A method and system for cross-industry data platform fusion based on semantic matching
By constructing a word order-structure joint vector through an industry-specific word segmentation dictionary and a multi-task joint pre-trained semantic coding network, the problem of word order and structure differences in field name matching in cross-industry data platforms is solved, and the accuracy and interpretability of cross-industry data fusion are achieved.
Patent Information
- Application Number
- CN202611627391.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-21
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods cannot effectively perceive the differences in the word order and combination structure of morphemes in field names in cross-industry data platforms, resulting in inaccurate semantic matching and an inability to identify the relationship between word order changes and semantic changes.
We adopt a cross-industry data platform fusion method based on semantic matching. We use industry-specific word segmentation dictionaries for word segmentation, construct word order-structure joint vectors, and use multi-task joint pre-trained semantic coding network and dependency syntax tree to calculate word order difference degree and structure difference index, and establish semantic mapping relationship between field names.
It achieves accurate quantification of word order changes and structural differences, improves the accuracy of cross-industry data fusion, solves the problem that existing methods cannot distinguish between semantically equivalent but different word orders, and reduces the uncertainty of black-box evaluation.
Smart Images

Figure CN123115258A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data fusion technology, and in particular to a method and system for cross-industry data platform fusion based on semantic matching. Background Technology
[0002] As enterprise informatization construction deepens, the coexistence of multi-source heterogeneous systems has become the norm. Cross-industry data fusion and data platform construction have become key links in enterprise digital transformation. During the data aggregation process, there are significant naming differences in field names from different data sources. The same business entity may be described using completely different field names in different systems. In cross-industry scenarios, the differences in terminology for the same business concept are even more significant. Although patient numbers in the medical industry and work order numbers in the manufacturing industry are both named as numbers, their business semantics belong to different domains. To achieve a unified view of cross-industry data, it is first necessary to establish semantic mapping relationships between field names from different data sources. Existing field mapping work usually adopts dictionary matching, rule mapping, or manual sorting methods, treating each morpheme in the field name as an independent lexical unit and determining the correspondence between field names by the degree of overlap at the lexical level.
[0003] However, when the above methods are applied to field name matching in cross-industry data platforms, they face the problem that the arrangement order and combination structure of morphemes within the field name cannot be effectively perceived. The semantics of a field name depends not only on the lexical meaning of the morphemes but also on their arrangement order and combination method. The business processes described have substantial differences. Existing methods based on the bag-of-words model treat each morpheme in the field name as an unordered set and cannot perceive semantic differences. Although methods based on edit distance can reflect some word order information, their measurement of word order changes lacks the ability to model the dependency and modification relationships between morphemes. When the arrangement order of each morpheme in the field name is exchanged or shifted, existing methods cannot distinguish between two essentially different situations, nor can they identify the potential semantic association between field names with different morphemes but similar structures. Therefore, how to achieve explicit perception and accurate quantification of differences in morpheme arrangement order and combination structure during field name matching, and establish the ability to distinguish the association between word order changes and semantic changes, is a technical problem that urgently needs to be solved in the field of cross-industry data platform integration. Summary of the Invention
[0004] This invention provides a method and system for cross-industry data platform fusion based on semantic matching to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides a cross-industry data platform fusion method based on semantic matching, comprising: S1: Based on the industry-specific word segmentation dictionary, perform Chinese word segmentation on the first and second field names to be matched in the cross-industry data platform to obtain the first word sequence and the second word sequence. S2: Perform multi-task joint pre-training on the first and second morpheme sequences to obtain the first word order-aware semantic vector and the second word order-aware semantic vector. Construct the first and second dependency syntax trees, fuse them to obtain the word order-structure joint vector, and evaluate the structural difference index of the two dependency syntax trees. S3: When the morpheme overlap rate between the first morpheme sequence and the second morpheme sequence exceeds a preset threshold, calculate the relative difference index of the common morpheme subsets in the first morpheme sequence and the second morpheme sequence, and perform cross-coupling calculation on the structural difference index and the relative difference index to obtain the word order difference degree. S4: Based on the word order difference, perform difference-aware correction on the similarity of the word order-structure joint vector, and establish a semantic mapping relationship between the first field name and the second field name to complete cross-industry data fusion.
[0006] To address the above problems, this invention also provides a cross-industry data platform fusion system based on semantic matching, the system comprising: The word segmentation module is used to perform Chinese word segmentation on the first and second field names to be matched in the cross-industry data platform according to the industry-specific word segmentation dictionary, so as to obtain the first word sequence and the second word sequence. The word order and structure awareness module is used to perform multi-task joint pre-training on the first and second word morpheme sequences to obtain the first and second word order awareness semantic vectors, construct the first and second dependency syntax trees, fuse them to obtain the word order-structure joint vector, and evaluate the structural difference index of the two dependency syntax trees. The word order difference quantification module is used to calculate the relative difference index of the common morpheme subsets in the first morpheme sequence and the second morpheme sequence when the morpheme overlap rate between the first morpheme sequence and the second morpheme sequence exceeds a preset threshold. The structural difference index and the relative difference index are cross-coupled and calculated to obtain the word order difference degree. The difference correction fusion module is used to perform difference-aware correction on the similarity of word order-structure joint vectors based on word order difference, and to establish a semantic mapping relationship between the first field name and the second field name to complete cross-industry data fusion.
[0007] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention enables the semantic encoding network to simultaneously acquire semantic invariant encoding capabilities and structural recombination sensitivity during training by jointly pre-training a word order invariant comparison task and a perturbation type hierarchical discrimination task. The word order invariant comparison task forces the encoder to output semantically similar embedding vectors for semantically equivalent field names with different word orders through contrastive learning. The perturbation type hierarchical discrimination task forces the encoder's intermediate hidden layers to retain structural sensitivity features that can distinguish between lightweight and heavyweight perturbations. The two tasks share the same set of intermediate hidden layer parameters of the encoder, and through gradient weighted superposition, these parameters are forced to optimize in two directions simultaneously: towards word order invariance to ignore surface order adjustments that do not affect semantics, and towards structural sensitivity to retain the ability to judge whether structural recombination changes semantics. This mechanism enables the encoder to accurately identify "convolutional neural network" and "neural network convolution" as semantically equivalent with only different word order, and to accurately identify the essential difference between "from Beijing to Shanghai" and "from Shanghai to Beijing" as semantically different due to different word order, thus solving the deficiency of existing methods in not being able to distinguish between the two cases.
[0008] 2. This invention integrates the relative word order difference index and the structural difference index through a cross-coupling calculation mechanism. The relative word order difference index quantifies the degree of change in the order of common morphemes in field names by distance, while the structural difference index quantifies the degree of change in the deep structure of the dependency syntax tree by the root node consistency coefficient and the subtree isomorphism ratio. Both describe the semantic impact of word order changes from different dimensions. Based on this, the structural difference index is used as the confidence weight of the relative word order difference index for cross-coupling. When the structural difference index exceeds a predetermined threshold, it is determined that the deep structural reorganization has reflected semantic differences, and the contribution of the relative word order difference is reduced to avoid double counting. When the structural difference index is below the predetermined threshold, it is determined that it is only a surface order adjustment, and its contribution is increased to emphasize the potential impact of the order change on semantics. This mechanism enables the calculation result of word order difference to simultaneously reflect the dual information of surface magnitude and deep structural impact, and provides a clear quantitative basis for difference perception correction, solving the problem of black box and uninterpretable evaluation of word order difference in existing methods. Attached Figure Description
[0009] Figure 1 A flowchart illustrating a cross-industry data platform fusion method based on semantic matching, provided in an embodiment of the present invention; Figure 2 A functional block diagram of a cross-industry data platform fusion system based on semantic matching, provided as an embodiment of the present invention.
[0010] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0011] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0012] This application provides a method for cross-industry data platform fusion based on semantic matching. The execution subject of the method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for cross-industry data platform fusion based on semantic matching can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.
[0013] Reference Figure 1 The diagram shown is a flowchart illustrating a cross-industry data platform fusion method based on semantic matching according to an embodiment of the present invention. In this embodiment, the cross-industry data platform fusion method based on semantic matching includes: S1: Based on the industry-specific word segmentation dictionary, perform Chinese word segmentation on the first and second field names to be matched in the cross-industry data platform to obtain the first word sequence and the second word sequence. In this embodiment of the invention, the field name samples of each data source in the cross-industry data platform are traversed character by character, and the co-occurrence frequency of each adjacent character combination is counted to obtain a candidate word set. It should be specifically noted that the cross-industry data platform stores multiple field names from different data sources. Chinese word segmentation of the field names requires the use of an industry-specific word segmentation dictionary. When constructing this dictionary, field name samples from each data source in the cross-industry data platform are first obtained. The number of field name samples should cover the major industry fields involved in the platform, such as manufacturing, medical, financial, and retail. The obtained field name samples are traversed character by character, and the frequency of any two adjacent character combinations in the field name samples is counted, i.e., co-occurrence frequency. The co-occurrence frequency reflects the probability that the adjacent character combination appears as a complete morpheme in the field name. The higher the co-occurrence frequency, the more likely the adjacent character combination is to be an independent industry term or business term. Adjacent character combinations with a co-occurrence frequency exceeding a preset frequency threshold are selected as candidate morphemes, and all candidate morphemes are stored in a candidate morpheme set.
[0014] The frequency of each morpheme in the candidate morpheme set is counted in various industry fields of the cross-industry data platform. The industry coverage weight is evaluated based on the frequency of occurrence. Candidate morphemes whose industry coverage weight exceeds a preset threshold are regarded as industry core morphemes and stored in the industry-specific dictionary. It should be noted that the frequency of the candidate morpheme in the manufacturing field is [number missing]. The frequency of occurrence in the medical industry field is The frequency of occurrence in the financial industry field is And so on. Based on the frequency distribution of the candidate word element in various industry sectors, its industry coverage weight is calculated. The industry coverage weight can be calculated in the following ways: after normalizing the frequency of the candidate word element in each industry sector, the number of industry sectors whose frequency exceeds a preset activity threshold is counted, and this number is used as the industry coverage weight; or, the variance or information entropy of the frequency of the candidate word element in each industry sector is calculated to reflect its breadth and balance of distribution across industries. The higher the industry coverage weight, the more active business meaning the candidate word element has in more industry sectors, and the more cross-industry universality and representativeness it has. Candidate word elements whose industry coverage weight exceeds a preset weight threshold are selected as industry core word elements, and the selected industry core word elements are stored in an industry-specific word segmentation dictionary. The industry-specific word segmentation dictionary constructed in the above way can effectively cover the core business terms of various industry sectors in the cross-industry data platform, providing an accurate dictionary foundation for subsequent field name word segmentation.
[0015] Using industry core morphemes as the longest matching benchmark, starting from the starting character position of the first field name or the second field name, the industry core morphemes of each industry-specific word segmentation dictionary are matched sequentially. When a match is successful, the corresponding industry core word is extracted as a word unit, and the matching continues from the next character position until all characters are traversed, completing the word segmentation process and obtaining the first word sequence and the second word sequence.
[0016] It needs to be specifically explained that, taking the segmentation of the first field name as an example, the industry core morphemes stored in the industry-specific segmentation dictionary are used as the longest matching benchmark. Starting from the starting character position of the first field name, the system attempts to match each industry core morpheme in the industry-specific segmentation dictionary in turn. Specifically, in each match, the longest character sequence that can be formed from the current starting position is taken and compared with each industry core morpheme in the industry-specific segmentation dictionary. If a match is successful, the longest successfully matched industry core morpheme is taken as an independent morpheme unit and truncated. After truncation, the matching continues from the next character position after the end position of the morpheme unit until all characters of the first field name are traversed. During the matching process, if no industry core morpheme can be matched at a certain character position, the single character is treated as an independent morpheme unit to ensure the integrity of the segmentation result. In this way, the segmentation of the first field name is completed, and the first morpheme sequence composed of multiple morpheme units arranged in the original order is obtained.
[0017] Using the exact same method, the second field name is segmented using an industry-specific word segmentation dictionary to obtain the second word sequence. The first and second word sequences each retain the original arrangement order of each word in the corresponding field name, providing basic data for the perception and quantification of word order differences in subsequent steps.
[0018] S2: Perform multi-task joint pre-training on the first and second morpheme sequences to obtain the first word order-aware semantic vector and the second word order-aware semantic vector. Construct the first and second dependency syntax trees, fuse them to obtain the word order-structure joint vector, and evaluate the structural difference index of the two dependency syntax trees. In this embodiment of the invention, the morpheme sequence is input into a pre-constructed semantic coding network, and each morpheme in the morpheme sequence is mapped to an initial embedding vector to obtain an initial embedding vector sequence. It should be specifically noted that the semantic encoding network adopts a deep neural network model based on the Transformer architecture. It takes a sequence of morphemes as input and the semantic embedding vectors corresponding to each morpheme as output. The first sequence of morphemes is input into the semantic encoding network. The semantic encoding network maps each morpheme in the first sequence of morphemes to an initial embedding vector of fixed dimension through its embedding layer, resulting in a first initial embedding vector sequence composed of multiple initial embedding vectors arranged in the original order of the morphemes in the first sequence of morphemes. Similarly, the second sequence of morphemes is input into the semantic encoding network, and each morpheme in the second sequence of morphemes is mapped to an initial embedding vector through the same embedding layer, resulting in a second initial embedding vector sequence. The dimension of each initial embedding vector is a preset fixed value. This value can be set by those skilled in the art according to the model size and computing resources. For example, it can be 768 dimensions or 1024 dimensions.
[0019] The absolute position of each vector in the initial embedded vector sequence is encoded to obtain the order encoding information. The order encoding information is then superimposed on each initial embedded vector to obtain the intermediate encoded vector sequence. It should be specifically noted that because the semantic encoding network of the Transformer architecture lacks a natural ability to perceive sequence order, positional encoding is required for each initial embedding vector to introduce the order information of morphemes in the sequence. Specifically, absolute positional encoding is performed on the arrangement of each initial embedding vector in the first initial embedding vector sequence. The absolute positional encoding uses a sine-cosine positional encoding function, where the position index is... Dimension index is The encoded components, whose encoded values are based on The absolute position encoding vector is calculated using the sine or cosine function. This vector is then superimposed onto the initial embedding vector at the corresponding position to obtain the first intermediate encoding vector sequence carrying the absolute order information of the morphemes. Similarly, the absolute position encoding is performed on the arrangement of each initial embedding vector in the second initial embedding vector sequence, and the encoding vector is superimposed onto the corresponding initial embedding vector to obtain the second intermediate encoding vector sequence. By superimposing absolute position encoding, the semantic encoding network can distinguish the same morphemes at different positions in the subsequent encoding process.
[0020] The word order sequence and its corresponding word order variant sequences are input into the semantic encoding network, and multi-task joint pre-training is performed to obtain the first word order-aware semantic vector and the second word order-aware semantic vector, including: It should be specifically noted that before pre-training, a word order variant library needs to be obtained. The word order variant library is pre-built and stores multiple word order variant sequences corresponding to the first and second word morpheme sequences, as well as perturbation type labels and perturbation degree values corresponding to each word order variant sequence. Each word order variant sequence corresponding to the first word morpheme sequence and each word order variant sequence corresponding to the second word morpheme sequence are extracted from the word order variant library. Each extracted word order variant sequence is then input into the semantic encoding network to perform multi-task joint pre-training. In each batch of pre-training, each word order variant sequence and its corresponding perturbation type labels and perturbation degree values are loaded simultaneously.
[0021] The perturbation level value corresponding to each word order variant sequence is used as the regression target for perturbation level regression training; It is necessary to explain the perturbation degree regression training in detail. After each word order variant sequence is input into the semantic encoding network, the semantic encoding network outputs the corresponding embedding vector. Then, the embedding vector is input into a regression prediction layer. The perturbation degree value corresponding to the variant sequence is used as the regression target. The embedding vector output by the semantic encoding network is trained to carry information related to the magnitude of word order change. The perturbation degree value is a value calculated based on the edit distance or Kendall-Tau distance of the word order variant sequence relative to the original word sequence. The larger the value, the greater the magnitude of word order change.
[0022] The morpheme sequence and its corresponding reverse morpheme sequence are used to form positive and negative sample pairs, and positive and reverse order discrimination training is performed. It needs to be specifically explained that the positive and negative order discrimination training involves forming positive and negative sample pairs by combining the original word sequence with the reversed word sequence of the same field name. The original word sequence is used as the positive sample and the reversed word sequence is used as the negative sample. The positive and negative sample pairs are input into the semantic encoding network, and the embedding vector output by the semantic encoding network is trained through a binary classification discriminant layer. The embedding vector carries the discrimination information of the absolute order of the word morphemes, enabling the semantic encoding network to distinguish whether the input sequence is in the original order.
[0023] Train on word order invariance by comparing semantically equivalent field names with semantically different field names; Specifically, the word order invariant contrastive training involves selecting semantically equivalent but differently ordered field name pairs from a word order variant library as positive samples. For example, "convolutional neural network" and "neural network convolution" form a positive sample pair. Conversely, field name pairs with different word orders and semantics are selected as negative samples. For example, "from Beijing to Shanghai" and "from Shanghai to Beijing" form a negative sample pair. The semantic encoding network is trained through contrastive learning, enabling it to output semantically similar embedding vectors for semantically equivalent field name pairs with different word orders, and semantically dissimilar embedding vectors for semantically dissimilar field name pairs with different word orders and semantics. The loss function for this contrastive training adopts the InfoNCE loss function. For each anchor sample, the encoding vector of its positive samples should have a high cosine similarity to the encoding vector of the anchor sample, while the encoding vectors of other negative samples within the batch should have a low cosine similarity to the encoding vector of the anchor sample.
[0024] For both lightly perturbated and heavily perturbated sequence variants, perturbation type hierarchical discrimination training is performed. Specifically, the perturbation type hierarchical discrimination training involves inputting the word order variant sequences obtained through lightweight perturbation and those obtained through heavyweight perturbation into the semantic encoding network. A hierarchical classifier is used to train the semantic encoding network, enabling the output embedding vectors to carry hierarchical discriminative features that can distinguish between surface word order changes and deep structural reorganizations. Lightweight perturbations include swapping adjacent morphemes and interchanging the positions of any two morphemes, while heavyweight perturbations include overall subtree shifting based on dependency syntactic relations. The hierarchical classifier employs a multilayer perceptron structure, and its output is a category label representing the perturbation type hierarchy. This training enables the semantic encoding network to acquire the ability to distinguish between surface word order adjustments and deep structural reorganizations.
[0025] The word order invariance contrast training and perturbation type hierarchical discrimination training share the intermediate hidden layer parameters of the semantic coding network; It should be specifically noted that the feature vector output by the intermediate hidden layer of the semantic encoding network serves as the input feature for both word order invariance comparison training and perturbation type hierarchical discrimination training. The backpropagation gradients of the two tasks are superimposed at the intermediate hidden layer parameters, forcing the same set of intermediate hidden layer parameters to be optimized in two directions simultaneously: towards word order invariance so that the encoder ignores surface word order changes that do not affect semantics, and towards hierarchical discrimination so that the encoder retains structurally sensitive features that can distinguish between lightweight perturbations and heavyweight perturbations. Through this training method with shared parameters, the semantic encoding network can ultimately distinguish between two fundamentally different situations: "semantic equivalence with only different word order" and "different word order leading to different semantics".
[0026] The embedding vectors output from perturbation degree regression training, forward and reverse order discrimination training, word order invariance comparison training, and perturbation type hierarchical discrimination training are fused to obtain the first word order-aware semantic vector and the second word order-aware semantic vector.
[0027] It should be specifically noted that the vector fusion can be achieved through concatenation, where the embedding vectors output by the four tasks are concatenated along their vector dimensions to form a higher-dimensional fused vector; or through weighted summation, where weight coefficients are assigned to the embedding vectors output by the four tasks, and then the vectors are summed. The fused vector is used as the first and second word order-aware semantic vectors. Through the aforementioned multi-task joint pre-training, the word order-aware semantic vector output by the semantic encoding network simultaneously carries information on the magnitude of word order changes, absolute morpheme order discrimination information, word order invariance information under semantic equivalence conditions, and perturbation type level discrimination information, providing rich word order-aware semantic features for subsequent field name matching.
[0028] Based on the part of speech of each morpheme in the morpheme sequence, determine the dependency relationship type between any two morphemes, establish a directed dependency arc, and combine each morpheme node and the directed dependency arc into a dependency syntax tree; It needs to be specifically explained that the construction of the first dependency syntax tree and the second dependency syntax tree specifically includes: performing part-of-speech tagging on each morpheme in the first morpheme sequence to obtain the first part-of-speech tagging sequence. The part-of-speech tagging can adopt a deep learning-based part-of-speech tagging model, such as a sequence tagging model using a bidirectional long short-term memory network combined with a conditional random field, to tag each morpheme with its part-of-speech category, such as noun, verb, adjective, adverb, etc. Using each morpheme in the first morpheme sequence as a node and the part-of-speech tagging of each morpheme in the first part-of-speech tagging sequence as the basis, a transition-based dependency parsing method or a graph-based dependency parsing method is used to determine the dependency relationship type between any two morphemes in the first morpheme sequence. The dependency relationship types include subject-predicate, verb-object, attributive-head, adverbial-head, and coordinate relationships. Based on the determined dependency relationship type, a directed dependency arc is established between two morpheme nodes with a dependency relationship. The directed dependency arc points from the core word to the dependent word, and the direction indicates the semantic relationship of modification or being modified. Each morpheme node and the directed dependency arc between them are combined to form the first dependency parsing tree. Using the same method, each morpheme in the second morpheme sequence is tagged with part of speech to determine the dependency relationship type between each morpheme and establish a directed dependency arc. Each morpheme node and the directed dependency arc are combined to form the second dependency parsing tree.
[0029] Using the directed dependency arcs in the dependency syntax tree as the edge structure and the corresponding morpheme semantic embedding vectors in the word order semantic perception vector as the initial features of the nodes, the neighbor node information aggregation operation is performed on each morpheme node to obtain the word order-structure joint vector. It needs to be specifically explained that in the graph attention mechanism, for each central morpheme node, the attention coefficient between it and each neighboring morpheme node is calculated. When calculating the attention coefficient, the dependency relationship type of the directed dependency arc connecting the central morpheme node and the neighboring morpheme nodes is used as a relation bias term and injected into the calculation of the attention coefficient. Specifically, the dependency relationship type is encoded as a relation embedding vector. This relation embedding vector, along with the feature vectors of the central morpheme node and the feature vectors of the neighboring morpheme nodes, is input into the attention coefficient calculation function, so that neighboring nodes with different dependency relationship types obtain different relation-aware attention weights. In this way, the central morpheme node obtains different relation-aware attention weights when aggregating neighboring node information with different dependency relationship types, thereby retaining the dependency relationship type information in the aggregation result. The updated feature vectors of each morpheme node obtained after information aggregation are combined into a first word order-structure joint vector according to the topological structure of the first dependency syntax tree. The second word order-aware semantic vector and the second dependency syntax tree are processed in the same way to obtain the second word order-structure joint vector.
[0030] The structural difference index is obtained by weighting and fusing the root node consistency coefficients of the two dependency syntax trees with the subtree isomorphism ratio.
[0031] It needs to be specifically explained that when evaluating the structural difference index of two dependency syntax trees, the root node morpheme and the main dependency paths from the root node morpheme to each leaf node are first extracted from the first dependency syntax tree. The root node morpheme is a morpheme node in the dependency syntax tree without any incoming edges, that is, a core word that is not modified by any other morpheme. The main dependency path is the sequence of nodes traversed from the root node morpheme to each leaf node morpheme. Similarly, the root node morpheme and each main dependency path are extracted from the second dependency syntax tree. The root node morpheme of the first dependency syntax tree and the second dependency syntax tree are compared to see if they are consistent. If they are consistent, the root node consistency coefficient is set to 1; if they are inconsistent, it is set to 0. The similarity ratio of node sequences on each main dependency path in the first dependency syntax tree and the second dependency syntax tree is compared. For a pair of main dependency paths, starting from the starting node of the path, the morphemes at corresponding positions are compared to see if they are the same. The proportion of the number of identical morphemes to the total number of nodes in the path is counted to obtain the value of the root node consistency coefficient. For the path isomorphism ratio, each backbone dependency path in the first dependency syntax tree is paired with each backbone dependency path in the second dependency syntax tree, and the path isomorphism ratio of each pair of paths is calculated. The average of all path isomorphism ratios is used as the subtree isomorphism ratio. The root node consistency coefficient and the subtree isomorphism ratio are weighted and fused. For example, the root node consistency coefficient is assigned a first weight and the subtree isomorphism ratio is assigned a second weight. The weighted sum is calculated as the structural difference index. In the above way, the structural difference index comprehensively reflects the degree of structural difference between the two dependency syntax trees from the two dimensions of root node consistency and subtree structural similarity.
[0032] S3: When the morpheme overlap rate between the first morpheme sequence and the second morpheme sequence exceeds a preset threshold, calculate the relative difference index of the common morpheme subsets in the first morpheme sequence and the second morpheme sequence, and perform cross-coupling calculation on the structural difference index and the relative difference index to obtain the word order difference degree. In this embodiment of the invention, common morphemes with the same content in the first morpheme sequence and the second morpheme sequence are sorted according to their arrangement order in the two morpheme sequences to obtain a first sequence sequence and a second sequence sequence. Specifically, the process involves extracting all morphemes from the first morpheme sequence to form the first morpheme set, and extracting all morphemes from the second morpheme sequence to form the second morpheme set. The intersection of the first and second morpheme sets is calculated, representing the number of morphemes with identical content in both sets. The union of the first and second morpheme sets is also calculated, representing the total number of unique morphemes in both sets. Dividing the intersection by the union yields the morpheme overlap rate, which reflects the degree of lexical overlap between the two field names. A higher overlap rate indicates more shared morphemes and greater comparability. The calculated overlap rate is compared to a preset threshold. If the overlap rate exceeds the preset threshold, the two field names are considered to have sufficient lexical overlap, and word order differences can be further calculated. If the overlap rate does not exceed the preset threshold, the lexical overlap between the two field names is insufficient, and further word order difference calculations are unnecessary. The preset threshold can be set by those skilled in the art based on actual application scenarios and accuracy requirements.
[0033] In one feasible implementation, the preset threshold for morpheme overlap rate is set to 0.3. That is, when the morpheme overlap rate between the first morpheme sequence and the second morpheme sequence is less than 0.3, it is determined that the lexical overlap of the two field names is too low, and there are not enough common morphemes to provide for the subsequent calculation of word order differences. At this time, S3 and subsequent matching steps are skipped directly, and the matching score is configured to the lowest value. The preset threshold of 0.3 is based on the fact that the expression of field names in different industries in cross-industry data platforms is quite different. For example, "patient number" in the medical industry and "work order number" in the manufacturing industry only overlap in the word "number", with an overlap rate of 0.33, which is slightly higher than the threshold. At this time, the number of common morphemes is 1, which still meets the basic conditions for calculating word order differences. However, when the overlap rate is less than 0.3, the proportion of common morphemes in the two field names is less than one-third. The comparison information at the morpheme level is insufficient to support the reliable quantification of subsequent differences. Rash calculation will lead to unreliable matching results. Therefore, using 0.3 as the critical value to distinguish between "comparable" and "incomparable" can reduce unnecessary calculation overhead while ensuring the validity of the calculation.
[0034] The first sequence and the second sequence contain the exact same set of morphemes, but the arrangement of each morpheme in the two sequences may be different. Starting from the first morpheme of the first sequence, each morpheme in the first sequence is taken as the target morpheme in turn. The arrangement position of the target morpheme in the second sequence is located, and the arrangement position of the target morpheme in the first sequence and the arrangement position in the second sequence are recorded to obtain the position correspondence of the target morpheme. The above positioning operation is performed on each morpheme in the first sequence to obtain the set of position correspondences of all common morphemes.
[0035] Starting from the first morpheme of the first sequence, each morpheme is taken as the target morpheme in turn, and the position of the target morpheme is located in the second sequence to obtain the positional correspondence of the target morphemes; Based on the positional correspondence, starting from the first morpheme of the second sequence, the proportion of morpheme pairs whose relative order of any two morphemes in the two sequences is inconsistent is used as the relative difference index.
[0036] It needs to be specifically explained that, starting from the first morpheme of the first sequential sequence, any two morphemes in the first sequential sequence are sequentially taken as a morpheme pair. For each morpheme pair, its first position in the first sequential sequence and its second position in the second sequential sequence are obtained. It is then determined whether the relative order of the two morphemes in the first sequential sequence is consistent with their relative order in the second sequential sequence: if the order is the same, the relative order is considered consistent; if it is reversed, they are considered inconsistent. The number of morpheme pairs with inconsistent relative orders and the total number of morpheme pairs requiring relative order comparison are counted. The ratio of the number of inconsistent morpheme pairs to the total number of morpheme pairs is used as the relative difference index. This index reflects the degree of difference in the order of the same morpheme set in the two field names; a higher index indicates a greater difference in word order between the two field names.
[0037] The structural difference index is used as a confidence weight to adjust the contribution ratio of the relative difference index; It should be specifically noted that the structural difference index reflects the degree of deep structural difference between the dependency syntactic trees of the two field names at the root node and subtree structure levels, while the relative difference index reflects the degree of difference in the surface arrangement order of common morphemes in the two field names. The structural difference index is used as a confidence weight to adjust the contribution ratio of the relative difference index, and the relative difference index and the structural difference index are cross-coupled and calculated.
[0038] When the structural difference index exceeds a predetermined threshold, the contribution ratio of the relative difference index is reduced; when the structural difference index is below a predetermined threshold, the contribution ratio of the relative difference index is increased. It should be specifically explained that the predetermined threshold is used to distinguish the main types of word order changes: when the structural difference index is low, it indicates that the dependency syntactic tree structures of the two field names are basically the same, and the word order change mainly manifests as the adjustment of the order of surface morphemes; when the structural difference index is high, it indicates that there are significant differences in the dependency syntactic tree structures of the two field names, and the word order change involves deep syntactic restructuring. Based on the comparison results, the contribution ratio of the relative difference index in the word order difference calculation is dynamically adjusted. When the structural difference index is lower than the predetermined threshold, it is determined that the word order difference between the first field name and the second field name is only an adjustment of the order of surface morphemes. At this time, the relative difference index is increased. The contribution weight of the difference index in the word order difference calculation is used to emphasize the potential impact of surface order changes on semantics. When the structural difference index exceeds a predetermined threshold, it is determined that the word order difference between the first field name and the second field name is mainly caused by the deep structural reorganization of the dependency syntax tree. At this time, the contribution weight of the relative difference index in the word order difference calculation is reduced to avoid the semantic differences already reflected by the deep structural reorganization being repeatedly included in the relative difference index. When the structural difference index is equal to the predetermined threshold, it is determined that the surface morpheme order adjustment and the deep structural reorganization are in a critical balance state. At this time, the original contribution weight of the relative difference index is maintained and no adjustment is made.
[0039] In one feasible implementation, a predetermined threshold for the structural difference index is set to 0.4. That is, when the structural difference index is below 0.4, the word order difference is determined to be merely a surface morpheme order adjustment, increasing the contribution weight of the relative difference index; when the structural difference index exceeds 0.4, the word order difference is determined to be mainly caused by deep structural reorganization, reducing the contribution weight of the relative difference index. The predetermined threshold of 0.4 is based on the following: the structural difference index is obtained by weighted fusion of the root node consistency coefficient (0.4) and the subtree isomorphism ratio (0.6). When the root nodes of two dependency syntactic trees are inconsistent, the root node consistency coefficient is 0, and this contribution is 0. In this case, the subtree isomorphism ratio needs to exceed 0.67 for the structural difference index to exceed 0.4. When the root nodes are consistent, the root node consistency coefficient is 0.4, requiring only a subtree isomorphism ratio of over 0.67. A subtree isomorphism ratio slightly above 0 is sufficient to cause the structural difference index to exceed 0.4. This design makes the consistency of the root node the primary indicator for determining the nature of structural reorganization. When the core words of two field names change, such as "customer name" and "name customer" both having "name" as the root node, the structural difference is mainly contributed by the surface order adjustment reflected by the subtree isomorphism ratio. However, when the root node changes, such as "delivery date" and "delivery date" with the root node "date" for the former and "delivery date" for the latter, the structural difference index will inevitably exceed 0.4, indicating deep structural reorganization. This accurately identifies the critical situation where word order changes lead to changes in core semantics. Using 0.4 as a threshold can effectively distinguish between word order differences of two different natures in field name matching: "order adjustment only" and "change in core semantics".
[0040] The relative difference index after adjusting the contribution ratio is weighted and fused with the structural difference index to obtain the word order difference degree.
[0041] Specifically, a first fusion weight is assigned to the relative difference index after adjusting the contribution ratio, and a second fusion weight is assigned to the structural difference index. The product of the relative difference index after adjusting the contribution ratio and the first fusion weight, and the product of the structural difference index and the second fusion weight are calculated. The two products are added together, and the result is used as the word order difference degree. The first fusion weight and the second fusion weight can be fixed values set in advance, or they can be dynamically adjusted according to the actual application scenario. Through the above cross-coupling calculation, the word order difference degree simultaneously integrates surface word order change information and deep structural reorganization information. It reflects both the degree of difference between the two field names in terms of morpheme arrangement and the degree of difference between the two field names at the syntactic structure level, providing a comprehensive quantitative basis for subsequent difference perception correction.
[0042] S4: Based on the word order difference, perform difference-aware correction on the similarity of the word order-structure joint vector, and establish a semantic mapping relationship between the first field name and the second field name to complete cross-industry data fusion.
[0043] In this embodiment of the invention, the vector similarity between two word order-structure joint vectors is used as the basic semantic matching score; It should be specifically noted that the vector similarity can be calculated using the cosine similarity method, that is, by calculating the cosine of the angle between two vectors. The closer this value is to 1, the more consistent the directions of the two vectors are and the more semantically similar they are. By calculating the cosine similarity between two word order-structure joint vectors, a basic semantic matching score reflecting the comprehensive semantic similarity of two field names after integrating word order and structural information can be obtained.
[0044] Obtain the morpheme overlap rate and word order difference between the first and second morpheme sequences; When the morpheme overlap rate is higher than the preset overlap qualification threshold, the word order difference is compared with the preset low-order difference threshold and high-order difference threshold, and the low-order difference threshold is less than the high-order difference threshold. It should be specifically noted that the morpheme overlap rate reflects the degree of overlap between the two field names at the lexical level, while the word order difference reflects the comprehensive difference between the two field names at the morpheme arrangement and syntactic structure levels. The morpheme overlap rate is compared with a preset overlap threshold. The overlap threshold is used to determine whether the two field names have sufficient lexical overlap for difference perception correction. When the morpheme overlap rate is higher than the overlap threshold, it indicates that the two field names share a sufficient number of morphemes, which meets the basic conditions for determining word order difference. At this time, the word order difference is further compared with the preset low-order difference threshold and high-order difference threshold. The low-order difference threshold is smaller than the high-order difference threshold. The low-order difference threshold is used to determine whether the word order difference is small enough that the two field names can be considered semantically equivalent with only a difference in word order. The high-order difference threshold is used to determine whether the word order difference is large enough that the difference in word order leads to semantic difference.
[0045] In one feasible implementation, the overlap qualification threshold is set to 0.2, the low-order difference threshold is set to 0.25, and the high-order difference threshold is set to 0.65. The preset threshold of 0.3 is used to filter field name pairs that meet the basic conditions for calculating word order differences, ensuring that the number of shared morphemes is sufficient to support the reliable quantification of the relative difference index. The overlap qualification threshold of 0.2 is used to determine whether difference perception correction is needed. When the overlap rate is between 0.2 and 0.3, although it is insufficient to perform quantitative calculation of word order differences, the correction step can still fine-tune the matching score based on the limited information of morpheme overlap rate combined with word order difference degree to improve the adaptability of the method in the low overlap rate range.
[0046] The word order difference thresholds of 0.25 (low-order) and 0.65 (high-order) divide the word order difference into three intervals. When the word order difference is below 0.25, it is judged as "semantically equivalent with only different word order" and a positive gain is applied. When the word order difference is above 0.65, it is judged as "different word order leads to different semantics" and a negative penalty is applied. When the word order difference is between 0.25 and 0.65, it is considered a critical state and no correction is applied. The values of 0.25 and 0.65 are based on the fact that the word order difference is weighted equally by the relative difference index and the structural difference index after adjusting the contribution ratio. The results show that when the word order difference is below 0.25, it means that both indices are at a low level and at least one index is significantly low. At this time, the word order difference is minimal, and the positive gain can effectively improve the matching score of semantically equivalent fields. When the word order difference is above 0.65, it means that at least one index is at a high level and the word order difference is significant. The negative penalty can effectively suppress mismatches of semantically different fields. The middle range of 0.25 to 0.65 corresponds to the situation where the word order difference is in an ambiguous zone. At this time, no correction is applied to avoid over-adjustment and the introduction of new errors.
[0047] When the word order difference is lower than the low-order difference threshold, a positive gain is applied to the basic semantic matching score to obtain the corrected semantic matching score. It should be specifically noted that when the word order difference is lower than the low-order difference threshold, the first field name and the second field name are determined to be a pair of field names that are "semantically equivalent but differ only in word order". In this case, although the word order of the two field names is different, their semantics are actually equivalent. For example, "convolutional neural network" and "neural network convolution" belong to this category. Since the basic semantic matching score is calculated based on the word order-structure joint vector, the existence of word order difference may lead to the underestimation of the basic semantic matching score. Therefore, it is necessary to apply a positive gain to the basic semantic matching score. The positive gain is achieved by superimposing a preset positive correction value on the basic semantic matching score. The score after superimposing the positive correction value is used as the corrected semantic matching score to correct the underestimation of the matching score caused by word order difference.
[0048] When the word order difference is higher than the high-order difference threshold, a negative penalty is applied to the basic semantic matching score to obtain the corrected semantic matching score.
[0049] It should be specifically noted that when the word order difference is higher than the high-order difference threshold, the first field name and the second field name are determined to be a pair of field names where "different word order leads to different semantics". In this case, the difference in the morpheme arrangement order of the two field names leads to a substantial change in semantics. For example, "from Beijing to Shanghai" and "from Shanghai to Beijing" belong to this category. Since the basic semantic matching score is calculated based on the word order-structure joint vector, the semantic change caused by word order difference may cause the basic semantic matching score to be overestimated. Therefore, it is necessary to apply a negative penalty to the basic semantic matching score. The negative penalty is achieved by deducting a preset negative correction value from the basic semantic matching score. The score after deducting the negative correction value is used as the corrected semantic matching score to correct the overestimation of the matching score caused by word order difference.
[0050] When the word order difference is between the low-order difference threshold and the high-order difference threshold, the word order difference is determined to be in a critical state. At this time, no positive gain or negative penalty is applied to the basic semantic matching score, and the basic semantic matching score is directly used as the corrected semantic matching score.
[0051] When the corrected semantic matching score exceeds the preset mapping establishment threshold, the first field name and the second field name are marked as a matching field name pair; It should be specifically noted that the mapping establishment threshold is used to determine whether the semantic similarity between two field names is sufficient to establish a semantic mapping relationship. When the final semantic matching score exceeds the mapping establishment threshold, it is determined that the first field name and the second field name have sufficient semantic similarity, and the first field name and the second field name are marked as a matching field name pair.
[0052] Based on the data type of the first and second fields in the matching field name pairs, determine the data type conversion rules between the matching field name pairs; It should be specifically noted that the field data types include, but are not limited to, string types, integer types, floating-point types, date and time types, and boolean types. The data type conversion rules include forward conversion methods from the first field data type to the second field data type, and reverse conversion methods from the second field data type to the first field data type. For example, when the first field data type is string type and the second field data type is integer type, the data type conversion rules may include parsing rules for converting strings to integers and formatting rules for converting integers to strings.
[0053] The semantic mapping relationship between the first field name and the second field name is used as the mapping connection, and the data type conversion rules are used as the basis for data conversion to establish a two-way data mapping link. It should be noted that the bidirectional data mapping link includes a first mapping direction from the first field name to the second field name and a second mapping direction from the second field name to the first field name. Both directions are configured with corresponding data type conversion rules.
[0054] Based on the bidirectional data mapping link, the data record corresponding to the first field name is converted according to the data type conversion rules and then written to the storage location corresponding to the second field name, thus completing the cross-industry data fusion.
[0055] It should be specifically explained that the data record corresponding to the second field name is converted according to the data type conversion rules and then written to the storage location corresponding to the first field name. Through the above bidirectional data mapping and data conversion operations, data interoperability and integration between the first field name and the second field name at the field level are realized, generating a unified cross-industry data view and completing cross-industry data integration.
[0056] like Figure 2 The diagram shown is a functional block diagram of a cross-industry data platform fusion system based on semantic matching provided in an embodiment of the present invention.
[0057] The semantic matching-based cross-industry data platform fusion system 100 of the present invention can be installed in an electronic device. Depending on the functions implemented, the semantic matching-based cross-industry data platform fusion system 100 may include a word segmentation processing module 101, a word order structure perception module 102, a word order difference quantification module 103, and a difference correction fusion module 104. The module of the present invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.
[0058] In this embodiment, the functions of each module / unit are as follows: The word segmentation processing module 101 is used to perform Chinese word segmentation processing on the first field name and the second field name to be matched in the cross-industry data platform according to the industry-specific word segmentation dictionary, so as to obtain the first word sequence and the second word sequence. The word order and structure awareness module 102 is used to perform multi-task joint pre-training on the first morpheme sequence and the second morpheme sequence to obtain the first word order awareness semantic vector and the second word order awareness semantic vector, construct the first dependency syntax tree and the second dependency syntax tree, fuse them to obtain the word order-structure joint vector, and evaluate the structural difference index of the two dependency syntax trees. The word order difference quantification module 103 is used to calculate the relative difference index of the common subset of morphemes in the first morpheme sequence and the second morpheme sequence when the morpheme overlap rate between the first morpheme sequence and the second morpheme sequence exceeds a preset threshold, and to perform cross-coupling calculation of the structural difference index and the relative difference index to obtain the word order difference degree. The difference correction fusion module 104 is used to perform difference-aware correction on the similarity of word order-structure joint vectors based on word order difference degree, and to establish a semantic mapping relationship between the first field name and the second field name to complete cross-industry data fusion.
[0059] In the several embodiments provided by the present invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of modules is merely a logical functional division, and there may be other division methods in actual implementation.
[0060] The modules described as separate components may or may not be physically separate. The components shown as modules 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 may be selected to achieve the purpose of this embodiment according to actual needs.
[0061] In addition, the functional modules in the various embodiments of the present invention 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 in the form of hardware plus software functional modules.
[0062] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.
[0063] The embodiments of this application can acquire and process relevant data based on an artificial intelligence technology. Artificial intelligence is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.
[0064] Finally, it should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.
Claims
1. A cross-industry data platform fusion method based on semantic matching, characterized in that, The method includes: S1: Based on the industry-specific word segmentation dictionary, perform Chinese word segmentation on the first and second field names to be matched in the cross-industry data platform to obtain the first word sequence and the second word sequence. S2: Perform multi-task joint pre-training on the first and second morpheme sequences to obtain the first word order-aware semantic vector and the second word order-aware semantic vector. Construct the first and second dependency syntax trees, fuse them to obtain the word order-structure joint vector, and evaluate the structural difference index of the two dependency syntax trees. S3: When the morpheme overlap rate between the first morpheme sequence and the second morpheme sequence exceeds a preset threshold, calculate the relative difference index of the common morpheme subsets in the first morpheme sequence and the second morpheme sequence, and perform cross-coupling calculation on the structural difference index and the relative difference index to obtain the word order difference degree. S4: Based on the word order difference, perform difference-aware correction on the similarity of the word order-structure joint vector, and establish a semantic mapping relationship between the first field name and the second field name to complete cross-industry data fusion.
2. The cross-industry data platform fusion method based on semantic matching as described in claim 1, characterized in that, The process involves performing Chinese word segmentation on the first and second field names to be matched in the cross-industry data platform based on an industry-specific word segmentation dictionary, resulting in a first word sequence and a second word sequence, including: Character traversal is performed on field name samples from various data sources in the cross-industry data platform to count the co-occurrence frequency of each adjacent character combination, thus obtaining a candidate word set; The frequency of each morpheme in the candidate morpheme set is counted in various industry fields of the cross-industry data platform. The industry coverage weight is evaluated based on the frequency of occurrence. Candidate morphemes whose industry coverage weight exceeds a preset threshold are regarded as industry core morphemes and stored in the industry-specific dictionary. Using industry core morphemes as the longest matching benchmark, starting from the starting character position of the first field name or the second field name, the industry core morphemes of each industry-specific word segmentation dictionary are matched sequentially. When a match is successful, the corresponding industry core word is extracted as a word unit, and the matching continues from the next character position until all characters are traversed, completing the word segmentation process and obtaining the first word sequence and the second word sequence.
3. The cross-industry data platform fusion method based on semantic matching as described in claim 1, characterized in that, The multi-task joint pre-training of the first and second morpheme sequences to obtain the first and second word order-aware semantic vectors includes: Input the morpheme sequence into a pre-constructed semantic encoding network, and map each morpheme in the morpheme sequence to an initial embedding vector to obtain an initial embedding vector sequence; The absolute position of each vector in the initial embedded vector sequence is encoded to obtain the order encoding information. The order encoding information is then superimposed on each initial embedded vector to obtain the intermediate encoded vector sequence. The word order sequence is input into the semantic coding network, and multi-task joint pre-training is performed to obtain the first word order-aware semantic vector and the second word order-aware semantic vector.
4. The cross-industry data platform fusion method based on semantic matching as described in claim 3, characterized in that, The multi-task joint pre-training process yields a first word order-aware semantic vector and a second word order-aware semantic vector, including: The perturbation level value corresponding to each word order variant sequence is used as the regression target for perturbation level regression training; The morpheme sequence and its corresponding reverse morpheme sequence are used to form positive and negative sample pairs, and positive and reverse order discrimination training is performed. Train on word order invariance by comparing semantically equivalent field names with semantically different field names; For both lightly perturbated and heavily perturbated sequence variants, perturbation type hierarchical discrimination training is performed. The word order invariance contrast training and perturbation type hierarchical discrimination training share the intermediate hidden layer parameters of the semantic coding network; The embedding vectors output from perturbation degree regression training, forward and reverse order discrimination training, word order invariance comparison training, and perturbation type hierarchical discrimination training are fused to obtain the first word order-aware semantic vector and the second word order-aware semantic vector.
5. The cross-industry data platform fusion method based on semantic matching as described in claim 4, characterized in that, The construction of the first and second dependency syntax trees, the fusion of these trees to obtain a word order-structure joint vector, and the evaluation of the structural difference index between the two dependency syntax trees include: Based on the part of speech of each morpheme in the morpheme sequence, determine the dependency relationship type between any two morphemes, establish a directed dependency arc, and combine each morpheme node and the directed dependency arc into a dependency syntax tree; Using the directed dependency arcs in the dependency syntax tree as the edge structure and the corresponding morpheme semantic embedding vectors in the word order semantic perception vector as the initial features of the nodes, the neighbor node information aggregation operation is performed on each morpheme node to obtain the word order-structure joint vector. The structural difference index is obtained by weighting and fusing the root node consistency coefficients of the two dependency syntax trees with the subtree isomorphism ratio.
6. The cross-industry data platform fusion method based on semantic matching as described in claim 1, characterized in that, When the morpheme overlap rate between the first morpheme sequence and the second morpheme sequence exceeds a preset threshold, the relative difference index of the common morpheme subsets in the first morpheme sequence and the second morpheme sequence is calculated, including: The common morphemes with the same content in the first morpheme sequence and the second morpheme sequence are sorted according to their arrangement order in the two morpheme sequences to obtain the first order sequence and the second order sequence. Starting from the first morpheme of the first sequence, each morpheme is taken as the target morpheme in turn, and the position of the target morpheme is located in the second sequence to obtain the positional correspondence of the target morphemes; Based on the positional correspondence, starting from the first morpheme of the second sequence, the proportion of morpheme pairs whose relative order of any two morphemes in the two sequences is inconsistent is used as the relative difference index.
7. The cross-industry data platform fusion method based on semantic matching as described in claim 6, characterized in that, The cross-coupling calculation of the structural difference index and the relative difference index to obtain the word order difference degree includes: The structural difference index is used as a confidence weight to adjust the contribution ratio of the relative difference index; When the structural difference index exceeds a predetermined threshold, the contribution ratio of the relative difference index is reduced; when the structural difference index is below a predetermined threshold, the contribution ratio of the relative difference index is increased. The relative difference index after adjusting the contribution ratio is weighted and fused with the structural difference index to obtain the word order difference degree.
8. The cross-industry data platform fusion method based on semantic matching as described in claim 1, characterized in that, The difference-aware correction of the similarity between word order and structure joint vectors based on word order difference includes: The vector similarity between the two word order-structure joint vectors is used as the basic semantic matching score; Obtain the morpheme overlap rate and word order difference between the first and second morpheme sequences; When the morpheme overlap rate is higher than the preset overlap qualification threshold, the word order difference is compared with the preset low-order difference threshold and high-order difference threshold, and the low-order difference threshold is less than the high-order difference threshold. When the word order difference is lower than the low-order difference threshold, a positive gain is applied to the basic semantic matching score to obtain the corrected semantic matching score. When the word order difference is higher than the high-order difference threshold, a negative penalty is applied to the basic semantic matching score to obtain the corrected semantic matching score.
9. The cross-industry data platform fusion method based on semantic matching as described in claim 8, characterized in that, The establishment of a semantic mapping relationship between the first field name and the second field name to complete cross-industry data fusion includes: When the corrected semantic matching score exceeds the preset mapping establishment threshold, the first field name and the second field name are marked as a matching field name pair; Based on the data type of the first and second fields in the matching field name pairs, determine the data type conversion rules between the matching field name pairs; The semantic mapping relationship between the first field name and the second field name is used as the mapping connection, and the data type conversion rules are used as the basis for data conversion to establish a two-way data mapping link. Based on the bidirectional data mapping link, the data record corresponding to the first field name is converted according to the data type conversion rules and then written to the storage location corresponding to the second field name, thus completing the cross-industry data fusion.
10. A cross-industry data platform fusion system based on semantic matching, characterized in that: The system for implementing the cross-industry data platform fusion method based on semantic matching as described in claim 1 includes: The word segmentation module is used to perform Chinese word segmentation on the first and second field names to be matched in the cross-industry data platform according to the industry-specific word segmentation dictionary, so as to obtain the first word sequence and the second word sequence. The word order and structure awareness module is used to perform multi-task joint pre-training on the first and second word morpheme sequences to obtain the first and second word order awareness semantic vectors, construct the first and second dependency syntax trees, fuse them to obtain the word order-structure joint vector, and evaluate the structural difference index of the two dependency syntax trees. The word order difference quantification module is used to calculate the relative difference index of the common morpheme subsets in the first morpheme sequence and the second morpheme sequence when the morpheme overlap rate between the first morpheme sequence and the second morpheme sequence exceeds a preset threshold. The structural difference index and the relative difference index are cross-coupled and calculated to obtain the word order difference degree. The difference correction fusion module is used to perform difference-aware correction on the similarity of word order-structure joint vectors based on word order difference, and to establish a semantic mapping relationship between the first field name and the second field name to complete cross-industry data fusion.
