A multi-source data fusion method, device, equipment and medium

Abstract

The present application relates to the technical field of artificial intelligence, and in particular to a multi-source data fusion method and device, equipment and medium. In the present application, multi-source data to be fused is obtained, and each source data is structured to obtain structured data in a unified format. Based on each structured data, each structured data is processed in blocks to obtain a plurality of data blocks corresponding to each structured data. Through the structured processing and block processing of the multi-source data, the original data is split and standardized, laying a foundation for subsequent precise matching and association of features. Then, the association information is determined through inter-block similarity matching, and only the data blocks with semantic association are fused, thereby avoiding ineffective integration of irrelevant information, better capturing deep semantic association between different source data, effectively improving the fusion effect of multi-source data, and making the finally obtained fusion data feature expression more comprehensive and robust.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device and medium for multi-source data fusion. Background Technology

[0002] With the rapid development of big data and natural language processing technologies, text data exhibits significant characteristics such as diverse sources, heterogeneous features, and dispersed dimensions. Multi-source data fusion technology aims to improve decision-making accuracy and reliability by integrating heterogeneous data from multiple sources. It is an integrated analysis technique for heterogeneous, multi-dimensional information, capable of uniformly standardizing and complementaryly fusing feature information from different sources, types, and representational dimensions. This breaks through the limitations of single-feature expression, achieving information redundancy cancellation and complementary advantages, thereby constructing a more comprehensive and robust feature representation system. In recent years, deep learning has been widely used to improve fusion performance; however, it lacks a deep understanding of the semantic relationships between heterogeneous data and relies on manually designed features or simple attention mechanisms. Therefore, the semantic gap between text and image / time-series data is difficult to bridge, resulting in poor fusion effects. Thus, improving fusion performance in the multi-source data fusion process has become an urgent problem to solve. Summary of the Invention

[0003] This invention provides a method, apparatus, device, and medium for multi-source data fusion to solve the problem of poor fusion effect in the process of multi-source data fusion.

[0004] Firstly, a multi-source data fusion method is provided, the multi-source data fusion method comprising: Acquire multi-source data to be merged, perform structured processing on each source data, and obtain structured data in a unified format; Based on each structured data, each structured data is divided into blocks to obtain multiple data blocks corresponding to each structured data. For any target data block in any structured data, calculate the similarity value between the target data block and other corresponding data blocks in other structured data, and determine the associated data blocks that are related to the target data block based on the similarity value. The target data block and the associated data block are merged to obtain the merged data.

[0005] Secondly, a multi-source data fusion device is provided, the multi-source data fusion device comprising: The acquisition module is used to acquire multi-source data to be merged, perform structured processing on each source data, and obtain structured data in a unified format. The block segmentation module is used to segment each structured data into blocks based on each structured data, resulting in multiple data blocks corresponding to each structured data. The calculation module is used to calculate the similarity value between any target data block in any structured data and other corresponding data blocks in other structured data for any target data block, and to determine the associated data blocks that are related to the target data block based on the similarity value. The fusion module is used to fuse the target data block with the associated data block to obtain fused data.

[0006] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described multi-source data fusion method.

[0007] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described multi-source data fusion method.

[0008] The advantages of this invention compared to the prior art are: In this application, multi-source data to be fused is acquired, and each source data is processed in a structured manner to obtain structured data in a unified format. Based on each structured data, it is further divided into blocks to obtain multiple data blocks corresponding to each structured data. By performing structured processing and block processing on the multi-source data separately, the original data is split and standardized, laying the foundation for subsequent accurate matching of associated features. Then, the association information is determined by inter-block similarity matching, and only data blocks with semantic relationships are fused in a targeted manner, avoiding the ineffective integration of irrelevant information and better capturing the deep semantic relationships between different source data. This effectively improves the fusion effect of multi-source data, making the final fused data feature expression more comprehensive and robust. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of an application environment for a multi-source data fusion method provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a multi-source data fusion method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a multi-source data fusion device provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0011] 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, not all, of the embodiments of the present invention. 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.

[0012] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0013] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0014] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0015] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0016] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0017] The embodiments of this invention can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0018] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0019] It should be understood that the sequence number of each step in the following embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0020] To illustrate the technical solution of the present invention, specific embodiments are described below.

[0021] The multi-source data fusion method provided in one embodiment of the present invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server receives multi-source data sent by the client, acquires other multi-source data, performs structured processing on each source data to obtain structured data in a unified format, and then divides each structured data into blocks to obtain multiple data blocks corresponding to each structured data. For any target data block in any structured data, the similarity between the target data block and other data blocks in other structured data is calculated. Based on the similarity value, associated data blocks with a relationship to the target data block are determined. The target data block and associated data blocks are then merged to obtain the merged data. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0022] Please see Figure 2 , Figure 2 This is a flowchart illustrating a multi-source data fusion method provided in an embodiment of the present invention, including the following steps: S21: Obtain the multi-source data to be merged, perform structured processing on each source data, and obtain structured data in a unified format.

[0023] In step S21, multi-source data refers to data collected from different data sources, and structured data refers to data obtained after standardizing the data format of multi-source data.

[0024] In this embodiment, multi-source data to be fused is acquired. The data source types corresponding to the multi-source data can include structured data, text data, and image data. For example, multi-source data can be various types of data in banking operations, including financial statements of corporate clients, such as various data in balance sheets, profit and loss statements, and cash flow statements; corporate transaction records at the bank, such as account statements and loan repayment records; unstructured text or image data generated by business personnel through written descriptions or photos during due diligence or visits; and unstructured text data such as online financial reports, corporate-related news reports, and industry trends.

[0025] It's important to note that within the banking system, data often originates from different branches. Firstly, different branches use different methods and dimensions to associate data, and their definitions of the same characteristic also differ. For instance, the naming and coding methods for legal entities and company names may not be entirely identical. For example, for the same company information table, branch A might contain a large amount of basic company information, while branch B might contain a large amount of basic information about the company's legal representative and transaction information. Similarly, the company name might be called "Company Name" in branch A's database and "Full Company Name" in branch B's database—seemingly inconsistent, but actually referring to the same entity. Secondly, different branches update data at different frequencies. Taking legal entity information as an example, some branches iterate slowly, while others iterate quickly, leading to inconsistencies in the data for the same instance. In such cases, it's necessary to use the most up-to-date data fusion method based on business needs to obtain the latest legal entity information and avoid information lag.

[0026] Each source data point undergoes structuring processing to obtain structured data in a unified format. For text data, key features are extracted and converted into structured data. This process involves segmenting the text according to pre-defined word segmentation rules, selecting core entity words relevant to the business scenario as key features, and then filling these features into fields with a pre-defined unified structure to complete the structuring transformation of the text data. For numerical data or source data already in a regular format, data can be directly mapped and matched to corresponding fields in a unified structure to complete the structuring processing. For image data, text information and feature content are first extracted from the image using a pre-defined image recognition model. The extracted feature content is then transformed using the same structuring processing method as text data and matched into corresponding fields in a pre-defined unified structure to complete the structuring processing of the image data. The network structure of the image recognition model can include convolutional layers, pooling layers, fully connected layers, and an output layer connected in sequence. The convolutional layer is used to extract features at different levels of the input image. The pooling layer performs dimensionality reduction processing on the feature map output by the convolutional layer, compressing the feature dimension while retaining key information and reducing the subsequent computation. The fully connected layer integrates the extracted features. Finally, the output layer outputs the recognized text information and feature content.

[0027] It should be noted that the fields in the pre-defined unified structure are standardized fields pre-defined according to business scenario requirements. Each field corresponds to a clear feature type and data storage specification, which can accommodate feature content of the same business attribute from different source data. This ensures that after structured processing, all key features of different source data can be stored in a unified and standardized framework, providing a standardized data foundation for subsequent fusion processing. For example, in the business scenario of personnel information integration, the pre-defined unified structure will pre-set standardized fields such as "name," "ID number," "contact information," and "date of employment." Information with the same name from resumes, employee registration forms, and paper scans from different sources will all be filled into the same standardized field, avoiding the problem of scattered storage and inconsistent formats of the same attribute information.

[0028] It should be noted that structured data exists in the form of structured data tables.

[0029] In this embodiment, after the above processing, each source data is converted into structured data in a unified and standardized format. All key features fall within the preset standardized framework, eliminating the problems of inconsistent data formats and storage standards from different sources, thus laying a solid data foundation for subsequent feature fusion processing.

[0030] Optionally, each source data is processed to obtain structured data in a uniform format, including: Determine the data source type for each source data; Obtain the preset conversion rules corresponding to each data source type, and convert each source data into structured data in a unified format based on the data source type and the preset conversion rules.

[0031] In this embodiment, the data source type for each source data is determined. The data source type includes structured data, text data, and image data, etc. The corresponding structured data is a structured data table. When converting each source data into structured data of a unified format, the corresponding source data is first converted into JSON format data using preset conversion rules, and then the JSON format data is converted into structured data.

[0032] The system retrieves the preset conversion rules for each data source type, which differ between data source types. Based on the data source type and the preset conversion rules, each source data is converted into structured data in a unified format. For example, if the data source type is text data, the preset conversion rules are: convert the text data to JSON format, and then convert the JSON format data to the corresponding structured data. If the data source type is image data, the preset conversion rules are: extract text from the image data, convert the text data to JSON format, and then convert the JSON format data to the corresponding structured data.

[0033] It should be noted that when converting text data to JSON format, the conversion logic is driven by a preset Prompt template, which clearly defines field extraction requirements, output format specifications, and redundant content filtering standards. Leveraging the semantic understanding and feature extraction capabilities of the large-scale model, the effective information in the unstructured text data is extracted, avoiding ambiguity and information mixing issues, and generating standard JSON format data strictly according to preset rules.

[0034] For example, the Prompt template is: Please convert the following sentence into JSON format, including the fields: name, age, city, occupation, and salary. Sentence: Mr. Zhang is 25 years old, a programmer in Beijing, and earns a monthly salary of 15,000 yuan. Accuracy optimization involves calling multiple large models to generate results and selecting the optimal solution through a voting mechanism. Performance optimization involves caching conversion results for frequently repeated sentences to reduce the number of API calls; time-consuming large model calls are placed in a message queue for asynchronous processing to avoid blocking the main process. Cost optimization involves using lightweight models for simple sentences and calling complex models for complex sentences; multiple sentences are merged into a single request (some APIs support batch input). The large models can be preset models such as GPT-4 and Claude models. The lightweight models can be GPT-3.5-Turbo models, and the complex models can be GPT-4 models, etc.

[0035] Based on the text data and using the Prompt template, corresponding JSON format data is generated and then converted into a structured data table. When generating the JSON data from the text data and Prompt template, a multi-precision generation and validation strategy can be employed. This involves calling multiple different large models based on the same Prompt template to generate JSON results for the input text data. A voting mechanism is used to compare the consistency of the results generated by each model, and the result with the most votes is selected as the final output, thereby improving the precision and accuracy of the generated data. Simultaneously, layered performance optimization is implemented: for frequently recurring text tasks, the corresponding conversion results are pre-cached, and when the same task is received again, the cached result is returned directly, reducing unnecessary API calls; long-running large model call requests are placed in a message queue for asynchronous processing to avoid blocking the main business process and ensure the overall service response stability. Furthermore, cost control can be implemented as needed: for text with clear format and simple content, a lightweight large model is called to complete the conversion; for text with complex logic and high information density, a more complex model with larger parameters is called to complete the generation; and for APIs that support batch processing, multiple texts to be converted are merged into a single request, further reducing call costs.

[0036] When converting JSON data into a structured data table, you can iterate through each key-value pair in the JSON data according to the preset mapping relationship between field names and data table column names, and insert the corresponding value into the corresponding field cell in the data table. If there are missing fields in the JSON format, they are marked as null in the corresponding data table position, thus completing the generation of the structured data table.

[0037] If the data source is image data, the text information of the image data is extracted through a preset model to obtain the text data corresponding to the image data. Using the above method, the text data is converted into JSON format data, and the JSON format data is converted into the corresponding structured data, i.e., a structured data table.

[0038] In this embodiment, by matching corresponding transformation rules to different types of source data, heterogeneous data from multiple sources are uniformly converted into a structured format. This eliminates the need for manual adjustments to the processing logic for each different data source, effectively improving the efficiency of multi-source data processing.

[0039] S22: Based on each structured data, perform block processing on each structured data to obtain multiple data blocks corresponding to each structured data.

[0040] In step S22, a data block is the smallest semantic unit of structured data.

[0041] In this embodiment, the structured data is the corresponding structured data table. When processing each piece of structured data in blocks, it can be done by column, extracting independent data blocks for each column to obtain multiple data blocks corresponding to each piece of structured data. Alternatively, it can be block-based, treating each record as a complete semantic unit. Or, it can be dynamically clustered based on the semantic relevance of fields, such as merging "name, age, city" into an identity block, and "occupation, salary, industry" into an occupation block, etc.

[0042] For example, a bank's customer information table contains six columns: "name, ID number, bank, account balance, transaction date, and transaction amount". These columns can be divided into six data blocks.

[0043] It should be noted that when processing each piece of structured data into blocks, the corresponding data block serves as the basic unit for merging with other data blocks. For example, if the data is divided into blocks by columns, then data merging is performed based on the corresponding columns.

[0044] In this embodiment, by processing in blocks, not only is fine-grained decoupling of data semantics achieved, but also subsequent on-demand combination and cross-source alignment are supported, improving the flexibility and accuracy of data fusion.

[0045] S23: For any target data block in any structured data, calculate the similarity value between the target data block and each other data block in other structured data, and determine the associated data blocks that are related to the target data block based on the similarity value.

[0046] In step S23, the associated data block is the corresponding data block in other structured data that is semantically consistent with or highly related to the target data block.

[0047] In this embodiment, for any target data block in any structured data, the similarity value between the target data block and each other data block in other structured data is calculated. When calculating the similarity, the field similarity between the target data block and each other data block can be calculated to obtain the similarity between the target data block and each other data block. When calculating the content similarity between the target data block and each other data block, a semantic vector encoding method based on a pre-trained language model can be used. First, the semantic vectors of the field names of the target data block and each other data block are extracted separately. Then, the similarity between the two vectors is calculated using cosine distance, and this similarity is used as the content similarity value between the corresponding data blocks. After obtaining all similarity values, other data blocks with similarity values ​​greater than a preset threshold are identified as associated data blocks of the current target data block.

[0048] It should be noted that the pre-trained language model can be BERT, RoBERTa, or ERNIE (specifically for Chinese). The pre-trained language model is a pre-trained model. During training, training samples are obtained, which include columns from different data tables and their corresponding semantic labels. Supervised learning is then used to train the model, resulting in a pre-trained language model.

[0049] In this embodiment, the similarity value between the target data block and each other data block is calculated to determine the associated data blocks, so as to accurately map the semantically consistent data blocks and thus perform data fusion and alignment.

[0050] Optionally, the similarity between the target data block and other data blocks in other structured data is calculated, including: Obtain the data types of the target data block and each other data block, as well as the mapping relationship between different data types. Based on the data types of the target data block and each other data block, and combined with the mapping relationship, determine the data type similarity between the target data block and each other data block. Semantic encoding is performed on the target data block and each other data block to obtain the target encoding vector of the target data block and the other encoding vectors of each other data block. Based on the target encoding vector and each other encoding vector, the semantic similarity between the target data block and each other data block is calculated. Obtain the preset distribution similarity calculation rules corresponding to different data types, and calculate the distribution similarity between the target data block and each other data block based on the data types of the target data block and each other data block, combined with the preset distribution similarity calculation rules. Based on data type similarity, semantic similarity, and distribution similarity, the similarity between the target data block and other data blocks in other structured data is calculated.

[0051] In this embodiment, the data types of the target data block and the corresponding data of each other data block are obtained, as well as the mapping relationships between different data types. Based on the data types of the target data block and the other data blocks, and combined with the mapping relationships, the data type similarity is determined. The data types can include basic types such as integer types, integer types represented as floating-point numbers, floating-point numbers, date and time types, and Boolean types. The mapping relationships between different data types represent the degree of similarity between them. Based on these mapping relationships, a data type similarity matrix is ​​constructed. This matrix represents the degree of similarity between each data type in a numerical quantification manner.

[0052] For example, regarding integer types, the similarity between an integer type and itself is 1.0, indicating they are completely identical; the similarity with integers represented in floating-point form is 0.9, although their storage formats differ, they are essentially both integer concepts with minor differences; the similarity with floating-point numbers is 0.7, because floating-point numbers include a decimal part, there is some difference compared to integers; the similarity with Boolean values ​​is 0.3, although Boolean values ​​can be represented by 0 and 1, their semantics are significantly different, hence the low similarity; and the similarity with other data types is 0, indicating incompatibility with other data types. For integers represented in floating-point form, the similarity with itself is 1.0, indicating they are completely identical; the similarity with integer types is 0.9, indicating they are essentially similar; the similarity with floating-point numbers is 0.8, compared to integers, it is closer to floating-point numbers in terms of storage format; and the similarity with other data types is 0, indicating incompatibility with other data types. For floating-point numbers, the similarity is 1.0 with itself, indicating identical data types; the similarity with integers is 0.7, as floating-point numbers include a decimal part, creating some difference; the similarity with integers represented in floating-point form is 0.8; and the similarity with other data types is 0, indicating incompatibility. For date / time types, the similarity is 1.0 with itself; and the similarity with date strings is 0.8, as date strings can be parsed and converted to date / time types, resulting in high similarity. For date strings, the similarity is 1.0 with itself; and the similarity with date / time types is 0.8. For Boolean types, the similarity is 1.0 with itself; and the similarity with integers is 0.3, as Boolean values ​​can be represented by 0 and 1, but the semantic differences are significant, resulting in low similarity. These similarity values ​​reflect the inherent connections and differences between different data types.

[0053] Based on the data types of the target data block and other data blocks, and combined with the mapping relationship, the data type similarity is determined. That is, based on the data types of the target data block and other data blocks, the corresponding similarity is found in the corresponding data type similarity matrix to obtain the data type similarity between the target data block and other data blocks.

[0054] Semantic encoding is performed on the target data block and all other data blocks to obtain the target encoding vector of the target data block and the other encoding vectors of each other data block. Based on the target encoding vector and the other encoding vectors, the semantic similarity between the target data block and each other data block is calculated. When semantically encoding the target data block and all other data blocks, a pre-trained language model can be used to extract context-aware semantic features. This pre-trained language model can be a Chinese-supporting model such as BERT, RoBERTa, or LLaMA. When calculating the semantic similarity between the target data block and each other data block based on the target encoding vector and the other encoding vectors, the cosine similarity formula can be used. Its value range is [-1, 1], with values ​​closer to 1 indicating greater semantic similarity.

[0055] The system obtains preset distribution similarity calculation rules corresponding to different data types. Based on the data types of the target data block and other data blocks, and in conjunction with the preset distribution similarity calculation rules, it calculates the distribution similarity between the target data block and other data blocks. Different data types correspond to different distribution similarity calculation rules.

[0056] For example, for numerical data, the KS (Kolmogorov-Smirnov) test and a combination of statistical measures can be used to calculate distribution similarity. The KS test is a commonly used non-parametric test method for comparing the similarity between two data distributions. It determines whether two data blocks come from the same population distribution by comparing the maximum difference between their empirical distribution functions. Specifically, the maximum difference between the empirical distribution functions of two data blocks is taken as the KS statistic. If the two data blocks come from the same population distribution, their empirical distribution functions should be very similar, i.e., the KS statistic should approach 0; conversely, if the difference is large, it indicates that the two data blocks may come from different population distributions, i.e., the KS statistic should approach 1.

[0057] Combinations of statistics quantify differences in central tendency and dispersion of data blocks by directly comparing statistical measures such as the mean and standard deviation of the values. The mean reflects the central location of the data, while the standard deviation describes the dispersion. Combining these two measures complements each other, improving the accuracy and reliability of similarity calculations. To calculate the similarity corresponding to a combination of statistics, the sum of the difference between normalized means and the difference between normalized standard deviations can be used. That is, calculate the difference between the mean of the target data block and the means of the corresponding other data blocks, and then normalize this difference to obtain the normalized mean difference; similarly, calculate the difference between the standard deviation of the target data block and the standard deviation of the corresponding other data blocks, and then normalize this difference to obtain the normalized standard deviation difference. Adding the normalized mean difference and the normalized standard deviation difference together yields the similarity corresponding to the combination of statistics.

[0058] The distribution similarity between a target data block and other data blocks is calculated based on the similarity of the KS statistic and its combination. This can be achieved by adding the first difference between the KS statistic and 1, and the second difference between the similarity of the statistic combination and 1. The larger the sum of these two differences, the more similar the distributions of the two data blocks. Weights can be assigned to the first and second differences, and their contribution ratio can be adjusted according to the specific application scenario. For example, if the distribution pattern of the data is more important, the weight of the KS statistic can be increased; if statistical characteristics are considered more important, the weight of the similarity of the statistic combination can be increased. A weighted sum based on these weights yields the distribution similarity.

[0059] For string data, frequency distribution cosine similarity is used to calculate distributional similarity. Frequency distribution cosine similarity considers the frequency of each category in categorical data, not just the category itself. This allows it to capture the similarity of two sequences in terms of category distribution, making it suitable for scenarios requiring comparison of category distributions. For example, to determine all possible categories in two string sequences of a target data block and various other data blocks, the union of all categories in the two string sequences is taken to form a list containing all possible categories. Based on this list of categories, a frequency vector is constructed for each string sequence, where each element corresponds to the frequency of a category in the sequence. Finally, the cosine similarity between these two frequency vectors is calculated as the distributional similarity of the string data.

[0060] For time-type data, the time is first extracted and processed, converted into a string, and then the frequency distribution cosine similarity is used to calculate the distribution similarity.

[0061] The similarity between the target data block and other data blocks in other structured data is calculated based on data type similarity, semantic similarity, and distribution similarity. The similarity can be obtained by adding the data type similarity, semantic similarity, and distribution similarity together. Alternatively, weights can be assigned to each of the data type similarity, semantic similarity, and distribution similarity, and then a weighted sum can be performed to flexibly adjust the contribution ratio of different feature dimensions in the final similarity evaluation. This embodiment does not impose limitations.

[0062] In this embodiment, by calculating data type similarity, semantic similarity, and distribution similarity, the similarity between a target data block and other data blocks can be comprehensively evaluated from three dimensions: structure, content, and statistical features. Data type similarity ensures the consistency of data structure, semantic similarity captures the semantic relationships of data content, and distribution similarity reflects the degree of similarity in statistical features. This multi-dimensional comprehensive evaluation can effectively improve the accuracy and robustness of data matching.

[0063] Optionally, the semantic similarity between the target data block and each other data block is calculated, including: Obtain a preset encoding model, and use the preset encoding model to encode the target data block and each other data block respectively to obtain the first encoding vector of the target data block and the second encoding vector of each other data block. Calculate the cosine similarity between the first encoding vector and each second encoding vector. Determine the longest continuous matching segment between the target data block and each other data block, and calculate the string matching degree between the target data block and each other data block based on the longest continuous matching segment; Calculate the edit distance between the target data block and all other data blocks, and based on the edit distance, calculate the distance value between the target data block and all other data blocks; Identify the overlapping words between the target data block and other data blocks, and calculate the overlap rate between the target data block and other data blocks based on the overlapping words. The semantic similarity between the target data block and each other data block is calculated based on cosine similarity, string matching degree, distance value and overlap rate.

[0064] In this embodiment, a preset encoding model is obtained, and the target data block and each other data block are encoded using the preset encoding model to obtain a first encoding vector for the target data block and second encoding vectors for each other data block. The preset encoding model can be a pre-trained language model, such as the bRoBERTa (Bidirectional Robustly Optimized BERT Pretraining Approach) pre-trained model. When using the bRoBERTa pre-trained model to encode the target data block and each other data block, tokenizer segmentation and encoding are combined with Pooling (CLS) to obtain the corresponding encoding vectors. The cosine similarity between the first encoding vector and each second encoding vector is calculated. The cosine similarity is calculated using the formula for cosine similarity; the closer the value is to 1, the more consistent the directions of the two data blocks in the semantic space, and the more similar their semantics.

[0065] The longest continuous matching segment between the target data block and all other data blocks is determined. Based on this longest continuous matching segment, the string matching degree between the target data block and all other data blocks is calculated. Specifically, it involves finding the longest continuous common substring between two strings and calculating the ratio of the length of this longest continuous matching segment to the total length of the two strings. This ratio serves as a quantitative indicator of the string matching degree. This indicator effectively reflects the degree of direct overlap between two data blocks at the character level.

[0066] Calculate the edit distance between the target data block and all other data blocks. Based on the edit distance, calculate the distance values ​​between the target data block and all other data blocks. The Levenstein distance algorithm can be used to calculate the minimum number of editing operations (insertion, deletion, and replacement) required to transform one string into another using dynamic programming. Then, the edit distance is normalized, converting it to a distance value between 0 and 1. The closer the distance value is to 1, the greater the difference between the two data blocks at the character level, i.e., the lower the similarity.

[0067] The process involves identifying overlapping words between the target data block and other data blocks, and then calculating the overlap rate based on these overlapping words. To determine the overlapping words, the text can be segmented, stop words removed, and the number of common words in both data blocks counted. This number is then divided by the union of the total number of words in both data blocks to quantify the degree of overlap at the lexical level. A higher overlap rate indicates a higher similarity.

[0068] When calculating the semantic similarity between a target data block and other data blocks based on cosine similarity, string matching degree, distance value, and overlap rate, a higher cosine similarity indicates higher semantic similarity; a higher string matching degree indicates higher semantic similarity; a lower distance value indicates higher semantic similarity; and a higher overlap rate indicates higher semantic similarity. Therefore, when calculating semantic similarity, the distance value is converted into a positive similarity contribution, i.e., (1 minus the distance value) is used for conversion, ensuring it maintains the same direction as cosine similarity, string matching degree, and overlap rate. Subsequently, the cosine similarity, string matching degree, (1 minus the distance value), and overlap rate are summed to obtain the semantic similarity score. Alternatively, a weighted fusion method can be used, assigning weight coefficients to cosine similarity, string matching degree, (1 minus the distance value), and overlap rate, and obtaining a comprehensive semantic similarity score through weighted summation.

[0069] In this embodiment, a multi-level semantic similarity evaluation system is constructed by integrating four dimensions of indicators: cosine similarity, string matching degree, distance value, and overlap rate. This system comprehensively measures the similarity between data blocks from four perspectives: deep semantics, character sequence, editing cost, and word overlap. It can more comprehensively evaluate the similarity between data blocks and improve the accuracy and robustness of semantic similarity calculation.

[0070] Optionally, the similarity between the target data block and other data blocks in other structured data is calculated, including: If the data type is string, then obtain the string length of the target data block and each other data block, and calculate the length similarity between the target data block and each other data block based on the string length of the target data block and each other data block. Based on data type similarity, semantic similarity, distribution similarity, and length similarity, the similarity between the target data block and other data blocks in other structured data is calculated.

[0071] In this embodiment, if the data type is string, the string lengths of the target data block and each other data block are obtained. Based on the string lengths of the target data block and each other data block, the length similarity between the target data block and each other data block is calculated. The length similarity can be quantified by calculating the ratio of the lengths of the two strings, using the following formula: in, For length similarity, The length of the string in the target data block. The string length of the other data blocks, This is the set of element length features corresponding to the string of the target data block. This is a set of element length features for the strings corresponding to other data blocks. The size of the intersection of the two sets. This represents the maximum number of elements in the two length feature sets, used to normalize the feature overlap. A higher length similarity indicates a higher similarity between the target data block and all other data blocks.

[0072] When calculating the similarity between a target data block and other data blocks in other structured data based on data type similarity, semantic similarity, distribution similarity, and length similarity, these factors can be summed to obtain the overall similarity between the target data block and other data blocks in other structured data. Alternatively, weights can be assigned to each of the data type similarity, semantic similarity, distribution similarity, and length similarity according to actual needs, and a weighted sum can be performed to flexibly adjust the contribution ratio of different dimensional features in the final similarity evaluation. This embodiment does not impose limitations.

[0073] In this embodiment, when calculating the length similarity between the target data block and other data blocks when the data type is string, the length similarity is used as a quantifiable indicator, taking into account the length differences of the strings themselves, thereby improving the accuracy of similarity calculation.

[0074] S24: Merge the target data block with the associated data block to obtain the merged data.

[0075] In step S24, the fused data is a new and complete data entity formed by merging the target data block and the associated data block.

[0076] In this embodiment, when merging the target data block and the associated data block, data alignment and matching can be performed. Using the target data block's field system as a unified benchmark, field mapping and dimension alignment are performed on the associated data block, correcting the field order, format, and value caliber of the associated data block to achieve a one-to-one correspondence between the two data blocks' fields and a completely unified structure. The content of the target data block and the associated data block is compared field by field and sequence by sequence to determine the fields that are missing in the target data block and validly exist in the associated data block, the fields that exist in both the target and associated data blocks, and the fields that conflict between the target and associated data blocks. For fields that are missing in the target data block and validly exist in the associated data block, information is supplemented. For fields that exist in both the target and associated data blocks, redundant information is deleted, retaining only one copy of the field information present in either the target or associated data block. For fields that conflict between the target and associated data blocks, conflict resolution is performed. Conflict resolution can combine multi-model voting results, knowledge base verification results, and manual review mechanisms to select the optimal and accurate data, replacing the conflicting field information with the selected optimal and accurate data.

[0077] It should be noted that when there are multiple associated data blocks related to the target data block, these associated data blocks will be merged with the target data block sequentially according to the above processing flow, gradually completing data completion, redundancy removal, and conflict resolution. When there are no associated data blocks with the target data block, it means that the target data block is an independent data block, and no merging processing is required; the original data can be directly retained.

[0078] In this embodiment, the step-by-step fusion process of aligning structures, filtering information, filling gaps, and resolving conflicts can effectively solve the problems of different sources, inconsistent structures, and redundant and conflicting content of multi-source data, thereby improving the integrity and accuracy of the fused data.

[0079] Optionally, the target data block and the associated data block are merged to obtain merged data, including: Identify the application scenarios for the target data blocks, and determine the fusion type based on these scenarios; Depending on the fusion type, the target data block is merged with the associated data block to obtain the merged data.

[0080] In this embodiment, the application scenario of the target data block is obtained, and the fusion type is determined based on the application scenario. The application scenario refers to the purpose of the target data block; for example, it may include data augmentation scenarios, intelligent decision-making scenarios, etc. Different application scenarios correspond to different fusion strategies. Based on the fusion type, the target data block is fused with related data blocks to obtain fused data. The fusion strategy is a pre-set strategy corresponding to the application scenario. Once the corresponding application scenario is determined, the corresponding fusion strategy can be invoked to complete the data fusion.

[0081] For example, if the application scenario is data augmentation, the fusion type is feature-level fusion. This involves weighted voting, retrieving the latest data, and concatenating data from the target and associated data blocks to generate richer feature dimensions. This expands the data capacity while preserving the original core information, meeting the data augmentation requirements for data scale and feature diversity. If the application scenario is intelligent decision-making, the semantic features of the target and associated data blocks are extracted separately. These features are then input into a task-specific model, such as a classifier or regression model, and the final fused data is output.

[0082] In this application, multi-source data to be fused is acquired, and each source data is processed in a structured manner to obtain structured data in a unified format. Based on each structured data, it is further divided into blocks to obtain multiple data blocks corresponding to each structured data. By performing structured processing and block processing on the multi-source data separately, the original data is split and standardized, laying the foundation for subsequent accurate matching of associated features. Then, the association information is determined by inter-block similarity matching, and only data blocks with semantic relationships are fused in a targeted manner, avoiding the ineffective integration of irrelevant information and better capturing the deep semantic relationships between different source data. This effectively improves the fusion effect of multi-source data, making the final fused data feature expression more comprehensive and robust.

[0083] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0084] In one embodiment, a multi-source data fusion apparatus is provided, which corresponds one-to-one with the multi-source data fusion method described in the above embodiments. For example... Figure 3 As shown, the multi-source data fusion device 30 includes: an acquisition module 31, a block segmentation module 32, a calculation module 33, and a fusion module 34.

[0085] The acquisition module 31 is used to acquire multi-source data to be merged, perform structured processing on each source data, and obtain structured data in a unified format.

[0086] The block segmentation module 32 is used to segment each structured data into blocks based on each structured data to obtain multiple data blocks corresponding to each structured data.

[0087] The calculation module 33 is used to calculate the similarity value between any target data block in any structured data and other corresponding data blocks in other structured data for any target data block, and to determine the associated data blocks that are related to the target data block based on the similarity value.

[0088] The fusion module 34 is used to fuse the target data block with the associated data block to obtain the fused data.

[0089] Optionally, the acquisition module 31 includes: The determination unit is used to determine the data source type for each source data.

[0090] The transformation unit is used to obtain the preset transformation rules corresponding to each data source type, and convert each source data into structured data in a unified format according to the data source type and the preset transformation rules.

[0091] Optionally, the computing module 33 includes: The first acquisition unit is used to acquire the data types of the target data block and each other data block, as well as the mapping relationship between different data types. Based on the data types of the target data block and each other data block, and combined with the mapping relationship, the data type similarity between the target data block and each other data block is determined.

[0092] The first computing unit is used to perform semantic encoding on the target data block and each other data block to obtain the target encoding vector of the target data block and the other encoding vectors of each other data block. Based on the target encoding vector and each other encoding vector, the semantic similarity between the target data block and each other data block is calculated.

[0093] The second calculation unit is used to obtain the preset distribution similarity calculation rules corresponding to different data types. Based on the data types of the target data block and each other data block, and in combination with the preset distribution similarity calculation rules, the distribution similarity between the target data block and each other data block is calculated.

[0094] The third calculation unit is used to calculate the similarity between the target data block and other data blocks in other structured data based on data type similarity, semantic similarity, and distribution similarity.

[0095] Optionally, the first computing unit includes: The sub-unit is used to obtain a preset encoding model. The preset encoding model is used to encode the target data block and each other data block respectively to obtain the first encoding vector of the target data block and the second encoding vector of each other data block. The cosine similarity between the first encoding vector and each second encoding vector is calculated.

[0096] The first calculation subunit is used to determine the longest continuous matching segment between the target data block and each other data block, and to calculate the string matching degree between the target data block and each other data block based on the longest continuous matching segment.

[0097] The third calculation subunit is used to calculate the edit distance between the target data block and each other data block. Based on the edit distance, the distance value between the target data block and each other data block is calculated.

[0098] The fourth calculation subunit is used to determine the overlapping words between the target data block and other data blocks, and to calculate the overlap rate between the target data block and other data blocks based on the overlapping words.

[0099] The fifth calculation subunit is used to calculate the semantic similarity between the target data block and each other data block based on cosine similarity, string matching degree, distance value and overlap rate.

[0100] Optionally, the computing module 33 further includes: The fourth calculation unit is used to obtain the string length of the target data block and each other data block if the data type is string, and to calculate the length similarity between the target data block and each other data block based on the string length of the target data block and each other data block.

[0101] The fifth calculation unit is used to calculate the similarity between the target data block and other data blocks in other structured data based on data type similarity, semantic similarity, distribution similarity, and length similarity.

[0102] Optionally, the fusion module 34 includes: The second acquisition unit is used to acquire the application scenario of the target data block and determine the fusion type based on the application scenario.

[0103] The fusion unit is used to merge the target data block with the associated data block according to the fusion type to obtain the fused data.

[0104] It should be noted that the information interaction and execution process between the above-mentioned units are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0105] Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Figure 4 As shown, the computer device of this embodiment includes: at least one processor ( Figure 4 (Only one is shown), a memory, and a computer program stored in the memory and executable on at least one processor, wherein the processor executes the computer program to implement the steps in any of the above embodiments of the multi-source data fusion method.

[0106] This computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that... Figure 4 The examples of computer devices are merely examples and do not constitute a limitation on computer devices. Computer devices may include more or fewer components than shown in the illustration, or combinations of certain components, or different components, such as network interfaces, displays, and input devices.

[0107] The processor referred to can be a CPU, but it can also be 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.

[0108] Memory includes readable storage media, internal memory, etc., wherein internal memory can be the RAM of a computer device, providing an environment for the operation of the operating system and computer-readable instructions stored in the readable storage media. The readable storage media can be the hard drive of a computer device, or in other embodiments, it can be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, memory can include both internal storage units and external storage devices of the computer device. Memory is used to store the operating system, applications, bootloader, data, and other programs, such as program code for computer programs. Memory can also be used to temporarily store data that has been output or will be output.

[0109] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the functions described above can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this invention. The specific working process of the units and modules in the above device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention can implement all or part of the processes in the methods of the above embodiments by instructing related hardware through a computer program. 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 above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium can include at least: any entity or device capable of carrying computer program code, a recording medium, a computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0110] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be accomplished by a computer program product. When the computer program product is run on a computer device, the computer device executes the steps in the above method embodiments.

[0111] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0112] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software 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 implementations should not be considered beyond the scope of this invention.

[0113] In the embodiments provided by this invention, it should be understood that the disclosed apparatus / computer devices and methods can be implemented in other ways. For example, the apparatus / computer device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0114] The units described as separate components may or may not be physically separate. 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 units can be selected to achieve the purpose of this embodiment according to actual needs.

[0115] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0116] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A multi-source data fusion method, characterized in that, The multi-source data fusion method includes: Acquire multi-source data to be merged, perform structured processing on each source data, and obtain structured data in a unified format; Based on each structured data, each structured data is divided into blocks to obtain multiple data blocks corresponding to each structured data. For any target data block in any structured data, calculate the similarity value between the target data block and other corresponding data blocks in other structured data, and determine the associated data blocks that are related to the target data block based on the similarity value. The target data block and the associated data block are merged to obtain the merged data.

2. The multi-source data fusion method as described in claim 1, characterized in that, The process of structuring each source data to obtain structured data in a unified format includes: Determine the data source type for each source data; Obtain the preset conversion rules corresponding to each data source type, and convert each source data into structured data in a unified format according to the data source type and the preset conversion rules.

3. The multi-source data fusion method as described in claim 1, characterized in that, The calculation of the similarity between the target data block and other data blocks in other structured data includes: Obtain the data types of the target data block and each other data block, as well as the mapping relationship between different data types. Based on the data types of the target data block and each other data block, and in conjunction with the mapping relationship, determine the data type similarity between the target data block and each other data block. Semantic encoding is performed on the target data block and each other data block to obtain the target encoding vector of the target data block and the other encoding vectors of each other data block. Based on the target encoding vector and each other encoding vector, the semantic similarity between the target data block and each other data block is calculated. Obtain the preset distribution similarity calculation rules corresponding to different data types, and calculate the distribution similarity between the target data block and each other data block based on the data types of the target data block and each other data block, combined with the preset distribution similarity calculation rules. Based on the data type similarity, semantic similarity, and distribution similarity, the similarity between the target data block and other data blocks in other structured data is calculated.

4. The multi-source data fusion method as described in claim 3, characterized in that, The calculation of semantic similarity between the target data block and each other data block includes: Obtain a preset encoding model, and use the preset encoding model to encode the target data block and each other data block respectively to obtain the first encoding vector of the target data block and the second encoding vector of each other data block. Calculate the cosine similarity between the first encoding vector and each second encoding vector. Determine the longest continuous matching segment between the target data block and each other data block, and calculate the string matching degree between the target data block and each other data block based on the longest continuous matching segment; Calculate the edit distance between the target data block and each other data block, and calculate the distance value between the target data block and each other data block based on the edit distance; Identify the overlapping words between the target data block and other data blocks, and calculate the overlap rate between the target data block and other data blocks based on the overlapping words. The semantic similarity between the target data block and each other data block is calculated based on the cosine similarity, the string matching degree, the distance value, and the overlap rate.

5. The multi-source data fusion method as described in claim 3, characterized in that, The calculation of the similarity between the target data block and other data blocks in other structured data includes: If the data type is a string, then obtain the string length of the target data block and each other data block, and calculate the length similarity between the target data block and each other data block based on the string length of the target data block and each other data block. Based on the data type similarity, semantic similarity, distribution similarity, and length similarity, the similarity between the target data block and other data blocks in other structured data is calculated.

6. The multi-source data fusion method as described in claim 1, characterized in that, The step of fusing the target data block with the associated data block to obtain the fused data includes: Obtain the application scenario of the target data block, and determine the fusion type based on the application scenario; According to the fusion type, the target data block and the associated data block are fused to obtain the fused data.

7. A multi-source data fusion device, characterized in that, The multi-source data fusion device includes: The acquisition module is used to acquire multi-source data to be merged, perform structured processing on each source data, and obtain structured data in a unified format. The block segmentation module is used to segment each structured data into blocks based on each structured data, resulting in multiple data blocks corresponding to each structured data. The calculation module is used to calculate the similarity value between any target data block in any structured data and other corresponding data blocks in other structured data for any target data block, and to determine the associated data blocks that are related to the target data block based on the similarity value. The fusion module is used to fuse the target data block with the associated data block to obtain fused data.

8. The multi-source data fusion device as described in claim 7, characterized in that, The acquisition module includes: The determination unit is used to determine the data source type for each source data. The conversion unit is used to obtain the preset conversion rules corresponding to each data source type, and convert each source data into structured data in a unified format according to the data source type and the preset conversion rules.

9. A computer device, characterized in that, The computer device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-source data fusion method as described in any one of claims 1 to 6.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the multi-source data fusion method as described in any one of claims 1 to 6.

Images