionary tree, and after the matching is completed, input other texts into the word segmentation model in the consumer protection field to generate annotations, and after completion, output the word segmentation with reference to the feature model;

[0023] S103.Keyword output

[0024] The output word segmentation is tagged with parts of speech and stop words are removed. The processed word segmentation and new documents are subjected to text preprocessing. The weights are calculated using the TE / IDF and LDA main models respectively. The weights calculated by the two are combined. After the combination is completed, the keywords are fragmented, the keywords of adjacent fragments are merged, and the key phrases are output.

[0025] Preferably, the S2 specifically includes the following process:

[0026] S201. Similar keyword extraction

[0027] Obtain key phrases from new documents and input the key phrases into the model for screening. After screening, calculate the similarity between keywords provided by Ping An. After calculation, sort the similar words.

[0028] S202. Calculate similarity and similar words

[0029] Obtain a key phrase list and a given keyword list, and output a candidate normalized keyword list through a similarity calculation method based on string fuzzy matching and a text classification structure, combined with a word extraction method based on similarity calculation;

[0030] S203. Key phrase and word segmentation output

[0031] The similarity calculation is performed based on the key phrases. After the calculation is completed, the key phrases and word segments are classified and output. The key phrases and word segments are specifically classified into the following categories:

[0032] ①Key phrase (length 4) + highest similarity;

[0033] ②Key phrases (length 4) + only consider key phrases containing the keywords given by Ping An + with the highest similarity;

[0034] ③Key phrases (length 6) + only consider key phrases containing the keywords given by Ping An + with the highest similarity;

[0035] ④Key phrase (length 4) + only consider the key phrase containing the keywords given by Ping An after splitting + the highest similarity:

[0036] ⑤ Word segmentation + deduplication + highest similarity;

[0037] ⑥Word segmentation + deduplication + only consider the results of splitting the key phrases of the keywords given by Ping An + the highest similarity.

[0038] Preferably, the S3 specifically includes the following process:

[0039] S301. Classifier establishment

[0040] By combining a variety of different algorithms, selecting the best one, and combining ensemble learning to synthesize the results of multiple weak classifiers, a classifier is established;

[0041] S302. Precision and recall trade-off

[0042] Sacrificing classification accuracy and prioritizing recall, the classification accuracy is reduced to about 93% and the recall is increased to more than 90%;

[0043] S303. Classification result output

[0044] Combine the results of multiple classifiers, handle the null value problem, and select the most likely equity when multiple classifiers do not output results, giving priority to ensuring that at least one equity classification is output.

[0045] Preferably, in S201, the key phrase screening selects the top 200 by weight, the phrase length is less than or equal to 4, and the similar words are sorted to select the top 10 by weight.

[0046] Preferably, in S201, keyword normalization follows the following principles: the weight in the key phrase is relatively large, the length is moderate, and it is as similar as possible to the keywords provided by Ping An, with the goal of extracting words related to the consumer protection field.

[0047] Preferably, in S202, the similarity calculation method based on string fuzzy matching is implemented by recursively searching for the longest continuous matching subsequence, the candidate word list is the candidate words arranged in order of similarity, the text classification result can be used as input for similarity calculation, and similar words under the specified equity tag are output.

[0048] Preferably, in S301, the multiple algorithms include SVM, decision tree, and Bayesian method.

[0049] The present invention provides a label marking model for the consumer protection business field of regulatory documents. It has the following beneficial effects:

[0050] 1. The present invention provides a labeling model for the consumer protection business field of regulatory documents. The model is based on a consumer protection vocabulary library and a consumer protection compliance judgment rule library, and adopts an AI document pre-examination with this model algorithm as the core, to achieve document parsing, stripping of key points, and refinement standards, solving the problems of too much material to check, too detailed key points to check, and too specialized content to check, thereby improving the breadth and depth of consumer protection work, improving timeliness and accuracy, and simplifying professional barriers. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A schematic diagram of the process of model training of the present invention;

[0052] Figure 2 A schematic diagram of a flow chart of the model application of the present invention;

[0053] Figure 3 A flowchart of keyword output of the present invention;

[0054] Figure 4 A schematic diagram of the process of similar keyword extraction of the present invention;

[0055] Figure 5 A schematic diagram of a process for calculating similarity and similar words of the present invention;

[0056] Figure 6 It is a flowchart of text classification of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] like Figure 1-6 As shown, the embodiment of the present invention provides a label marking model for the consumer protection business field of regulatory documents, which specifically includes the following steps:

[0059] S1. Keyword extraction

[0060] S101. Model training

[0061] Obtain text files in the field of consumer protection, input the text files in the field of consumer protection into the first-order CRF model, train the CRF model through the ADF training method, and after training, convert the CRF model into a word segmentation model in the field of consumer protection for output, which is conducive to providing a basis for the subsequent use of the model;

[0062] S102. Model application

[0063] Input the custom dictionary into the word segmentation model in the field of consumer protection, match the proprietary words in the custom dictionary with the word segmentation model in the field of consumer protection through the dictionary tree, and after the matching is completed, input other texts into the word segmentation model in the field of consumer protection to generate annotations. After completion, output the word segmentation with reference to the feature model, which is conducive to realizing the use function of the model;

[0064] S103.Keyword output

[0065] The output word segmentation is tagged with parts of speech and stop words are removed. The processed word segmentation and new documents are preprocessed, and weights are calculated using the TE / IDF and LDA main models respectively. The weights calculated by the two are combined. After the combination is completed, the keywords are fragmented, and the keywords of adjacent fragments are merged to output key phrases to ensure that the model can accurately output keywords.

[0066] S2. Keyword Normalization

[0067] S201. Similar keyword extraction

[0068] Obtain key phrases from new documents and input the key phrases into the model for screening. After screening, calculate the similarity of words based on the keywords provided by Ping An. After the calculation is completed, sort the similar words to facilitate subsequent similarity calculations.

[0069] S202. Calculate similarity and similar words

[0070] Obtain a list of key phrases and a given keyword list, and output a candidate normalized keyword list through a similarity calculation method based on string fuzzy matching and a text classification structure, combined with a word extraction method based on similarity calculation, to facilitate subsequent extraction of key phrases;

[0071] S203. Key phrase and word segmentation output

[0072] The similarity calculation is performed based on the key phrases. After the calculation is completed, the key phrases and word segments are classified and output. The key phrases and word segments are specifically classified into the following categories:

[0073] ①Key phrase (length 4) + highest similarity;

[0074] ②Key phrases (length 4) + only consider key phrases containing the keywords given by Ping An + with the highest similarity;

[0075] ③Key phrases (length 6) + only consider key phrases containing the keywords given by Ping An + with the highest similarity;

[0076] ④Key phrase (length 4) + only consider the key phrase containing the keywords given by Ping An after splitting + the highest similarity:

[0077] ⑤ Word segmentation + deduplication + highest similarity;

[0078] ⑥ Segmentation + deduplication + only consider the results of key phrase splitting with the keywords given by Ping An + the highest similarity, so that the model can output key phrases and segmentation in different forms according to classification;

[0079] S3. Text classification

[0080] S301. Classifier establishment

[0081] By combining a variety of different algorithms, selecting the best one, and combining ensemble learning to integrate the results of multiple weak classifiers, a classifier is established, which is conducive to providing a basis for subsequent classification;

[0082] S302. Precision and recall trade-off

[0083] Sacrificing classification accuracy and giving priority to recall, the classification accuracy is reduced to about 93%, and the recall is increased to more than 90%, which is conducive to improving the recall while ensuring the classification accuracy;

[0084] S303. Classification result output

[0085] By integrating the results of multiple classifiers and handling the null value problem, when multiple classifiers do not output results, the most likely equity is selected, and at least one equity classification is output first, so as to realize the model's function of marking labels.

[0086] In S201, the key phrase screening selects the top 200 with a phrase length of less than or equal to 4, and the similar words sorting selects the top 10 with a weight, which is conducive to screening out the key phrases that best meet the requirements. The similarity calculation method based on string fuzzy matching is implemented by recursively finding the longest continuous matching subsequence. The candidate word list is the candidate words arranged in order of similarity. The text classification result can be used as the input of the similarity calculation, and the similar words under the specified equity label are output, which is conducive to realizing the main usage functions of the model. In S301, multiple algorithms including SVM, decision tree, and Bayesian method are conducive to improving the algorithm effect. The classification accuracy of the algorithm is increased from 87.6% to 96%.

[0087] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. Labeling model for consumer protection business areas of regulatory documents, It is characterized in that The specific steps include: S1. Keyword extraction S101. Model training; S102. Model application; S103. Keyword output; S2. Keyword Normalization S201. Similar keyword extraction; S202. Calculate similarity and similar words; S203. Output of key phrases and word segmentation; S3. Text classification S301. Classifier establishment; S302. Accuracy and recall trade-off; S303. Output classification results.

2. The consumer protection business field labeling model for regulatory documents according to claim 1, Features: In S1, specifically The process includes: S101. Model training Obtain text files in the field of consumer protection, input the text files in the field of consumer protection into the first-order CRF model, train the CRF model using the ADF training method, and after training, convert the CRF model into a word segmentation model in the field of consumer protection for output; S102. Model application Input the custom dictionary into the word segmentation model in the consumer protection field, match the proprietary words in the custom dictionary with the word segmentation model in the consumer protection field through the dictionary tree, and after the matching is completed, input other texts into the word segmentation model in the consumer protection field to generate annotations, and after completion, output the word segmentation with reference to the feature model; S103.Keyword output The output word segmentation is tagged with parts of speech and stop words are removed. The processed word segmentation and new documents are subjected to text preprocessing. The weights are calculated using the TE / IDF and LDA main models respectively. The weights calculated by the two are combined. After the combination is completed, the keywords are fragmented, the keywords of adjacent fragments are merged, and the key phrases are output.

3. The consumer protection business field labeling model for regulatory documents according to claim 1, Features: The S2 specifically includes the following processes: S201. Similar keyword extraction Obtain key phrases from new documents and input the key phrases into the model for screening. After screening, calculate the similarity between keywords provided by Ping An. After calculation, sort the similar words. S202. Calculate similarity and similar words Obtain a key phrase list and a given keyword list, and output a candidate normalized keyword list through a similarity calculation method based on string fuzzy matching and a text classification structure, combined with a word extraction method based on similarity calculation; S203. Key phrase and word segmentation output The similarity calculation is performed based on the key phrases. After the calculation is completed, the key phrases and word segments are classified and output. The key phrases and word segments are specifically classified into the following categories: ①Key phrase (length 4) + highest similarity; ②Key phrases (length 4) + only consider key phrases containing the keywords given by Ping An + with the highest similarity; ③Key phrases (length 6) + only consider key phrases containing the keywords given by Ping An + with the highest similarity; ④Key phrase (length 4) + only consider the key phrase containing the keywords given by Ping An after splitting + the highest similarity: ⑤ Word segmentation + deduplication + highest similarity; ⑥Word segmentation + deduplication + only consider the results of splitting the key phrases of the keywords given by Ping An + the highest similarity.

4. The consumer protection business field labeling model for regulatory documents according to claim 1, Features: In S3, specifically The process includes: S301. Classifier establishment By combining a variety of different algorithms, selecting the best, and combining ensemble learning to synthesize the results of multiple weak classifiers, a classifier is established; S302. Precision and recall trade-off Sacrificing classification accuracy and giving priority to recall, the classification accuracy is reduced to about 93% and the recall is increased to more than 90%; S303. Classification result output Combine the results of multiple classifiers, handle the null value problem, and select the most likely equity when multiple classifiers do not output results, giving priority to ensuring that at least one equity classification is output.

5. The consumer protection business field labeling model for regulatory documents according to claim 3, Features: In S201, the key phrase screening selects the top 200 by weight, the phrase length is less than or equal to 4, and the similar words are sorted to select the top 10 by weight.

6. The consumer protection business field labeling model for regulatory documents according to claim 3, Features: In S201, keyword normalization follows the following principles: the weight in the key phrase is relatively large, the length is moderate, and it is as similar as possible to the keywords provided by Ping An, with the goal of extracting words related to the consumer protection field.

7. The consumer protection business field labeling model for regulatory documents according to claim 3, Features: In S202, the similarity calculation method based on string fuzzy matching is implemented by recursively searching for the longest continuous matching subsequence. The candidate word list is the candidate words arranged in order of similarity. The text classification result can be used as the input of the similarity calculation, and the similar words under the specified equity tag are output.

8. The consumer protection business field labeling model for regulatory documents according to claim 4, Features: In S301, the multiple algorithms include SVM, decision tree, and Bayesian method.

Images