A feature extraction method, device, equipment, medium and program product

Abstract

This application provides a feature extraction method, apparatus, device, medium, and program product, relating to the field of artificial intelligence technology, for improving the accuracy of extracting features reflecting regional development priorities. The specific technical solution is as follows: acquiring management regulation information of a target city; performing information extraction processing on the management regulation information to obtain a semantic feature description text set and a keyword set corresponding to the management regulation information; determining a first feature set corresponding to the target city based on the semantic feature description text set and the keyword set; the first feature set includes at least one feature vector; based on the correlation between the feature vectors in the first feature set and the target region, determining a second feature set from the first feature set whose correlation with the target region is greater than or equal to a first threshold; the second feature set includes at least one feature vector. This application is applicable to scenarios such as urban management and smart city assessment.

Description

Technical Field

[0001] This application relates to the field of smart city technology, and in particular to a feature extraction method, apparatus, device, medium, and program product. Background Technology

[0002] In fields such as smart city assessment and regional development analysis, accurately extracting features that reflect the key points of regional development from urban management regulations is a crucial prerequisite for conducting assessments.

[0003] Currently, text mining methods such as Term Frequency-Inverse Document Frequency (TF-IDF) statistics and Latent Dirichlet Allocation (LDA) topic models can be used to extract features from urban management regulations. Specifically, TF-IDF statistics measure the importance of a word in a document and are commonly used in text mining, keyword extraction, text similarity, and search engines. LDA topic models, which employ Latent Dirichlet Distribution, are unsupervised text topic mining algorithms used to automatically identify hidden topics.

[0004] However, while the aforementioned text mining methods can extract features reflecting regional development priorities to some extent—for example, keyword or topic distribution—they primarily rely on surface-level vocabulary statistics, resulting in relatively poor accuracy in extracting features reflecting regional development priorities. Summary of the Invention

[0005] This application provides a feature extraction method, apparatus, device, medium, and program product to improve the accuracy of extracting features that reflect regional development priorities.

[0006] In a first aspect, embodiments of this application provide a feature extraction method, which includes: acquiring management regulation information of a target city; determining the target city based on at least one of the following: city size, richness of management regulation information, and city development level; performing information extraction processing on the management regulation information to obtain a semantic feature description text set and a keyword set corresponding to the management regulation information; the semantic feature description text set includes M semantic feature description texts; the keyword set includes N keywords and the word frequency statistics corresponding to each keyword; M and N are both positive integers; determining a first feature set corresponding to the target city based on the semantic feature description text set and the keyword set; the first feature set includes at least one feature vector; determining a second feature set from the first feature set whose correlation with the target region is greater than or equal to a first threshold based on the correlation degree between the feature vectors in the first feature set and the target region; the second feature set includes at least one feature vector.

[0007] The technical solution provided in this application brings at least the following beneficial effects: It can first obtain the semantic feature description text set and keyword set corresponding to the city's management standard information, and then determine the feature set strongly related to the target area based on the semantic feature description text set and keyword set, thereby enabling a deep understanding of the semantics and context of the management standard information, thus improving the accuracy of extracting features that reflect the key points of regional development.

[0008] One possible implementation involves the above-mentioned information extraction processing of management standard information to obtain a set of semantic feature description texts and a set of keywords corresponding to the management standard information. This includes: processing the management standard information using a large language model to obtain M semantic feature description texts, and determining the M semantic feature description texts as a set of semantic feature description texts; performing word segmentation and stop word removal on the management standard information to obtain at least one candidate keyword; obtaining the word frequency statistics corresponding to the at least one candidate keyword; determining the candidate keywords whose word frequency statistics are greater than or equal to a second threshold as N keywords, and determining the N keywords and their corresponding word frequency statistics as a set of keywords.

[0009] Another possible implementation, as described above, involves determining the first feature set corresponding to the target city based on the semantic feature description text set and the keyword set. This includes: determining the similarity between each semantic feature description text in the semantic feature description text set and each keyword in the keyword set using a word vector similarity algorithm or a large language model; determining Q feature vectors based on the similarity; each of the Q feature vectors includes: a semantic feature description text, P keywords, and the word frequency statistics of each keyword in the P keywords; Q is a positive integer, and P is a positive integer less than N; and determining the above Q feature vectors as the first feature set.

[0010] Another possible implementation involves determining a second feature set from the first feature set whose correlation with the target region is greater than or equal to a first threshold, based on the correlation between the feature vectors in the first feature set and the target region. This includes: using a clustering algorithm to determine a third feature set from the first feature set whose correlation with the target region is greater than or equal to the first threshold, based on the correlation between the feature vectors in the first feature set and the target region; the third feature set includes at least one feature vector cluster; and determining a feature vector cluster with a purity greater than or equal to the third threshold from the at least one feature vector cluster, and defining the determined feature vector cluster as the second feature set.

[0011] Another possible implementation is to determine a second feature set from the first feature set whose correlation with the target region is greater than or equal to a first threshold, based on the correlation between the feature vectors in the first feature set and the target region. This includes: when the matching degree between the predicted regional label of the first feature set and the real label of the target city is greater than or equal to a fourth threshold, determining a second feature set from the first feature set whose correlation with the target region is greater than or equal to a first threshold, based on the correlation between the feature vectors in the first feature set and the target region.

[0012] Secondly, embodiments of this application provide a feature extraction device, comprising: an acquisition module, a processing module, and a determination module; the acquisition module is used to acquire management regulation information of a target city; the target city is determined based on at least one of the following: city size, richness of management regulation information, and city development level; the processing module is used to perform information extraction processing on the management regulation information to obtain a semantic feature description text set and a keyword set corresponding to the management regulation information; the semantic feature description text set includes M semantic feature description texts; the keyword set includes N keywords and the word frequency statistics corresponding to each keyword; M and N are both positive integers; the determination module is used to determine a first feature set corresponding to the target city based on the semantic feature description text set and the keyword set; the first feature set includes at least one feature vector; and based on the degree of correlation between the feature vectors in the first feature set and the target area, a second feature set with a degree of correlation with the target area greater than or equal to a first threshold is determined from the first feature set; the second feature set includes at least one feature vector.

[0013] One possible implementation is that the aforementioned processing module is specifically used to process the management specification information through a large language model to obtain M semantic feature description texts, and to determine the M semantic feature description texts as a set of semantic feature description texts; and to perform word segmentation and stop word removal on the management specification information to obtain at least one candidate keyword; and to obtain the word frequency statistics value corresponding to the at least one candidate keyword; and to determine the candidate keywords whose word frequency statistics value is greater than or equal to a second threshold as N keywords, and to determine the N keywords and the word frequency statistics value corresponding to the N keywords as a set of keywords.

[0014] Another possible implementation is that the aforementioned determining module is specifically used to determine the similarity between each semantic feature description text in the semantic feature description text set and each keyword in the keyword set through a word vector similarity algorithm or a large language model; and based on the similarity, determine Q feature vectors; each of the Q feature vectors includes: a semantic feature description text, P keywords, and the word frequency statistics of each keyword in the P keywords; Q is a positive integer, and P is a positive integer less than N; the Q feature vectors are determined as the first feature set.

[0015] Another possible implementation is that the aforementioned determining module is specifically used to determine, based on the correlation between the feature vectors in the first feature set and the target region, a clustering algorithm is used to determine a third feature set from the first feature set whose correlation with the target region is greater than or equal to a first threshold; the third feature set includes at least one feature vector cluster; from the at least one feature vector cluster, a feature vector cluster with a purity greater than or equal to the third threshold is determined, and the determined feature vector cluster is determined as the second feature set.

[0016] Another possible implementation is that the aforementioned determining module is specifically used to determine a second feature set from the first feature set whose correlation with the target region is greater than or equal to the first threshold, based on the correlation between the feature vectors in the first feature set and the target region, when the matching degree between the predicted regional label of the first feature set and the real label of the target city is greater than or equal to the fourth threshold.

[0017] Thirdly, this application provides an electronic device comprising: a processor and a memory; the memory stores a program or instructions executable on the processor, wherein the program or instructions, when executed by the processor, implement the method of the first aspect described above.

[0018] Fourthly, this application provides a readable storage medium on which a program or instructions are stored, which, when executed by a computer, implement the method of the first aspect described above.

[0019] Fifthly, this application provides a computer program product stored in a storage medium, which, when executed by a computer, implements the method described in the first aspect.

[0020] In a sixth aspect, embodiments of this application provide a chip including a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the method described in the first aspect.

[0021] The beneficial effects of the second to sixth aspects mentioned above are described in the corresponding description of the first aspect and will not be repeated here. Attached Figure Description

[0022] Figure 1 A schematic diagram of a network architecture for the application of a feature extraction method provided in this application embodiment;

[0023] Figure 2 A flowchart illustrating a feature extraction method provided in an embodiment of this application;

[0024] Figure 3 A flowchart illustrating another feature extraction method provided in an embodiment of this application;

[0025] Figure 4 A flowchart illustrating another feature extraction method provided in an embodiment of this application;

[0026] Figure 5 A flowchart illustrating another feature extraction method provided in an embodiment of this application;

[0027] Figure 6 A flowchart illustrating another feature extraction method provided in this application embodiment;

[0028] Figure 7 This is a schematic diagram of the structure of a feature extraction device provided in an embodiment of this application;

[0029] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0031] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0032] The terms "at least one," "at least one of," etc., used in the specification and claims of this application refer to any one, any two, or a combination of two or more of the included items. For example, at least one of a, b, and c can mean: "a," "b," "c," "a and b," "a and c," "b and c," and "a, b, and c," where a, b, and c can be single or multiple. Similarly, "at least two" refers to two or more items, and its meaning is similar to that of "at least one."

[0033] In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0034] The feature extraction method, apparatus, device, medium, and program product provided in this application embodiment can be applied to scenarios such as urban management and smart city assessment.

[0035] In fields such as smart city assessment and regional development analysis, accurately extracting features that reflect the key points of regional development from urban management regulations is a crucial prerequisite for conducting assessments.

[0036] Currently, text mining methods such as TF-IDF term frequency analysis and LDA topic modeling can be used to extract features from urban management regulations. Specifically, TF-IDF term frequency analysis measures the importance of a word in a document and is commonly used in text mining, keyword extraction, text similarity analysis, and search engines. LDA topic modeling, which uses the Latent Dirichlet Distribution, is an unsupervised text topic mining algorithm used to automatically identify hidden topics.

[0037] There are three main types of existing feature extraction methods: methods based on statistical indicators, such as directly using "current city's per capita GDP > national per capita GDP" as the feature of "economically developed"; methods based on traditional text mining, such as TF-IDF word frequency statistics and LDA topic model, which can extract vocabulary or topic distribution from management and regulatory information; and semantic methods based on deep learning, such as using models like Bidirectional Encoder Representations from Transformers (BERT) to extract document vectors, which can capture deep semantics.

[0038] However, while the aforementioned statistical indicator-based methods are direct, they are too simplistic and fail to capture rich semantic information from management regulations, such as the government's response strategies and development priorities. The traditional text mining methods suffer from two major problems: first, the extracted "features" have weak semantic connections with the specific management regulations, resulting in poor interpretability; second, they cannot automatically and quantitatively assess the relative importance of each feature to the region. Although the deep learning-based semantic methods can capture deep semantics, they typically lack optimization for "regional comparisons," and the generated abstract vectors are difficult to directly map to understandable and quantifiable features in specific management regulations.

[0039] Therefore, a method for "analyzing management regulations information by combining keywords and semantic vectors" has emerged in the existing technology. This method first selects keywords through expert experience or simple word frequency, and then uses a word vector model to calculate the semantic similarity of the keywords for classification.

[0040] However, the initial screening of keywords relies on subjective experience or simple statistics, which may miss important semantic concepts; and it has failed to build an end-to-end automated and quantitative analysis framework from text to features, and then to the importance weight of features; thus, the accuracy of extracting features that reflect the key points of regional development is poor.

[0041] Furthermore, the aforementioned methods suffer from several drawbacks: region segmentation is subjective or simplistic; many methods fail to use objective and unified quantitative standards for region pre-classification, leading to a chaotic baseline for subsequent feature extraction; feature extraction and quantification are disconnected; traditional methods can only extract feature lists or generate abstract vectors, making it difficult to simultaneously output quantitative results indicating "what feature it is" and "how important the feature is"; interpretability and automation are contradictory; highly automated methods (such as deep learning) have poor interpretability; and highly interpretable methods (such as human interpretation) have low automation and poor efficiency.

[0042] To address the aforementioned technical problems, embodiments of this application provide a feature extraction method, apparatus, device, medium, and program product. This method first acquires a set of semantic feature description texts and a set of keywords corresponding to the city's management regulations information. Then, based on these semantic feature description texts and keyword sets, it determines a set of features strongly correlated with the target area. This allows for a deeper understanding of the semantics and context of the management regulations information, thereby improving the accuracy of extracting features reflecting the region's development priorities.

[0043] The feature extraction methods, apparatus, devices, media, and program products provided in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0044] Figure 1 This illustration shows a network architecture for an application of a feature extraction method provided in an embodiment of this application. For example... Figure 1As shown, the network architecture includes a feature extraction device 101 and a terminal device 102. The feature extraction device 101 and the terminal device 102 are interconnected.

[0045] In some embodiments, the feature extraction device 101 may be a server, a computer, or a processor or processing unit within a server or computer. The server may be a single server or a server cluster consisting of multiple servers. It should be noted that the embodiments of this application do not limit the specific device form of the feature extraction device 101. Figure 1 The feature extraction device 101 is used as an example of a single server.

[0046] In some embodiments, the terminal device may be a mobile phone, tablet computer, laptop computer, handheld computer, in-vehicle electronic device, mobile internet device (MID), augmented reality (AR) / virtual reality (VR) device, robot, wearable device, personal computer (PC), ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc., and the embodiments of this application do not specifically limit it. Figure 1 The example shown is a mobile phone, with terminal device 102 as an example.

[0047] It should be noted that the network architecture described in the embodiments of this application is for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and does not constitute a limitation on the technical solutions provided in the embodiments of this application. As network architectures evolve, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0048] See Figure 2 This is a flowchart illustrating a feature extraction method provided in an embodiment of this application. Figure 2 As shown, the feature extraction method provided in this application embodiment can be implemented by the above-mentioned feature extraction device, specifically including the following steps 201 to 204.

[0049] Step 201: The feature extraction device acquires the management regulations information of the target city.

[0050] The target cities mentioned above are determined based on at least one of the following: city size, richness of management and regulatory information, and level of urban development.

[0051] In some embodiments, the target city can be a stratified sampling city determined by its size, the richness of its management regulations, and its level of development, ensuring that representative cities are sampled at each stratum.

[0052] In some embodiments, the aforementioned management information is used to reflect the characteristics of regional development priorities, including: strategies for addressing these characteristics, development priorities, and other information.

[0053] In some embodiments, city size can be determined by indicators such as total GDP, number of permanent residents, and city area.

[0054] In some embodiments, the richness of management regulations information can be calculated based on factors such as the length, update frequency, and keyword diversity of management regulations information documents for each city.

[0055] In some embodiments, the level of urban development may include indicators such as GDP per capita and fiscal revenue.

[0056] For example, the target city can sort and determine the cities in each region based on the city size, the richness of management and regulatory information, and the level of urban development. The number of sampling layers (usually 3-5 layers) is dynamically determined according to the total number of cities in the region and the preset sampling ratio (e.g., 30%). The cities in each layer have similar comprehensive characteristics, ensuring that representative cities are sampled in each layer.

[0057] In some embodiments, the regions to be evaluated are pre-labeled based on objective standards, providing a clear and consistent basis for constructing positive and negative samples for subsequent comparative learning. Each city is labeled as a "population inflow area" or a "population outflow area" according to predefined rules (e.g., resident population > registered population). The regional definition here can be extended to other dimensions, such as "economically developed area" or "severely aging area," as long as objective statistical standards are met. The input to this module is the resident and registered population data of the target city set; the output is a list of cities with clear regional labels (e.g., the original city list is...). The list of cities in Region A is as follows: The list of cities in Region_B is as follows This label is the fundamental basis for constructing contrastive learning sample pairs.

[0058] In some embodiments, in addition to population flow data, the above-mentioned regional labeling may use other single or composite indicators such as "economically developed / underdeveloped", "main industries are primary / secondary / tertiary industries", and "multi-ethnic areas" as objective standards for regional classification.

[0059] In some embodiments, this application proposes a multi-dimensional stratified sampling method. First, sampling reference dimensions are selected, including: city size dimension, using indicators such as total GDP, resident population, and city area, normalized to obtain a size score; management regulation information text richness dimension, calculating the management regulation information text richness score based on the length, update frequency, and keyword diversity of each city's management regulation information documents; and development level dimension, considering indicators such as per capita GDP and fiscal revenue. Then, stratification is performed, including: ranking cities in each region based on the above dimensions; dynamically determining the number of sampling layers (usually 3-5 layers) based on the total number of cities in the region and a preset sampling ratio (e.g., 30%); ensuring that cities within each layer have similar comprehensive characteristics, guaranteeing that representative cities are sampled in each layer. Finally, stratified sampling is performed, proportionally selecting cities from each layer, with a slightly higher sampling ratio for higher-level cities (cities with higher comprehensive scores), as their contribution to regional characteristics is usually greater; simultaneously, a balance in development levels is considered to ensure that the sampled cities have representative samples at different development stages.

[0060] In some embodiments, the above-mentioned multi-dimensional hierarchical city sampling strategy can be to determine multiple sampling reference dimensions (such as city size, richness of management norms information text, and development level); to conduct a comprehensive evaluation and hierarchical analysis of cities within the region based on the above dimensions; and to extract city samples proportionally from each level, with a higher sampling proportion for higher-level cities.

[0061] In some embodiments, this application generates regional labels based on regional pre-labeling of objective statistical indicators; constructs positive and negative sample pairs for comparative learning based on the aforementioned regional labels; and trains a comparative learning model using the aforementioned positive and negative sample pairs to extract management regulation information features with regional discriminative characteristics.

[0062] Thus, by using objective regional classification (non-textual features) as a driving signal to guide the comparative learning of textual features, the extracted features can explicitly reflect the differences between regions, solving the problems of existing methods lacking "regional comparison" optimization and weak feature discriminativeness. By adopting a multi-dimensional hierarchical sampling method, the problems of high computational cost of full analysis and insufficient representativeness of random sampling are solved, ensuring that the sample set used for feature extraction is representative in terms of scale, management standard information activity, and development stage, thereby improving the comprehensiveness and reliability of subsequent feature extraction.

[0063] Step 202: The feature processing device performs information extraction processing on the management specification information to obtain a set of semantic feature description texts and a set of keywords corresponding to the management specification information.

[0064] The semantic feature description text set includes M semantic feature description texts; the keyword set includes N keywords and the word frequency statistics corresponding to each keyword; M and N are both positive integers.

[0065] In some embodiments, the semantic feature description text set described above can be a set of M structured, natural language-described semantic features.

[0066] In some embodiments, M and N may be the same or different.

[0067] In some embodiments, a semantic feature description text is used to describe a semantic feature corresponding to the aforementioned management specification information.

[0068] In some embodiments, the above N keywords may include at least one of the following: keywords that appear in the above management specification information with a frequency greater than or equal to a frequency threshold, keywords that are more important than or equal to a degree threshold in the above management specification information, and keywords that have practical significance in the above management specification information.

[0069] In some embodiments, the word frequency statistics value corresponding to a keyword is used to characterize the number of times the keyword appears in the aforementioned management specification information.

[0070] In some embodiments, combined with Figure 2 ,like Figure 3 As shown, step 202 above can be implemented through steps 202a and 202d.

[0071] Step 202a: The feature extraction device processes the management specification information through a large language model to obtain M semantic feature description texts, and determines the M semantic feature description texts as a set of semantic feature description texts.

[0072] In some embodiments, the semantic understanding channel described above can use different LLMs (such as the Generative Pre-trained Transformer (GPT) series, Claude, and domestic large models) or fine-tune domain-specific models to complete feature induction.

[0073] Step 202b: The feature extraction device performs word segmentation and stop word removal on the management specification information to obtain at least one candidate keyword.

[0074] In some embodiments, the above-described word segmentation and stop word removal processes may include word segmentation and stop word removal processes.

[0075] In some embodiments, the above word segmentation process can be to break down a continuous text into meaningful words according to semantics.

[0076] In some embodiments, the above-described stop word removal process is used to remove stop words from the text; the stop words can be words in the text that have no actual meaning and only serve a grammatical function.

[0077] In some embodiments, the candidate keywords may be high-frequency keywords or meaningful management and regulatory information words.

[0078] Step 202c: The feature extraction device obtains the word frequency statistics corresponding to at least one candidate keyword.

[0079] In some embodiments, the above-mentioned word frequency statistics can be the number of candidate keywords appearing in the management specification information.

[0080] Step 202d: The feature extraction device determines N keywords as candidates whose word frequency statistics are greater than or equal to the second threshold among at least one candidate keyword, and determines the N keywords and the word frequency statistics corresponding to the N keywords as a keyword set.

[0081] In some embodiments, the second threshold may be a system default or may be arbitrarily set by the user according to actual usage needs; this application does not impose any limitations on this.

[0082] In some embodiments, the feature extraction device may use a dual-channel approach to extract features from management regulations information. For each city management regulations information document, the following two channels are processed in parallel:

[0083] Channel 1 (Semantic Understanding Channel) employs a Large Language Model (LLM). Through carefully designed prompts, it guides the model to directly summarize and infer several semantic feature descriptions from the document. For example, for information on elderly care management regulations, the LLM might output: "Feature 1: Emphasizing the construction of community-based home care models," "Feature 2: Focusing on introducing social capital to participate in the operation of elderly care institutions," etc. The input to this module is the preprocessed management regulation information document; the output is a set of p structured, natural language-described semantic features. .

[0084] Channel Two (Statistical Quantification Channel) analyzes the same batch of management standard information documents through statistical quantification. First, it performs word segmentation and removes stop words; then, it calculates the term frequency (TF) or TF-IDF value of all words; finally, it filters out high-frequency keywords or a list of meaningful management standard information terms based on thresholds. Its word frequency statistics are as follows .

[0085] In some embodiments, in addition to term frequency (TF) and TF-IDF, the above statistical quantification channels can use graph algorithms such as TextRank to extract keywords, or use indicators such as mutual information (MI) to filter feature words.

[0086] In some embodiments, the input to this module is the same batch of management specification information documents from the semantic understanding channel; the output is a list of keywords and their corresponding word frequency statistics.

[0087] In some embodiments, the above dual-channel mechanism can be a dual-channel feature fusion mechanism of "LLM semantic understanding + traditional statistical quantization". The first channel (LLM semantic understanding) extracts structured semantic feature descriptions from the management specification information text; the second channel (word frequency statistics) extracts keywords and their frequencies from the same management specification information text; the semantic relevance between the above semantic feature descriptions and the above keywords is calculated, and the keywords and their frequencies are associated and bound to the corresponding semantic features; the initial quantization weight of the above semantic features is calculated based on the frequency of the bound keywords.

[0088] Thus, since the aforementioned semantic feature description text set can achieve a deep understanding of semantics, and the aforementioned keyword set can filter out high-frequency keywords or a list of meaningful management standard information terms based on thresholds, processing management standard information through these two methods not only allows us to understand the semantics of management standard information, but also allows us to obtain important information from management standard information.

[0089] Step 203: The feature extraction device determines the first feature set corresponding to the target city based on the semantic feature description text set and the keyword set.

[0090] The first feature set mentioned above includes at least one feature vector.

[0091] In some embodiments, combined with Figure 2 ,like Figure 4 As shown, step 203 above can be implemented through steps 203a and 203c.

[0092] Step 203a: The feature extraction device determines the similarity between each semantic feature description text in the semantic feature description text set and each keyword in the keyword set by using a word vector similarity algorithm or a large language model.

[0093] In some embodiments, the word vector similarity algorithm described above can determine the semantic similarity between two word vectors by calculating the distance between them in space.

[0094] In some embodiments, feature-term association analysis is first performed: the semantic relevance of each semantic feature description text to all keywords and the word frequency corresponding to each keyword is calculated.

[0095] In some embodiments, the feature extraction device can be implemented through word vector similarity or by using LLM to determine the matching degree. (Correlation coefficient) The calculation formula is shown in formula (1) below:

[0096] ; (1)

[0097] Among them, doc2vec(∙) is the text vectorization function.

[0098] Step 203b: The feature extraction device determines Q feature vectors based on similarity.

[0099] Each of the Q feature vectors mentioned above includes: a semantic feature description text, P keywords, and the word frequency statistics of each of the P keywords.

[0100] Where Q is a positive integer and P is a positive integer less than N.

[0101] In some embodiments, for any Choose the one with the highest relevance. Bind and mark as The scene feature text and keyword pairs are as follows:

[0102]

[0103] in Other keywords can still be bound, forming the following binding relationships.

[0104]

[0105] The binding relationships of all semantic features are as follows:

[0106]

[0107]

[0108] Then, the initial weights of the features are assigned: for each semantic feature, the word frequency statistics of all s keywords bound to it are weighted, summed, and normalized to serve as the initial weight of the semantic feature. This weight reflects the explicit frequency of the feature in the management specification information document. The weight formula is shown in formula (2) below:

[0109] ; (2)

[0110] In some embodiments, in addition to word vector similarity, the above association analysis can use co-occurrence analysis or train a small classifier to determine the association between keywords and semantic features.

[0111] Step 203c: The feature extraction device determines the Q feature vectors as the first feature set.

[0112] In some embodiments, the first feature set mentioned above can be a structured city feature set, which organizes all the processing results of each city in a region to form the feature set of that city: This module takes as input a set of semantic feature description texts and a set of keywords; the output is a set of features for the city.

[0113] In this way, by binding each semantic feature description text in the semantic feature description text set with each keyword in the keyword set and the word frequency statistics of each keyword through similarity, the initial weight of the semantic feature is obtained, which reflects the explicit frequency of the feature in the management and regulatory information. This creatively combines the deep semantic induction ability of the large language model with the objective quantification ability of traditional word frequency statistics to generate a feature set that is both interpretable (semantic description) and quantifiable (weight), thus solving the problems of the separation between semantic and statistical information and the disconnect between feature extraction and quantification.

[0114] Step 204: Based on the correlation between the feature vectors in the first feature set and the target region, the feature extraction device determines a second feature set from the first feature set whose correlation with the target region is greater than or equal to a first threshold.

[0115] The second feature set mentioned above includes at least one feature vector.

[0116] In some embodiments, combined with Figure 2 ,like Figure 5 As shown, step 204 above can be implemented through steps 204a and 204b.

[0117] Step 204a: Based on the correlation between the feature vectors in the first feature set and the target region, the feature extraction device uses a clustering algorithm to determine a third feature set from the first feature set whose correlation with the target region is greater than or equal to a first threshold.

[0118] The third feature set includes at least one cluster of feature vectors.

[0119] In some embodiments, the above clustering algorithm can automatically divide unlabeled feature vectors into multiple groups based on similarity, making the features within a group more similar and the features between groups less similar.

[0120] In some embodiments, cluster analysis is used to automatically identify the feature combinations that best distinguish region labels. A subset of features strongly correlated with region labels is selected from all features, and an optimal feature space partition is learned that maximizes the distance between positive and negative sample labels in the feature space, thereby optimizing the feature partitioning method.

[0121] Based on the obtained city samples Tagging +1 represents region A, and -1 represents region B.

[0122] Each city There is a set of p semantic feature vectors Each vector has the same region label. .

[0123] Perform unsupervised clustering on all feature vector samples. Clustering all semantic feature vectors can be done using spectral clustering algorithms (alternatives include K-Means / DBSCAN), forming M feature vector clusters. M can be selected using methods such as profile coefficient and elbow rule.

[0124] Step 204b: The feature extraction device determines a feature vector cluster with a purity greater than or equal to a third threshold from at least one feature vector cluster, and defines the determined feature vector cluster as the second feature set.

[0125] In some embodiments, the third threshold may be 80% or 85%, etc., and is not limited in this application.

[0126] In some embodiments, the feature extraction device performs high-purity cluster screening on the clustering results. Each cluster is calculated. The proportion of regional labels, The percentage is , The percentage is , The consistency of labels within a cluster is filtered using region labels. If the labels of most (over 80%) feature vectors within a cluster are consistent, then... If the cluster is classified as "Class A high-purity cluster", then all feature vectors labeled Region_B in the cluster will be removed; similarly, if the majority of the labels are... If the labels are distributed evenly (e.g., between 40% and 60%), then the vectors from Region_A are removed. If the labels constitute the majority (60%–80%), then the vectors from that cluster with the largest proportion are retained. If the labels are evenly distributed (e.g., between 40% and 60%), then all vectors from that cluster are retained. This achieves sample vector purification.

[0127] In some embodiments, the feature extraction device performs discriminative feature training based on contrastive learning. Using purified samples, feature samples from the "Class A high-purity cluster" that are 0.5 standard deviations away from the mean center are selected. Marked as an anchor point. Other samples labeled in the same area as the anchor point sample. This is considered a positive sample. Samples with labels in regions other than the anchor point samples are also considered positive samples. These are used as negative samples. A contrastive learning loss function (such as InfoNCE Loss) is used for training, with the goal of: narrowing the distance between the anchor point and positive samples in the feature space; and widening the distance between the anchor point and negative samples. The loss function is as follows:

[0128]

[0129] In some embodiments, the contrastive learning framework may employ other contrastive loss functions such as Triplet Loss or SupCon Loss, or introduce a "hard sample mining" strategy during sample construction.

[0130] In some embodiments, the method for purifying contrastive learning samples based on cluster purity screening can be as follows: In contrastive learning training, unsupervised clustering is performed on the initial feature vectors; the proportion of different region labels in each cluster is calculated; based on a preset purity threshold (e.g., 80%), high-purity clusters are selected, and feature vectors within the clusters that are inconsistent with the dominant label are removed, so as to obtain purified samples for contrastive loss calculation.

[0131] Thus, by introducing clustering and label purity-based screening steps before contrastive learning, we can automatically identify and strengthen feature patterns that are strongly correlated with region labels, filter noise, and improve the effectiveness of contrastive learning and the discriminativeness of the feature space.

[0132] The feature extraction method provided in this application can first obtain the semantic feature description text set and keyword set corresponding to the city's management norms information, and then determine the feature set strongly related to the target area based on the semantic feature description text set and keyword set. This allows for a deep understanding of the semantics and context of the management norms information, thereby improving the accuracy of extracting features that reflect the key points of regional development.

[0133] In some embodiments, combined with Figure 2 ,like Figure 6 As shown, after step 204 above, the feature extraction method provided in this application embodiment may further include the following step 205.

[0134] Step 205: If the matching degree between the predicted regional label of the first feature set and the real label of the target city is greater than or equal to the fourth threshold, the feature extraction device determines a second feature set from the first feature set that has a matching degree with the target region that is greater than or equal to the first threshold, based on the correlation degree between the feature vectors in the first feature set and the target region.

[0135] In some embodiments, the above-mentioned regional label prediction results can be results with clear regional labels such as "population inflow area", "population outflow area", "economically developed area", "severely aging area".

[0136] In some embodiments, the management regulations information text of the new city is obtained, and its feature vector set is obtained by a feature extraction method; the feature vector is input into a trained encoding network to obtain its regional label prediction result;

[0137] Feature vector set is only considered as a valid feature of the city under the label of that region if the predicted label matches the city's true label (if any), thus achieving feature extraction.

[0138] Thus, since the feature vector set is only considered as a valid feature of the city under the label of the region when the predicted label matches the city's true label (if any), feature extraction is achieved. Therefore, through the above discrimination process, valid features can be automatically identified and feature extraction can be achieved.

[0139] The feature extraction method of this application will be described below through specific embodiments.

[0140] In some embodiments, the implementation process of the feature extraction method provided in this application includes the following S1 to S6:

[0141] S1: Pre-labeling of the regions to be evaluated based on objective standards provides a clear and consistent basis for constructing positive and negative samples for subsequent comparative learning. Each city is labeled as a "population inflow area" or "population outflow area" according to predefined rules (e.g., resident population > registered population). The regional definition here can be extended to other dimensions, such as "economically developed area" or "severely aging area," as long as objective statistical standards are met. The input to this module is the resident and registered population data of the target city set; the output is a list of cities with clear regional labels (e.g., the original city list is...). The list of cities in Region_A is The list of cities in Region_B is as follows This label is the fundamental basis for constructing contrastive learning sample pairs.

[0142] S2: Multi-dimensional hierarchical urban sampling

[0143] To address the issues of high computational costs in full-sample analysis and insufficient representativeness in simple random sampling, this invention proposes a multi-dimensional stratified sampling method. First, sampling reference dimensions are selected, including: city size dimension, using indicators such as total GDP, resident population, and city area, normalized to obtain a size score; management regulation information text richness dimension, calculating the management regulation information text richness score based on the length, update frequency, and keyword diversity of each city's management regulation information documents; and development level dimension, considering indicators such as per capita GDP and fiscal revenue. Then, stratification is performed, including: ranking cities in each region based on the above dimensions; dynamically determining the number of sampling strata (usually 3-5 strata) based on the total number of cities in the region and a preset sampling ratio (e.g., 30%); ensuring that cities within each stratum have similar comprehensive characteristics, guaranteeing that representative cities are sampled in each stratum. Finally, stratified sampling is performed, proportionally selecting cities from each stratum, with a slightly higher proportion selected from higher-level strata (cities with higher comprehensive scores), as they typically contribute more to regional characteristics; simultaneously, a balance in development levels is considered to ensure that the sampled cities have representative samples at different stages of development.

[0144] S3: Dual-channel management specification information text feature extraction. For each city management specification information document, the following two channels are processed in parallel:

[0145] Channel 1 (Semantic Understanding Channel) employs a Large Language Model (LLM). Through carefully designed prompts, it guides the model to directly summarize and infer several semantic feature descriptions from the document. For example, for information on elderly care management regulations, LLM might output: "Feature 1: Emphasizing the construction of community-based home care models," "Feature 2: Focusing on introducing social capital to participate in the operation of elderly care institutions," etc. The input to this module is the preprocessed management regulation information document; the output is a set of p structured, natural language-described semantic features. .

[0146] Channel Two (Statistical Quantification Channel): Analyze the same batch of management standard information documents. First, segment the data into words and remove stop words; then, calculate the term frequency (TF) or TF-IDF value of all words; finally, filter out high-frequency keywords or a list of meaningful management standard information terms based on thresholds. Its word frequency statistics are as follows The module takes as input the same batch of management specification information documents from Channel 1; the output is a list of keywords and their corresponding word frequency statistics.

[0147] S4: Feature Fusion, Association Analysis, and Preliminary Quantification. First, feature-term association analysis is performed: calculating the semantic relevance of each semantic feature description text (from Channel 1) to all keywords (from Channel 2). This can be achieved through word vector similarity or by using LLM to determine the matching degree. Relevance coefficient. The calculation formula is shown in formula (1) above, and will not be repeated in this application.

[0148] For any Choose the one with the highest relevance. Bind and mark as The scene feature text and keyword pairs are as follows:

[0149]

[0150] in Other keywords can still be bound, forming the following binding relationships.

[0151]

[0152] The binding relationships of all semantic features are as follows:

[0153]

[0154]

[0155] Then, the initial weights of the features are assigned: for each semantic feature, the word frequency statistics of all s keywords bound to it are weighted, summed, and normalized to serve as the initial weight of the semantic feature. This weight reflects the explicit frequency of the feature in the management and regulatory information text. The weight formula is shown in formula (2) above, and will not be repeated in this application.

[0156] Finally, a structured city feature set is formed: all the processing results for each city in the region are organized to form the feature set for that city. The module takes the results from Channel 1 and Channel 2 as input and outputs the feature set of the city.

[0157] S5: Discriminative Feature Extraction Based on Cluster-Enhanced Contrastive Learning. Cluster analysis is used to automatically identify the feature combinations that best distinguish region labels. From all features, a subset of features strongly correlated with region labels is selected, and an optimal feature space partitioning is learned to maximize the distance between positive and negative sample labels in the feature space, thereby optimizing the feature partitioning method.

[0158] City samples obtained based on S1 Enter for labeling +1 represents region A, and -1 represents region B.

[0159] Each city C_k has a set of p semantic feature vectors. Each vector has the same region label. .

[0160] Perform unsupervised clustering on all feature vector samples. Clustering all semantic feature vectors can be done using spectral clustering algorithms (alternatives include K-Means / DBSCAN), forming M feature vector clusters. M can be selected using methods such as profile coefficient and elbow rule.

[0161] Perform high-purity cluster screening on the clustering results. Calculate the purity of each cluster. The proportion of regional labels, The percentage is , The percentage is , The consistency of labels within a cluster is screened using region labels. If the label of most (over 80%) feature vectors within a cluster is y_k=1, then the cluster is considered a "class A high-purity cluster," and all feature vectors labeled Region_B are removed. Similarly, if the majority of labels are... If the labels are distributed evenly (e.g., between 40% and 60%), then the vectors from Region_A are removed. If the labels constitute the majority (60%–80%), then the vectors from that cluster with the largest proportion are retained. If the labels are evenly distributed (e.g., between 40% and 60%), then all vectors from that cluster are retained. This achieves sample vector purification.

[0162] Discriminative feature training based on contrastive learning is performed. Using purified samples, feature samples F_An from the "high-purity cluster A" that are 0.5 standard deviations away from the mean center are labeled as anchor points. Other samples F_A with the same label region as the anchor point are designated as positive samples. Samples F_B with a different label region than the anchor point are designated as negative samples. A contrastive learning loss function (such as InfoNCE Loss) is used for training, with the goal of: narrowing the distance between the anchor point and positive samples in the feature space; and widening the distance between the anchor point and negative samples. The loss function is as follows:

[0163]

[0164] S6: New Feature Recognition Process

[0165] The management regulations information text of the new city is obtained and processed through S2~S4 to obtain its feature vector set; the feature vector is input into the trained encoding network to obtain its regional label prediction result;

[0166] Feature vector set is only considered as a valid feature of the city under the label of that region if the predicted label matches the city's true label (if any), thus achieving feature extraction.

[0167] In this way, we can first obtain the semantic feature description text set and keyword set corresponding to the city's management regulations information, and then determine the feature set that is strongly related to the target area based on the semantic feature description text set and keyword set. This allows for a deep understanding of the semantics and context of the management regulations information, thereby improving the accuracy of extracting features that reflect the key points of regional development.

[0168] Furthermore, the data foundation of this application is complete: the required demographic data and management regulations are readily available from the website; the technology stack is mature: the large language model API, word segmentation and word frequency statistics tools, and contrastive learning framework are all mature technical components in the current AI field.

[0169] Furthermore, this application clearly defines its application scenarios: in urban management, it is used to automatically and quantitatively monitor and compare the focus of management regulations information in different regions, assisting in the formulation and evaluation of management regulations information; in smart city platforms, it serves as a core analysis module, providing dynamic feature inputs for urban health checks and development assessments; in business consulting, it is used for regional market entry strategy analysis, identifying management regulations information opportunities in different regions; and in terms of modularity and scalability, the system's modules are loosely coupled, easy to replace or upgrade (e.g., by replacing with a more powerful LLM), and can be easily extended to the analysis of other management regulations information topics such as education, environmental protection, and industry.

[0170] Therefore, this method is planned to be applied to the smart society development level monitoring and evaluation platform.

[0171] It should be noted that the descriptions of each step S1 to S6 in this embodiment can be found in the descriptions in the above embodiments, and will not be repeated here.

[0172] It should be noted that the above-described method embodiments, or the various possible implementations of the method embodiments, can be executed individually, or, provided there is no conflict, they can be combined with each other. The specific implementation can be determined according to actual usage requirements, and this application embodiment does not impose any restrictions on this.

[0173] As can be seen, the above mainly describes the solutions provided by the embodiments of this application from a methodological perspective. To achieve the above functions, the embodiments of this application provide corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the modules and algorithm steps of the various examples described in the embodiments disclosed herein, the embodiments of this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0174] This application embodiment can divide the feature extraction device into functional modules according to the above method example. For example, each function can be divided into its own functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. Optionally, the module division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0175] In some embodiments, this application also provides a feature extraction apparatus. The feature extraction apparatus may include one or more functional modules for implementing the feature extraction method of the above method embodiments.

[0176] For example, Figure 7 This is a schematic diagram of a feature extraction device provided in an embodiment of this application. Figure 7 As shown, the feature extraction device 900 includes: an acquisition module 901, a processing module 902, and a determination module 903.

[0177] The acquisition module 901 is used to acquire management regulation information of the target city; the target city is determined based on at least one of the following: city size, richness of management regulation information, and city development level; the processing module 902 is used to extract and process the management regulation information to obtain a semantic feature description text set and a keyword set corresponding to the management regulation information; the semantic feature description text set includes M semantic feature description texts; the keyword set includes N keywords and the word frequency statistics corresponding to each keyword; M and N are both positive integers; the determination module 903 is used to determine a first feature set corresponding to the target city based on the semantic feature description text set and the keyword set; the first feature set includes at least one feature vector; and based on the correlation between the feature vectors in the first feature set and the target area, a second feature set with a correlation greater than or equal to a first threshold is determined from the first feature set; the second feature set includes at least one feature vector.

[0178] The feature extraction device provided in this application can first obtain the semantic feature description text set and keyword set corresponding to the management norms information of the city, and then determine the feature set strongly related to the target area based on the semantic feature description text set and keyword set. This allows for a deep understanding of the semantics and context of the management norms information, thereby improving the accuracy of extracting features that reflect the key development areas of the region.

[0179] In some embodiments, the processing module 902 is specifically used to process the management specification information through a large language model to obtain M semantic feature description texts, and to determine the M semantic feature description texts as a set of semantic feature description texts; and to perform word segmentation and stop word removal on the management specification information to obtain at least one candidate keyword; and to obtain the word frequency statistics value corresponding to the at least one candidate keyword; and to determine the candidate keywords whose word frequency statistics value is greater than or equal to a second threshold as N keywords, and to determine the N keywords and the word frequency statistics value corresponding to the N keywords as a set of keywords.

[0180] In other embodiments, the determining module 903 is specifically used to determine the similarity between each semantic feature description text in the semantic feature description text set and each keyword in the keyword set through a word vector similarity algorithm or a large language model; and based on the similarity, determine Q feature vectors; each of the Q feature vectors includes: a semantic feature description text, P keywords, and the word frequency statistics of each keyword in the P keywords; Q is a positive integer, P is a positive integer less than N; and determine the Q feature vectors as the first feature set.

[0181] In some other embodiments, the determination module 903 is specifically used to determine, based on the degree of association between the feature vectors in the first feature set and the target region, a clustering algorithm is used to determine a third feature set from the first feature set whose degree of association with the target region is greater than or equal to a first threshold; the third feature set includes at least one feature vector cluster; and from the at least one feature vector cluster, a feature vector cluster with a purity greater than or equal to the third threshold is determined, and the determined feature vector cluster is determined as the second feature set.

[0182] In some other embodiments, the determination module 903 is specifically used to determine a second feature set from the first feature set whose correlation with the target region is greater than or equal to the first threshold, based on the correlation between the feature vectors in the first feature set and the target region, when the matching degree between the predicted regional label of the first feature set and the real label of the target city is greater than or equal to the fourth threshold.

[0183] It should be noted that the feature extraction device can implement all the processes implemented in the above method embodiments and achieve the same beneficial effects. To avoid repetition, it will not be described again here.

[0184] In the case where the functions of the integrated modules described above are implemented in hardware, this application provides a possible structural schematic diagram of the electronic device involved in the above embodiments. For example... Figure 8 As shown, the electronic device 90 includes: a processor 92, a communication interface 93, and a bus 94. Optionally, the electronic device 90 may also include a memory 91.

[0185] Processor 92 may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 92 may be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 92 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0186] Communication interface 93 is used to connect with other devices via a communication network. This communication network can be Ethernet, wireless access network, wireless local area network (WLAN), etc.

[0187] The memory 91 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0188] As one possible implementation, the memory 91 can exist independently of the processor 92. The memory 91 can be connected to the processor 92 via a bus 94 and is used to store instructions or program code. When the processor 92 calls and executes the instructions or program code stored in the memory 91, it can implement the feature extraction method provided in the embodiments of this application.

[0189] In another possible implementation, memory 91 can also be integrated with processor 92.

[0190] Bus 94 can be an Extended Industry Standard Architecture (EISA) bus, etc. Bus 94 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 8 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0191] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the service calling device can be divided into different functional modules to complete all or part of the functions described above.

[0192] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above feature extraction method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0193] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0194] This application also provides a readable storage medium storing a program or instructions that, when executed by a computer, implement the feature extraction method provided in the above embodiments. It is understood that all or part of the processes in the above method embodiments can be executed by computer instructions instructing related hardware; the readable storage medium can be any of the foregoing embodiments or memory; the readable storage medium can also be an external storage device of the service invocation device, such as a plug-in hard drive, Smart MediaCard (SMC), Secure Digital (SD) card, flash card, etc., equipped on the service invocation device. Further, the readable storage medium can include both internal storage units of the service invocation device and external storage devices. The readable storage medium is used to store the computer program and other programs and data required by the service invocation device. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0195] This application also provides a computer program product, which is stored in a storage medium and implements the feature extraction method provided in the above embodiments when the computer program product is executed by a computer.

[0196] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0197] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0198] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A feature extraction method, characterized in that, include: Obtain management regulations information for the target city; the target city is determined based on at least one of the following: city size, richness of management regulations information, and city development level. The management specification information is processed by information extraction to obtain a set of semantic feature description texts and a set of keywords corresponding to the management specification information; the set of semantic feature description texts includes M semantic feature description texts; the set of keywords includes N keywords and the word frequency statistics value corresponding to each keyword; M and N are both positive integers; Based on the semantic feature description text set and the keyword set, a first feature set corresponding to the target city is determined; the first feature set includes at least one feature vector. Based on the correlation between the feature vectors in the first feature set and the target region, a second feature set is determined from the first feature set whose correlation with the target region is greater than or equal to a first threshold; the second feature set includes at least one feature vector.

2. The feature extraction method according to claim 1, characterized in that, The process of extracting information from the management specification information yields a set of semantic feature description texts and a set of keywords corresponding to the management specification information, including: The management specification information is processed by a large language model to obtain the M semantic feature description texts, and the M semantic feature description texts are determined as the semantic feature description text set. The management specification information is segmented and stop words are removed to obtain at least one candidate keyword; Obtain the word frequency statistics corresponding to the at least one candidate keyword; The candidate keywords whose word frequency statistics are greater than or equal to the second threshold among the at least one candidate keywords are determined as the N keywords, and the N keywords and the word frequency statistics corresponding to the N keywords are determined as the keyword set.

3. The feature extraction method according to claim 1, characterized in that, The step of determining the first feature set corresponding to the target city based on the semantic feature description text set and the keyword set includes: The similarity between each semantic feature description text in the semantic feature description text set and each keyword in the keyword set is determined by using a word vector similarity algorithm or a large language model. Based on the similarity, Q feature vectors are determined; each of the Q feature vectors includes: a semantic feature description text, P keywords, and the word frequency statistics of each of the P keywords; Q is a positive integer, and P is a positive integer less than N; The Q feature vectors are determined as the first feature set.

4. The feature extraction method according to claim 1, characterized in that, The step of determining a second feature set whose correlation with the target region is greater than or equal to a first threshold, based on the correlation between feature vectors in the first feature set and the target region, includes: Based on the correlation between the feature vectors in the first feature set and the target region, a clustering algorithm is used to determine a third feature set from the first feature set whose correlation with the target region is greater than or equal to the first threshold; the third feature set includes at least one feature vector cluster. From the at least one feature vector cluster, a feature vector cluster with a purity greater than or equal to a third threshold is determined, and the determined feature vector cluster is defined as the second feature set.

5. The feature extraction method according to any one of claims 1 to 4, characterized in that, The step of determining a second feature set whose correlation with the target region is greater than or equal to a first threshold, based on the correlation between feature vectors in the first feature set and the target region, includes: If the matching degree between the predicted regional label of the first feature set and the real label of the target city is greater than or equal to the fourth threshold, based on the correlation degree between the feature vectors in the first feature set and the target region, a second feature set with a correlation degree greater than or equal to the first threshold is determined from the first feature set.

6. A feature extraction apparatus, characterized in that, include: Acquisition module, processing module, and determination module; The acquisition module is used to acquire management regulation information of the target city; the target city is determined based on at least one of the following: city size, richness of management regulation information, and city development level; The processing module is used to perform information extraction processing on the management specification information to obtain a semantic feature description text set and a keyword set corresponding to the management specification information; the semantic feature description text set includes M semantic feature description texts; the keyword set includes N keywords and the word frequency statistics value corresponding to each keyword; M and N are both positive integers; The determining module is used to determine a first feature set corresponding to the target city based on the semantic feature description text set and the keyword set; the first feature set includes at least one feature vector; and based on the degree of association between the feature vectors in the first feature set and the target region, to determine a second feature set from the first feature set whose degree of association with the target region is greater than or equal to a first threshold; the second feature set includes at least one feature vector.

7. The feature extraction apparatus according to claim 6, characterized in that, The processing module is specifically used to process the management specification information through a large language model to obtain the M semantic feature description texts, and to determine the M semantic feature description texts as the semantic feature description text set; and to perform word segmentation and stop word removal on the management specification information to obtain at least one candidate keyword; and to obtain the word frequency statistics value corresponding to the at least one candidate keyword; The candidate keywords whose word frequency statistics are greater than or equal to the second threshold among the at least one candidate keywords are determined as the N keywords, and the N keywords and their corresponding word frequency statistics are determined as the keyword set.

8. The feature extraction apparatus according to claim 6, characterized in that, The determining module is specifically used to determine the similarity between each semantic feature description text in the semantic feature description text set and each keyword in the keyword set by using a word vector similarity algorithm or a large language model; and to determine Q feature vectors based on the similarity. Each of the Q feature vectors includes: a semantic feature description text, P keywords, and the word frequency statistics of each of the P keywords; Q is a positive integer, P is a positive integer less than N; and the Q feature vectors are determined as the first feature set.

9. The feature extraction apparatus according to claim 6, characterized in that, The determining module is specifically used to determine, based on the correlation between the feature vectors in the first feature set and the target region, a third feature set from the first feature set whose correlation with the target region is greater than or equal to the first threshold using a clustering algorithm; the third feature set includes at least one feature vector cluster; and determines a feature vector cluster from the at least one feature vector cluster whose purity is greater than or equal to the third threshold, and determines the determined feature vector cluster as the second feature set.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a program or instructions that can run on the processor, the program or instructions being executed by the processor to implement the feature extraction method as described in any one of claims 1-5.

11. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a computer, implement the feature extraction method as described in any one of claims 1-5.

12. A computer program product, characterized in that, The computer program product is stored in a storage medium, and when executed by a computer, the computer program product implements the feature extraction method as described in any one of claims 1-5.

Images