Structured data generation model based on natural language processing

Through the structured data generation model based on the BERT language model, the problems of data bias and insufficient interpretability in natural language processing are solved, more efficient data analysis and semantic understanding are achieved, and manual repetitive labor and legal risks are reduced.

CN118940719AActive Publication Date: 2024-11-12GUANGDONG POWER GRID CO LTD +1

Patent Information

Application Number
CN202410854943.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-28
Publication Date
2024-11-12
Estimated Expiration
2044-06-28

AI Technical Summary

Technical Problem

The existing structured data generation model of natural language processing has problems of data bias, unfairness and insufficient interpretability.

Method used

Using a structured data generation model based on the BERT language model, including keyword extraction module, similarity comparison module and scoring module, documents are identified and analyzed through the BERT language model, structured data is generated, and semantics are combined with RNN recurrent neural network. Understand, capture the context information of words, perform data fusion and analysis, identify similar paragraphs in the document and highlight different clauses, improve the real-time and dynamic nature of data analysis, and reduce data bias and manual repetition.

Benefits of technology

It improves the real-time and dynamic nature of the system's data analysis methods, reduces the limitations of semantic understanding and data deviations, reduces the time cost of manual mistakes and repeated labor, and enhances the credibility and accuracy of data analysis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940719A_ABST
    Figure CN118940719A_ABST
Patent Text Reader

Abstract

The invention discloses a structured data generation model based on natural language processing. The structured data generation model is used for solving the problems of limitation of semantic understanding, language diversity, data deviation and unfairness, insufficient interpretability and the like in the prior art. Comprising a keyword extraction module, a similarity comparison module, a scoring module, a text preprocessing module, a recognition processing module, a credibility judgment module and a report generation module. Intelligent processing and demand analysis of big data in a specific working environment are achieved, meanwhile, the time cost wasted by manual work in mechanical repeated work is reduced, meanwhile, legal risks caused by manual errors are reduced, and the problem of low efficiency of repeated work is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of natural language processing, and in particular to a structured data generation model based on natural language processing. Background Art

[0002] Natural language processing (NLP) is an interdisciplinary subject in the fields of artificial intelligence and linguistics. It studies various theories and methods that can achieve effective communication between humans and computers using natural language. The applications of natural language processing include machine translation, public opinion monitoring, automatic summarization, opinion extraction, text classification, question answering, text semantic comparison, speech recognition, Chinese OCR, etc. The applications of natural language processing are very wide, including intelligent customer service, search engine optimization, social media analysis, intelligent writing assistants, etc.

[0003] Natural language processing has been widely and deeply applied in management and production. It provides enterprises with more efficient and accurate data processing and analysis tools, thereby optimizing management and production processes. For example, natural language processing technology can perform sentiment analysis on text data such as corporate brand reputation, product evaluation, and customer feedback. This analysis helps enterprises to understand market feedback in a timely manner, improve and optimize the company, thereby improving customer satisfaction and brand loyalty. At the same time, natural language processing can automatically extract specific types of information from text, such as human resources information, marketing information, economic data, etc. This enables enterprises to obtain the latest and most comprehensive information at the first time, greatly improving decision-making efficiency.

[0004] Although NLP has been widely used in management and production, there are still some challenges and limitations, such as limitations in semantic understanding, language diversity issues, data bias and unfairness, lack of interpretability, etc. Therefore, when applying NLP technology, enterprises need to fully consider these factors and make choices and adjustments based on their actual conditions. Summary of the invention

[0005] The purpose of the present invention is to propose a structured data generation model based on natural language processing in order to solve the problems of data bias and unfairness, lack of interpretability, etc. in the existing structured data generation model for natural language processing.

[0006] The purpose of the present invention can be achieved through the following technical scheme: a structured data generation model based on natural language processing, including a business subsystem, a customer service platform subsystem, a demand retrieval module and a database; wherein the business subsystem includes a keyword extraction module, a similarity comparison module and a scoring module, and the customer service platform subsystem includes a text preprocessing module, a recognition processing module, a credibility determination module, and a report generation module.

[0007] The keyword extraction module classifies the document and uses the BERT language model to identify the entire document. It captures type judgment elements in the document. Type judgment elements include keywords such as "rights", "fees", "terms", "confidentiality", "effectiveness", and change images in the change order. The frequency N of occurrence of various type judgment elements in the entire document is recorded. Each type judgment element corresponds to a preset judgment weight λ. The number of occurrences of various type judgment elements is multiplied by the judgment weight and then summed to obtain the type judgment parameter α. Each document type corresponds to a specific type judgment parameter range. Document types include bids, response documents, business contracts, change orders, and administrative contracts.

[0008] For each document type, the corresponding regular expression is used to extract keywords from the document, and the keywords are combined into data vectors to generate structured data.

[0009] For example, extract keywords about "time", "project name", "construction scale" and "contract estimated price" in the bidding documents, extract keywords about certificate requirements, professional title requirements, qualification requirements and reputation requirements in the bidder qualification part, extract keywords about "the maximum bid limit price compiled by the tenderer", and form all keywords into a data vector. The keyword extraction module scans several bidding documents, generates several data vectors, and forms a structured data table.

[0010] The similarity comparison module compares the full text of documents of the same document type, marks the semantically different words in similar paragraphs as difference characters, uses the BERT language model to analyze the meaning of the difference characters and match a reference character, generates the difference character meaning vector X = (x1, x2, x3, ..., xn) and the reference character meaning vector Y = (y1, y2, y3, ..., yn) where n is the number of dimensions contained in the coordinate system where the meaning vector is located, and the value of n depends on the learning depth of the BERT language model; analyze the words of the difference characters Parts of speech include proper nouns, general nouns, verbs, adjectives, modal particles, function words, etc. Each part of speech corresponds to a preset part of speech factor e1; analyze the location of the difference text, the location includes the document directory, announcement chapter, appendix chapter, instructions chapter, general provisions, text, etc. The BERT language model uses the RNN recurrent neural network to learn and obtain the accurate location category. Each location corresponds to a location factor e2; analyze the context and context of the difference text, and obtain a context factor e3 based on the BERT language model analysis. Through the formula The difference score E is obtained, where λ1, λ2, λ3, and λ4 are preset weight factors. The difference characters with a difference score E greater than 100 are highlighted, and the difference score E is added to the data vector itself.

[0011] The scoring module analyzes the data vector generated by the keyword extraction module and presets a scoring factor for each element in the data vector. For example, a scoring factor 1 is preset for the project name element. When the element contains city A, the scoring factor 1 is a1, and when the element contains city B, the scoring factor 1 is a2. A scoring factor 2 is preset for the contract estimate price element. The scoring factor 2 for elements greater than 1 million is b2, and the scoring factor 2 for elements less than 1 million is b2, and so on. All scoring factors are summed to obtain a scoring parameter T, and the scoring parameter of each data vector is also added to the data vector itself to become a new element, completing the scoring step 1. The scoring module is based on the difference score of each data vector obtained by the similarity comparison module, and the difference score is summed to obtain a difference parameter K. The difference parameter of each data vector is also added to the data vector itself to become a new element, completing the scoring step 2.

[0012] The scoring module inputs all data vectors into the database after scoring step one and scoring step two are completed.

[0013] The text preprocessing module preprocesses the customer service feedback information such as letters, complaints, and complaints received by the customer service platform subsystem. First, the information source of the customer service feedback information is obtained, the name of the sender and the time of delivery are obtained, and the name of the sender is encrypted using the encryption algorithm AES. Then, each Chinese character and English character in the customer service feedback information is split into separate strings, and the XLNet natural regression pre-trained language model is used to obtain the word probability P of each single character and the two adjacent characters before and after. When the sum of the word probabilities of the two adjacent characters is greater than the preset probability threshold When , the two characters are combined into a word, and the meaning vector of the word is obtained in the language model. If the meaning vector of the word belongs to the category of defamatory and extremely negative adjectives, the word is marked as useless information, the number of characters contained in the useless information is recorded as the information shielding parameter k, and the characters of the useless information are replaced by *. Then the pre-processed customer service feedback information is sent to the recognition processing module in the form of a string.

[0014] The recognition processing module uses the character string sent by the text preprocessing module to obtain forum replies and comments on the Internet as training data for training. Based on the natural processing (NLP) technology, the emotional tendencies contained in each sentence are analyzed to form a segmented emotional dictionary, and the emotional tendencies are divided into three first-level emotional tendencies: positive, negative, and neutral. Each first-level emotional tendency contains segmented second-level emotional tendencies. For example, positive emotional tendencies can be segmented into gratitude, excitement, hope, etc. Each emotional tendency corresponds to a preset emotional degree parameter. The emotional degree parameter of positive emotional tendencies is a positive number, the emotional degree parameter of negative emotional tendencies is a negative number, and the emotional degree parameter of neutral emotional tendencies is 0. The recognition processing module extracts the meaning vector of each word in the character string, understands the word's part of speech, meaning, and emotional tendency based on the BERT language model, obtains the emotional degree parameter of each word, and sums the emotional degree parameters of each customer service feedback information to obtain the emotional index k1.

[0015] The recognition module obtains the key parameters of each word in each customer service feedback information based on the BERT language model. The key parameters of proper nouns and words such as "suggestion", "best", and "optimization" are the highest. All key parameters of each customer service feedback information are summed to obtain the key index k2. The recognition module then outputs k, k1, and k2 to the credibility judgment module.

[0016] Credibility determination module uses formula The credibility index F of the customer service feedback information is obtained, and the encrypted sender's name Na, delivery time Ti, the original text of the customer service feedback information Te and the credibility index F are combined to generate the data vector (Na, Ti, Te, F) of the customer service feedback information and send it to the report generation module.

[0017] After receiving the data vector of customer service feedback information, the report generation module extracts the credibility index F of each data vector and conducts targeted analysis on the data vectors whose credibility index F is greater than the threshold. Then, the report generation module understands the original text of the data vector based on the BERT language model, extracts the questions raised in the original text, records them as value question data, and adds them to the data vector itself.

[0018] The report generation module obtains information reports from the customer service platform on the Internet as report samples, annotates and trains a large number of report samples based on the BERT language model, and learns how to generate high-quality reports based on the extracted information. After the model training is completed, data report generation can begin. The NLP-based method can integrate and output the elements in the data vector according to pre-defined templates and rules, and automatically generate reports that meet the specifications. The generated report can cover various data indicators, analysis results, and suggestions.

[0019] The report generation module evaluates the indicators of the generated report by comparing it with the standard report template, including the report accuracy indicator h1, completeness indicator h2 and readability indicator h3, and obtains the comprehensive evaluation indicator H by weighted summation of indicators h1, h2 and h3. When the comprehensive evaluation indicator is less than the preset threshold, the report is regenerated, otherwise the report is added as an element to the data vector itself, and finally the report generation module sends all data vectors to the database.

[0020] After receiving the data vector sent by the business subsystem, the database adds a new element number 1 in front of the data vector; after receiving the data vector sent by the customer service platform subsystem, it adds a new element number 2 in front of the data vector, and archives all data vectors so that they can be retrieved and called up by the demand retrieval module at any time.

[0021] After receiving the query statement, the demand retrieval module understands the meaning of each word in the query instruction, extracts the keywords and performs word replacement processing according to the similarity, searches in the database, and then extracts elements in the data vector to display on the display end. For example, the query instruction "What are the projects that require a second-level and above registered construction engineer certificate for water conservancy and hydropower engineering in yesterday's bidding documents" is input to the demand retrieval module, and the demand of the instruction is understood as a query based on the BERT language model, and then the keywords "yesterday", "bidding documents", "second-level and above registered construction engineer certificate for water conservancy and hydropower engineering" are extracted. First, the date of the query is obtained through the Internet, 20xx, x month, x day, and the search element 1 is obtained by calculating the time one day forward: the search word "20y year y month y day" in the time element, and then the search element 2 is identified from the keyword "second-level and above registered construction engineer certificate for water conservancy and hydropower engineering" in the positive number requirement. The search is then searched in the database, and the element "project name" in the data vector that meets the search element 1 and the search element 2 is output to the display end. The preferred embodiment of the present invention disclosed above is only used to help explain the present invention. The preferred embodiments do not describe all details in detail, nor do they limit the invention to only specific embodiments.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. A structured data generation model based on natural language processing trains the language model based on a neural network, while capturing the contextual information of words in semantic understanding, improving the real-time and dynamic nature of the system's data analysis method and reducing the limitations of semantic understanding;

[0024] 2. A structured data generation model based on natural language processing integrates and analyzes the captured data, and analyzes and quantifies the credibility in a specific context, reducing data bias and unfairness in data analysis;

[0025] 3. A structured data generation model based on natural language processing identifies similar paragraphs in documents based on semantic analysis, highlights tampered clauses and fields, reduces the time cost wasted on mechanical repetitive work, and also reduces the legal risks caused by human errors, solving the inefficiency problem of repetitive work. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0027] Figure 1 It is a principle block diagram of the present invention.

[0028] Figure 2 It is a schematic diagram of a programming language implementation of a regular expression in an example of the present invention. DETAILED DESCRIPTION

[0029] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. 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.

[0030] The usage scenario of this system is preset as the analysis of huge amounts of information for business and customer service work in production and life.

[0031] See also Figure 1 As shown, a structured data generation model based on natural language processing includes a business subsystem, a customer service platform subsystem, a demand retrieval module and a database; wherein the business subsystem includes a keyword extraction module, a similarity comparison module and a scoring module, and the customer service platform subsystem includes a text preprocessing module, a recognition processing module, a credibility determination module and a report generation module.

[0032] The business subsystem is used to process various documents such as bids, response documents, business contracts, change orders and administrative reports in daily business work. The business subsystem identifies, checks and extracts keywords from the entire document; interprets the meaning of the document through natural language processing, obtains key fields and enters them into a structured database. The business subsystem includes a keyword extraction module, a similarity comparison module and a scoring module.

[0033] The keyword extraction module classifies the document and uses the BERT language model to identify the entire document. It captures type judgment elements in the document. Type judgment elements include keywords such as "rights", "fees", "terms", "confidentiality", "effectiveness", and change images in the change order. The frequency N of occurrence of various type judgment elements in the entire document is recorded. Each type judgment element corresponds to a preset judgment weight λ. The number of occurrences of various type judgment elements is multiplied by the judgment weight and then summed to obtain the type judgment parameter α. Each document type corresponds to a specific type judgment parameter range. Document types include bids, response documents, business contracts, change orders, and administrative contracts.

[0034] See also Figure 2 As shown, each document type uses the corresponding regular expression to extract keywords from the document, organizes the keywords into data vectors, and generates structured data.

[0035] For example, extract keywords about "time", "project name", "construction scale" and "contract estimated price" in the bidding documents, extract keywords about certificate requirements, professional title requirements, qualification requirements and reputation requirements in the bidder qualification part, extract keywords about "the maximum bid limit price compiled by the tenderer", and form all keywords into a data vector. The keyword extraction module scans several bidding documents, generates several data vectors, and forms a structured data table.

[0036] The similarity comparison module compares the full text of documents of the same document type, marks the semantically different words in similar paragraphs as difference characters, uses the BERT language model to analyze the meaning of the difference characters and match a reference character, generates the difference character meaning vector X = (x1, x2, x3, ..., xn) and the reference character meaning vector Y = (y1, y2, y3, ..., yn) where n is the number of dimensions contained in the coordinate system where the meaning vector is located, and the value of n depends on the learning depth of the BERT language model; analyze the words of the difference characters Parts of speech include proper nouns, general nouns, verbs, adjectives, modal particles, function words, etc. Each part of speech corresponds to a preset part of speech factor e1; analyze the location of the difference text, the location includes the document directory, announcement chapter, appendix chapter, instructions chapter, general provisions, text, etc. The BERT language model uses the RNN recurrent neural network to learn and obtain the accurate location category. Each location corresponds to a location factor e2; analyze the context and context of the difference text, and obtain a context factor e3 based on the BERT language model analysis. Through the formula The difference score E is obtained, where λ1, λ2, λ3, and λ4 are preset weight factors. The difference characters with a difference score E greater than 100 are highlighted, and the difference score E is added to the data vector itself.

[0037] The scoring module analyzes the data vector generated by the keyword extraction module and presets a scoring factor for each element in the data vector. For example, a scoring factor 1 is preset for the project name element. When the element contains city A, the scoring factor 1 is a1, and when the element contains city B, the scoring factor 1 is a2. A scoring factor 2 is preset for the contract estimate price element. The scoring factor 2 for elements greater than 1 million is b2, and the scoring factor 2 for elements less than 1 million is b2, and so on. All scoring factors are summed to obtain a scoring parameter T, and the scoring parameter of each data vector is also added to the data vector itself to become a new element, completing the scoring step 1. The scoring module is based on the difference score of each data vector obtained by the similarity comparison module, and the difference score is summed to obtain a difference parameter K. The difference parameter of each data vector is also added to the data vector itself to become a new element, completing the scoring step 2.

[0038] The scoring module inputs all data vectors into the database after scoring step one and scoring step two are completed.

[0039] The text preprocessing module preprocesses the customer service feedback information such as letters, complaints, and complaints received by the customer service platform subsystem. First, the information source of the customer service feedback information is obtained, the name of the sender and the time of delivery are obtained, and the name of the sender is encrypted using the encryption algorithm AES. Then, each Chinese character and English character in the customer service feedback information is split into separate strings, and the XLNet natural regression pre-trained language model is used to obtain the word probability P of each single character and the two adjacent characters before and after. When the sum of the word probabilities of the two adjacent characters is greater than the preset probability threshold When , the two characters are combined into a word, and the meaning vector of the word is obtained in the language model. If the meaning vector of the word belongs to the category of defamatory and extremely negative adjectives, the word is marked as useless information, the number of characters contained in the useless information is recorded as the information shielding parameter k, and the characters of the useless information are replaced by *. Then the pre-processed customer service feedback information is sent to the recognition processing module in the form of a string.

[0040] The recognition processing module uses the character string sent by the text preprocessing module to obtain forum replies and comments on the Internet as training data for training. Based on the natural processing (NLP) technology, the emotional tendencies contained in each sentence are analyzed to form a segmented emotional dictionary, and the emotional tendencies are divided into three first-level emotional tendencies: positive, negative, and neutral. Each first-level emotional tendency contains segmented second-level emotional tendencies. For example, positive emotional tendencies can be segmented into gratitude, excitement, hope, etc. Each emotional tendency corresponds to a preset emotional degree parameter. The emotional degree parameter of positive emotional tendencies is a positive number, the emotional degree parameter of negative emotional tendencies is a negative number, and the emotional degree parameter of neutral emotional tendencies is 0. The recognition processing module extracts the meaning vector of each word in the character string, understands the word's part of speech, meaning, and emotional tendency based on the BERT language model, obtains the emotional degree parameter of each word, and sums the emotional degree parameters of each customer service feedback information to obtain the emotional index k1.

[0041] The recognition module obtains the key parameters of each word in each customer service feedback information based on the BERT language model. The key parameters of proper nouns and words such as "suggestion", "best", and "optimization" are the highest. All key parameters of each customer service feedback information are summed to obtain the key index k2. The recognition module then outputs k, k1, and k2 to the credibility judgment module.

[0042] Credibility determination module uses formula The credibility index F of the customer service feedback information is obtained, and the encrypted sender's name Na, delivery time Ti, the original text of the customer service feedback information Te and the credibility index F are combined to generate the data vector (Na, Ti, Te, F) of the customer service feedback information and send it to the report generation module.

[0043] After receiving the data vector of customer service feedback information, the report generation module extracts the credibility index F of each data vector and conducts targeted analysis on the data vectors whose credibility index F is greater than the threshold. Then, the report generation module understands the original text of the data vector based on the BERT language model, extracts the questions raised in the original text, records them as value question data, and adds them to the data vector itself.

[0044] The report generation module obtains information reports from the customer service platform on the Internet as report samples, annotates and trains a large number of report samples based on the BERT language model, and learns how to generate high-quality reports based on the extracted information. After the model training is completed, data report generation can begin. The NLP-based method can integrate and output the elements in the data vector according to pre-defined templates and rules, and automatically generate reports that meet the specifications. The generated report can cover various data indicators, analysis results, and suggestions.

[0045] The report generation module evaluates the indicators of the generated report by comparing it with the standard report template, including the report accuracy indicator h1, completeness indicator h2 and readability indicator h3, and obtains the comprehensive evaluation indicator H by weighted summation of indicators h1, h2 and h3. When the comprehensive evaluation indicator is less than the preset threshold, the report is regenerated, otherwise the report is added as an element to the data vector itself, and finally the report generation module sends all data vectors to the database.

[0046] After receiving the data vector sent by the business subsystem, the database adds a new element number 1 in front of the data vector; after receiving the data vector sent by the customer service platform subsystem, it adds a new element number 2 in front of the data vector, and archives all data vectors so that they can be retrieved and called up by the demand retrieval module at any time.

[0047] After receiving the query statement, the demand retrieval module understands the meaning of each word in the query instruction, extracts the keywords and performs word replacement processing according to the similarity, searches in the database, and then extracts elements in the data vector to display on the display end. For example, the query instruction "What are the projects that require a second-level and above registered construction engineer certificate for water conservancy and hydropower engineering in yesterday's bidding documents" is input to the demand retrieval module, and the demand of the instruction is understood as a query based on the BERT language model, and then the keywords "yesterday", "bidding documents", "second-level and above registered construction engineer certificate for water conservancy and hydropower engineering" are extracted. First, the date of the query is obtained through the Internet, 20xx, x month, x day, and the search element 1 is obtained by calculating the time one day forward: the search word "20y year y month y day" in the time element, and then the search element 2 is identified from the keyword "second-level and above registered construction engineer certificate for water conservancy and hydropower engineering" in the positive number requirement. The search is then searched in the database, and the element "project name" in the data vector that meets the search element 1 and the search element 2 is output to the display end. The preferred embodiment of the present invention disclosed above is only used to help explain the present invention. The preferred embodiments do not describe all details in detail, nor do they limit the invention to only specific embodiments.

[0048] Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and describes these embodiments in detail in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A structured data generation model based on natural language processing, comprising a keyword extraction module, a similarity comparison module, a scoring module, a text preprocessing module, a recognition processing module, a credibility determination module and a report generation module, characterized in that: The keyword extraction module classifies the documents received by the business subsystem, uses the BERT language model to identify the entire document, and captures the type judgment elements in the document to determine the document type; The similarity comparison module compares the full texts of documents of the same document type to mark the different characters, obtains the difference score based on the characteristic analysis of the different characters, and highlights the characters with the difference score greater than the threshold; The scoring module scores the document quality based on keywords and difference scores, and then generates structured data from all the data in the business subsystem and sends it to the database; The text preprocessing module splits the customer service feedback information received by the customer service platform subsystem into Chinese characters, encrypts sensitive information, and shields useless information; The recognition processing module quantifies the credibility index in customer service feedback information based on the language model; The credibility determination module obtains the credibility index of the customer service feedback information by calculating the credibility index, and after comprehensively encrypting the sender's name, delivery time, original text of the customer service feedback information and the credibility index, generates a data vector of the customer service feedback information and sends it to the report generation module; The report generation module extracts keywords from the original text of the customer service feedback information and generates a data report for the original text, and then generates structured data from all the data of the customer service platform subsystem and sends it to the database.

2. The structured data generation model based on natural language processing according to claim 1, characterized in that: It also includes a demand retrieval module and database; The database is used to store structured data sent by the business subsystem and customer service platform subsystem; The demand retrieval module uses natural language processing technology to analyze the proposed retrieval requirements, obtain structured data from the database and display it.

3. The structured data generation model based on natural language processing according to claim 1, characterized in that: The process of keyword extraction module classifying and extracting keywords from documents received by the business subsystem is as follows: Use the BERT language model to identify the entire document, capture type judgment elements in the document, and record the frequency of occurrence of various type judgment elements in the entire document. Each type judgment element corresponds to a preset judgment weight. The value of the type judgment parameter is obtained by calculation and the document type is determined; each document type uses the corresponding regular expression to extract keywords from the document, and the keywords are combined into data vectors to generate structured data.

4. The structured data generation model based on natural language processing according to claim 1, characterized in that: The specific process of the similarity comparison module marking the different characters and highlighting them is as follows: Compare the full texts of documents of the same document type, mark the semantically different words in similar paragraphs as difference characters, use the BERT language model to analyze the meaning of the difference characters and match a reference character, generate the meaning vector of the difference characters and the meaning vector of the reference character; analyze the part of speech of the difference characters to obtain the preset part of speech factor; Analyze the position of the difference text to obtain the corresponding position factor; analyze the context and language context of the difference text to obtain a context factor, obtain the difference score through calculation, highlight the difference characters whose difference score is greater than the preset threshold, and add the difference score to the data vector itself.

5. The structured data generation model based on natural language processing according to claim 1, characterized in that: The specific process of the scoring module to score document quality based on keywords and difference scores is as follows: Based on the analysis of the data vector generated by the keyword extraction module, a scoring factor is preset for each element in the data vector, the elements are matched to obtain the scoring factor, all the scoring factors are summed to obtain the scoring parameter, the scoring parameter of each data vector is also added to the data vector itself to become a new element, and the scoring step one is completed; based on the difference score of each data vector obtained by the similarity comparison module, the difference score is summed to obtain the difference parameter, the difference parameter of each data vector is also added to the data vector itself to become a new element, and the scoring step two is completed; after the scoring steps one and two are completed, all data vectors are input into the database.

6. The structured data generation model based on natural language processing according to claim 1, characterized in that: The specific process of the text preprocessing module for Chinese character splitting, sensitive information encryption and useless information shielding is as follows: First, the information source of the customer service feedback information is obtained, the name of the sender and the delivery time are obtained, and the encryption algorithm AES is used to encrypt the name of the sender. Then, each Chinese character and English character in the customer service feedback information is split into separate strings. The XLNet natural regression pre-trained language model is used to obtain the word probability of each single character and the two adjacent characters before and after. When the sum of the word probabilities of two adjacent characters is greater than the preset probability threshold, the two characters are combined into a word and the meaning vector of the word is obtained in the language model. If the meaning vector of the word belongs to a defamatory or extremely negative adjective, the word is marked as useless information, the number of characters contained in the useless information is recorded as the information shielding parameter, and the characters of the useless information are replaced by *.

7. The structured data generation model based on natural language processing according to claim 1, characterized in that: The process of the recognition processing module quantifying the credibility index in customer service feedback information based on the language model is as follows: The recognition processing module uses the character string sent by the text preprocessing module to analyze the emotional tendency parameters contained in each sentence based on the natural language processing (NLP) technology, and sums up the emotional parameters of each customer service feedback information to obtain the emotional index; based on the BERT language model, it obtains the criticality parameters of each word in each customer service feedback information, and sums up all the criticality parameters of each customer service feedback information to obtain the criticality index; The recognition module then outputs the emotional tendency parameter, the criticality index and the information shielding parameter to the credibility determination module.

8. The structured data generation model based on natural language processing according to claim 1, characterized in that: The specific process of the report generation module extracting keywords from the original text of the customer service feedback information and generating a data report for the original text is as follows: Based on the BERT language model, the original text of the data vector is understood, the questions raised in the original text are extracted, recorded as value question data and added to the data vector itself, and a report that complies with the standards is automatically generated based on the NLP method. The report that complies with the standards is added as an element to the data vector itself. Finally, the report generation module sends all data vectors to the database.

9. The structured data generation model based on natural language processing according to claim 2, characterized in that: The specific process of the database storing the structured data sent by the business subsystem and the customer service platform subsystem and the demand retrieval module obtaining the structured data from the database and displaying it is as follows: After receiving the data vector sent by the business subsystem, the database adds a marker element 1 in front of the data vector; after receiving the data vector sent by the customer service platform subsystem, the database adds a marker element 2 in front of the data vector, and archives all data vectors so that they can be retrieved and called up by the demand retrieval module at any time; After receiving the query statement, the demand retrieval module understands the meaning of each word in the query instruction, extracts the keywords and replaces the words according to the similarity, searches in the database, and then extracts the elements in the data vector and displays them on the display end.

Citation Information

Patent Citations

  • Method and system for identifying text difference content in combination with semantic recognition

    CN113051869A

  • Method and system for processing investment and research report in futures field

    CN115358201A

  • Customer marketing scene data analysis system and method based on NLP algorithm

    CN116663664A

  • Contract term risk check method, system and device and storage medium

    CN116843175A

  • Short video content label knowledge base quick retrieval method based on natural language processing

    CN117009461A

Cited By

  • Clinical test scheme deviation automatic checking method and device and computer equipment

    CN120257967A