Medical data document classification and tagging system

Abstract

The present invention relates to the field of medical data management technology, specifically to a medical data document classification and marking system, the system includes a document structure analysis module, a semantic information recognition module, a classification and marking module, and a document intelligent label module. In the present invention, by analyzing the document content, extracting the title, text and appendix, improving the document visualization and retrieval efficiency, by identifying key information and semantic relationships, accurately extracting and annotating medical terms, drug research data and medical records, enhancing the data parsability and operability, combining classification and labeling, optimizing storage with priority, clarifying the clinical value and applicable scenarios of the document, refining the document management process, and forming a management compatible with the clinical data format through structured storage and classification label generation of labeled data, improving storage standardization and retrieval efficiency, meeting the needs of accurate information call in clinical trials, improving document management capabilities, and providing support for clinical decision-making and research efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data management, and in particular to a medical data document classification and marking system. Background Art

[0002] The field of medical data management technology refers to various technologies and methods related to the collection, storage, processing, analysis and sharing of medical data. With the continuous development of digital medicine, the amount and complexity of medical data have increased significantly. Medical institutions need effective tools and systems to manage data. This field includes data standardization, data protection and privacy, data mining and analysis, etc., aiming to improve the accuracy, completeness and availability of data to support clinical decision-making, scientific research, patient care and other needs. Medical data management technology not only focuses on the structure and content of data, but also involves how to improve data processing efficiency and quality through intelligent means to promote the digital transformation of the medical industry.

[0003] Among them, the medical data document classification and labeling system is an important application in the field of medical data management technology, especially for the massive data documents generated in clinical trials. The main purpose of the system is to intelligently classify and label medical data documents through automation technology to facilitate rapid retrieval, analysis and utilization of key information. By reducing the workload of manual labeling and searching, it can significantly improve the efficiency of document management and ensure timely and accurate acquisition of required information during clinical trials. The system can also prioritize according to the clinical value and urgency of the document content, further improving the accuracy of clinical decision-making and trial efficiency.

[0004] When processing massive medical data documents, existing technologies lack the ability to automatically extract and hierarchically analyze the document content structure, which easily leads to a decrease in retrieval efficiency when the complexity of the document structure increases. Existing methods rely on manual operations to extract and label medical data fields, and lack the ability to automatically parse medical terms, drug research data, and medical records, resulting in inefficient data use and a high risk of missing key information. The classification and labeling process is not refined enough, making it difficult to quickly layer document content of different priorities, affecting the applicability of document content in different application scenarios. Existing systems lack standardization of the storage format of labeled data, resulting in poor compatibility during retrieval and increased complexity of data calls. The shortcomings are particularly evident in clinical trials, which not only prolongs the time for information retrieval and analysis, but also has an adverse impact on the accuracy and efficiency of trial decisions. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a medical data document classification and marking system.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme: The medical data document classification and marking system comprises:

[0007] The document structure analysis module is based on medical research documents, extracts document content, identifies the structural parts in the document, distinguishes the differentiated type information areas, extracts medical data fields, marks key symbols and terms in the document, obtains document structure recognition results, hierarchical weights, and the total number of words in paragraphs, extracts label information for each part of the text, performs classification storage, and obtains text label data sets;

[0008] The semantic information recognition module performs semantic recognition based on the text label data set, identifies key information, terms and phrases in the document, extracts semantic relationships and medical-specific vocabulary in sentences, annotates diseases, drugs and research program information, generates semantic recognition results, identifies potential relevance, medical relevance weight and medical relevance semantic criticality score in the text, and obtains semantic relevance data;

[0009] The classification and tagging module classifies the document content based on the semantic association data, identifies the document category, tags it in combination with the text priority, tags each document part according to the classification rules of drugs, diseases, and experimental progress, obtains document assignment label data, refines the classification information of each document through the tag content, identifies the field compensation parameters and the preliminary matching score, and obtains the document content tag data;

[0010] The document intelligent tag module generates document tags based on the document content tag data and the tag data, stores the tags in a structured form, identifies the distribution deviation in the semantic feature dimension through the clinical data tag format, analyzes and calculates the feature occurrence frequency weight and feature context distribution weight, and generates document classification tagging results.

[0011] As a further solution of the present invention, the step of obtaining the document structure recognition result is specifically:

[0012] Based on medical research documents, remove extra spaces and irrelevant symbols, unify text formats, clean up grammatical and punctuation errors, and generate standardized document text;

[0013] According to the standardized document text, the text hierarchy is analyzed line by line, and the title, paragraph and appendix information are separated. By counting the title keyword frequency and context relevance, the formula is adopted:

[0014] ;

[0015] Calculate the hierarchical weights of text paragraphs and generate a hierarchically marked document structure;

[0016] in, represents the hierarchical weight of the ith paragraph, is the frequency of the jth keyword, is the relevance of the jth keyword, is the total number of words in the paragraph, is the smoothing parameter, Indicates the total number of keywords;

[0017] According to the hierarchically marked document structure, a medical term list is used for matching, content that conforms to the medical data field pattern is marked, key medical data fields in titles and paragraphs are extracted, and a document structure recognition result is generated.

[0018] As a further solution of the present invention, the steps of acquiring the text label data set are specifically as follows:

[0019] Based on the document structure recognition result, parse the text blocks and corresponding label contents of each part of the document, extract the labels and text contents according to the specified start and end marks of the text paragraphs, verify the extracted contents, remove the non-compliant and duplicate label records, and generate a label matching data set;

[0020] Using the label matching data set, classify each label item, extract the corresponding text according to the classification standard, rearrange the text content and store it in groups, allocate storage paths according to categories, record the index and label mapping relationship in the association table, and obtain the classified storage data set;

[0021] Based on the classified storage data set, the storage content is retrieved item by item according to the label category, the internal fields of each storage content are parsed and the text is extracted, the redundant tags and redundant data are removed, the data is reorganized according to the standard format, and a text label data set is generated.

[0022] As a further solution of the present invention, the step of obtaining the semantic recognition result is specifically:

[0023] Parsing sentence-level information from the text label data set, extracting sentence components, calling basic semantic units, including the subject, predicate, object and modifiers of the sentence, marking the grammatical category and function of words, and generating a preliminary semantic parsing structure;

[0024] Using the preliminary semantic analysis structure, medical semantic matching is performed on the marked subject, predicate, object and modifying components, and the professional vocabulary associated with the sentence is located through the medical terminology database, using the formula:

[0025] ;

[0026] Calculate the medical-related semantic criticality score of the sentence, screen the medical-meaningful vocabulary and semantic relations in the sentence, and generate the semantic relation structure of the sentence;

[0027] in, represents the medical-related semantic criticality score of the sentence, is the frequency of occurrence of medical terms in sentences, is the medical relevance weight of the term, is the field compensation parameter, is the total number of words in the sentence, Indicates the total number of terms;

[0028] The semantic relationship structure of the sentence is called, and the semantic relationship between medical vocabulary and sentences is analyzed. The disease, drug and research plan information are annotated according to the strength of the association, and the semantic recognition result is generated in combination with the semantic information of the sentence.

[0029] As a further solution of the present invention, the step of acquiring the semantic association data is specifically as follows:

[0030] Based on the semantic recognition results, parse each text and decompose it into sentence structures, identify the core semantic elements in the sentences, extract keywords and phrases sentence by sentence, mark the direct association between the elements, establish semantic pairs corresponding to keywords and phrases, and obtain the semantic infrastructure;

[0031] Based on the semantic infrastructure, the context content is analyzed item by item according to the semantics, and the potential semantic chain between them is expanded by parsing the association logic of each pair of elements, the indirect association in the semantic chain is identified, and the multi-level association relationship between phrases and keywords is integrated to obtain the extended semantic relationship data;

[0032] Based on the extended semantic relationship data, the extended semantic chain is verified, invalid chains are eliminated, and the remaining chains are structured according to semantic logic and reorganized into a standardized semantic object set to obtain semantic association data.

[0033] As a further solution of the present invention, the step of obtaining the document assignment tag data is specifically as follows:

[0034] Based on the semantic association data, semantic features are extracted according to the document content, and a set of keywords in the fields of drugs, diseases and experimental progress is called to semantically segment and preliminarily mark the content in the document. Each document and classification rules are analyzed to optimize document segmentation. The formula is used:

[0035] ;

[0036] Calculate the classification matching degree of each piece of content and obtain the preliminary label assignment results of the document;

[0037] in, Indicates the classification matching degree, and Respectively represent the preliminary matching scores of documents under drug and disease classification, is the distribution deviation on the semantic feature dimension, and is the priority weight for drugs and disease categories;

[0038] The preliminary label assignment result of the document is called, the semantic features of each type of label are aggregated, the semantic similarity between the classification labels is analyzed, the distribution rule of the marking labels is optimized, the document is content-classified using the optimized rule, and the optimized label assignment result is obtained through the semantic distance;

[0039] Based on the optimized label assignment results, each document content is classified and marked according to the classification standards in the fields of drugs, diseases and experimental progress, duplicate tags are merged and the marking rules of the fuzzy boundary parts are adjusted, the semantic position relationship of the document paragraphs is analyzed, and the document assignment label data is obtained.

[0040] As a further solution of the present invention, the step of obtaining the document content mark data is specifically as follows:

[0041] Based on the document, label data is assigned, each tagged content is parsed and located in the document, document content is extracted paragraph by paragraph according to classification attributes, field consistency between the tag and paragraph content is compared, erroneous tags are checked, and tag information corresponding to each paragraph content is recorded to obtain document classification parsing data;

[0042] Based on the document classification and parsing data, the document content is segmented and reorganized according to the classification attributes, a mapping relationship is established between each paragraph of content and the hierarchical information of the corresponding mark, the order of the document paragraphs is rearranged according to the mark hierarchy, the classified core content of each paragraph is extracted to structure the data, and the document classification and reorganization data is obtained;

[0043] Based on the document classification and reorganization data, the content of each type of document is refined according to the marking rules, the hierarchical tags in the reorganized content are extracted and the tag inconsistencies are corrected to obtain the document content tag data.

[0044] As a further solution of the present invention, the step of obtaining the document classification and marking result is specifically as follows:

[0045] By analyzing each tag item in the document content tag data, extracting key feature information of the tag, and analyzing the occurrence frequency of each feature and the distribution characteristics in the document context, a preliminary tag set of feature frequency and distribution weight is generated;

[0046] Based on the preliminary tag set of the feature frequency and distribution weight, the tag set is subjected to weight normalization analysis according to the document classification requirements, using the formula:

[0047] ;

[0048] Calculate the tag weight value and obtain the feature weight data;

[0049] in, Represents the tag weight value, is the frequency weight of the feature, represents the feature context distribution weight, is the relevance weight between the feature and the document classification requirement, is the adjustment factor;

[0050] According to the feature weight data, features with weight values ​​higher than a threshold are screened as document labels, and are stored in a structured form to obtain document classification labeling results.

[0051] Compared with the prior art, the advantages and positive effects of the present invention are:

[0052] In the present invention, by analyzing the document content, the structural parts of the document, including the title, text and appendix, are accurately extracted, a hierarchical document relationship is established, and information areas are effectively distinguished, the visualization and retrieval efficiency of the document hierarchy and content are improved, key information and semantic relationships are identified based on the content, and the comprehensibility and operability of medical data are enhanced by accurately extracting and annotating medical terms, drug research data and medical records. Combined with the classification and labeling of the content, the classified storage of documents is optimized based on priority, the clinical value and applicable scenarios of the document content are clarified, and the management process of document content is refined from multiple dimensions. Through the structured storage and classification label generation of the marked data, an intelligent document management system compatible with the clinical data format is formed, which greatly improves the storage standardization and retrieval efficiency of the data, meets the needs of efficient and accurate call of key information in the clinical trial process, thereby improving the overall ability of document management and analysis, and laying a technical foundation for improving clinical decision-making and research efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a system flow chart of the present invention;

[0054] Figure 2 A flowchart of the document structure recognition result in the present invention;

[0055] Figure 3 This is a flowchart of the text label data set in the present invention;

[0056] Figure 4 It is a flow chart of the semantic recognition result in the present invention;

[0057] Figure 5 It is a flow chart of semantic association data in the present invention;

[0058] Figure 6 A flowchart for assigning tag data to documents in the present invention;

[0059] Figure 7 A flowchart for marking data for document content in the present invention;

[0060] Figure 8 This is a flowchart of the document classification and labeling results in the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0062] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0063] See also Figure 1 , the medical data document classification and tagging system includes:

[0064] The document structure analysis module is based on medical research documents, extracts document content, identifies the structural parts in the document, distinguishes the differentiated type information areas, extracts medical data fields, marks key symbols and terms in the document, obtains document structure recognition results, hierarchical weights, and the total number of words in paragraphs, extracts label information for each part of the text, performs classification storage, and obtains text label data sets;

[0065] The semantic information recognition module performs semantic recognition based on the text label data set, identifies key information, terms and phrases in the document, extracts semantic relationships and medical-specific vocabulary in sentences, annotates diseases, drugs and research program information, generates semantic recognition results, identifies potential relevance, medical relevance weight and medical relevance semantic criticality score in the text, and obtains semantic relevance data;

[0066] The classification and tagging module classifies the document content based on the semantic association data, identifies the document category, tags it in combination with the text priority, tags each document part according to the classification rules of drugs, diseases, and experimental progress, obtains document assignment label data, refines the classification information of each document through the tag content, identifies the field compensation parameters and the preliminary matching score, and obtains the document content tag data;

[0067] The document intelligent tag module generates document tags based on the document content tag data and the tag data, stores the tags in a structured form, identifies the distribution deviation in the semantic feature dimension through the clinical data tag format, analyzes and calculates the feature occurrence frequency weight and feature context distribution weight, and generates document classification tagging results.

[0068] The results of document structure recognition include title classification, text structure division, and appendix hierarchy extraction. The text labeling data set includes document part labels, key field tags, and information area identifiers. The results of semantic recognition include medical term extraction, disease information annotation, and semantic relationship recognition. Semantic association data include drug association tags, disease association tags, and research data association tags. Document assignment label data include document category classification, priority tags, and experimental progress classification. Document content labeling data include structure label tags, classification rule tags, and content priority information. Document classification labeling results include label storage format, classification label generation, and document label organization.

[0069] See also Figure 2 , the specific steps for obtaining the document structure recognition results are:

[0070] Based on medical research documents, remove extra spaces and irrelevant symbols, unify text formats, clean up grammatical and punctuation errors, and generate standardized document text;

[0071] Based on the original text content of the document, the cleaning process requires checking the grammar and punctuation of each line of text, including locating excess spaces in the document and removing them uniformly, calling sentence splitting rules to segment sentences, using standardized punctuation sets to re-match irregular punctuation, and automatically supplementing missing punctuation. By scanning the redundant symbols and redundant content in the text such as blank lines and meaningless characters, character matching rules are executed to remove irrelevant symbols, and text rules are used to detect ungrammatical phrases and reorganize and correct grammatical errors, clearing text segments with disordered formats, and further uniformly adjusting the structural format of the entire document to generate standardized document text.

[0072] According to the standardized document text, the text hierarchy is analyzed line by line, and the title, paragraph and appendix information are separated. By counting the title keyword frequency and context relevance, the formula is adopted:

[0073] ;

[0074] Calculate the hierarchical weights of text paragraphs and generate a hierarchically marked document structure;

[0075] in, represents the hierarchical weight of the ith paragraph, is the frequency of the jth keyword, is the relevance of the jth keyword, is the total number of words in the paragraph, is the smoothing parameter, Indicates the total number of keywords;

[0076] The formula is beneficial in that it combines keyword frequency and relevance, and uses smoothing parameters to adjust the impact of low-frequency keywords, thereby more accurately evaluating paragraph-level weights and facilitating structural tagging;

[0077] The number of times a keyword appears in a paragraph is obtained by counting the frequency of specific keywords in the current paragraph;

[0078] The importance relevance of keywords is calculated by the relative distribution ratio of keywords in the title and the text;

[0079] is the total number of words in the paragraph, obtained by counting the total number of words in each paragraph;

[0080] It is a smoothing parameter, which is set according to the overall mean value of the distribution of low-frequency keywords in the document and is used to reduce the impact of low-frequency keywords on weight calculation;

[0081] Setting data: keyword frequency , correlation , total number of words in the paragraph , smoothing parameter ;

[0082] Calculate the weight product and the sum of the keywords: ;

[0083] Add smoothing parameters: ;

[0084] Divide by the total number of words: ;

[0085] Calculate the level weight by taking the square root: ;

[0086] This result shows that the hierarchical weight of the i-th paragraph is 0.293, which indicates the hierarchical importance it occupies in the document structure. The numerical value is used to further label the hierarchical document structure.

[0087] According to the hierarchically marked document structure, the medical term list is used for matching, the content that conforms to the medical data field pattern is marked, the key medical data fields in the title and paragraph are extracted, and the document structure recognition result is generated;

[0088] According to the result data of the hierarchical marked document structure, the medical terminology table is called to match the key terms in the marked paragraph one by one according to the keyword matching rules. The identified medical fields are verified according to the predefined medical phrase format in the terminology table, and the field content that meets the semantic pattern is extracted and sorted. The content that does not meet the matching rules of the medical terminology table is marked as noise. The fields with insufficient semantic integrity are eliminated by setting the threshold for field pattern matching. At the same time, the semantic relevance of the matched fields is analyzed in combination with the context, and the high-frequency terminology model is called to calculate the word frequency distribution characteristics of the fields. Finally, the medical data fields contained in the identified document structure are generated to form the document structure recognition results.

[0089] See also Figure 3 , the specific steps for obtaining the text label dataset are:

[0090] Based on the document structure recognition results, parse the text blocks and corresponding label contents of each part of the document, extract the labels and text contents according to the specified text paragraph start and end marks, verify the extracted contents, remove the non-compliant and duplicate label records, and generate a label matching data set;

[0091] First, the overall structure of the document is decomposed, and the start and end ranges of the text blocks are located by analyzing them segment by segment. The label items marked in each paragraph and their corresponding text content are extracted. During the extraction process, the labels are parsed in turn, and the identifiers are matched with the text content and recorded in the initial data table. Duplicate verification rules are set, and the recorded label items are compared with the current parsed labels item by item. Redundant label records are eliminated based on uniqueness and format specifications. Through normalization verification, the integrity of the text corresponding to each label is confirmed, including ensuring that the length of the label is within the specified range and that the text content does not contain illegal characters. After eliminating data that does not meet the requirements, a label matching data set is generated in sequence. The data set needs to be verified for consistency to ensure that the logical association between the label and the content is correct and provide a formatting basis for subsequent operations.

[0092] Using the label matching dataset, classify each label item, extract the corresponding text according to the classification standard, rearrange the text content and store it in groups, assign storage paths according to categories, record the index and label mapping relationship in the association table, and obtain the classified storage dataset;

[0093] Compare the labels in the data set with the preset classification standards one by one, extract the corresponding text content according to the attribute value of the label, rearrange the classified labels and text content, and generate a new group storage format. During the grouping process, number each group of data to ensure the uniqueness of the storage path and label matching. Combined with the classification standards, an independent storage path is assigned to each type of data, and the mapping relationship between the grouped index value and the label is recorded in the association table to ensure that each index and label have a unique correspondence. After the classification is completed, the grouped data is written to the corresponding storage path, and a consistency check is performed on the stored data to verify the degree of matching between the text content and the classification standard, avoid omissions or storage path errors during the classification process, and form a classified storage data set for subsequent analysis.

[0094] Based on the classified storage data set, the storage content is retrieved item by item according to the label category, the internal fields of each storage content are parsed and the text is extracted, the redundant tags and redundant data are removed, and the data is reorganized according to the standard format to generate a text label data set;

[0095] Load the classification files under the storage path one by one, parse the internal fields in each record, extract the corresponding text content, and remove irrelevant marks such as line breaks, extra spaces, and non-essential symbols when extracting text. Eliminate redundant data items such as repeated fields and format errors, and reorganize the remaining data in a standard format to ensure that the data content is structured and compliant with specifications. To further optimize the results, establish data reorganization rules, adjust the order of text content through the logical relationship between fields, and ensure the rationality and readability of field arrangement. After the reorganization is completed, generate text label data sets by category. Each record contains labels, classification indexes, and reorganized text content, and outputs them to the specified storage path to ensure the efficiency and traceability of the entire process.

[0096] See also Figure 4 , the specific steps for obtaining semantic recognition results are:

[0097] Parse sentence-level information from text label datasets, extract sentence components, call basic semantic units, including the subject, predicate, object, and modifiers of the sentence, mark the grammatical category and function of words, and generate a preliminary semantic parsing structure;

[0098] Parse sentence-level information from the text label dataset. During the parsing process, the sentences need to be divided into paragraphs first, and the sentences in each paragraph need to be separated into independent units. Then, each sentence needs to be analyzed grammatically in turn. The grammatical parsing rules are called to group the subject, predicate, object and modifying components in the sentence, and each component after grouping is marked. During the marking process, the grammatical rationality of the sentence is checked in combination with the grammatical structure model, and the existing grammatical conflicts are adjusted. After the grammatical analysis is completed, the part-of-speech tagging rules are called to assign grammatical category labels to each word, such as nouns, verbs, adjectives and other categories. The grammatical components are recombined according to the part-of-speech tags and grammatical rules to ensure the integrity and accuracy of the sentence structure. Then, the semantic extensibility in the sentence is analyzed in combination with the context information, and the contextual connection relationship between sentences is processed. Finally, the analysis results are output as a preliminary semantic parsing structure.

[0099] Using the preliminary semantic analysis structure, the marked subject, predicate, object and modifiers are matched with medical semantics, and the professional vocabulary associated with the sentence is located through the medical terminology database, using the formula:

[0100] ;

[0101] Calculate the medical-related semantic criticality score of the sentence, screen the medical-meaningful vocabulary and semantic relations in the sentence, and generate the semantic relation structure of the sentence;

[0102] in, represents the medical-related semantic criticality score of the sentence, is the frequency of occurrence of medical terms in sentences, is the medical relevance weight of the term, is the field compensation parameter, is the total number of words in the sentence, Indicates the total number of terms;

[0103] The formula is beneficial in that it can better identify information related to the medical field and improve the accuracy of the overall semantic relationship by combining the frequency and relevance weights of terms, adjusting the field compensation parameters, and improving the pertinence of semantic screening;

[0104] is the frequency of occurrence of medical terms in sentences, obtained by counting the number of occurrences of the term in the current sentence;

[0105] is the medical relevance weight of the term, obtained by scoring the importance of the term in medical literature;

[0106] It is a domain compensation parameter used to balance the impact of different terms and is calculated and set based on the average distribution of medical terms.

[0107] is the total number of words in the sentence, obtained by counting the number of all words in the current sentence;

[0108] Data description: Setting the frequency of terms in a sentence , weight , total number of words in the sentence , field compensation parameters ;

[0109] Calculate the weight contribution of a term: ;

[0110] Added field compensation parameters: ;

[0111] Square root of the denominator: ;

[0112] Calculate semantic importance: ;

[0113] The result shows that the medical-related semantic criticality score of the sentence is 1.216, indicating that it has a high relevance in the medical field semantics and can be further used to generate semantic relationship structures.

[0114] Call the semantic relationship structure of the sentence, analyze the semantic relevance between sentences based on the semantic relationship between medical vocabulary and sentences, annotate the disease, drug and research plan information according to the strength of the association, and generate semantic recognition results based on the semantic information of the sentence;

[0115] The semantic relationship structure is called, and the grammatical components in the parsing results are matched in the medical field in combination with the medical terminology database. The matching process first screens the subject, predicate, object and modifying components in the parsing structure, and matches them one by one with the standard terms in the medical terminology database. The matching rules are used to verify whether each word in the sentence is associated with the entry in the terminology database, analyze the semantic contextual relationship of the words and identify terminology phrases. In the matching process, the recognition scope is expanded according to the semantic context information, and phrases that may be related to medicine are marked as candidate terms. The relevance of the terms and sentences is further verified according to the contextual distribution pattern of the terms. The overall semantic information of the sentence is screened and reorganized through the hierarchical semantic analysis method, and the vocabulary that conforms to the medical semantic rules is extracted. The extracted medical information is annotated according to categories such as diseases, drugs and research plans. Finally, the semantic recognition results are generated to complete the medical semantic annotation work.

[0116] See also Figure 5 ,The specific steps for obtaining semantically associated data are:

[0117] Based on the semantic recognition results, parse each paragraph of text and decompose it into sentence structures, identify the core semantic elements in the sentence, extract keywords and phrases sentence by sentence, mark the direct association between elements, establish semantic pairs corresponding to keywords and phrases, and obtain the semantic infrastructure;

[0118] First, each sentence is segmented and the subject, predicate and object in the sentence are marked. By analyzing the parts of speech of words in the sentence and the logical connections between adjacent words, a preliminary grammatical tree is generated, core semantic elements are extracted, and each core semantic element is marked. By giving priority to key words such as nouns, verbs, adjectives, etc. in the sentence, keywords and phrases are extracted, and significant core phrases are obtained by statistically analyzing the co-occurrence frequency and position distribution. Based on the marked core semantic elements and their word frequency characteristics, logical deduction and hierarchical analysis methods are used to mark the direct relationship between each keyword and related words, and clarify their dependency relationships, such as subject-predicate structure or verb-object collocation relationship. Each keyword and its related words are integrated into a semantic pair and stored in a structured form. By sorting out the multi-directional associations between keywords and phrases, a semantic infrastructure is formed.

[0119] Based on the semantic infrastructure, the context content is analyzed item by item according to the semantics. By parsing the association logic of each pair of elements, the potential semantic chain between them is expanded, the indirect association in the semantic chain is identified, and the multi-level association relationship between phrases and keywords is integrated to obtain the extended semantic relationship data;

[0120] First, the two end elements in the semantic pair are parsed, and their semantic distribution and position characteristics in the context are calculated. The potential semantic chain between them is further expanded by combining the grammatical rules and logical dependencies in the semantic pair. For each semantic pair, the context information is combined to determine whether it can form a potential semantic chain with logical significance, and the indirect association relationship contained in each chain is identified. The intermediary elements in the potential chain are located through continuous matching and path search, and their credibility is evaluated. By mining the multi-level association relationships between phrases and keywords, the semantic chains with high association are rearranged and merged. The semantic chains are verified in combination with the context content to obtain the extended semantic relationship data. The integrated semantic chains can cover a wider semantic range.

[0121] Based on the extended semantic relationship data, the extended semantic chain is verified, invalid chains are eliminated, and the remaining chains are structured according to semantic logic, and reorganized into a standardized semantic object set to obtain semantic association data;

[0122] The logical dependencies of each semantic pair in the chain are analyzed one by one to determine whether they conform to the semantic logic in the context. When invalid chains are eliminated, the misjudged chains are further analyzed and corrected by comprehensively considering the repetition rate of elements in the chain, the consistency of semantic direction, and the strength of associated logic. For the retained chains, structured processing is performed based on semantic logic, the elements in the chain are arranged in layers, and the chains are prioritized based on the degree of association. After verifying the multi-level association relationships contained in the sorted chains, the semantic chains are integrated into a standardized set of semantic objects, and the key features in the context are annotated for each semantic object to enhance its interpretability and accuracy in subsequent applications, and to obtain semantic association data to support the efficient execution of subsequent processing tasks.

[0123] See also Figure 6 , the specific steps for obtaining document assignment label data are:

[0124] Based on semantically associated data, semantic features are extracted according to the document content, and a set of keywords in the fields of drugs, diseases, and experimental progress is called to semantically segment and preliminarily mark the content in the document. Each document and classification rules are analyzed to optimize document segmentation. The formula is used:

[0125] ;

[0126] Calculate the classification matching degree of each piece of content and obtain the preliminary label assignment results of the document;

[0127] in, Indicates the classification matching degree, and Respectively represent the preliminary matching scores of documents under drug and disease classification, is the distribution deviation on the semantic feature dimension, and is the priority weight for drugs and disease categories;

[0128] The benefit of the formula is that by introducing the distribution bias of the semantic feature dimension and the category priority weight, the classification matching measurement is made more comprehensive, and the comprehensive measurement ability of the semantic association of document categories is enhanced;

[0129] : Classification matching degree, which indicates the matching degree between the current document content and the classification rules;

[0130] and : The semantic matching scores of the document content in drug classification and disease classification are calculated by the similarity of semantic vectors, which are 0.8 and 0.6 respectively;

[0131] : Distribution deviation in the semantic feature dimension, representing the standard deviation of feature distribution of document content in the three fields of drugs, diseases, and experimental progress, which are 0.3, 0.4, and 0.2 respectively;

[0132] and : The priority weights for drug and disease categories are set to 0.7 and 0.5, respectively, based on the document classification strategy;

[0133] Calculation derivation process:

[0134] Calculate the sum of squared deviations of the semantic feature dimension distribution in the denominator: ;

[0135] Calculate the numerator of the classification match: ;

[0136] Comprehensive calculation of matching degree:

[0137] ;

[0138] The results show that the comprehensive matching degree of document content under the drug and disease categories is 0.319. Combined with the optimized classification rules, the preliminary label assignment results of the document can be generated for the next step of content classification optimization.

[0139] Call the preliminary label assignment results of the document, aggregate the semantic features of each type of label, analyze the semantic similarity between the classification labels, optimize the distribution rules of the marked labels, use the optimized rules to classify the content of the document, and obtain the optimized label assignment results through semantic distance;

[0140] Call the document's preliminary classification tags, extract the keyword set corresponding to the classification rules based on the semantic association of the classification tags, analyze the matching degree of each tag with the document's semantic features, calculate the similarity of each content segment with all classification tags based on the semantic vector, optimize the tag distribution rules by sorting and adjusting the semantic similarity values, and in the tag optimization process, reallocate the classification tag priority weights through multiple rounds of iterative analysis in combination with contextual semantic features. When performing priority allocation, set the tag mutual exclusion adjustment rules based on the impact of key sentences in the document on its tag priority, and ensure that the classification weight coefficient is dynamically adjusted based on the contextual conditions. After the classification rules are generated, re-check the consistency of the tag optimization degree and priority allocation, and finally re-assign and update the tags to obtain the optimized tag allocation result.

[0141] Based on the optimized label assignment results, each document content is classified and labeled according to the classification standards of drugs, diseases and experimental progress, duplicate tags are merged and the tagging rules of the fuzzy boundary parts are adjusted, the semantic position relationship of the document paragraphs is analyzed, and the document assignment label data is obtained;

[0142] Based on the optimized label assignment results, the field classification standards of drugs, diseases and experimental progress are called, and the content boundary analysis is performed on the final classification and marking of each document. By merging the repeated content of the tagged tags, the fuzzy boundary parts are reclassified according to the upper and lower semantic relationship of the document paragraphs. When reclassifying, the classification labels are sorted according to the matching scores of the classification labels and the document content. Combined with the deviation value of the semantic feature distribution, the classification labels are reassigned to the paragraphs with large deviation values. The classification results are cross-validated through dynamic label correction rules, and the boundary adjustment part is determined after the matching relationship of the context labels is optimized twice. After the final label assignment is completed, the labeling results are checked for consistency, and the classification label hierarchical relationship is readjusted in combination with the overall semantic level of the document to ensure that the classification label distribution meets the final classification standard and generate document assignment label data.

[0143] See also Figure 7 , the specific steps for obtaining document content tag data are:

[0144] Based on the document assigned tag data, parse each tagged content and locate its position in the document, extract the document content paragraph by paragraph according to the classification attribute, compare the field consistency of the tag and paragraph content, check for incorrect tags, and record the tag information corresponding to each paragraph content to obtain document classification and analysis data;

[0145] By scanning the correspondence between the tag information and the document content, the attributes of the tag content and the paragraph content in the document are extracted paragraph by paragraph, and a field-level comparison is performed one by one. The core elements contained in the tag field, such as keywords, numbers or classification attributes, are matched with the related parts in the document paragraphs. Logical judgment is used to analyze whether the tag is missing, redundant or erroneous. When the tag is inconsistent with the paragraph content, the specific inconsistency is recorded, such as the tag field omitting important content or mislabeling the paragraph, and the paragraph number and tag index in the document are recorded at the same time. For the tag content with consistent fields, it is marked as a valid tag, and further verified whether its tag attributes are correctly attributed to the classification attributes. The paragraph number, paragraph content and corresponding tag information of the document are formed into a corresponding relationship table, and the document classification analysis data is obtained to ensure the accuracy of the tag and content matching.

[0146] Based on document classification and parsing data, the document content is segmented and reorganized according to classification attributes, and a mapping relationship is established between each paragraph of content and the hierarchical information of the corresponding mark. The order of the document paragraphs is rearranged according to the mark level, and the classified core content of each paragraph is extracted to structure the data and obtain document classification and reorganization data;

[0147] Extract the paragraph content and its marking attributes in the parsed data one by one, classify the content with the same classification attributes into the same group by parsing the hierarchical information and classification labels in the tags, and rearrange the paragraphs in each group of content according to the hierarchical order in the tags to ensure the logical coherence between paragraphs. In the reorganization process, extract the key phrases or core content in each paragraph, compare the key information with the classification tags to ensure the consistency of the content and the tags, check and complete the missing marking parts in the paragraphs, record the index, marking level and classification label of each paragraph involved in the reorganization, group and arrange the reorganized paragraph content according to the classification attributes, obtain document classification and reorganization data, and ensure that the document content logic is clear and meets the classification requirements.

[0148] Reorganize the data based on document classification, refine the content of each type of document according to the tagging rules, extract the hierarchical tags in the reorganized content and correct the tag inconsistencies to obtain the document content tag data;

[0149] Parse the paragraph content in the reorganized data paragraph by paragraph, extract the hierarchical tags and compare the correspondence between the tagged content and the actual content of the document one by one, check whether there are inconsistencies in the hierarchical tags, for example, correct the tagged content item by item for items with missing, repeated or incorrectly sorted tags, and record the correction results in a new tag data table. For the paragraph content of each type of document, further analyze the core attributes contained in its hierarchical tags, such as classification labels, sub-classification numbers or content descriptions, to ensure their accuracy and improve the details missed in the tags. Finally, form a standardized semantic pair with the corrected tagged data and the paragraph content, and integrate the tag information of all classified content into structured document content tag data to ensure the consistency of tagging rules and the integrity of content classification, providing accurate document tagging basic data for subsequent applications.

[0150] See also Figure 8 , the specific steps for obtaining the document classification and tagging results are:

[0151] By analyzing each tag item in the document content tag data, extracting key feature information of the tag, and analyzing the occurrence frequency of each feature and the distribution characteristics in the document context, a preliminary tag set of feature frequency and distribution weight is generated;

[0152] First, the tagged items are disassembled and classified one by one, and the main feature information of the tags is extracted, including the contextual dependency, frequency of occurrence and relevance weight of the features. The paragraphs and sentence structures of the document and the specific sources of each tagged data need to be analyzed, and they are classified and tagged according to their characteristic properties. The paragraph density and coverage of the tagged data distributed in the entire document are counted. Then, based on the statistical results of the paragraph density and the weight distribution value of the coverage, a frequency matrix of the feature information appearing in the document is constructed. At the same time, the distribution of high-frequency tagged items and low-frequency tagged items in the frequency matrix is ​​recorded. Further analysis is performed on whether there is a significant grouping trend in the contextual relationship. The importance of the feature information is ranked by weight threshold screening and dependency density checking. The sorted tagged data is organized into a feature set grouped by paragraph dependency weights to form a preliminary tag set of feature frequencies and distribution weights.

[0153] Based on the preliminary tag set of feature frequency and distribution weight, the tag set is weighted and normalized according to the document classification requirements, using the formula:

[0154] ;

[0155] Calculate the tag weight value and obtain the feature weight data;

[0156] in, Represents the tag weight value, is the frequency weight of the feature, represents the feature context distribution weight, is the relevance weight between the feature and the document classification requirement, is the adjustment factor;

[0157] The benefit of the formula is that by introducing context weight, relevance weight and adjustment coefficient, it comprehensively considers feature frequency and context relevance, balances the weight distribution between local feature density and global relevance, and effectively improves the accuracy of feature weight calculation and the interpretability of results;

[0158] : Obtained by counting the number of times each feature appears in the document and dividing it by the total number of words in the document. For example, if feature A appears 45 times and the total number of words in the document is 5000, then ;

[0159] : It is obtained by analyzing the density of context distribution of the labeled data. For example, feature A appears in three key paragraphs, and its context distribution weight is 0.7;

[0160] :It is obtained by calculating the correlation weight between the feature and the classification target, for example, the correlation score is 0.85;

[0161] Adjustment factor : Adjust according to the feature importance and context distribution difference range, for example, take the values ​​as 0.5 and 1.5 respectively;

[0162] Enter the formula to calculate:

[0163] ;

[0164] ;

[0165] ;

[0166] The result shows that the weight value of feature A in the document is 0.2614, which reflects its importance in the context and its relevance to the classification target, and provides a scientific basis for subsequent classification labeling.

[0167] According to the feature weight data, the features with weight values ​​higher than the threshold are selected as document labels, and stored in a structured form to obtain the document classification labeling results;

[0168] According to the feature frequency and context distribution weight obtained through calculation, the feature weight values ​​need to be arranged in descending order, and each feature weight value needs to be gradually compared to see if it meets the screening criteria. Features with higher weight values ​​are retained according to the screening threshold, and features with weight values ​​below the threshold are eliminated. In this process, the context relevance of the weight value and the importance score of the document topic need to be combined to ensure that the retained features have high-quality classification and differentiation capabilities. The filtered feature set is reorganized into a structured data format according to its context dependency weight and classification requirements, and further matched with the classification target requirements, and the document classification labeling results are output, which include the context dependency weight of each feature, the distribution density and the corresponding relationship between the classification labels in the document structure.

[0169] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A medical data document classification and marking system, characterized in that: The system comprises: The document structure analysis module is based on medical research documents, extracts document content, identifies the structural parts in the document, distinguishes the differentiated type information areas, extracts medical data fields, marks key symbols and terms in the document, obtains document structure recognition results, hierarchical weights, and the total number of words in paragraphs, extracts label information for each part of the text, performs classification storage, and obtains text label data sets; The semantic information recognition module performs semantic recognition based on the text label data set, identifies key information, terms and phrases in the document, extracts semantic relationships and medical-specific vocabulary in sentences, annotates diseases, drugs and research program information, generates semantic recognition results, identifies potential relevance, medical relevance weight and medical relevance semantic criticality score in the text, and obtains semantic relevance data; The classification and tagging module classifies the document content based on the semantic association data, identifies the document category, tags it in combination with the text priority, tags each document part according to the classification rules of drugs, diseases, and experimental progress, obtains document assignment label data, refines the classification information of each document through the tag content, identifies the field compensation parameters and the preliminary matching score, and obtains the document content tag data; The document intelligent tag module generates document tags based on the document content tag data and the tag data, stores the tags in a structured form, identifies the distribution deviation in the semantic feature dimension through the clinical data tag format, analyzes and calculates the feature occurrence frequency weight and feature context distribution weight, and generates document classification tagging results.

2. The medical data document classification and marking system according to claim 1, characterized in that: The steps for obtaining the document structure recognition result are specifically as follows: Based on medical research documents, remove extra spaces and irrelevant symbols, unify text formats, clean up grammatical and punctuation errors, and generate standardized document text; According to the standardized document text, the text hierarchy is analyzed line by line, and the title, paragraph and appendix information are separated. By counting the title keyword frequency and context relevance, the formula is adopted: ； Calculate the hierarchical weights of text paragraphs and generate a hierarchically marked document structure; in, represents the hierarchical weight of the ith paragraph, is the frequency of the jth keyword, is the relevance of the jth keyword, is the total number of words in the paragraph, is the smoothing parameter, Indicates the total number of keywords; According to the hierarchically marked document structure, a medical term list is used for matching, content that conforms to the medical data field pattern is marked, key medical data fields in titles and paragraphs are extracted, and a document structure recognition result is generated.

3. The medical data document classification and marking system according to claim 2, characterized in that: The steps for obtaining the text label dataset are specifically as follows: Based on the document structure recognition result, parse the text blocks and corresponding label contents of each part of the document, extract the labels and text contents according to the specified start and end marks of the text paragraphs, verify the extracted contents, remove the non-compliant and duplicate label records, and generate a label matching data set; Using the label matching data set, classify each label item, extract the corresponding text according to the classification standard, rearrange the text content and store it in groups, allocate storage paths according to categories, record the index and label mapping relationship in the association table, and obtain the classified storage data set; Based on the classified storage data set, the storage content is retrieved item by item according to the label category, the internal fields of each storage content are parsed and the text is extracted, the redundant tags and redundant data are removed, the data is reorganized according to the standard format, and a text label data set is generated.

4. The medical data document classification and marking system according to claim 3, characterized in that: The steps for obtaining the semantic recognition result are specifically as follows: Parsing sentence-level information from the text label data set, extracting sentence components, calling basic semantic units, including the subject, predicate, object and modifiers of the sentence, marking the grammatical category and function of words, and generating a preliminary semantic parsing structure; Using the preliminary semantic analysis structure, medical semantic matching is performed on the marked subject, predicate, object and modifying components, and the professional vocabulary associated with the sentence is located through the medical terminology database, using the formula: ； Calculate the medical-related semantic criticality score of the sentence, screen the medical-meaningful vocabulary and semantic relations in the sentence, and generate the semantic relation structure of the sentence; in, represents the medical-related semantic criticality score of the sentence, is the frequency of occurrence of medical terms in sentences, is the medical relevance weight of the term, is the field compensation parameter, is the total number of words in the sentence, Indicates the total number of terms; The semantic relationship structure of the sentence is called, and the semantic relationship between medical vocabulary and sentences is analyzed. The disease, drug and research plan information are annotated according to the strength of the association, and the semantic recognition result is generated in combination with the semantic information of the sentence.

5. The medical data document classification and marking system according to claim 4, characterized in that: The steps for acquiring the semantically associated data are specifically as follows: Based on the semantic recognition results, parse each text and decompose it into sentence structures, identify the core semantic elements in the sentences, extract keywords and phrases sentence by sentence, mark the direct association between the elements, establish semantic pairs corresponding to keywords and phrases, and obtain the semantic infrastructure; Based on the semantic infrastructure, the context content is analyzed item by item according to the semantics, and the potential semantic chain between them is expanded by parsing the association logic of each pair of elements, the indirect association in the semantic chain is identified, and the multi-level association relationship between phrases and keywords is integrated to obtain the extended semantic relationship data; Based on the extended semantic relationship data, the extended semantic chain is verified, invalid chains are eliminated, and the remaining chains are structured according to semantic logic and reorganized into a standardized semantic object set to obtain semantic association data.

6. The medical data document classification and marking system according to claim 5, characterized in that: The steps for obtaining the document assignment tag data are specifically as follows: Based on the semantic association data, semantic features are extracted according to the document content, and a set of keywords in the fields of drugs, diseases and experimental progress is called to semantically segment and preliminarily mark the content in the document. Each document and classification rules are analyzed to optimize document segmentation. The formula is used: ； Calculate the classification matching degree of each piece of content and obtain the preliminary label assignment results of the document; in, Indicates the classification matching degree, and Respectively represent the preliminary matching scores of documents under drug and disease classification, is the distribution deviation on the semantic feature dimension, and is the priority weight for drugs and disease categories; The preliminary label assignment result of the document is called, the semantic features of each type of label are aggregated, the semantic similarity between the classification labels is analyzed, the distribution rule of the marking labels is optimized, the document is content-classified using the optimized rule, and the optimized label assignment result is obtained through the semantic distance; Based on the optimized label assignment results, each document content is classified and marked according to the classification standards in the fields of drugs, diseases and experimental progress, duplicate tags are merged and the marking rules of the fuzzy boundary parts are adjusted, the semantic position relationship of the document paragraphs is analyzed, and the document assignment label data is obtained.

7. The medical data document classification and marking system according to claim 6, characterized in that: The steps for obtaining the document content mark data are specifically as follows: Based on the document, label data is assigned, each tagged content is parsed and located in the document, document content is extracted paragraph by paragraph according to classification attributes, field consistency between the tag and paragraph content is compared, erroneous tags are checked, and tag information corresponding to each paragraph content is recorded to obtain document classification parsing data; Based on the document classification and parsing data, the document content is segmented and reorganized according to the classification attributes, a mapping relationship is established between each paragraph of content and the hierarchical information of the corresponding mark, the order of the document paragraphs is rearranged according to the mark hierarchy, the classified core content of each paragraph is extracted to structure the data, and the document classification and reorganization data is obtained; Based on the document classification and reorganization data, the content of each type of document is refined according to the marking rules, the hierarchical tags in the reorganized content are extracted and the tag inconsistencies are corrected to obtain the document content tag data.

8. The medical data document classification and marking system according to claim 7, characterized in that: The steps for obtaining the document classification and marking results are specifically as follows: By analyzing each tag item in the document content tag data, extracting key feature information of the tag, and analyzing the occurrence frequency of each feature and the distribution characteristics in the document context, a preliminary tag set of feature frequency and distribution weight is generated; Based on the preliminary tag set of the feature frequency and distribution weight, the tag set is subjected to weight normalization analysis according to the document classification requirements, using the formula: ； Calculate the tag weight value and obtain the feature weight data; in, Represents the tag weight value, is the frequency weight of the feature, represents the feature context distribution weight, is the relevance weight between the feature and the document classification requirement, is the adjustment factor; According to the feature weight data, features with weight values ​​higher than a threshold are screened as document labels, and are stored in a structured form to obtain document classification labeling results.

Images