Information identification method, system and device based on keywords and deep learning

Abstract

The invention discloses an information identification method, system and device based on keywords and deep learning. The method comprises the steps of obtaining a personal data set and a public data set; constructing a special dictionary library based on the personal data set and the public data set; extracting keywords, specific symbols and topics based on the public data set; judging whether the personal data set is matched with the keyword or the specific symbol; if not, determining a matching dictionary by using the theme and the special dictionary library; after the matching dictionary is determined, judging whether the matching dictionary is matched with the personal data set or not; if not, based on the personal data set, training through a language representation model to obtain an embedded vector; after an embedded vector is obtained, inputting the embedded vector into a sequence neural network model for modeling to obtain a hidden state sequence; and mapping the dimension of the hidden state sequence into a target dimension, and inputting the hidden state sequence mapped to the target dimension into a conditional random field algorithm to obtain a personal information identification result.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an information recognition method, system and device based on keywords and deep learning. Background Art

[0002] With the rapid development of information technology, personal information identification has become a key factor in promoting the healthy development of the digital economy. In today's digital age, a large amount of personal information flows in various network platforms and information systems. How to accurately identify this information is particularly important. In terms of government administration, through accurate personal information identification, government departments can more efficiently handle various social management affairs such as population data, household registration management, social security information, etc., which can improve administrative efficiency and ensure the accuracy and timeliness of public services. For example, in household registration management, accurate personal information identification can help government departments quickly verify the identity of residents, simplify the process of handling affairs, reduce human errors, and improve the overall management level; in terms of public security, accurate personal information identification can help public security organs and other law enforcement departments to timely discover and prevent criminal activities and maintain social security and stability. For example, through accurate identification and analysis of information of entry and exit personnel, relevant departments can effectively monitor and manage potential security threats and ensure public safety; in terms of e-commerce, personal information identification can help merchants better understand user needs and provide customized product recommendations and services. At the same time, it helps to protect user privacy, prevent information leakage and abuse, and enhance users' trust in online services.

[0003] In the existing technology, personal information identification mainly relies on two methods: open rule identification and manual review. Open rule identification is a method based on pre-set rules. These rules are usually derived from the standardized understanding of specific data types. For example, when processing household registration information, the system may use fixed administrative division code rules to verify the legitimacy of the address, and confirm the accuracy of personal identity based on the naming conventions of the name. The core advantage of the open rule identification method is that it has a high degree of automation, can quickly process a large amount of information, and to a certain extent reduce the need for human intervention; manual review is a method that relies on the professional knowledge and experience of manual reviewers to verify and confirm information. Its advantage lies in its flexibility and adaptability, and its ability to handle complex and unstructured data.

[0004] However, although public rule recognition performs well in processing structured and standardized data, with the diversification of data types and formats, it is difficult for a single rule to adapt to the rapidly changing data environment, especially when facing new data formats and unstructured data, the applicability and flexibility of the rules are particularly insufficient; although the manual review method has good flexibility, it is difficult to meet the real-time processing needs of massive data, and is easily affected by the subjective judgment of the reviewer, resulting in inconsistent recognition results. Both of these recognition methods have unavoidable limitations in personal information recognition, resulting in low efficiency in personal information recognition, which seriously affects the quality of recognition results. Summary of the invention

[0005] In order to solve the above technical problems, the present application provides an information recognition method, system and device based on keywords and deep learning.

[0006] The technical solution provided in this application is described below: The first aspect of the present application provides an information recognition method based on keywords and deep learning, the method comprising: Access personal and public datasets; Building a dedicated dictionary library based on the personal dataset and the public dataset; Extracting keywords, specific symbols, and themes based on the public dataset; Determining whether the personal data set matches the keyword or the specific symbol; If not, determining a matching dictionary using the subject and the dedicated dictionary library; After determining the matching dictionary, determining whether the matching dictionary matches the personal data set; If not, then obtaining an embedding vector through language representation model training based on the personal data set; After the embedding vector is obtained, the embedding vector is input into a sequence neural network model for modeling to obtain a hidden state sequence; The dimension of the hidden state sequence is mapped to the target dimension, and the hidden state sequence mapped to the target dimension is input into the conditional random field algorithm to obtain the personal information recognition result.

[0007] Optionally, after acquiring the personal data set and the public data set and before constructing a dedicated dictionary based on the personal data set and the public data set, the method further includes: The personal dataset and the public dataset are subjected to quality filtering and redundancy removal.

[0008] Optionally, the quality filtering includes language filtering, indicator filtering, statistical feature filtering and keyword filtering; The redundancy removal includes sentence level removal, document level removal, and data set level removal.

[0009] Optionally, the constructing a dedicated dictionary library based on the personal dataset and the public dataset includes: Building an initial dictionary based on the public dataset, and analyzing the personal dataset to generate candidate sequences; Generate a classification scoring model based on the initial dictionary library and the candidate sequence; The classification and scoring model is used to expand the initial dictionary library to obtain a special dictionary library.

[0010] Optionally, the use of the classification and scoring model to expand the initial dictionary library to obtain a dedicated dictionary library includes: Using the classification scoring model to evaluate the candidate sequence; generating a language representation scorer based on the scored candidate sequences and the initial dictionary library; Using the language representation scorer to perform quality scoring on the candidate sequence; The initial dictionary library is expanded based on the quality score to obtain a dedicated dictionary library.

[0011] Optionally, the obtaining an embedding vector by training a language representation model based on the personal data set includes: Obtaining a specification document and a requirement characteristic, and performing entity definition on the personal data set according to the specification document and the requirement characteristic; The language representation model is used to train a personal data set that has completed the entity definition to obtain an embedding vector.

[0012] Optionally, mapping the dimension of the hidden state sequence to a target dimension, and inputting the hidden state sequence mapped to the target dimension into a conditional random field algorithm to obtain a personal information identification result includes: Mapping the dimension of the hidden state sequence to a target dimension, and inputting the hidden state sequence mapped to the target dimension into a conditional random field algorithm; The conditional random field algorithm labels the hidden state sequence after being mapped to the target dimension and calculates a score for each of the labels; The target tag with the highest score is obtained, and the personal information recognition result is obtained based on the target tag.

[0013] A second aspect of the present application provides an information recognition system based on keywords and deep learning, the system comprising: An acquisition unit, used to acquire personal data sets and public data sets; A construction unit, configured to construct a dedicated dictionary library based on the personal data set and the public data set; An extraction unit, configured to extract keywords, specific symbols, and themes based on the public data set; A first determination unit, configured to determine whether the personal data set matches the keyword or the specific symbol; a determining unit, for determining a matching dictionary by using the subject and the dedicated dictionary library if no; a second judgment unit, configured to, after determining a matching dictionary, judge whether the matching dictionary matches the personal data set; A training unit, configured to, if not, obtain an embedding vector by training a language representation model based on the personal dataset; A modeling unit, configured to input the embedding vector into a sequence neural network model for modeling after obtaining the embedding vector, so as to obtain a hidden state sequence; The identification unit is used to map the dimension of the hidden state sequence to the target dimension, and input the hidden state sequence mapped to the target dimension into the conditional random field algorithm to obtain the personal information identification result.

[0014] The third aspect of the present application provides an information recognition device based on keywords and deep learning, the device comprising: Processor, memory, input-output unit, and bus; The processor is connected to the memory, the input and output unit, and the bus; The memory stores a program, and the processor calls the program to execute any one of the methods.

[0015] A fourth aspect of the present application provides a computer-readable storage medium, on which a program is stored. When the program is executed on a computer, the program executes the first aspect and any optional method in the first aspect.

[0016] It can be seen from the above technical solutions that this application has the following advantages: The information recognition method based on keywords and deep learning achieves more efficient and accurate recognition of personal information. Through fast matching based on rules such as keywords, specific symbols and themes, combined with advanced learning technologies such as language representation models, sequence neural network models and conditional random field algorithms, it can respond quickly when performing personal information recognition, and ensure high efficiency and high quality recognition when facing complex and diverse data situations. In addition, the multi-level judgment process of this application can flexibly select recognition strategies according to data conditions, further improving the accuracy and flexibility of recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solution in the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A flowchart of an embodiment of the information identification method based on keywords and deep learning provided in this application; Figure 2 A flowchart of another embodiment of the information identification method based on keywords and deep learning provided by the present application; Figure 3 A schematic diagram of the structure of an embodiment of the information recognition system based on keywords and deep learning provided by the present application; Figure 4 A schematic diagram of the structure of an embodiment of an information recognition device based on keywords and deep learning provided in this application. DETAILED DESCRIPTION

[0019] The present application provides an information identification method based on keywords and deep learning, which is used to identify personal information with high efficiency and quality. It should be noted that the information identification method based on keywords and deep learning of the present application is applied to a terminal.

[0020] See also Figure 1 , the present application first provides an embodiment of an information identification method based on keywords and deep learning, the embodiment comprising: S101. Obtain personal data sets and public data sets; In this embodiment, to obtain personal data sets and public data sets, it is necessary to determine and screen the source channels of the data sets. For example, under the premise of complying with relevant laws and regulations and after the user's explicit authorization, the user's name, age, contact information browsing history, and behavior records and other relevant personal information are obtained from the user registration platform, personal information management system and other related systems as personal data sets. For the acquisition of public data sets, they can be obtained from authoritative data repositories, such as government affairs disclosure data platforms, encyclopedia knowledge bases, government affairs disclosure data and other sources.

[0021] The personal data set obtained in this embodiment can reflect the unique characteristics and behavior patterns of an individual. In the subsequent steps, the personal data set is an indispensable foundation. For example, based on the personal data set, the user's preferences can be deeply understood, so as to accurately recommend products or content that meet the user's interests. Compared with personal data sets, public data sets are widely representative and universal, and can provide macro background information or standard data patterns in the field. In research and applications in many fields, public data sets can serve as an important reference. For example, in the construction of disease prediction models, public medical data sets can provide macro information such as disease incidence and epidemic trends. Combined with the individual characteristics of personal data sets, individual disease risks can be predicted more accurately.

[0022] S102, building a dedicated dictionary library based on personal data sets and public data sets; In this embodiment, the acquired personal data set and public data set are first integrated and analyzed, representative attribute nouns such as name, age, contact information and other related words are extracted from the personal data set, and attribute values ​​such as administrative district name, unit name, and general identity identifier are parsed from the public data set. Then, a dedicated dictionary library is constructed based on these attribute nouns and attribute values ​​obtained from different sources.

[0023] This embodiment constructs a dedicated dictionary based on personal data sets and public data sets with higher accuracy and comprehensiveness. Since personal data sets and public data sets integrate various data sources, they can cover more types of vocabulary, including both personal information vocabulary and widely representative public information vocabulary, so that personal information can be more accurately identified. For example, when identifying special personal information such as ID card numbers and bank card numbers, the relevant rules and vocabulary in the dictionary can quickly match them. When performing subsequent personal information identification, a comprehensive and accurate dedicated dictionary can greatly shorten the matching time, improve identification efficiency, and reduce the possibility of misjudgment.

[0024] S103, extracting keywords, specific symbols, and themes based on the public data set; In this embodiment, the public data set is comprehensively combed, and data mining, data analysis and other technologies are used to extract keywords from the public data set. For example, the frequency of each word in the public data set is counted, and words with high frequency of occurrence and related to personal information identification are screened out as keywords, such as "identity card", "bank card", "name" and other words that are common in personal information; for the extraction of specific symbols, symbols with special meanings are usually searched in the public data set, such as separators in identity card numbers, specific identification symbols in bank card numbers, etc.; for topic extraction, the text content of the public data set needs to be classified according to different topics, for example, content related to personal identity is classified into one topic, and content related to financial information is classified into another topic, and then the topic is determined by analyzing the semantic structure, vocabulary distribution and other features of the text.

[0025] This embodiment extracts keywords from a public data set. The obtained keywords help improve the accuracy and efficiency of subsequent personal information identification and can be used as an important basis for subsequent matching. When identifying personal information, the text part that may contain personal information can be quickly located; special characters are extracted from the public data set, and character strings can be accurately distinguished based on the special characters, avoiding misjudgment and improving recognition accuracy; topics are extracted from the public data set, and the public data set can be classified and managed, which helps to perform more targeted personal information identification work within different subject areas.

[0026] S104, determining whether the personal data set matches the keyword or specific symbol; In this embodiment, the personal data set needs to be formatted first so that the data structure of the personal data set is easy to compare with the keywords and specific symbols. For example, if the personal data set is in text form, it is processed according to a unified encoding format, and then for the keywords or specific symbols, it is determined whether there is content in the personal data set that is exactly the same as the keywords or specific symbols. If it is determined that it is, the personal information identification result is directly obtained. If not, step 105 is executed.

[0027] The operation of judging whether a personal data set matches a keyword and a specific symbol in this embodiment helps to quickly filter data and can directly locate the personal data content containing the keyword or the specific symbol, thereby improving the accuracy and efficiency of recognition and laying the foundation for subsequent personal information recognition.

[0028] S105, determining a matching dictionary using the subject and the special dictionary library; In this embodiment, when it is determined that the personal data set does not match the keyword or specific symbol, the subject and the special dictionary library are used to determine the matching dictionary. For example, if the subject is financial information, then the special dictionary library part related to finance is focused on. At the same time, according to the semantic category of the subject, a vocabulary subset with a high relevance to the subject is screened from the special dictionary library to determine and form a matching dictionary.

[0029] This embodiment uses the subject and the special dictionary library to determine the matching dictionary, which can improve the specificity of recognition. When the initial match with the keyword or specific symbol fails, the scope of the matching dictionary is narrowed by the subject, which can avoid blind searching in the special dictionary library and reduce the waste of computing resources. In addition, by determining the matching dictionary based on the subject and the special dictionary library, the matching dictionary can be more in line with actual needs and the accuracy of subsequent matching can be improved.

[0030] S106, after the matching dictionary is determined, determining whether the matching dictionary matches the personal data set; In this embodiment, after determining the matching dictionary, the words in the matching dictionary are compared with the personal data set one by one. For each word in the matching dictionary, a character-level comparison method can be used, starting from the first character of the word, and comparing each character in turn to see if they are the same. If it is text data, the case of the characters, full-width and half-width, etc. can also be used for comparison and judgment. If it is yes, the personal information recognition result is directly obtained. If not, step 107 is executed.

[0031] This embodiment makes matching judgments between the matching dictionary and the personal data set, which helps to more accurately determine personal information, and when faced with large-scale personal data sets and matching dictionaries, it can quickly complete the matching task, reducing the possibility of false matching, thereby further improving recognition accuracy and efficiency.

[0032] S107, based on the personal data set, obtain the embedding vector through language representation model training; In this embodiment, when it is determined that the matching dictionary does not match the personal data set, a suitable language representation model is first selected, and the personal data set is preprocessed according to the input requirements of the language representation model, such as performing preprocessing operations such as word segmentation and tokenization on the text data, and then the preprocessed personal data set is input into the language representation model. The language representation model will learn and train the personal data set. During the learning and training process, the language representation model will generate a corresponding embedding vector for each word or data unit based on factors such as the co-occurrence relationship of the words. The embedding vector can capture the potential connection between the words. For example, some words do not match on the surface, but may have similarities in the semantic space. The embedding vector can reflect this similarity. In contrast, traditional matching methods cannot identify some complex semantic relationships.

[0033] This embodiment can mine deeper semantic information in the personal data set by training the language representation model based on the personal data set to obtain the embedding vector, and input the personal data set as the training parameter of the language representation model. The quality of the embedding vector can be optimized to more accurately represent the characteristics of the personal data set, provide richer and more valuable information for subsequent modeling and personal information identification, and improve the accuracy and comprehensiveness of personal information identification.

[0034] S108, after the embedding vector is obtained, the embedding vector is input into the sequence neural network model for modeling to obtain a hidden state sequence; In this embodiment, after the embedding vector is obtained, it is input into the sequential neural network model. First, the sequential neural network model needs to be initialized and the initial parameters of the model, such as the number of neurons in the hidden layer, the activation function, etc., are set. Then, the embedding vector enters the sequential neural network model in sequence. The sequential neural network model processes each embedding vector. During the processing, the sequential neural network model generates a hidden state sequence, mainly to capture the sequential relationship between the embedded vectors, such as the chronological relationship in the time series or the semantic logical relationship.

[0035] The sequence neural network model in this embodiment can effectively process embedded vectors with sequential characteristics and mine the long-term dependencies therein. In the process of processing each embedded vector, the sequence neural network model generates a hidden state sequence, which can better represent the complex structure and semantic information of the personal data set. At the same time, the hidden state sequence can be used as an intermediate representation, which can better reflect the overall characteristics of personal information than the original embedded vector, and provide a more effective data basis for subsequent dimensional mapping and personal information identification, which helps to improve the accuracy of personal information identification and the ability to process complex data structures.

[0036] S109, mapping the dimension of the hidden state sequence to the target dimension, and inputting the hidden state sequence mapped to the target dimension into the conditional random field algorithm to obtain the personal information recognition result.

[0037] In this embodiment, the target dimension is first determined. The target dimension is set according to the needs of the subsequent personal information identification task and the requirements of the model. This embodiment does not make specific limitations. After the target dimension is determined, the dimension of the hidden state sequence is mapped to the target dimension by linear mapping or nonlinear mapping. For example, if the original dimension of the hidden state sequence is high and the target dimension is low, the linear dimensionality reduction method of principal component analysis can be used for mapping. After the dimension mapping is completed, the hidden state sequence mapped to the target dimension is input into the conditional random field algorithm. The conditional random field algorithm calculates each element in the hidden state sequence and finally obtains the personal information identification result.

[0038] This embodiment maps the dimension of the latent state sequence to the target dimension, which helps to reduce the complexity of the data and the amount of computation. The mapped latent state sequence is input into the conditional random field algorithm. The conditional random field algorithm can make full use of the relationship between elements to identify personal information, and fully consider the contextual relationship of the entire sequence. In personal information identification, it can better process personal data with a sequence structure, accurately identify the various components of personal information and their relationships, and improve the performance of the entire personal information identification system.

[0039] See also Figure 2 , Figure 2 Another embodiment of an information identification method based on keywords and deep learning is provided for the present application, and the embodiment includes: S201. Obtaining personal data sets and public data sets; In this embodiment, step S201 is similar to step S101 in the above embodiment and will not be described again here.

[0040] S202: Perform quality filtering and redundancy removal on the personal dataset and the public dataset, wherein the quality filtering includes language filtering, indicator filtering, statistical feature filtering, and keyword filtering, and the redundancy removal includes sentence level removal, document level removal, and dataset level removal.

[0041] In this embodiment, the personal data set and the public data set are subjected to quality filtering and redundancy removal. The quality filtering includes language filtering, index filtering, statistical feature filtering and keyword filtering. The redundancy removal includes sentence level removal, document level removal and data set level removal. Language filtering is to detect the language type in the personal data set and the public data set, and remove the content of non-target languages. For example, if the target is to process Chinese personal information, then the texts in other languages ​​such as English and French are removed; index filtering is to filter according to the set index range, for example, the text length within a certain interval is retained, and the text that is too long or too short is removed; statistical feature filtering is to remove the parts that do not meet the statistical characteristics according to the statistical characteristics of the data, such as word frequency distribution, character distribution, etc.; keyword filtering is to remove the content containing specific keywords, such as noise words; sentence level removal is to analyze the similarity between sentences and remove repeated sentences; document level removal is to compare the content of different documents and remove documents with basically the same content; data set level removal is to evaluate the personal data set and the public data set as a whole, and remove data that is repeated in large quantities with other data or has no value for the overall analysis.

[0042] The quality filtering and redundancy removal in this embodiment can improve the quality and availability of data, wherein language filtering ensures the language consistency of data, indicator filtering can remove data of abnormal length that may affect the analysis results, so that the data structure is more in line with the analysis requirements, statistical feature filtering can exclude outliers in the data and improve the standardization of data, keyword filtering can reduce the interference of irrelevant information, redundancy removal reduces the duplication of data at different levels, and can reduce the cost of data storage and processing, sentence-level removal makes data more refined, document-level removal can avoid unnecessary calculations and analysis, and data set level can remove useless data, thereby improving the value density of the entire data set and helping to improve the efficiency and accuracy of subsequent analysis, modeling and other operations.

[0043] S203, constructing an initial dictionary based on the public data set, and analyzing the personal data set to generate candidate sequences; In this embodiment, when constructing the initial dictionary, it is necessary to first perform lexical analysis on the public data set, decompose the text into individual words, and then select representative, frequently occurring words related to personal information to construct the initial dictionary based on the frequency of occurrence, part of speech and other information of each word. For example, select words such as "disease name", "symptoms" and "treatment methods" from the public medical data set to construct the initial medical-related dictionary. For analyzing the personal data set to generate the candidate sequence, the personal data set is divided, and each divided unit is marked to generate the candidate sequence. For example, each work experience in the personal resume is marked as a semantic unit to generate the candidate sequence. This embodiment builds an initial dictionary based on a public data set, which helps to provide a basic vocabulary reference framework for subsequent operations. By screening relevant words in the public data set, it is possible to focus on important words related to personal information, providing a basis for subsequent classification, scoring and other operations. Analyzing the personal data set to generate a candidate sequence can convert a complex personal data set into a structured sequence form, which can better interact with other structures of the initial dictionary, and can compare and analyze each element in the candidate sequence with the initial dictionary, further improving the efficiency and accuracy of processing the personal data set.

[0044] S204, generating a classification scoring model based on the initial dictionary library and the candidate sequence; In this embodiment, features are first extracted from the words in the initial dictionary library and the elements in the candidate sequence, and then an input vector of the classification and scoring model is constructed based on the features. For example, the semantic feature vector of the words and the matching feature vector of the candidate sequence elements are combined into a new input vector. After the input vector of the classification and scoring model is constructed, a suitable classification algorithm is selected, and the input vector is input into the classification algorithm as training data to train the classification and scoring model. After the training is completed, a classification and scoring model is generated.

[0045] This embodiment constructs a classification and scoring model based on the initial dictionary library and candidate sequences, which can effectively classify and score candidate sequences. Through feature extraction, complex vocabulary and candidate sequence elements can be converted into quantifiable feature vectors. At the same time, the generated classification and scoring model can provide an objective and quantifiable standard for the subsequent evaluation of candidate sequences, which helps to screen out candidate sequence elements with a high degree of relevance to personal information and improve the efficiency and accuracy of the entire personal information processing system.

[0046] S205, using a classification scoring model to estimate the score of the candidate sequence; In this embodiment, each element in the candidate sequence is input into a trained classification scoring model, and the classification scoring model calculates each element based on the learned features. For example, if the classification scoring model is built based on a support vector machine, then the score will be scored based on the relationship between the element's feature vector and the decision boundary of the support vector machine. The score calculated by the classification scoring model can reflect factors such as the degree of matching between the element and the vocabulary in the initial dictionary library and its importance in personal information.

[0047] This embodiment uses a classification scoring model to estimate the scores of the candidate sequences. The estimated scores can intuitively reflect the degree to which each element in the candidate sequence is related to the personal information. Through the quantified scores, the elements in the candidate sequence can be conveniently sorted and screened. This type of scoring method is based on the learning and judgment of the classification scoring model, has a certain degree of objectivity and accuracy, and avoids the uncertainty of manual subjective judgment.

[0048] S206, generating a language representation scorer based on the estimated candidate sequences and the initial dictionary library; In this embodiment, it is first necessary to perform a statistical analysis on the scores based on the distribution of element scores in the estimated candidate sequence, and then combine the vocabulary information in the initial dictionary library to determine the language features related to the scores. For example, if high-scoring elements are all related to a specific vocabulary category, then the vocabulary category is an important language feature. Finally, a language representation scorer is constructed based on these language features.

[0049] This embodiment constructs a language representation scorer based on the scored candidate sequences and the initial dictionary library, which can further refine the evaluation of the candidate sequences, and can deeply understand the overall characteristics of the candidate sequence elements, and mine the language features related to the scores. These language features reflect the relationship between the candidate sequence and the personal information at the language level. At the same time, constructing a language representation scorer can perform more accurate quality scoring on the candidate sequences based on these language features, which can provide more comprehensive and detailed scoring results, and help to more accurately screen out high-quality candidate sequence elements.

[0050] S207, using a language representation scorer to perform a quality score on the candidate sequence; In this embodiment, each element in the candidate sequence is input into a language representation scorer, which calculates each element. The quality score obtained by the calculation reflects the quality of each element in the candidate sequence at the language representation level, including factors such as semantic relevance to personal information and grammatical normativity.

[0051] This embodiment uses a language representation scorer to perform quality scoring on candidate sequences, which can more comprehensively evaluate the quality of candidate sequence elements from the perspective of language representation. At the same time, the quality score can provide an effective basis for subsequent expansion of the initial dictionary library, thereby improving the quality and effectiveness of the dictionary library after the expansion is completed.

[0052] S208. Expand the initial dictionary library based on the quality score to obtain a dedicated dictionary library.

[0053] In this embodiment, the quality scores of the sequence elements are first obtained, the sequence elements are sorted according to the quality scores, and then the sequence elements with higher quality scores are selected, and the sequence elements with higher scores are expanded to the initial dictionary library. It should be noted that during the addition process, the newly added sequence elements need to be checked for duplicates and semantically integrated. For example, if the newly added sequence elements are semantically basically the same as the existing elements in the dictionary library, they are merged or their weights are adjusted. However, if the newly added sequence elements are completely new elements, they are directly expanded. By continuously selecting high-quality sequence elements and expanding the initial dictionary library, a dedicated dictionary library is finally obtained.

[0054] This embodiment selects sequence elements according to quality scores to expand the initial dictionary library, which can ensure the quality of the expanded and added sequence elements. The special dictionary library obtained after the expansion can provide more accurate and effective vocabulary references in subsequent personal information processing operations, making it more suitable for specific personal information processing tasks.

[0055] S209, extracting keywords, specific symbols, and themes based on the public data set; In this embodiment, step S209 is similar to step S103 in the above embodiment and will not be described again here.

[0056] S210, determining whether the personal data set matches the keyword or specific symbol; In this embodiment, step S210 is similar to step S104 in the above embodiment and will not be described again here.

[0057] S211, determining a matching dictionary using a subject and a special dictionary library; In this embodiment, step S211 is similar to step S15 in the above embodiment and will not be described again here.

[0058] S212, after the matching dictionary is determined, determining whether the matching dictionary matches the personal data set; In this embodiment, step S212 is similar to step S106 in the above embodiment, and will not be described in detail here. It should be noted that if the determination is yes, the personal information identification result is obtained directly, and if not, step 213 is executed.

[0059] S213, obtaining specification documents and requirement characteristics, and performing entity definition of the personal data set according to the specification documents and requirement characteristics; In this embodiment, when the matching dictionary does not match the personal data set, a specification file and a requirement feature are obtained. The specification file refers to a document of predefined rules and standards for personal information processing, such as data format specifications, information classification standards, etc., and the requirement feature is a feature that personal information should have according to specific business needs. According to the definition and requirement feature in the specification file, the personal data set is physically defined.

[0060] In the case of a matching failure, this embodiment can ensure that the personal data set meets business requirements and standards by defining the entity according to the specification file and the demand characteristics, which helps to improve the standardization and availability of the personal data set, and also provides a high-quality data foundation for subsequent operations such as model training. The specification file provides a unified framework and standard for entity definition, avoiding the irregularity of the definition, while the demand characteristics make the entity definition more targeted and can meet the needs of specific businesses.

[0061] S214. Train the personal data set for which entity definition has been completed through a language representation model to obtain an embedding vector.

[0062] In this embodiment, the personal data set with completed entity definition needs to be first input into the language representation model. In the language representation model, each word or entity in the personal data set will be mapped to a high-dimensional vector space, and the language representation model will perform feature learning on the personal data set. For example, the language representation model will learn the semantic relationship between words, the logical relationship between entities, etc. Through multiple iterative training, the embedding vector of the personal data set under the language representation model is obtained.

[0063] In this embodiment, the personal data set is trained through the language representation model to obtain an embedding vector, which can capture the rich semantic and logical information in the personal data set. The obtained embedding vector can more compactly represent the characteristics of the personal data set and is more efficient than the original text form.

[0064] S215, after the embedding vector is obtained, the embedding vector is input into the sequence neural network model for modeling to obtain a hidden state sequence; In this embodiment, step S215 is similar to step S108 in the above embodiment and will not be described again here.

[0065] S216, mapping the dimension of the hidden state sequence to the target dimension, and inputting the hidden state sequence mapped to the target dimension into the conditional random field algorithm; In this embodiment, step S216 is similar to step S109 in the above embodiment and will not be described again here.

[0066] S217, the conditional random field algorithm labels the hidden state sequence mapped to the target dimension and calculates the score of each label; In this embodiment, after the conditional random field algorithm receives the mapped hidden state sequence, it labels each element in the hidden state sequence according to the probability model and characteristic function within the conditional random field algorithm. While labeling the labels, the conditional random field algorithm calculates the score of each label according to a predefined score calculation rule. The calculated score reflects the rationality and possibility of the label in the current hidden state sequence.

[0067] This embodiment uses the conditional random field algorithm to label and score the hidden state sequence, which is a key step in determining whether the final personal information identification of the personal data set is accurate. It helps to improve the accuracy and reliability of personal information identification and provides an important basis for ultimately obtaining accurate personal information identification results. The label marking can clearly classify the elements in the hidden state sequence into different entity types, which helps to convert the personal data set into more meaningful structured information. The score calculation provides a quantitative evaluation standard for each label, which can distinguish the rationality of different labels in the current data.

[0068] S218. Obtain the target tag with the highest score, and obtain the personal information recognition result based on the target tag.

[0069] In this embodiment, after the conditional random field algorithm calculates the score of each label, the target label with the highest score is obtained from all labels. The target label represents the correct entity type in the current personal data set processing. Finally, the personal information identification result is obtained based on the target label. For example, if the target label is "occupation label" and the corresponding entity is "doctor", then the personal information identification result is that the individual's occupation is doctor. By selecting the label with the highest score, we can ensure that the personal information identification results obtained have high accuracy and reliability. The score-based selection method can select the result that best matches the data characteristics when there are multiple possibilities. Obtaining the target label with the highest score and obtaining the personal information identification result further improves the accuracy and practicality of processing personal information.

[0070] The following is a detailed description of the information recognition system based on keywords and deep learning provided by this application. Figure 3 , Figure 3 Another embodiment of the information recognition system based on keywords and deep learning provided by the present application includes: An acquisition unit 301 is used to acquire a personal data set and a public data set; A construction unit 302, configured to construct a dedicated dictionary library based on the personal data set and the public data set; An extraction unit 303, configured to extract keywords, specific symbols, and themes based on a public data set; A first determination unit 304 is used to determine whether the personal data set matches the keyword or the specific symbol; A determination unit 305, for determining a matching dictionary using the subject and a special dictionary library if no; A second judgment unit 306 is used to judge whether the matching dictionary matches the personal data set after determining the matching dictionary; A training unit 307, configured to, if not, obtain an embedding vector by training a language representation model based on the personal data set; A modeling unit 308 is used to input the embedded vector into a sequence neural network model for modeling after the embedded vector is obtained, so as to obtain a hidden state sequence; The identification unit 309 is used to map the dimension of the hidden state sequence to the target dimension, and input the hidden state sequence mapped to the target dimension into the conditional random field algorithm to obtain the personal information identification result.

[0071] Optionally, a processing unit 310 is further included, which is specifically configured to: Personal and public datasets are filtered for quality and redundancy removal.

[0072] Optionally, quality filtering includes language filtering, indicator filtering, statistical feature filtering, and keyword filtering; Redundancy removal includes sentence level removal, document level removal, and dataset level removal.

[0073] Optionally, the construction unit 302 is further configured to: Build an initial dictionary based on public datasets and analyze personal datasets to generate candidate sequences; Generate a classification scoring model based on the initial dictionary library and candidate sequences; The classification and scoring model is used to expand the initial dictionary library to obtain a special dictionary library.

[0074] Optionally, the construction unit 302 is further configured to: Use the classification scoring model to estimate the scores of candidate sequences; Generate a language representation scorer based on the scored candidate sequences and the initial dictionary library; Using a language representation scorer to score the quality of candidate sequences; The initial dictionary library is expanded based on the quality score to obtain a dedicated dictionary library.

[0075] Optionally, the training unit 307 is further configured to: Obtaining specification documents and requirement characteristics, and performing entity definition of the personal data set according to the specification documents and requirement characteristics; The embedding vector is obtained by training the language representation model on a personal dataset with completed entity definitions.

[0076] Optionally, the identification unit 309 is further configured to: Map the dimension of the hidden state sequence to the target dimension, and input the hidden state sequence mapped to the target dimension into the conditional random field algorithm; The conditional random field algorithm labels the hidden state sequence after it has been mapped to the target dimension and calculates the score of each label; Get the target tag with the highest score, and obtain the personal information recognition result based on the target tag.

[0077] This application also provides an information recognition device based on keywords and deep learning, please refer to Figure 4 , Figure 4 An embodiment of an information recognition device based on keywords and deep learning provided in this application includes: Processor 401, memory 402, input and output unit 403, bus 404; The processor 401 is connected to the memory 402, the input and output unit 403 and the bus 404; The memory 402 stores a program, and the processor 401 calls the program to execute any of the above methods.

[0078] The present application also relates to a computer-readable storage medium, on which a program is stored. When the program is run on a computer, the computer executes any of the above methods.

[0079] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0080] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0081] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0082] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0083] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, read-only memory), random access memory (RAM, random access memory), disk or optical disk and other media that can store program code.

Claims

1. An information recognition method based on keywords and deep learning, characterized in that: The method comprises: Access personal and public datasets; Building a dedicated dictionary library based on the personal dataset and the public dataset; Extracting keywords, specific symbols, and themes based on the public dataset; Determining whether the personal data set matches the keyword or the specific symbol; If not, determining a matching dictionary using the subject and the dedicated dictionary library; After determining the matching dictionary, determining whether the matching dictionary matches the personal data set; If not, then obtaining an embedding vector through language representation model training based on the personal data set; After the embedding vector is obtained, the embedding vector is input into a sequence neural network model for modeling to obtain a hidden state sequence; The dimension of the hidden state sequence is mapped to the target dimension, and the hidden state sequence mapped to the target dimension is input into the conditional random field algorithm to obtain the personal information recognition result: Mapping the dimension of the hidden state sequence to the target dimension, and inputting the hidden state sequence mapped to the target dimension into the conditional random field algorithm to obtain the personal information recognition result includes: Mapping the dimension of the hidden state sequence to a target dimension, and inputting the hidden state sequence mapped to the target dimension into a conditional random field algorithm; The conditional random field algorithm labels the hidden state sequence after being mapped to the target dimension and calculates a score for each of the labels; The target tag with the highest score is obtained, and the personal information recognition result is obtained based on the target tag.

2. The method according to claim 1, characterized in that After acquiring the personal data set and the public data set, and before building a dedicated dictionary based on the personal data set and the public data set, the method further includes: The personal dataset and the public dataset are subjected to quality filtering and redundancy removal.

3. The method according to claim 2, characterized in that The quality filtering includes language filtering, indicator filtering, statistical feature filtering and keyword filtering; The redundancy removal includes sentence level removal, document level removal, and data set level removal.

4. The method according to claim 1, characterized in that The constructing a dedicated dictionary library based on the personal data set and the public data set includes: Building an initial dictionary based on the public dataset, and analyzing the personal dataset to generate candidate sequences; Generate a classification scoring model based on the initial dictionary library and the candidate sequence; The classification and scoring model is used to expand the initial dictionary library to obtain a special dictionary library.

5. The method according to claim 4, characterized in that The use of the classification and scoring model to expand the initial dictionary library to obtain a dedicated dictionary library includes: Using the classification scoring model to evaluate the candidate sequence; generating a language representation scorer based on the scored candidate sequences and the initial dictionary library; Using the language representation scorer to perform quality scoring on the candidate sequence; The initial dictionary library is expanded based on the quality score to obtain a dedicated dictionary library.

6. The method according to claim 1, characterized in that The obtaining of an embedding vector by training a language representation model based on the personal data set includes: Obtaining a specification document and a requirement characteristic, and performing entity definition on the personal data set according to the specification document and the requirement characteristic; The language representation model is used to train a personal data set that has completed the entity definition to obtain an embedding vector.

7. An information recognition system based on keywords and deep learning, characterized in that: The system comprises: An acquisition unit, used to acquire personal data sets and public data sets; A construction unit, configured to construct a dedicated dictionary library based on the personal data set and the public data set; An extraction unit, configured to extract keywords, specific symbols, and themes based on the public data set; A first determination unit, configured to determine whether the personal data set matches the keyword or the specific symbol; a determining unit, for determining a matching dictionary by using the subject and the dedicated dictionary library if no; a second judgment unit, configured to, after determining a matching dictionary, judge whether the matching dictionary matches the personal data set; A training unit, configured to, if not, obtain an embedding vector by training a language representation model based on the personal dataset; A modeling unit, configured to input the embedding vector into a sequence neural network model for modeling after obtaining the embedding vector, so as to obtain a hidden state sequence; An identification unit, used to map the dimension of the hidden state sequence to a target dimension, and input the hidden state sequence mapped to the target dimension into a conditional random field algorithm to obtain a personal information identification result; The identification unit is also used for: Mapping the dimension of the hidden state sequence to a target dimension, and inputting the hidden state sequence mapped to the target dimension into a conditional random field algorithm; The conditional random field algorithm labels the hidden state sequence after being mapped to the target dimension and calculates a score for each of the labels; The target tag with the highest score is obtained, and the personal information recognition result is obtained based on the target tag.

8. An information recognition device based on keywords and deep learning, characterized in that: The device comprises: Processor, memory, input-output unit, and bus; The processor is connected to the memory, the input and output unit, and the bus; The memory stores a program, and the processor calls the program to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a program stored thereon, wherein the program, when executed on a computer, performs the method according to any one of claims 1 to 6.

Images