Intellectual property data acquisition method and system based on multi-source data fusion

Abstract

The invention relates to the technical field of intellectual property data processing, and discloses an intellectual property data collection method and system based on multi-source data fusion, and the method comprises the steps: obtaining an original information set, and carrying out the clustering and grouping processing of the original information set, and obtaining a classification data set; performing feature verification and matching analysis on the classification data set to obtain a feature matching score; if the feature matching score is higher than a preset analysis threshold value, performing structured analysis on the classification data set to obtain a preliminary analysis field; performing standardized mapping processing on the preliminary analysis field to obtain data in a unified format; performing associated value mining on the data in the unified format to obtain high-value associated information; and fusing the high-value associated information to a preset overall data set to obtain a target intellectual property data set. According to the method, efficient integration and deep value mining of multi-source heterogeneous data can be realized, and the acquisition precision and standardization level of intellectual property data are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intellectual property data processing technology, and in particular to an intellectual property data acquisition method and system based on multi-source data fusion. Background Technology

[0002] Currently, intellectual property data has become a core strategic resource for enterprises' technological innovation and market competition. Faced with massive amounts of patent documents, trademark information, scientific journals, and publicly available online data, efficiently acquiring and integrating valuable information from scattered, heterogeneous, and dynamically changing data sources has become a key direction for the application of big data collection and processing technologies in the field of intellectual property. Accurate and comprehensive intellectual property data is not only the foundation for enterprises to conduct technological planning, but also an important basis for avoiding infringement risks and uncovering potential technical intelligence.

[0003] In one existing technology, the acquisition of intellectual property data mainly relies on a single database interface call or a simple web crawler tool based on preset keywords. This type of method typically stores text, image, and tabular data obtained from different channels directly, and subsequent processing is often limited to basic deduplication or format conversion, lacking in-depth analysis and adaptation of the characteristics of multi-source data. For example, when processing patent documents containing complex formulas or engineering drawings, existing systems often cannot effectively parse unstructured data, and lack intelligent fusion and verification mechanisms when faced with data conflicts from different sources.

[0004] Therefore, existing technologies suffer from low acquisition and parsing efficiency when dealing with multi-source heterogeneous data. Summary of the Invention

[0005] This invention provides a method and system for acquiring intellectual property data based on multi-source data fusion, in order to solve the technical problem of low acquisition and parsing efficiency when dealing with multi-source heterogeneous data in the prior art.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for acquiring intellectual property data based on multi-source data fusion, comprising: Obtain the original information set and perform clustering and grouping processing on the original information set to obtain classified data groups; Feature verification and matching analysis are performed on the classified data set to obtain a feature matching score; If the feature matching score is higher than the preset parsing threshold, then the classified data group is subjected to structured parsing to obtain preliminary parsing fields; The initially parsed fields are standardized and mapped to obtain data in a uniform format; By performing correlation value mining on the unified format data, high-value correlation information can be obtained; The high-value related information is integrated into a pre-defined overall dataset to obtain the target intellectual property dataset.

[0007] Secondly, the present invention provides an intellectual property data acquisition system based on multi-source data fusion, comprising: The data acquisition and grouping module is used to acquire the original information set and perform clustering and grouping processing on the original information set to obtain classified data groups; The feature matching module is used to perform feature verification and matching analysis on the classified data set to obtain a feature matching score. The structure parsing module is used to perform structured parsing on the classified data group to obtain preliminary parsing fields if the feature matching score is higher than a preset parsing threshold. The format unification module is used to perform standardized mapping processing on the initially parsed fields to obtain data with a unified format. The value mining module is used to perform correlation value mining on the unified format data to obtain high-value correlation information. The data fusion module is used to fuse the high-value related information into a preset overall dataset to obtain the target intellectual property dataset.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains the original information set and performs clustering and grouping, and then performs structured parsing and standardized mapping based on feature matching and scoring to filter the data; this big data processing and format normalization mechanism for multi-source heterogeneous data effectively solves the problem of chaotic structure and incompatible format of intellectual property data from different sources (such as text and images), realizes the accurate conversion of unstructured information into a unified standard format, and lays a high-quality data foundation for subsequent analysis.

[0009] (2) This invention performs correlation value mining on data in a unified format, uses correlation rule algorithms to deeply analyze the hidden patterns between data and accurately locate value points; this deep mining technology breaks through the limitations of traditional collection methods that only perform simple storage, and can keenly capture potential technical correlations and high-value intelligence from massive messy information, significantly improving the utilization depth of intellectual property data and its ability to support business decisions.

[0010] (3) The present invention integrates the high-value related information mined into the preset overall dataset and performs deep integration operations for duplicate elements and conflicts; this closed-loop data integration strategy not only effectively eliminates data redundancy, but also ensures the integrity and logical consistency of the final generated target intellectual property dataset, and realizes the efficient accumulation and dynamic updating of intellectual property data assets. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the intellectual property data acquisition method based on multi-source data fusion provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an intellectual property data acquisition system based on multi-source data fusion provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a method for acquiring intellectual property data based on multi-source data fusion, comprising the following steps: S11, Obtain the original information set, and perform clustering and grouping processing on the original information set to obtain classified data groups; S12, Perform feature verification and matching analysis on the classified data group to obtain a feature matching score; S13, if the feature matching score is higher than the preset parsing threshold, then the classified data group is subjected to structured parsing to obtain preliminary parsing fields; S14, Perform standardized mapping processing on the initially parsed fields to obtain data in a unified format; S15, perform correlation value mining on the unified format data to obtain high-value correlation information; S16, the high-value related information is integrated into a preset overall dataset to obtain the target intellectual property dataset.

[0014] In step S11, the original information set is obtained, and the original information set is clustered and grouped to obtain categorized data groups, including: Obtain the original information set, and perform data cleaning and deduplication on the original information set to obtain the sorted information data; The sorted information data is subjected to source feature extraction and grouping processing to obtain grouped source feature data; The grouped source feature data is subjected to anomaly isolation and reclassification processing to obtain the classified data group.

[0015] It should be noted that the acquisition of the raw information set is achieved through a distributed web crawler cluster deployed in the cloud and a dedicated database interface adapter. The web crawler cluster performs full data scraping from publicly available patent databases, academic resource websites, and industry technical forums on the Internet, while the interface adapter accesses commercial intellectual property databases through a pre-authorized API channel. The raw information set includes unstructured HTML web pages, semi-structured XML / JSON metadata, and unstructured PDF document streams.

[0016] It should be noted that the data cleaning and deduplication of the original information set is implemented using Locality Sensitive Hashing (LSH) technology based on the SimHash algorithm. This process aims to identify and remove highly repetitive redundant data. First, each original information item is segmented using the TF-IDF algorithm, word frequencies are counted, and inverse document frequencies are calculated to obtain the weight value of each feature word. The top-64 feature words with the highest weights are extracted, and a 64-bit binary hash value for each feature word is generated using the MD5 algorithm. Then, a 64-dimensional weighted vector V is initialized, and the hash value of each feature word is iterated. If the i-th bit of the hash value is 1, then... Add the weight of the word; if it is 0, subtract the weight. Next, perform dimensionality reduction projection (binarization) on the weighted vector V. If the i-th bit of the final fingerprint is set to 1, otherwise it is set to 0, thus generating a unique 64-bit SimHash fingerprint for this information item. Finally, the fingerprint of the current information item is XORed with the fingerprints in the pre-established database of stored fingerprints, and the number of "1"s in the result is counted. This count is the Hamming Distance. If the Hamming Distance is less than a preset similarity threshold, it is determined to be duplicate data and physical deletion is performed. The only valid data retained is the sorted information data.

[0017] It is worth noting that the preset similarity threshold (e.g., 3) is determined based on statistical analysis of the Hamming distance distribution between duplicate and non-duplicate document pairs in the historical intellectual property dataset. Specifically, positive and negative sample sets are constructed and probability density curves of the distance distribution are plotted. The distance value at the intersection of the two curves (i.e., the point with the lowest classification error rate) is selected as the preset value to ensure that version differences are preserved while deduplication.

[0018] It should be noted that the source feature extraction and grouping processing of the organized information data is achieved using feature vectorization based on structure density and the K-Means clustering algorithm. First, the system parses the organized information data to extract key features that characterize its source. Specifically, the DOM tree depth is obtained by traversing the DOM tree structure of HTML and XML and calculating the maximum level depth; the text tag ratio is obtained by calculating the ratio of plain text characters to tag nodes; and the metadata field completeness is obtained by statistically analyzing the proportion of non-empty metadata fields to the total standard fields. Subsequently, these three feature values ​​are respectively subjected to min-max normalization and mapped to three-dimensional feature vectors. Then, these feature vectors are input into a K-Means clusterer, with a preset number of clusters K. By iteratively calculating the Euclidean distance from each sample point to the K cluster center vectors, the sample is assigned to the nearest cluster, and the formula is used... .

[0019] The algorithm recalculates the new center of each cluster until the displacement of the cluster center is less than a preset convergence threshold (e.g., 0.001), at which point the algorithm converges. The cluster label (Cluster ID) assigned to each data item is its source group identifier, and the data set with this identifier is the source feature data of that group.

[0020] It is worth noting that the determination of the preset number of clusters K is based on the Elbow Method analysis. Specifically, the sum of squared errors (SSE) within all clusters is calculated for different K values ​​(e.g., K=2 to 10), a curve of SSE versus K is plotted, and the K value at the point of maximum curvature (e.g., 5) is selected as the preset value. This value typically corresponds to the number of natural categories of the data (e.g., patents, trademarks, papers, standards, others). The determination of the preset convergence threshold is based on the resolution statistical analysis of the feature space. Specifically, in the normalized feature space, the Euclidean distance distribution between all sample points and their nearest neighbor sample points is calculated, and the 1st percentile of this distance distribution is selected as the convergence threshold. This setting ensures that the movement of the cluster center is less than the minimum distinguishable difference between samples, thereby ensuring clustering accuracy while avoiding unnecessary invalid iterations.

[0021] It should be noted that the anomaly isolation and reclassification processing of the grouped source feature data aims to correct outliers and misclassified items in the clustering process. This processing consists of two stages. The first stage is anomaly isolation, implemented using the Isolation Forest algorithm. This algorithm constructs a preset number (e.g., 100) of binary isolation trees (iTrees). When constructing each tree, for the dataset contained in the root node, a feature dimension q is randomly selected, along with a split point p between the maximum and minimum values ​​of that dimension in the current data, dividing the data into... and Two subsets are used, and this process is performed recursively until each data point is isolated or the tree height limit is reached. Then, the average path length of each data point x across all trees is calculated. And using the formula, .

[0022] The first stage calculates an anomaly score. If the anomaly score exceeds a preset anomaly threshold, it is identified as an outlier and temporarily isolated. The second stage is reclassification, implemented using the K-Nearest Neighbors (KNN) voting algorithm. For each isolated outlier, the system finds its K nearest neighbors in the unisolated normal dataset, counts the cluster labels of these neighbors, and assigns the label with the highest frequency to the outlier, thus reclassifying it into the most likely correct category. The corrected full dataset is the classified data set.

[0023] It is worth noting that the anomaly detection threshold is determined based on the contamination rate of historical data. Specifically, the proportion of noisy points in the dataset (e.g., 5%) is estimated, and the quantile value corresponding to the anomaly score distribution is used as the threshold. The value of K in the KNN algorithm is determined based on the square root of the sample size of the normal dataset to ensure statistical robustness.

[0024] For example, the system collects a patent document Doc-A to be processed and a document Doc-B already in the database. After calculating the SimHash fingerprint of Doc-A, it is found that its Hamming distance with the fingerprint of Doc-B is 2. Since 2 is less than the preset similarity threshold of 3, the system determines that Doc-A is a redundant duplicate and removes it. Subsequently, the system extracts features from the remaining processed information data. The feature vector of document Doc-C shows that its DOM tree depth is 15 layers and the text label ratio is 0.6. The K-Means algorithm assigns it to the patent data cluster. In the anomaly detection stage, the Isolation Forest algorithm calculates the anomaly score of document Doc-D as 0.85 (greater than the threshold of 0.6), and marks it as an anomaly. The KNN algorithm (K=7) then finds the 7 nearest neighbors of Doc-D in the normal data, and finds that 6 of these neighbors belong to the technical standard cluster. Based on this, the system corrects Doc-D from the anomaly state and classifies it into the technical standard category, finally outputting a correctly classified data group.

[0025] In step S12, feature verification and matching analysis are performed on the classified data set to obtain a feature matching score, including: The classified data set is subjected to structural feature extraction and normalization processing to obtain normalized feature data; The update frequency parameters are extracted from the classified data set to obtain the update frequency parameters; The standardized feature data is compared with a preset feature library to obtain preliminary matching results; If the initial matching result is lower than the preset matching threshold, the normalized feature data is adaptively adjusted using the update frequency parameter to obtain corrected feature data. The modified feature data is subjected to a second matching calculation to obtain the feature matching score.

[0026] It should be noted that the structural feature extraction and normalization of the categorized data set were achieved using multimodal feature encoding and Z-Score normalization techniques. First, for the text content (abstract, claims), a BERT-Base pre-trained model based on the Transformer architecture was used for extraction. This model was obtained from an open-source model hosting platform (such as Hugging Face Model Hub). Before use, the model was further pre-trained using a masked language model (MLM) with Chinese Wikipedia and public patent literature datasets (such as Google Patents Public Data). The pre-training process involved using a whole-word masking strategy, randomly replacing 15% of the input tokens with [MASK] tags, and iteratively training with an AdamW optimizer (learning rate set to 1e-4) until the loss function converged. During use, the text was truncated or padded to a length of 512 tokens, input into the model, and the 768-dimensional vector corresponding to the [CLS] tag output from the last layer was extracted as the semantic feature vector. For the image content (patent illustrations), a ResNet-50 deep residual network was used for extraction. The network was obtained using a deep learning framework (such as the TorchVision library) and pre-trained on the ImageNet dataset. The fine-tuning process involved replacing the last fully connected layer with a classification layer whose output dimension equals the number of patent technology fields (e.g., 128 classes), freezing the weights of the first three residual blocks, and training only the fourth residual block and the classification layer. Parameter updates were performed using a cross-entropy loss function and a stochastic gradient descent (SGD) optimizer until the classification accuracy stabilized. In practice, the image was resized to 224x224 pixels and input into the fine-tuned network. The last classification layer was removed, and the 2048-dimensional vector output from the global average pooling layer was extracted as the visual feature vector. One-Hot encoding was used to generate the structured metadata (IPC classification numbers). First, a unique value dictionary of all IPC main classification numbers was constructed, establishing an index mapping. Then, an all-zero vector was initialized, setting the element at the index position corresponding to the current data IPC classification number to 1, generating a sparse feature vector. Then, a vector concatenation operation is used to connect the semantic vector, visual vector, and sparse feature vector in a fixed order, merging them into a single high-dimensional original feature vector. Finally, using the formula, Z-score normalization is performed on each dimension of the feature vector, where and These are the arithmetic mean and standard deviation of the corresponding feature dimensions in the historical full intellectual property dataset, respectively, thereby eliminating the dimensional differences between different features and obtaining the normalized feature data.

[0027] It should be noted that the update frequency parameter extraction for the aforementioned categorized data groups is achieved through a timestamp differential statistical method. The system accesses a pre-established intellectual property lifecycle database and, based on the unique identifier of the data item (such as the application number), extracts the timestamp sequence of the data and its related family data from the database's audit log field for the most recent n (e.g., 10) changes (including status changes such as publication, substantive examination, authorization, and transfer). Using the formula, ; Calculate the average number of updates F per unit time, if If F is expressed in months, then F is the number of updates per month, which is the update frequency parameter. This parameter objectively reflects the activity level and technological iteration speed of the intellectual property data.

[0028] It is worth noting that the extraction strategy for the most recent n changes can be optimized in practical applications to a method based on a fixed observation time window. Specifically, a backtracking time window is set (e.g., tracing back 24 months from the current date), and the timestamps of all state change events that occurred in the data and its related family data within this time window are extracted to form a timestamp sequence. The setting of the time window length (24 months) is determined based on industry experience values ​​of intellectual property technology iteration cycle and legal status update activity, which can balance the timeliness and statistical stability of the data.

[0029] It is worth noting that the construction process of the intellectual property full life cycle database is as follows: the system regularly synchronizes official patent gazettes, legal status announcements, and transfer registration information through data interfaces; a relational database containing a basic information table, a legal status change table, and a family association table is established using the unique application number as the primary key; and an audit log record with a precise timestamp is generated for each status change (such as publication, substantive examination, authorization, invalidation, transfer, pledge) as the data source for frequency extraction.

[0030] It should be noted that the similarity comparison between the standardized feature data and the preset feature library is achieved using the Weighted Cosine Similarity method. The system first loads a preset feature library. The library is constructed by selecting the top 5% of patents cited in historical data, as well as patents involved in invalidation requests or infringement lawsuits, as a high-value seed set; extracting feature vectors from the seeds using the aforementioned multimodal feature encoding method; and then using the K-Means++ clustering algorithm to cluster and remove duplicates from the feature vectors, retaining the vectors of each cluster center as standard reference vectors. The data is stored in a vector database. During comparison, the current normalized feature vector V is compared with the standard vector in the feature library. Using formulas ; Calculate similarity, where The feature weight vector contains the weights of all dimensions in the initial state. All values ​​are set to 1.0. The maximum similarity obtained is the preliminary matching result.

[0031] It should be noted that the number of citations can be obtained by accessing commercial patent analysis databases (such as the DerwentCitation Index) or parsing citation information from official patent gazettes; the patent information involving invalidation requests or infringement lawsuits can be obtained by querying decisions issued by the Patent Reexamination and Invalidation Department of the State Intellectual Property Office and other public judicial channels such as China Judgments Online; the system regularly synchronizes information from these authoritative data sources to ensure the accuracy and timeliness of the high-value seed set.

[0032] It is worth noting that the preset matching threshold (e.g., 0.75) is determined based on ROC curve analysis of historical mismatch cases. Specifically, the true positive rate (TPR) and false positive rate (FPR) under different similarity thresholds are calculated, and the similarity value corresponding to the maximum point of the Youden Index is selected as the threshold to minimize the risk of misclassification while ensuring recall.

[0033] It should be noted that the adaptive weight adjustment of the normalized feature data using the update frequency parameter aims to introduce the dynamic activity of the data as confidence compensation. When the initial matching result is lower than the preset matching threshold, the system believes that the static structural features of the data may deviate from the standard vectors in the preset feature library due to novelty or non-standard writing, but its high-frequency update characteristic suggests its potential high value. Therefore, a weight adjustment formula is constructed using a Sigmoid activation function with a normalization factor. ; in, These are the initial weights for the feature dimensions; The adjusted weights; F is the update frequency parameter for the current data; This is the frequency baseline mean; The standard deviation of the frequency reference; This is the maximum gain factor (e.g., 0.5), used to control the maximum amplification of the weights; This is a slope control factor (a dimensionless pure number, e.g., 1.0) used to control the sensitivity of the activation function; this parameter is typically tuned on the validation set using a grid search to achieve the optimal balance between precision and recall; for example, it can be set to... The search is performed in increments of 0.1 within the range of 0.5 to 2.0.

[0034] It is worth noting that, This constitutes the dimensionless Z-Score standard score, eliminating the frequency unit. The influence of this ensures that the input to the exponential function e is a pure numerical value. This formula non-linearly maps the activity deviation of the data to the gain weights of the feature vector, thereby generating the corrected feature data with dynamic weight information. In one implementation, the frequency reference mean... with standard deviation The determination is based on the statistical distribution of update frequency of all active data in the intellectual property lifecycle database within the most recent year. The system periodically (e.g., weekly) calculates the update frequency of all data and extracts its arithmetic mean as the update frequency. Extract its standard deviation as This is to construct a dynamic statistical benchmark. Maximum gain coefficient. The determination of this is based on the test settings for the model's generalization ability. Different [specific parameters] are tested on the validation set. The impact of the value (0.1 to 1.0) on the final recall rate was determined by selecting the inflection point value that improves the recall rate without significantly reducing the precision rate.

[0035] For example, the system extracts the normalized feature vector of a patent application document. Its initial cosine similarity with the standard vectors in the preset feature library is 0.65, which is lower than the preset matching threshold of 0.75, posing a risk of being filtered out. However, the system detected 12 family update records for this patent in the past 6 months and calculated the update frequency parameter. Times / month. Consulting statistical benchmarks reveals the historical average frequency. Times / month, standard deviation Times / month. Substitute into the formula to calculate the standard score. (That is, activity level is 3 standard deviations above the mean). Let... The Sigmoid part is approximately The final weight gain coefficient is The system uses this coefficient to enhance the weights of key dimensions such as technical field classification in the feature vector. Based on the adjusted weight vector, a secondary matching calculation is performed, resulting in a feature matching score of 0.78, successfully passing the threshold verification and ensuring the effective collection of this highly active data.

[0036] In step S13, if the feature matching score is higher than a preset parsing threshold, then the classified data set is subjected to structured parsing to obtain preliminary parsing fields, including: If the feature matching score is higher than the preset parsing threshold, the element type of the classified data group is separated to obtain a set of text elements; Field location and attribute detection are performed on the text element set to obtain field attribute parameters; If the field attribute parameter does not meet the preset length threshold, then the text element set is subjected to boundary expansion and neighborhood concatenation to obtain the optimized field position; The text element set is captured based on the optimized field position to obtain the preliminary parsed field.

[0037] It should be noted that if the feature matching score is higher than the preset parsing threshold, it means that the score calculated in S12 has passed the quality gate. The element type separation of the classified data group is achieved using a layout analysis method based on morphological gradients. This method first converts the original image data into grayscale images, using the following conversion formula: Then, the morphological gradient of the image is calculated to highlight edges. Specifically, a 3x3 structuring element is used to perform dilation and erosion operations on the grayscale image, and the dilation result is subtracted from the erosion result to obtain the gradient map. The dilation operation involves sliding the 3x3 structuring element on the image and taking the maximum grayscale value of the pixels within the structuring element's coverage area as the new value of the center pixel. The erosion operation involves taking the minimum grayscale value of the pixels within the structuring element's coverage area as the new value of the center pixel. Next, a rectangular structuring element with an aspect ratio of the morphological structuring element (e.g., 10:1) is applied to the gradient map to perform morphological dilation, merging adjacent text characters into connected regions in the horizontal direction, while non-text characters (such as images and table lines) form independent block regions. Finally, the image is traversed using the Two-Pass Algorithm, assigning a unique label to each connected pixel region, and calculating the minimum bounding rectangle of each region to extract geometric attributes (aspect ratio, fill density). If the aspect ratio of a region is greater than a preset text ratio threshold and the fill density is higher than a preset density threshold, it is determined to be a text region and included in the text element set.

[0038] It is worth noting that the preset parsing threshold is based on logistic regression analysis of feature matching scores and parsing success rates. Specifically, matching scores and parsing result labels (success / failure) of historical samples are collected, and a logistic regression model is constructed. ; The score value X corresponding to a predicted probability P reaching 98% was selected as the threshold. The preset ratio of the morphological structural elements (10:1) was set based on the statistical ratio of the average character width to the character spacing in Chinese and English documents, ensuring that characters can connect into lines horizontally while keeping lines separated vertically. The preset text ratio threshold and the preset density threshold were derived from the statistical distribution of geometric features of plain text areas and non-text areas (images, tables) in historical documents. The 5th percentile of the text area distribution was selected as the lower limit threshold to filter out irregularly shaped noise.

[0039] It should be noted that the field localization and attribute detection of the text element set are achieved using a deep text recognition model based on CRNN (Convolutional Recurrent Neural Network) and CTC (Connection-Temporal Classification) architecture, combined with keyword anchoring technology. The system first normalizes each region image in the text element set to a fixed height (e.g., 32 pixels) and inputs it into the CRNN model. The CRNN model structure is as follows: the feature extraction layer uses a VGG-16 architecture (containing 7 convolutional layers, 3x3 kernel size, stride 1, ReLU activation function, and 2x2 max pooling layers after layers 2, 4, and 6); the sequence modeling layer uses a bidirectional long short-term memory network (Bi-LSTM), containing 2 hidden layers, each with 256 hidden units; the transcription layer uses CTC Loss. The training process uses a publicly available synthetic text dataset (Synth90k) for pre-training and fine-tuning using an annotated real patent image dataset until the character recognition accuracy converges. After the model outputs a character sequence, the system loads a pre-built intellectual property keyword library. This library is constructed by statistically analyzing the frequency of occurrence of field titles (such as abstract, claims, and applicant) in historical standard documents, selecting the top-50 most frequent keywords and their common variations. The system uses regular expression matching to locate the identified text lines, traversing them to find anchor words that match the keyword library, and returning the coordinates of the bounding rectangle of the text line containing the anchor word. For each identified potential field, its geometric attributes (width w, height h) and semantic attributes (number of characters N) are extracted. These data constitute the field attribute parameters.

[0040] It should be noted that if the field attribute parameters do not meet the preset length threshold, the text element set undergoes boundary expansion and neighborhood concatenation, which is a repair mechanism for the common line breakage problem during parsing. When the number of characters N of a key field (such as the summary content) is detected to be less than the preset length threshold, the system triggers a neighborhood concatenation operation. This operation iterates through the candidate boxes for that field. All surrounding text blocks Calculate them with vertical distance and horizontal overlap rate . The calculation formula is as follows: (Assuming) exist (below). The calculation formula is as follows: ; If the conditions are met ( (where the line spacing threshold is) and ( (for alignment thresholds), then determine for The broken section. The system performs a boundary expansion operation, merging the two rectangles into a new bounding box. The formula for calculating its coordinates is: , The final bounding box coordinates determined after recursive merging are the optimized field positions.

[0041] It is worth noting that the preset length threshold is determined based on the minimum character length distribution of each key field in historical standard documents, selecting the 5th percentile of this distribution as the threshold. For example, the threshold for the abstract field is set relatively high (e.g., 50 characters), while the threshold for the inventor field is set relatively low (e.g., 2 characters). The line spacing threshold... The threshold is determined based on document layout analysis and statistics, selecting the 99th percentile of the line spacing distribution in the main text as the threshold. The preset parsing threshold is determined based on logistic regression analysis of feature matching scores and parsing success rates, selecting the score value corresponding to a parsing success rate of 98%. The alignment threshold... (For example, 0.5) is based on typesetting standards. Usually, when text lines in the same paragraph are left-aligned and justified, their horizontal overlap is extremely high. Setting it to 0.5 can effectively distinguish aligned text lines from offset side notes and page numbers.

[0042] It should be noted that capturing the content of the text element set based on the optimized field position refers to using the adjusted... The coordinates are used to crop the target region (ROI) on the original high-resolution image. In the original image matrix, the coordinates are used to crop the ROI. For the starting line, To end the line, with For the starting column, To conclude the analysis, a submatrix is ​​extracted to obtain a local high-pass-rate image patch; subsequently, this image patch undergoes secondary recognition. The secondary recognition employs the high-precision mode of the aforementioned CRNN model, maintaining the input image resolution without downsampling, directly inputting it into the network for inference, and concatenating the recognized text stream in a top-down, left-to-right order, removing newline characters and extra whitespace, ultimately outputting structured key-value pair data (e.g., key="Abstract", value="This invention relates to..."), which is the preliminary parsed field.

[0043] For example, when processing a scanned patent document, the system first locates the text block Box-A in the abstract field based on anchor words. The system loads a preset length threshold (50 characters) for the abstract field. Upon inspection, the content of Box-A is only "This invention relates to a kind of," with 7 characters, which is less than the specific threshold and is judged as a risk of line breakage. The system triggers a neighbor concatenation operation, and the search finds another text block Box-B located 10 pixels directly below Box-A, with the content "Based on multi-source data fusion...". The vertical distance between the two is calculated. pixels smaller than the line spacing threshold (15 pixels); Horizontal overlap rate Greater than the alignment threshold (0.5). The system determined that the two belonged to the same paragraph and performed boundary expansion to merge Box-A and Box-B into a large rectangle. Based on this optimized position, a second recognition was performed, successfully capturing the complete summary content "This invention relates to an intellectual property data acquisition method based on multi-source data fusion...", thus avoiding information truncation.

[0044] In step S14, the initially parsed fields are standardized and mapped to obtain data in a unified format, including: The preliminary parsed fields are subjected to structural alignment analysis with the preset standard template to obtain field offset parameters; If the field offset parameter exceeds the preset offset threshold, the field distribution of the initially parsed field is reconstructed to obtain the reconstructed field distribution; The reconstructed field distribution is subjected to template mapping and logical consistency verification to obtain verification result data; The data is formatted based on the verification results to obtain the unified format data.

[0045] It should be noted that the structural alignment analysis of the preliminary parsed fields with the preset standard template is achieved using a geometric matching method based on anchor point features. First, a preset standard template is loaded. This template is constructed during the system initialization phase. The construction process involves collecting a large number (e.g., 100,000) of historical documents conforming to industry standard layouts, extracting the centroid coordinates of key fixed fields (such as application numbers and invention name tags), and calculating the average position of these centroids in the standard page coordinate system. This forms a standard layout topology diagram. During analysis, the system extracts the actual centroid coordinates of the corresponding key fields from the preliminary analysis fields. Using the Euclidean distance formula, Calculate the displacement of each anchor point and take the arithmetic mean of the displacements of all anchor points. This average value is the field offset parameter.

[0046] It is worth noting that the identification of the key fixed fields can be defined in conjunction with the internationally accepted structure of patent documents (e.g., according to the WIPO ST.36 standard or the document components specified in the Chinese Patent Examination Guidelines). These typically include titles or fixed identifier text for fields such as "Application Number," "Invention Title," "Applicant," and "Abstract." When constructing the template, the system can cluster the layout analysis results of multiple sample documents, exclude documents with abnormal layout styles, and perform statistical analysis on the coordinates of the same key field in the remaining documents. The median, rather than the arithmetic mean, is used as its standard position to reduce the influence of extreme values, thereby obtaining a more robust standard layout topology.

[0047] It should be noted that if the field offset parameter exceeds a preset offset threshold, the field distribution of the initially parsed fields is reconstructed. This process involves using the RANSAC (Random Sample Consensus) algorithm to solve for the optimal affine transformation matrix. This operation first assumes that the document undergoes overall rotation or translation deformation, and then randomly selects three pairs of anchor points from the initially parsed fields to construct the affine transformation equation. Parameters are solved iteratively Find the transformation matrix that maximizes the number of interior points (i.e., the number of points whose distance from the transformed coordinates to the template coordinates is less than a preset interior point determination threshold). Use this matrix to perform coordinate correction transformation on the coordinates of all fields in the preliminary analysis fields. Specifically, directly use the solved optimal parameters to transform the original coordinate vector of each field. Substitute into the formula The corrected new coordinates are calculated; where M is the linear transformation matrix of the affine transformation, responsible for geometric operations such as rotation, scaling, and shearing; T is the translation vector of the affine transformation, responsible for the overall displacement operation. This layout information, reverting to the standard template coordinate system, constitutes the reconstructed field distribution.

[0048] It is worth noting that the preset offset threshold is determined based on statistical analysis of printing deviations in historical document scans. Specifically, the distribution of mechanical deviations caused by normal scanning and printing in historical data is statistically analyzed, and the 95th percentile of this distribution (e.g., 20 pixels) is calculated and set as the allowable natural error limit, i.e., the preset offset threshold. The preset interior point determination threshold is determined based on image registration accuracy requirements, typically selecting the maximum allowable pixel error (e.g., 3 pixels) to ensure that only high-confidence anchor points participate in model calculations.

[0049] It should be noted that the template mapping and logical consistency verification of the reconstructed field distribution are implemented using a location-index-based mapping mechanism and a rule engine. First, based on the reconstructed coordinates, the content of each text block is directly mapped to the schema slots defined in the standard template (e.g., mapping the text in the upper right corner of the page to the publication number). Then, a pre-built logical rule library is invoked for verification. Rules include, but are not limited to, date logic verification (e.g., the publication date must be later than the application date) and format regular expression verification (e.g., the application number must conform to the CN\d{13}\.? format). Fields that fail verification are marked as exceptions pending processing. The verification status and mapping content of all fields constitute the verification result data.

[0050] It is worth noting that the construction process of the rule base involves analyzing intellectual property-related laws and regulations (such as the Examination Guidelines of the Patent Law) and industry data dictionaries, extracting legal logical constraints (such as date order) and data format constraints (such as numbering rules) between fields, and writing them into executable regular expressions.

[0051] It should be noted that the format normalization processing based on the verification results is implemented using an ETL pipeline. For the data that passes verification, the system performs a series of atomic transformation operations, specifically including date normalization (converting "2023.01.01" and "Jan 1,2023" into "2023-01-01"), full-width / half-width conversion (converting full-width numbers into half-width numbers), and null padding (filling non-mandatory blank fields with "N / A"). The structured data generated after cleaning, conforming to the target database storage specifications, is the unified format data.

[0052] For example, the system processes a skewed patent document. During structural alignment analysis, the centroid of the application number label is found to be offset by 50 pixels relative to the standard template, exceeding the preset offset threshold of 20 pixels. The system triggers field distribution reconstruction, using the RANSAC algorithm to calculate a 3-degree clockwise rotation and a translation of (10, -5) in the document. After applying a correction transformation matrix, the field coordinates return to normal. Subsequently, the system maps the text "2024.05.20" at (200, 300) to the application date. After logical verification, the format normalization module converts it to the standard format "2024-05-20" and finally stores it in the unified format database.

[0053] In step S15, the unified format data is subjected to association value mining to obtain high-value association information, including: The distribution characteristics of the uniformly formatted data are analyzed to obtain the field association strength. If the correlation strength of the field is lower than the preset strength threshold, the feature dimensions of the unified format data are expanded and completed to obtain the reorganized correlation pattern. Logical verification is performed on the recombined association pattern to obtain the verification hidden pattern; The high-value related information is obtained by matching and analyzing the verification concealment pattern with the preset key intelligence feature mapping table.

[0054] It should be noted that the distribution feature analysis of the uniform format data is implemented using the Frequent Itemset Mining (FP-Growth) method. The system first converts the uniform format data into a transaction dataset, where each data record is considered a transaction, and the field values ​​in the record (such as "Applicant: Company A" and "Technology Field: Field B") are considered items. Then, an FP-tree (Frequent Pattern Tree) is constructed to compress the transaction data. The specific mining steps are as follows: 1) Calculate the frequency (support) of all items; 2) Remove items with support below the minimum and sort them in descending order of frequency; 3) Traverse the transaction set to construct the root node and prefix path of the FP-tree; 4) Starting from the bottom item of the FP-tree, construct a Conditional Pattern Base and build a Conditional FP-tree based on it; 5) Recursively mine frequent itemsets on the Conditional FP-tree until the tree is empty. For each set of association rules mined... Using formulas Calculate its confidence weighted value, which is the correlation strength of the field.

[0055] It is worth noting that the determination of the preset strength threshold is based on statistical analysis of historical high-value intelligence data. Specifically, it involves retrospectively analyzing association rules that have been verified as key intelligence in the past (such as strong associations between specific applicants and emerging technology terms), calculating the probability distribution of their association strength, and selecting the 15th percentile of this distribution as the threshold to ensure that weak signals in their nascent or concealed stages can be captured.

[0056] It should be noted that if the correlation strength of the field is lower than the preset strength threshold, feature dimension expansion and completion are performed on the unified format data. This is an operation aimed at introducing external contextual information to enhance data correlation. This operation is achieved by loading a pre-set industry knowledge graph. The construction process of the industry knowledge graph is as follows: entities (applicants, inventors, IPC classification numbers, technical keywords) are extracted from the full patent database and business information database as nodes, and the relationships between entities (joint applications, citations, transfers, parent and subsidiary companies, synonyms) are extracted as edges and stored in a graph database (such as Neo4j). The system maps the entities in the current data to nodes in the knowledge graph and extracts the K-hop neighbor node information of the node in the graph (such as parent company, cooperative institutions, and superior concepts). This external knowledge is then concatenated into the original data as new feature dimensions to form a high-dimensional feature vector. Subsequently, the mutual information between features is recalculated in the expanded high-dimensional space, and the calculation formula is as follows: in For joint probability distribution, and The marginal probability distribution is defined as follows: feature pairs whose mutual information exceeds a preset mutual information threshold are identified as the recombination association pattern.

[0057] It is worth noting that the preset mutual information threshold is determined based on the feature independence test. Specifically, unrelated feature pairs are randomly selected from the historical dataset to calculate mutual information, a noise distribution is constructed, and the 99th percentile of this distribution (corresponding to the significance level) is selected as the threshold.

[0058] It should be noted that the logical verification of the aforementioned reorganization association patterns is implemented using a rule engine based on temporal causal constraints. The system checks whether the event nodes in the association patterns violate physical time logic. For example, for the pattern "Technology A derives from Technology B", the system verifies whether the earliest publication time of Technology A is earlier than the application time of Technology B. Simultaneously, it checks the legal status logic between entities (e.g., transfer must occur after authorization). Patterns that pass verification are the verified hidden patterns.

[0059] It should be noted that the matching analysis between the verification hiding pattern and the preset key intelligence feature mapping table is implemented using a weighted scoring matrix. First, the preset key intelligence feature mapping table is loaded. This table is constructed based on the system's periodic extraction of national industry guidance catalogs, industry white papers, and industry technical guidance documents. High-frequency strategic keywords are extracted using NLP technology and clustered into high-value tags (such as competitors' core patents, bottleneck technologies, and overseas deployments). The system calculates the semantic similarity between the keywords or entities contained in the verification hiding pattern and these high-value tags. Specifically, a pre-trained Word2Vec model is used to map words to vectors, and the cosine similarity between the vectors is calculated. If the similarity exceeds a preset semantic similarity threshold, the corresponding tag is assigned to the pattern, and a value score is calculated. (in For tag weight, (Similarity). Patterns with scores higher than a preset value score baseline are extracted and constitute the high-value association information.

[0060] It is worth noting that the preset semantic similarity threshold is determined based on clustering analysis of the thesaurus, calculating the average cosine similarity of synonym pairs (e.g., 0.6), and using this as the lower limit for determining semantic relevance. The preset value scoring baseline is determined based on the statistical distribution of scores from historically adopted intelligence, selecting the 90th percentile of the score distribution as the baseline to ensure the high value attributes of the output information. The label weights are determined by quantifying the administrative authority level of the source documents. Specifically, if the keywords originate from national policy documents or laws and regulations, the weight is set to 1.0; if they originate from industry association standards or technical white papers, the weight is set to 0.8; and if they originate from general business research reports or news information, the weight is set to 0.6.

[0061] For example, system analysis revealed that the direct correlation strength between "Applicant: Company A" and "Technical term: Graphene" was 0.3, lower than the preset strength threshold of 0.5 (indicating that Company A did not directly apply for graphene patents). The system triggered feature dimension expansion, discovering through a knowledge graph that Company A's wholly-owned subsidiary, Company B, held a large number of graphene patents, and that Company A had recently recruited Company B's core inventor. The system introduced two new dimensions: subsidiary holdings and inventor mobility, reconstructing the correlation pattern of "Company A is strategically positioning itself in the graphene sector through talent acquisition," increasing the mutual information to 0.8 (higher than the preset mutual information threshold of 0.6). After verification using time logic (recruitment time prior to the new patent application) and confirming its accuracy, this pattern matched the potential competitor activity tags in the business value table (similarity 0.85), ultimately outputting high-value correlation information.

[0062] In step S16, the high-value related information is integrated into a preset overall dataset to obtain a target intellectual property dataset, including: Redundancy analysis and quantitative calculation are performed on the high-value correlation information and the preset overall dataset to obtain redundancy statistical indicators; If the redundant statistical indicators exceed the preset threshold, the high-value related information is deduplicated, fused, and conflict-resolved to obtain a refined classification result. The target intellectual property dataset is obtained by merging the data based on the refined classification results.

[0063] It should be noted that the redundancy analysis and quantitative calculation of the high-value related information and the preset overall dataset are achieved using a fast retrieval technology based on an inverted index. The system first loads the preset overall dataset. This dataset originates from a persistent repository of historically accumulated, cleaned, and standardized multi-source intellectual property data (including official gazette data, commercial database procurement data, and web crawled data). The system establishes a primary key index (such as application number and unified social credit code) for this dataset. Then, it iterates through each data item in the high-value related information, querying whether its primary key already exists in the index. If it does, it further calculates the Jaccard Similarity Coefficient for key fields (such as abstract text and legal status), the formula of which is... Where A and B represent the feature word sets after word segmentation in the two data records, respectively. The number of elements in the intersection of the two sets. This represents the number of elements in the union of two sets. The system counts the number of completely duplicate items (identical primary key and content) and conflicting items (identical primary key but different content) using the formula. Calculate the overlap rate; this ratio is the redundancy statistic. This represents the sum of the number of completely duplicate items and conflicting items. This represents the total number of high-value related information records to be integrated this time.

[0064] It is worth noting that the predetermined threshold (e.g., an overlap rate of 5%) is determined based on a load balancing analysis of system storage I / O overhead and data update concurrency. Specifically, the linear relationship between the amount of new data and system I / O load during historical data fusion operations is statistically analyzed, and the overlap rate corresponding to the critical load before the I / O bottleneck occurs is selected as the threshold to avoid invalid high-frequency write operations blocking the database.

[0065] It should be noted that if the redundant statistical indicators exceed a preset threshold, the deduplication, fusion, and conflict resolution of the high-value related information are achieved using a truth discovery method based on confidence priority. This method loads a pre-defined data source-field confidence matrix for conflicting items (i.e., the same entity described differently in different sources). The matrix is ​​constructed as follows: the system periodically samples and evaluates the data quality of each data source, calculating weighted scores for three dimensions: accuracy, completeness, and timeliness (e.g., accuracy 0.5, timeliness 0.3, completeness 0.2). These scores are then normalized and used as the authority weight of the data source for a specific field. This matrix defines the authority weights of different sources (e.g., official announcements vs. third-party scraping) for a specific field. For conflicting fields, the algorithm selects the source value with the highest weight as the truth value; if the weights are the same, the latest data version is retained, following the timestamp update priority principle. The processed data is labeled with three categories: new records, updated records, and discarded records, which is the refined classification result.

[0066] It should be noted that the data merging operation based on the refined classification results is implemented using the database's batch Upsert transaction mechanism. The system constructs a transaction batch, performing INSERT operations on data marked as new records; for data marked as updated records, performing UPDATE operations only on the changed fields; and ignoring discarded records. After the transaction is committed, the index is automatically rebuilt, and the final updated full database is the target intellectual property dataset.

[0067] For example, the system compares a batch of newly mined high-value information (100 records) with the overall dataset. Redundancy analysis shows that the primary keys of 80 records already exist, with a redundancy statistic (overlap rate) of 80%, far exceeding the preset threshold of 5%. The system then enters the conflict resolution phase, discovering that the legal status of one record, "Patent A," is "public" in the overall dataset, but "substantive examination" in the new information. Checking the confidence matrix, the source of the new information (official API, weight 0.95) is higher than the source of the old data (web crawler, weight 0.6), and the timestamp is updated. The system resolves the conflict, retains the new status, and marks the record as an updated record. Finally, the system merges the remaining 20 newly added records with the updated record, completing the dynamic update of the dataset.

[0068] In summary, this invention, by constructing a full-chain processing flow from raw information acquisition and clustering, intelligent parsing and standardized mapping based on feature matching scoring, to deep correlation value mining and multi-source data fusion, achieves accurate collection and efficient integration of multi-source heterogeneous intellectual property data. It solves the technical problems of complex data sources, inconsistent formats, and difficulty in deep mining of value information in existing technologies, and significantly improves the automation level, data quality, and intelligence analysis capabilities of big data processing in the field of intellectual property.

[0069] Reference Figure 2 The second embodiment of the present invention provides an intellectual property data acquisition system based on multi-source data fusion, comprising: The data acquisition and grouping module is used to acquire the original information set and perform clustering and grouping processing on the original information set to obtain classified data groups; The feature matching module is used to perform feature verification and matching analysis on the classified data set to obtain a feature matching score. The structure parsing module is used to perform structured parsing on the classified data group to obtain preliminary parsing fields if the feature matching score is higher than a preset parsing threshold. The format unification module is used to perform standardized mapping processing on the initially parsed fields to obtain data with a unified format. The value mining module is used to perform correlation value mining on the unified format data to obtain high-value correlation information. The data fusion module is used to fuse the high-value related information into a preset overall dataset to obtain the target intellectual property dataset.

[0070] It should be noted that the intellectual property data acquisition system based on multi-source data fusion provided in this embodiment of the invention is used to execute all the process steps of the intellectual property data acquisition method based on multi-source data fusion in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0071] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an intellectual property data acquisition program based on multi-source data fusion. When the processor executes the computer program, it implements the steps described in the various embodiments of the intellectual property data acquisition method based on multi-source data fusion, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the data acquisition and grouping module.

[0072] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0073] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0074] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0075] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0076] If the modules / units integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0077] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0078] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for acquiring intellectual property data based on multi-source data fusion, characterized in that, include: Obtain the original information set and perform clustering and grouping processing on the original information set to obtain classified data groups; Feature verification and matching analysis are performed on the classified data set to obtain a feature matching score; If the feature matching score is higher than the preset parsing threshold, then the classified data group is subjected to structured parsing to obtain preliminary parsing fields; The initially parsed fields are standardized and mapped to obtain data in a uniform format; By performing correlation value mining on the unified format data, high-value correlation information can be obtained; The high-value related information is integrated into a pre-defined overall dataset to obtain the target intellectual property dataset.

2. The intellectual property data acquisition method based on multi-source data fusion according to claim 1, characterized in that, The process of obtaining the original information set and performing clustering and grouping processing on the original information set to obtain categorized data groups includes: Obtain the original information set, and perform data cleaning and deduplication on the original information set to obtain the sorted information data; The sorted information data is subjected to source feature extraction and grouping processing to obtain grouped source feature data; The grouped source feature data is subjected to anomaly isolation and reclassification processing to obtain the classified data group.

3. The intellectual property data acquisition method based on multi-source data fusion according to claim 1, characterized in that, The step of performing feature verification and matching analysis on the classified data set to obtain a feature matching score includes: The classified data set is subjected to structural feature extraction and normalization processing to obtain normalized feature data; The update frequency parameters are extracted from the classified data set to obtain the update frequency parameters; The standardized feature data is compared with a preset feature library to obtain preliminary matching results; If the initial matching result is lower than the preset matching threshold, the normalized feature data is adaptively adjusted using the update frequency parameter to obtain corrected feature data. The modified feature data is subjected to a second matching calculation to obtain the feature matching score.

4. The intellectual property data acquisition method based on multi-source data fusion according to claim 1, characterized in that, If the feature matching score is higher than a preset parsing threshold, then the classified data group is subjected to structured parsing to obtain preliminary parsing fields, including: If the feature matching score is higher than the preset parsing threshold, the element type of the classified data group is separated to obtain a set of text elements; Field location and attribute detection are performed on the text element set to obtain field attribute parameters; If the field attribute parameter does not meet the preset length threshold, then the text element set is subjected to boundary expansion and neighborhood concatenation to obtain the optimized field position; The text element set is captured based on the optimized field position to obtain the preliminary parsed field.

5. The intellectual property data acquisition method based on multi-source data fusion according to claim 1, characterized in that, The standardization mapping process performed on the initially parsed fields to obtain data in a unified format includes: The preliminary parsed fields are subjected to structural alignment analysis with the preset standard template to obtain field offset parameters; If the field offset parameter exceeds the preset offset threshold, the field distribution of the initially parsed field is reconstructed to obtain the reconstructed field distribution; The reconstructed field distribution is subjected to template mapping and logical consistency verification to obtain verification result data; The data is formatted based on the verification results to obtain the unified format data.

6. The intellectual property data acquisition method based on multi-source data fusion according to claim 1, characterized in that, The process of performing correlation value mining on the unified format data to obtain high-value correlation information includes: The distribution characteristics of the uniformly formatted data are analyzed to obtain the field association strength. If the correlation strength of the field is lower than the preset strength threshold, the feature dimensions of the unified format data are expanded and completed to obtain the reorganized correlation pattern. Logical verification is performed on the recombined association pattern to obtain the verification hidden pattern; The high-value related information is obtained by matching and analyzing the verification concealment pattern with the preset key intelligence feature mapping table.

7. The intellectual property data acquisition method based on multi-source data fusion according to claim 1, characterized in that, The process of integrating the high-value related information into a preset overall dataset to obtain a target intellectual property dataset includes: Redundancy analysis and quantitative calculation are performed on the high-value correlation information and the preset overall dataset to obtain redundancy statistical indicators; If the redundant statistical indicators exceed the preset threshold, the high-value related information is deduplicated, fused, and conflict-resolved to obtain a refined classification result. The target intellectual property dataset is obtained by merging the data based on the refined classification results.

8. An intellectual property data acquisition system based on multi-source data fusion, characterized in that, include: The data acquisition and grouping module is used to acquire the original information set and perform clustering and grouping processing on the original information set to obtain classified data groups; The feature matching module is used to perform feature verification and matching analysis on the classified data set to obtain a feature matching score. The structure parsing module is used to perform structured parsing on the classified data group to obtain preliminary parsing fields if the feature matching score is higher than a preset parsing threshold. The format unification module is used to perform standardized mapping processing on the initially parsed fields to obtain data with a unified format. The value mining module is used to perform correlation value mining on the unified format data to obtain high-value correlation information. The data fusion module is used to fuse the high-value related information into a preset overall dataset to obtain the target intellectual property dataset.

Images