A health policy automatic classification method based on multi-dimensional feature fusion

Abstract

The present application relates to the technical field of policy text recognition and classification, and proposes a health policy automatic classification method based on multi-dimensional feature fusion. The system preprocesses the policy original text, extracts deep semantic features using the BERT encoder, constructs domain ontology features combined with TF-IDF, and extracts information such as document level and policy tools to generate structural features. Through the attention module, the three types of features are mapped to the same space and the weights are adaptively allocated, and the comprehensive features are fused. The multi-label classification module presets 12 labels and deduces in parallel, and outputs the subdivided classification results. The present application overcomes the defects of traditional methods such as poor fine-grained recognition and lack of structured data mining, can clearly distinguish health science and technology policies from regular documents, and effectively improves the efficiency of policy archiving, intelligent retrieval and government decision-making assistance.

Description

Technical Field

[0001] The present invention relates to the field of policy text recognition and classification, in particular to an automatic classification method for health policies based on multi-dimensional feature fusion, which can be applied to the scenario of automatic classification of health science and technology policy texts. Background Art

[0002] In recent years, with the rapid development and wide application of natural language processing technology and artificial intelligence, the artificial intelligence-based automatic text classification technology has received extensive attention in various fields. At present, the automatic classification oriented to health policy texts remains a highly challenging subject, involving multiple disciplines such as natural language processing, knowledge engineering, policy analysis, and machine learning, and has broad application prospects in the fields of policy knowledge management, intelligent retrieval, auxiliary decision-making, and policy effect evaluation.

[0003] Policy text classification is a basic link in the research of policy text information analysis. At present, most research directions in policy text classification are classification based on a single text feature, for example, using term frequency statistics (TF-IDF) or a general pre-trained language model (BERT) to extract semantic features, whose main purpose is to realize coarse-grained archiving and classification of texts. The focus of classification research on health policy texts is multi-dimensional fine-grained annotation, that is, one policy text needs to be annotated with multiple dimensional labels such as strategic planning, industrial innovation, clinical research, data management, project management, achievement transformation, talent management, base and platform management, ethical supervision, integrity supervision, new technology supervision, and biosafety at the same time. Since health policy texts involve a large number of professional terms, and "science and technology-related health policies" and "ordinary medical administrative documents" are often highly similar in text expression, it is difficult to effectively distinguish them only by a single text feature; at the same time, policies at different issuing levels have significant differences in expression style and term use, and policy texts also contain rich structured information such as issuing level, style and format, and policy tools, which are not fully utilized by existing methods. In addition, most of the existing feature fusion methods are simple splicing or static weighting, lacking the ability to adaptively learn the importance of features in various dimensions, and all these factors increase the difficulty of automatic classification of health policy texts. At present, comparative policy research mainly relies on manual reading and qualitative analysis, which has low efficiency and limited coverage.

[0004] Therefore, an automatic classification method for health policies based on multi-dimensional feature fusion is invented. This method constructs a classification framework that integrates textual semantic features, domain ontology features, and policy structure features. It utilizes an attention mechanism to adaptively learn the importance weights of each dimension's features, thereby achieving high-precision multi-label automatic classification of health science and technology policy texts. This method can effectively distinguish between fine-grained categories such as "science and technology-related health policies" and "ordinary medical administrative documents," which is of great significance for improving the management efficiency, knowledge retrieval accuracy, and decision-making support capabilities of health policies. Summary of the Invention

[0005] This invention provides an automatic classification method for health policies based on multi-dimensional feature fusion. Addressing the problem that existing text classification technologies rely on single text features and have difficulty distinguishing fine-grained policy categories, this invention constructs a classification framework that integrates three dimensions: text semantic features, domain ontology features, and policy structure features. By combining an attention mechanism to adaptively learn the importance weights of each dimension of features, this invention achieves high-precision multi-label automatic classification of health science and technology policy texts.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: Step 1: Collect and aggregate the original text information of health policies to be classified. The steps are as follows: Step A1: The system receives the original text of health policies input by the user or imported in batches through the interface. The text includes, but is not limited to, fields such as policy document name, issuing agency, release date, and policy text. Step A2: Perform integrity verification on the input text, including uniform conversion of text encoding format (unified to UTF-8 encoding), cleaning of non-text characters (such as special control characters and redundant whitespace characters), and text length compliance check, to ensure that the input data in subsequent processing flows is standardized and complete.

[0007] Step 2: Preprocess the text, including Chinese word segmentation, stop word removal, dependency parsing, etc., the steps of which are as follows: Step B1: Chinese word segmentation. The domain-enhanced jieba word segmentation tool is used, and a custom dictionary specifically for the health policy field is loaded to segment the policy text to ensure that terms in fields such as "major science and technology projects", "technology transfer", and "medical-research-enterprise collaboration" are not incorrectly segmented. Step B2: Stop word removal. Based on the preset general stop word list and the policy field stop word list, remove high-frequency functional words, punctuation marks and meaningless characters that do not contribute to the classification. Step B3: Word form normalization, converting traditional Chinese characters to simplified Chinese characters, unifying full-width and half-width characters, and standardizing the expression of numbers, dates, etc. Step B4: Dependency parsing. Using syntactic analysis tools, dependency relations are annotated on the segmented text to extract the subject-verb-object structure and modification relations of the sentences, providing structured input for subsequent semantic feature extraction.

[0008] Step 3: Input the preprocessed text into the BERT encoder to extract the deep semantic representation of the text. The specific steps are as follows: C1: The preprocessed text is serialized into a set of token sequences and converted into corresponding word vectors, position vectors, and segment vectors, which are then input into the pre-trained BERT encoder. The BERT encoder consists of multiple stacked Transformer encoders that encode the token sequences layer by layer using a multi-head self-attention mechanism and feedforward neural network layers, capturing the bidirectional contextual dependencies between words. C2: Take the output vector corresponding to the [CLS] label of the last hidden layer of the BERT encoder as the deep semantic representation vector s of the entire policy text, where , For semantic features, [CLS] is an extra special symbol added at the beginning of a text. After the full policy text is fed into BERT for calculation, the string of numbers corresponding to the [CLS] position in the last network layer is directly extracted. This string of numbers condenses the overall theme, core ideas, and deeper meaning of the full policy text, and can be used for tasks such as policy similarity comparison, classification, clustering, and retrieval.

[0009] Step 4: Match the preprocessed text with a specialized thesaurus for the health and medical technology field, and calculate the TF-IDF weights of terms in each category. The specific steps are as follows: D1: Based on a pre-built thesaurus of specialized terms in the field of health and medical technology, terminology matching is performed on the pre-processed text. The thesaurus covers standard terms under 12 thematic categories: strategic planning, industrial innovation, clinical research, data management, project management, technology transfer, talent management, base and platform management, ethical supervision, integrity supervision, new technology supervision, and biosafety.

[0010] D2: For each matched term, calculate its term frequency (TF) and inverse document frequency (IDF) in the current document to obtain the TF-IDF weight value of each term. Aggregate the TF-IDF weights of terms under all topic categories by category to form the domain ontology feature vector. ,in For domain feature dimensions, .

[0011] Step 5: Extract the issuing level, document format, policy tools, and implementation stage features from the structured metadata of the policy text. The steps are as follows: E1: Based on the level of the issuing organization, the document level is encoded as a one-hot vector, including four levels: national, provincial / ministerial, municipal, and county / district. E2: Identify policy document types (strategic planning, industrial innovation, clinical research, data management, project management, technology transfer, talent management, base / platform management, ethical regulation, integrity regulation, new technology regulation, biosafety), and encode them as discrete vectors; E3: Identify the types of supply-side, demand-side, and environmental policy tools used in policy texts based on keyword rule bases, count the frequency of mention of each type of policy tool and normalize it into a weight vector; E4: Based on the time nodes and implementation progress information in the policy text, mark the implementation stage of the policy (planning, implementation, supervision and evaluation, summary and acceptance). Concatenate the above four types of structured features and generate a policy structure feature vector after linear mapping. ,That For structural feature dimension, .

[0012] Step 6: Calculate the importance weights of each dimension of features based on the input text features using a self-attention mechanism. The steps are as follows: F1: Map the three-dimensional feature vectors s, o, and u obtained in steps 3 to 5 to a unified attention space using linear projection matrices to obtain the projected features. , , All three dimensions are ; F2: Calculate the attention score for each projected feature. ,in For learnable parameters, The projected feature vector, i.e. T represents a summary of policy categories.

[0013] Step 7: Perform weighted fusion of the features in the three dimensions according to the attention weights, and output the fused feature vector f.

[0014] Step 8: Input the fused features into the multi-label classifier and output the label probabilities of each classification dimension in parallel; the steps are as follows: G1: Input the fused feature vector f into a multi-label classifier. The classifier contains K parallel independent classification sub-networks, each of which corresponds to a classification dimension. G2: For K classification dimensions, the prediction probability of each candidate javelin is calculated using the sigmoid activation function. ,in, Let be the weight vector of this subnetwork. This is the bias term, where k is the number of categories (k=1,2,...,K,K=12). Using the Sigmoid function, each sub-network independently outputs the probability value of each label in that dimension, enabling parallel prediction of multi-dimensional labels.

[0015] Step 9: Based on the set thresholds, output the final classification labels for each classification dimension; the steps are as follows: H1: For the label probabilities of each dimension output in step 8, a threshold strategy is adopted, and a probability threshold is preset for each dimension k. When the predicted probability of a certain label The label is determined to be a positive class if it is positive, otherwise it is a negative class.

[0016] Preferably, when performing step 2, a custom dictionary specifically for the field of health and wellness policies is constructed (such as: major science and technology projects, achievement transformation, medical research and enterprise collaboration, pharmaceutical price reform, hierarchical diagnosis and treatment system, hierarchical diagnosis and treatment, national immunization program, Class I vaccine program, infectious disease prevention, integration of traditional Chinese and Western medicine, equal emphasis on traditional Chinese and Western medicine, smart hospitals, digital hospitals, hospital information construction, etc.) to solve the problem that general word segmentation dictionaries cannot recognize health and wellness policy-specific terms, abbreviations, institution names, policy document numbers, disease standards, and medical reform-specific concepts.

[0017] Preferably, during text preprocessing, a standardized Chinese character lookup table is used to convert traditional Chinese characters to simplified Chinese, and a whitelist of proper nouns is set to avoid misconversion of names and technical terms. Character encoding conversion is used to uniformly convert full-width numbers, letters, and symbols to half-width format, while Chinese punctuation retains its original style. Regular expression matching and date parsing tools are used to convert uppercase and lowercase Chinese numbers to Arabic numerals, and all abbreviations and Chinese dates are standardized to YYYY-MM-DD format. After processing, content validation is performed, and abnormal data is logged for review and optimization. The final output is standardized text with uniform format and standard terminology, adaptable to the needs of text retrieval, semantic analysis, and other business scenarios.

[0018] Preferably, when performing step 3, the text semantic vector output by the native BERT is usually a high-dimensional floating-point dense vector. Due to the high dimensionality of the data and the enormous size of the vector library when storing massive amounts of policy documents, this paper employs PCA (Principal Component Analysis) to linearly reduce the dimensionality of the original high-dimensional BERT semantic vectors and introduces cosine similarity constraints to calibrate the vector space distribution, compressing the dense vectors to 256 dimensions. This eliminates redundant features and subtle noise in the high-dimensional vectors, significantly reducing the storage space of a single vector while strictly controlling semantic representation loss. Simultaneously, it reduces the computational overhead of cosine similarity calculation and vector nearest neighbor retrieval, effectively improving the storage efficiency and retrieval throughput of the massive policy text vector library.

[0019] Preferably, in step 7, to improve the accuracy of policy document classification, the text semantic features, domain ontology features, and policy structure features are weighted and fused to obtain a fused feature vector f. , ,in , , Let be the learnable parameters of the fusion layer, where the attention weights are . .

[0020] Preferably, when executing step 9, the policy text types are divided into a total of 12 categories (strategic planning, industrial innovation, clinical research, data management, project management, technology transfer, talent management, base platform management, ethical supervision, integrity supervision, new technology supervision, and biosafety), and the predicted label for the k-th dimension is... ,in is the classification threshold for the k-th dimension.

[0021] This invention belongs to the field of intelligent recognition and classification technology of health policy texts, specifically a method for automatic classification of health policies based on multi-dimensional feature fusion. It aims to overcome the industry challenge of existing policy classification technologies relying on single text features for computation, which cannot adapt to the fine-grained, multi-label annotation requirements of health policies. This invention constructs a three-dimensional feature fusion system encompassing text semantics, domain ontology, and policy structure. Relying on an attention mechanism, it adaptively adjusts the weights of features across various dimensions, solving many pain points of traditional methods, such as difficulty in distinguishing between science and technology-related health policies and routine medical administrative documents, failure to fully extract structured policy information, rigid feature fusion models, and low efficiency of manual judgment. This provides solid technical support and reliable reference for intelligent management, accurate retrieval, policy judgment, and effect evaluation of health policies. The classification results of this solution are highly accurate, with excellent environmental adaptability, operational stability, and scalability, fully meeting the practical needs of various business scenarios in the health field, including policy archiving, intelligent retrieval, and auxiliary administrative decision-making. Attached Figure Description

[0022] Figure 1 Flowchart for automatic classification of health policies based on multi-dimensional feature fusion; Figure 2 A flowchart of text preprocessing; Figure 3 Flowchart for dimensionality reduction of semantic vectors. Detailed Implementation

[0023] When researchers conduct health science and technology policy assessments and academic research, they often need to carry out specialized sorting, in-depth analysis, and comparative verification of specific policy categories. However, at present, the volume of health policies continues to expand and the text content is complex. Existing classification methods mostly rely on single text features for rough classification, making it difficult to simultaneously label strategic planning, industrial innovation, clinical research, and other multi-dimensional fine-grained tags. In addition, the textual expression of science and technology health policies is highly similar to that of ordinary medical administrative documents, and the structured information such as policy level and document format has not been effectively mined and utilized. Relying on manual reading and qualitative analysis is not only time-consuming and laborious with low work efficiency, but also prone to classification errors, making it difficult to meet the needs of refined and specialized policy research.

[0024] Based on the aforementioned practical pain points and application limitations, this invention selects a multi-dimensional feature fusion approach to build a three-dimensional feature classification framework that integrates text semantics, domain ontology, and policy structure. It also introduces an attention mechanism to adaptively allocate the weights of features in each dimension, thereby achieving high-precision multi-label automatic classification of health science and technology policy texts.

[0025] The technical solutions in the embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings.

[0026] See Figure 1 Automatic classification of health policies based on multi-dimensional feature fusion mainly includes policy text collection, text processing, feature extraction, weight calculation, feature fusion, and label classification.

[0027] The policy text collection process involves user input and access to the policy document data interface. The collected text dimensions include, but are not limited to, fields such as policy document name, issuing agency, release date, and policy text, in order to process and classify the text.

[0028] See Figure 1 , Figure 2 The text preprocessing module includes filtering text content, normalizing text content, standardizing text content, validating text content, and annotating text dependencies.

[0029] This process involves conducting pre-processing data integrity and format verification, unifying all text encoding to UTF-8 to avoid garbled characters; batch cleaning of dirty data such as invisible control characters, redundant line breaks and spaces, and invalid special symbols, with corresponding pre-processing filtering in the flowchart: removing invisible special characters and control characters; simultaneously verifying text length thresholds, filtering blank fragments and excessively short or invalid content, verifying field integrity and filling in missing basic fields; after pre-processing cleaning and verification, the neat and clean raw data officially flows into the refined pre-processing chain.

[0030] Step I: The text refinement preprocessing stage involves sequentially executing word segmentation, stop word removal, lexical normalization, and syntactic parsing. The lexical normalization stage strictly follows the flowchart's sequence, and its steps are as follows: Step I1: Domain-specific Chinese word segmentation. The optimized version of the jieba word segmentation framework is selected, and a professional dictionary dedicated to health and wellness policies is imported to include specialized terms in medical research, medical management, and public health. This ensures that professional terms such as "major science and technology projects," "technology transfer," "medical-research-enterprise collaboration," and "grassroots medical and health institutions" are completely segmented, and eliminates errors in terminology segmentation. Step I2: Two-layer stop word filtering, which integrates a general stop word library and a health policy-customized stop word library to remove function words, modal particles, redundant punctuation, and meaningless conjunctions, thereby simplifying redundant information in the text and reducing the interference of invalid data on subsequent semantic operations; Step I3: Multi-dimensional word form normalization processing. First, batch conversion of traditional Chinese characters to simplified Chinese characters is completed. Then, full-width numbers, letters, and common punctuation are uniformly converted to half-width characters, while retaining the Chinese punctuation format. Standardization and unification of variant words are carried out. Finally, the standardization of numbers and dates is completed, with Chinese uppercase and lowercase numbers uniformly converted to Arabic numerals, and various scattered dates standardized into a fixed YYYY-MM-DD format, thus unifying the writing standards of text characters in all aspects. After the conversion is completed, the results are checked and corrected to identify conversion deviations and residual format issues. Abnormal data is logged for iterative optimization. Step I4: Dependency parsing. Based on the natural language parsing model, the dependency relationships between words in the normalized text after verification are marked. The main components of the subject-verb-object structure and the modifier logic of the attributive and adverbial clauses are decomposed. The hierarchical structure of the sentences is sorted out, and structured semantic data is output. Finally, a standardized document is formed, which provides standardized feature materials for subsequent upper-level business such as text classification, topic extraction, and policy semantic comparison. The preprocessing process is closed.

[0031] See Figure 1 The special extraction process involves vectorizing text semantic features, domain ontology features, and policy structure features. The steps to obtain text semantic features are as follows: Step J1: Obtain text semantic feature data, select the Chinese BERT-base pre-trained model, and the maximum input token length of the model is 512; count the number of tokens in the cleaned text segmentation. If the total number of tokens exceeds 510, truncate the excess tokens from the end of the text, and prioritize retaining key content such as policy general principles and core subsidy clauses; if the total number of tokens is less than 510, fill in placeholders at the end of the text to ensure that the length of the input sequence is uniform. Step J2: Token segmentation and serialization: Call the built-in WordPiece Chinese word segmenter in BERT to segment the preprocessed policy text, forcibly insert a special classification tag [CLS] at the beginning of the sequence, add a separator tag [SEP] at the end of the text, and generate a complete token sequence; Step J3: Three types of embedding vector mapping: pre-trained BERT built-in word vector table, position encoding table, and segment encoding table, independently generating three sets of 768-dimensional real vectors for each token in the sequence; Step J4: A 12-layer stacked Transformer encoder is used to form the main body of the BERT encoder. Each layer of the encoder performs multi-head self-attention calculation and feedforward neural network transformation in sequence. The feature matrix output by the first layer of the Transformer is used as the input of the second layer, and so on until the 12th layer (the last hidden layer). The shallow Transformer mainly extracts the lexical and syntactic surface features, while the deep layers aggregate the overall logic, policy guidance, and core constraint information of the whole text layer by layer. The final output size is still the feature matrix of the last hidden layer [512, 768]. Step J5: Locate the output corresponding to [CLS] in the feature matrix. The first element of the token sequence is marked with [CLS], corresponding to the vector in the 0th row of the feature matrix of the last hidden layer. This vector is a 768-dimensional one-dimensional real number array, which is the global deep semantic representation vector s of the entire policy text. , This is the semantic feature dimension.

[0032] See Figure 1 , Figure 3 Principal Component Analysis (PCA) with cosine similarity constraints is used to linearly reduce the 768-dimensional high-dimensional policy semantic dense vectors output by the BERT model. While accurately preserving the relative semantic relationships of the texts, the high-dimensional vectors are uniformly compressed to 256 dimensions, effectively reducing the storage space of the vector library, reducing the computational overhead of cosine similarity, and significantly improving the computational efficiency of downstream tasks such as policy text retrieval, clustering, and classification. Moreover, the overall semantic loss is controllable. The specific steps are as follows: Step K1: Vector standardization preprocessing. Assume the experimental dataset contains N policy texts. Each text is encoded by a BERT encoder, which outputs a 768-dimensional semantic vector. All vectors together constitute the input feature matrix. N represents the total number of policy text samples; R represents the real number field; and the matrix dimension is... Each row represents one text entry, and each column represents one semantic feature. To eliminate the interference of vector magnitude on semantic similarity calculation, cosine distance is uniformly used to represent text semantic differences, and L2 normalization is performed on the original vectors. The core formula is: , Let i be the original 768-dimensional BERT semantic vector of the i-th policy text. This is the normalized standard semantic vector; This is an L2 number operation used to calculate the vector magnitude. After this processing, the magnitude of all vectors is unified to 1, ensuring that subsequent semantic similarity is determined solely by the angle between the vector spaces.

[0033] Step K2: Standard PCA Principal Component Analysis. The normalized vector dataset is centered to eliminate dimensionality shift errors, and then the global covariance matrix is ​​calculated. C is The semantic covariance matrix is ​​used to characterize the degree of dispersion of each semantic dimension and the correlation between dimensions. Let be the transpose of the normalized vector.

[0034] Step K3: Perform eigenvalue decomposition on the covariance matrix. , The matrix is ​​a diagonal eigenvalue matrix, where the values ​​on the diagonal are eigenvalues. The magnitude of the value corresponds to the semantic information content of the principal component; the larger the value, the stronger the semantic discriminative ability. V is an orthogonal eigenvector matrix, where each column corresponds to a new set of orthogonal semantic projection directions. All eigenvalues ​​are sorted in descending order of magnitude, and the top 256 principal components with the highest information content are selected as the core effective features for dimensionality reduction.

[0035] Step K4: The cosine similarity constraint is used to calibrate the vector space. The native PCA algorithm, which optimizes by maximizing the variance of the Euclidean space, only focuses on data dispersion and easily disrupts the original semantic proximity relationships of high-dimensional vectors, leading to distortion of the text cosine similarity after dimensionality reduction. To address this issue, this paper introduces a cosine similarity constraint. By minimizing the deviation of the inner product matrix of vectors before and after dimensionality reduction, the semantic space relationships are locked, and a loss objective function is constructed: , The inner product matrix of the original high-dimensional vectors represents the semantic similarity relationship between the original texts. The inner product matrix of the reduced-dimensional vectors. This is the square of the Frobenius norm, used to calculate the overall error of the matrix.

[0036] Step K5: Linear projection completes dimensionality reduction. Based on the selected 256-dimensional optimal principal components, a dimensionality transformation projection matrix is ​​constructed. The matrix's row dimensions correspond to the original 768-dimensional features, and its column dimensions correspond to the target 256-dimensional features. Dimensional compression is achieved through a linear projection formula. See Figure 1 The steps to obtain domain ontology feature data are as follows: Step L1: A specialized thesaurus for the health and medical technology field is pre-built offline. The dictionary uses a hierarchical structured storage. The top layer is divided into 12 mutually exclusive primary thematic categories that fully cover industry policies: strategic planning, industrial innovation, clinical research, data management, project management, technology transfer, talent management, base / platform management, ethical regulation, integrity regulation, new technology regulation, and biosafety. Each primary thematic category includes a secondary standard terminology database. Terminologies are categorized into three types: basic nouns, verb-object policy phrases, and proprietary institutional terms. Each term includes a unique category identifier, term length, industry semantic tags, and synonym extensions. The dictionary is deployed in a dual format: a structured database and an in-memory retrieval dictionary (Trie prefix tree). The database persistently stores complete terms and metadata, while the Trie tree is loaded into memory for high-speed text matching. Step L2: Perform multi-granularity precise term matching according to specified rules. Prioritize matching longer compound terms from the dictionary, then match shorter single-word terms to avoid semantic loss caused by shorter terms covering longer terms. For example, if the text contains "clinical research ethics supervision," prioritize matching the long compound term and avoid splitting it into two shorter terms, "clinical research" and "ethics supervision," which would be counted repeatedly. If the text segmentation is completely consistent with a synonym expansion entry of a certain standard term, it is automatically mapped to the corresponding standard term and assigned to the first-level topic category bound to that standard term for unified classification and statistics, eliminating interference from differences in text wording.

[0037] Step L3: For each successfully matched term, three types of metadata are recorded simultaneously: ① the matched term text; ② the unique category code of the term's 12 major categories; ③ the offsets of all occurrences of the term within the current policy document. Term segments that do not match any term in the dictionary are discarded and do not participate in subsequent weight calculations. After completing the full-text traversal matching, a structured matching data table is generated. Each record in the table is bound to "Document ID + Standard Term + Topic Category + Coordinates of All Occurrences in the Text," serving as the basic data source for TF and IDF calculations in step D2.

[0038] Step L4: For the matching result set of the current single policy document, calculate the local term frequency (TF) of each term category and each item, using the original term frequency calculation method. , The word frequency of term t in the current document d. The total number of occurrences of term t in document d that match the document. This represents the total number of all entries in document d that match domain dictionary terms. Step L5: DF is pre-calculated offline based on a pre-defined global corpus of policy documents and updated periodically. The global corpus contains a set D of all health and medical science and technology policy documents to be analyzed, with a total document count of |D|. The IDF calculation formula uses a smoothed logarithmic form to avoid the problem of a denominator of 0 caused by terms appearing only in a single document. doc_num(t) is the total number of documents in the global corpus that match term t; Step L6: TF-IDF weight fusion calculation: Multiply the single-document TF value by the globally stored IDF value to obtain the final weight of the term in the current document. ; Step L7: Aggregate the TF-IDF weights of terms under all topic categories by category to form the domain ontology feature vector. ,in For domain feature dimensions, .

[0039] See Figure 1 The steps to obtain policy structure characteristic data are as follows: Step M1: Publish the construction of the rule base for determining the level of the organization, and pre-build a dictionary for matching the administrative organization level. The dictionary contains keyword sets for national, provincial and ministerial, municipal and county-level organizations. Step M2: One-hot vector encoding is implemented, with 4 hierarchical dimensions set. The encoding dimension order is fixed as [national level, provincial / ministerial level, municipal level, county / district level]. The corresponding hierarchical position is assigned a value of 1, and the other dimensions are assigned a value of 0, generating a 4-dimensional discrete feature vector of the posting level. Step M3: Abnormal fallback mechanism: If no level of institution is matched (draft for comments or draft without a clear issuing unit), set a fallback vector [0,0,0,0] and mark the missing level label to facilitate the subsequent model to identify missing samples; Step M4: Discrete Vector Encoding Rules: Construct a one-dimensional discrete feature vector of length 12. The order of the vector dimensions corresponds to the 12 text styles mentioned above. Only the dimension corresponding to the main text style is assigned a value of 1, and all others are assigned a value of 0, thus generating a 12-dimensional text style format feature vector. Step M5: Policy tool classification and keyword rule base construction, dividing policy tools into three categories: supply-side, demand-side, and environment-side, with each category having a built-in subdivided keyword library; Step M6: Based on the time nodes and implementation progress information in the policy text, mark the implementation stage of the policy (planning, implementation, supervision and evaluation, summary and acceptance). Concatenate the above four types of structured features and generate a policy structure feature vector through linear mapping. ,That For structural feature dimension, .

[0040] See Figure 1 After obtaining the vector data, attention weight calculation, adaptive feature fusion, and multi-label classification are performed. The specific steps are as follows: Step N1: Attention weight calculation. First, project the features from the three dimensions onto the same dimension. ,in , , ,in , , The projection matrix is ​​learnable. For bias terms; after projection ; Step N2: Calculate the attention score for each projected feature. ,in For learnable parameters, The projected feature vector, i.e. T represents a summary of policy categories; Step N3: Obtain attention weights through Softmax normalization. , We obtain a fused feature vector f by weighting and fusing text semantic features, domain ontology features, and policy structure features. , ,in , , These are the learnable parameters for the fusion layer.

[0041] See Figure 1 Multi-label classification was performed, with a total of 12 labels, corresponding to the data in the categories of (strategic planning, industrial innovation, clinical research, data management, project management, technology transfer, talent management, base platform management, ethical supervision, integrity supervision, new technology supervision, and biosafety). The predicted label for the k-th dimension is... ,in is the classification threshold for the k-th dimension.

[0042] When using this invention, The above description only lists preferred embodiments of the present invention and is used solely for illustrating the technical mechanism of the present invention; it does not limit the scope of protection of the present invention. Modifications, equivalent substitutions, and optimizations performed by those skilled in the art without departing from the core principles and purpose of the present invention are all included within the scope of protection of the present invention.

Claims

1. An automatic classification method for health policies based on multi-dimensional feature fusion, characterized in that, Includes the following steps: Step 1: Receive the original text of the health policy and complete the text integrity verification, including unifying the encoding, removing invalid characters, and verifying the text length; Step 2: Perform preprocessing on the text, using domain-enhanced jieba segmentation and loading a health policy-specific dictionary for segmentation, removing general and domain-specific stop words, completing word form normalization and date standardization, and using dependency parsing to annotate the subject-verb-object structure of the sentences; Step 3: Input the preprocessed text into the BERT encoder to extract deep semantic vectors, and use the PCA algorithm with cosine similarity constraints to reduce the 768-dimensional original semantic vectors to 256-dimensional vectors; Step 4: Match text terms based on the 12 major categories of health science and technology topic dictionaries, calculate the TF-IDF weights of the terms and aggregate them by category to generate domain ontology feature vectors; The 12 areas are strategic planning, industrial innovation, clinical research, data management, project management, technology transfer, talent management, base and platform management, ethical regulation, integrity regulation, new technology regulation, and biosafety. Step 5: Extract policy structured metadata, encode the document level into a one-hot vector and the document style into a discrete vector, calculate the weight of policy tools, mark the policy implementation stage, and concatenate and map to obtain the policy structure feature vector; Step 6: Project the three types of features—text semantics, domain ontology, and policy structure—to a unified space using a self-attention mechanism, and calculate and normalize the attention weights of each dimension of features; Step 7: Adaptively weighted fuse the three types of features based on attention weights to generate a fused feature vector; 8: Input the fused features into a multi-label classifier containing 12 parallel sub-networks, and output the predicted probabilities of 12 domain labels through the Sigmoid function; Step 9: Set probability thresholds for each dimension and determine the output multi-dimensional classification labels for health policies.

2. The automatic classification method for health policies based on multi-dimensional feature fusion according to claim 1, characterized in that, In step 2, the domain-enhanced jieba word segmentation loads a custom dictionary specifically for health and wellness policies, which includes exclusive terms, institutional terms, and industry abbreviations in fields such as major science and technology projects, achievement transformation, medical research and enterprise collaboration, pharmaceutical price reform, and hierarchical diagnosis and treatment; word form normalization includes traditional to simplified conversion, full-width to half-width conversion, unification of variant words, and standardization of Chinese numbers and dates.

3. The automatic classification method for health policies based on multi-dimensional feature fusion according to claim 1, characterized in that, In step 3, the maximum length of the input text token of the BERT encoder is 512. For excessively long texts, redundant content at the end is truncated, and for insufficient texts, placeholders are added to the end. The [CLS] token is extracted and output as the full-text semantic vector. Before PCA dimensionality reduction, L2 normalization is performed on the original BERT vector, and cosine similarity loss is introduced to constrain the semantic proximity relationship of the text.

4. The automatic classification method for health policies based on multi-dimensional feature fusion according to claim 1, characterized in that, The domain topic dictionary in step 4 adopts a dual mode of Trie prefix tree in-memory retrieval + database persistent storage. Term matching follows the long term priority rule, and synonyms are automatically mapped to standard terms. IDF is calculated using a smooth logarithm formula to avoid the problem of the denominator being 0 for single-document terms.

5. The automatic classification method for health policies based on multi-dimensional feature fusion according to claim 1, characterized in that, The document issuance level in step 5 is divided into four categories: national level, provincial / ministerial level, municipal level, and county / district level. When there is no clearly defined issuing agency, a zero-based catch-all vector is filled in. Policy tools are divided into supply-oriented, demand-oriented, and environment-oriented types. The implementation stages are divided into four categories: planning and formulation, organization and implementation, supervision and evaluation, and summary and acceptance.

6. The automatic classification method for health policies based on multi-dimensional feature fusion according to claim 1, characterized in that, In step 6, the three types of features are mapped to the same dimension through a learnable linear projection matrix, and the attention weights of each dimension are obtained by normalization using the Softmax function. The weights are used for feature weighted fusion calculation in step 7.

7. The automatic classification method for health policies based on multi-dimensional feature fusion according to claim 1, characterized in that, In step 8, each classification sub-network independently corresponds to one health science and technology policy domain, and outputs the single domain label probability in parallel; in step 9, when the predicted probability of a certain domain is greater than a preset threshold, it is determined that the policy contains the corresponding domain label, and multiple labels can be output for a single policy at the same time.

8. An automatic classification system for health policies based on multi-dimensional feature fusion, characterized in that, The automatic classification method for executing any one of claims 1-7 includes: a text acquisition module, used to receive batches or single original texts of health policies and to complete encoding, dirty data, and text length integrity verification; The text preprocessing module incorporates a health-related word segmentation dictionary, a two-layer stop word library, a normalization rule base, and syntactic analysis tools to output standardized text. The three-dimensional feature parallel extraction module includes semantic feature extraction, domain ontology feature extraction, and policy structure feature extraction sub-modules. The semantic feature extraction sub-module is equipped with a BERT encoder and a PCA dimensionality reduction unit with cosine constraints. The domain ontology feature extraction sub-module stores 12 major categories of health-related topic dictionaries and a TF-IDF calculation unit. The policy structure feature extraction sub-module incorporates coding rules for document level, style, policy tools, and implementation stage. The attention weight calculation module projects the three types of features to a unified dimension and calculates normalized attention weights. The adaptive feature fusion module generates fused feature vectors based on attention weight weights. The multi-label classification output module incorporates 12 sets of parallel classification sub-networks and a threshold determination unit to output multi-dimensional classification labels for 12 domains.

9. The automatic classification system for health policies based on multi-dimensional feature fusion according to claim 8, characterized in that, The semantic feature extraction submodule also includes a vector storage unit for storing the 256-dimensional policy semantic vector after dimensionality reduction, supporting vector retrieval, clustering, and similarity comparison.

10. The automatic classification system for health policies based on multi-dimensional feature fusion according to claim 8, characterized in that, The domain ontology feature extraction submodule is equipped with a Trie prefix tree high-speed retrieval unit to achieve rapid matching of multi-granular terms in text and regularly update the IDF value of the global corpus offline.

Images