A multi-modal data adaptation and feature association method for intelligent insight

Abstract

The present application relates to the technical field of data analysis, and particularly relates to a multi-modal data adaptation and feature correlation method for intelligent insight, comprising the following steps: standardizing and arranging structured data fields in data sources of different systems to form a structured field set, identifying field drift patterns, and generating a structure drift label set; guiding semantic flux analysis on content related to the structured field in unstructured modal data, and extracting core semantic candidate units; finally screening out a standardized target semantic unit set; performing graph structure alignment on the standardized target semantic unit set and the structured field set to obtain a field evolution graph, and performing dynamic adaptation based on field inheritance relationships and time sequence evolution paths in the field evolution graph. The present application effectively excludes pseudo-related expressions that are only close in static semantics but have inconsistent contexts, and improves the authenticity and business effectiveness of semantic unit screening.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a multimodal data adaptation and feature association method for intelligent insights. Background Technology

[0002] In cross-system, cross-source multimodal data analysis and intelligent insight applications, such as foreign trade supply chain management, cross-border risk control, and enterprise knowledge modeling, the processing of mixed structured fields and unstructured text data is often involved. Data sources include government systems, business systems, free text reports, and industry documents. Due to system development time, standard differences, and business evolution, the same business entity or attribute field often exhibits naming differences, structural evolution, and semantic drift in different systems, leading to difficulties in downstream semantic understanding, rule matching, and model building.

[0003] Most mainstream field fusion and alignment methods currently rely on manually setting thesaurus, regular expression rules, or mapping tables to unify different fields into standard field names. While this approach is applicable to small-scale, static systems, it is difficult to maintain in large-scale, multi-variable systems and cannot handle semantically evolving or context-dependent naming variations. Field matching algorithms based on string similarity or edit distance, such as Jaccard and Levenshtein, are used to identify surface similarities between field names, but they cannot capture hidden semantic equivalence relationships. Their recognition ability significantly decreases, especially when faced with redundant modifiers, abbreviations, transliterations, or historical naming contamination.

[0004] Furthermore, existing research generally overlooks the semantic interference between structured fields and unstructured text. In particular, unstructured modalities often contain historical field aliases, noisy fragments, or ambiguous expressions resulting from cross-domain migrations, which can easily lead to semantic recognition biases. Moreover, even if field alignment is achieved, the temporal evolution of field semantics during business development is often ignored, resulting in distorted feature representations and limiting the accuracy and usability of multimodal fusion analysis in practical intelligent insight tasks. Summary of the Invention

[0005] This invention provides a multimodal data adaptation and feature association method for intelligent insight, which is an adaptation method for intelligent insight scenarios with the ability to model field semantic evolution and multimodal feature fusion, in order to solve the bottleneck problems of current systems in terms of structural drift perception, field alignment accuracy and semantic consistency modeling.

[0006] A multimodal data adaptation and feature association method for intelligent insights includes the following steps:

[0007] S1, Construct a structured field set and perform cross-source structure drift identification: Standardize and organize the structured data fields from data sources from different systems in foreign trade business scenarios to form a structured field set; Based on the preset domain mapping dictionary and naming trajectory history, extract field naming variants, semantic heterogeneous forms and frequency features, identify field drift patterns, and generate a structure drift annotation label set;

[0008] S2 completes the redundant semantic stripping and core semantic unit stability screening of unstructured modalities: the structural drift label set output from S1 is used as a fuzzy interference factor library to guide the semantic flux analysis of content related to structured fields in unstructured modal data, and calculates content consistency and expression compactness to extract core semantic candidate units; at the same time, a pseudo-feedback backtracking mechanism is introduced to inject the core semantic candidate units into the preset insight model to evaluate their response stability, and finally screen out the standardized target semantic unit set;

[0009] S3, based on the field evolution graph, performs semantic unit alignment and dynamic feature fusion: the standardized target semantic unit set obtained in S2 is aligned with the structured field set constructed in S1 to obtain the field evolution graph, and dynamic adaptation is performed based on the field inheritance relationship and temporal evolution path in the field evolution graph to generate a fusion feature vector set for subsequent intelligent insight tasks.

[0010] Optionally, the construction of the structured field set includes: cleaning the original data fields from different business systems in the foreign trade business scenario, removing prefixes, suffixes and special delimiters, and normalizing the field format based on a unified data type template; dividing the cleaned fields into multiple types of data according to the entity objects and business attributes they describe, forming an initial structured field set.

[0011] Optionally, based on a preset domain mapping dictionary, the field names in the initial structured field set are matched. The domain mapping dictionary defines standard field names and their corresponding synonyms, abbreviations, and common spelling variations. At the same time, a naming trajectory history database is invoked. The naming trajectory history database records the historical name change sequence of the same field in each data source as the system version iterates. By comparing the current field name with the historical name sequence, field naming variations and implicit semantic heterogeneous forms are extracted.

[0012] The frequency of each naming variant and semantic heterogeneous form in historical data and current multi-source data is statistically analyzed, and its frequency distribution and time decay weight are calculated. Based on the frequency distribution characteristics and the naming trajectory history, the field drift patterns are identified and labeled as: stable standard type, high-frequency synonym drift type, low-frequency outdated alias type, and cross-system semantic bifurcation type.

[0013] Optionally, for each field in the initial structured field set, associate it with its identified drift pattern type, all extracted named variants and their confidence and frequency feature vectors to generate a machine-readable structured drift annotation label set.

[0014] Optionally, the fuzzy interference factor library includes the standard names of each structured field, their corresponding naming variants, and noise keywords corresponding to the low-frequency outdated alias type and cross-system semantic bifurcation drift pattern; for unstructured modal data, text segmentation and basic cleaning are first performed, and based on the fuzzy interference factor library, text segments are scanned and matched to identify and label all text fragments that mention or are associated with structured fields and their variants; for each labeled text fragment, its content consistency and expression compactness are calculated:

[0015] The content consistency is obtained by calculating the average vectorized semantic similarity of the same entity or attribute described in different text fragments.

[0016] The expression compactness is determined by analyzing the keyword density and sentence complexity in the text fragment, and eliminating lengthy expressions containing a large amount of modifying, vague or irrelevant information;

[0017] Text segments with both content consistency and expression tightness scores above the preset threshold are extracted as core semantic candidate units.

[0018] Optionally, the pseudo-feedback backtracking mechanism includes taking the core semantic candidate unit as input and injecting it into a preset insight model for foreign trade scenarios for multiple rounds of inference testing; by perturbing the input context or mixing other candidate units, observing and recording the fluctuation of the model's response to the information represented by the semantic unit, and calculating its response stability score.

[0019] Optionally, based on the response stability score and combined with the S content consistency and expression compactness scores, the core semantic candidate units are comprehensively ranked and screened; the screened units are semantically standardized and redundant are eliminated to form a standardized target semantic unit set for feature alignment.

[0020] Optionally, the construction of the field evolution graph includes taking each standardized field in the structured field set and its structure drift label set as the master node, and establishing directed edges between field nodes with synonym, inheritance or replacement relationships according to the naming trajectory history and drift pattern. The directed edges are accompanied by time stamps and relationship weights to form a field evolution graph that represents the semantic evolution relationship of the fields.

[0021] Optionally, the execution graph structure alignment includes taking each semantic unit in the standardized target semantic unit set as a node to be aligned, calculating its semantic similarity with each main node in the field evolution graph; based on semantic similarity threshold and relational path analysis, mapping each semantic unit node to the most relevant single or multiple structured field main nodes, and establishing bidirectional alignment edges to complete the expansion and alignment of the graph structure; relying on the structure of the aligned subgraph in the field evolution graph, traversing the temporal evolution paths between key field nodes and their associated semantic unit nodes; for each temporal evolution path, introducing a temporal decay factor to weight the features of historical nodes, and based on the graph attention mechanism, aggregating the feature information of multi-hop neighbor nodes and along the evolution path to dynamically generate a context-enhanced feature representation for each node.

[0022] Optionally, S3 further includes generating a fusion feature vector set, specifically including extracting the standardized features, context-enhanced features, and type encoding of the drift mode of each node in the dynamically adapted alignment subgraph; concatenating and reducing the dimensionality of the three types of features to generate a fusion feature vector that integrates multimodal alignment and evolution information for each original data instance, thus forming a fusion feature vector set for subsequent intelligent insight tasks.

[0023] The beneficial effects of this invention are:

[0024] This invention systematically introduces a field evolution graph modeling mechanism, incorporating the naming variants, semantic migration history, and cross-system semantic bifurcation relationships of structured fields into a unified modeling framework to form a field evolution graph with temporal information and semantic weights. By constructing a semantic alignment channel, unstructured semantic fragments are accurately mapped to standard field nodes, compensating for the semantic fragmentation problem caused by structural drift. This enables consistent alignment of synonymous fields in cross-modal and cross-system data, significantly improving the robustness and interpretability of field fusion and downstream modeling.

[0025] This invention differs from traditional text matching methods based on rules or static models. It constructs a pseudo-feedback reasoning mechanism, injects candidate semantic units into an intelligent insight model for multiple rounds of contextual perturbation testing, and quantifies stability based on the output variance of the model response. This effectively eliminates pseudo-related expressions that are only similar in static semantics but inconsistent in context, improving the authenticity and business effectiveness of semantic unit selection. It is suitable for high-accuracy labeling and extraction of structured fields in actual business logic.

[0026] This invention, by fusing original field features, evolutionary path contextual enhancement features, and structural drift type encoding, forms a unified fused feature vector set. This set can be directly input into downstream AI models for classification, clustering, and prediction, supporting unified information expression and intelligent processing across multimodal data. This fused feature vector possesses temporal sensitivity, semantic consistency, and structural evolution traceability, empowering complex intelligent insight tasks such as data asset value mining, field lifecycle prediction, and cross-system question-answering matching, thus promoting the construction and application of multi-source heterogeneous data value systems for enterprises. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram illustrating the redundant semantic stripping and core semantic unit stability screening of unstructured modalities in an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram illustrating semantic unit alignment and dynamic feature fusion based on field evolution graphs in an embodiment of the present invention. Detailed Implementation

[0031] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0032] like Figures 1-3 As shown, a multimodal data adaptation and feature association method for intelligent insights includes the following steps:

[0033] S1, Construct a structured field set and perform cross-source structure drift identification: Standardize and organize the structured data fields from data sources from different systems in foreign trade business scenarios to form a structured field set; Based on the preset domain mapping dictionary and naming trajectory history, extract field naming variants, semantic heterogeneous forms and frequency features, identify field drift patterns, and generate a structure drift annotation label set.

[0034] S11, Constructing a Structured Field Set: In actual foreign trade operations, various systems, such as CRM systems, ERP systems, and logistics management systems, may use different names and formats for the same business field. For example, the field might be "cus id" in system A and "customer code" in system B. System A might use the YYYY / MM / DD date format, while system B might use DD-MM-YYYY. Some field names may also include prefixes like "sysA_" or suffixes like "_temp," or use special characters such as underscores or hyphens to connect words. Therefore, the first step is to clean these field names and formats to make them consistent in form, facilitating subsequent classification and analysis. This involves cleaning the raw data fields from different business systems in a foreign trade scenario. The steps include:

[0035] Remove system-specific prefixes, suffixes, and special delimiters from the field;

[0036] Field formats are normalized based on a unified data type template, such as DateTime, String, Currency, and Boolean.

[0037] Based on the entity objects and business attribute dimensions described by the fields, the fields are divided into four categories: Order, Customer, Logistics, and Product. The approach uses entity objects + business attribute dimensions, that is:

[0038] 1. Order Category: These fields are usually related to a transaction order itself, such as order number, order time, order status, and total amount; these fields describe the basic attributes of an order, and are therefore classified as order categories.

[0039] 2. Customer Class: These fields are related to the customer information that initiated the order, such as customer ID, customer name, customer type, and region. They all revolve around the customer entity and are therefore classified as customer class.

[0040] 3. Logistics: These fields reflect information about the goods during transportation, such as shipping time, arrival time, carrier name, and waybill number. These fields describing the logistics process are classified as logistics.

[0041] 4. Product Category: These fields are related to the specific product itself, such as product number, product name, category, and unit price; they express the core attributes of the product entity, and therefore belong to the product category.

[0042] The final initial structured field set is denoted as: .

[0043] S12, Extracting Field Naming Variations and Semantic Heterogeneities: In multi-source foreign trade systems, the same business field may be named differently due to different developers, different system language habits, or version updates. Such naming variations or semantic heterogeneity can cause difficulties in subsequent data fusion and analysis. Therefore, the goal is to identify all field expressions with inconsistent names but consistent semantics, laying the foundation for subsequent unification and fusion.

[0044] S121, based on a preset domain mapping dictionary ,right For each field name, a matching process is performed to extract synonyms, abbreviations, and spelling variations corresponding to the standard field name, forming a set of explicit naming variations. ;

[0045] The domain mapping dictionary is a pre-built professional knowledge base used to describe the standard expressions and various naming variations of common fields in foreign trade operations. The purpose of this dictionary is to solve the problem of the same field having different names in different systems. It contains multiple standard field names, each associated with possible synonyms, abbreviations, spelling variations, and historically common misspellings. For example, for the field "customer number," some systems might call it "customer_id," others "cust_code," and still others "client_number." The domain mapping dictionary aims to aggregate these different names and unify them to a single standard field name. This dictionary is manually annotated by experts, who have compiled naming conventions for common fields in different systems based on their experience. In practical applications, when encountering a field name, this dictionary is consulted first to see if it belongs to a variation of a standard field. If a correspondence is found, it can be included in a unified standard field category for subsequent structural adaptation and semantic analysis.

[0046] S122, simultaneously, call the named trajectory history database. This database records the name evolution sequence of fields across different system versions, for each field. Compare with the current name Its naming trajectory Extracting implicit semantic heterogeneous forms: ;in, Indicates passage Extracted synonymous naming variants, Indicates passage The semantic heterogeneous forms obtained by comparison.

[0047] The naming history database is a repository used to record historical information about how field naming conventions have changed over time and with system evolution. Its main purpose is to address the issue of the same field's name changing during system upgrades or replacements. Each record in this database describes the historical name of a field used at different points in time and in different system versions. For example, a field might be called `partner_code` in an older system, `customer_id` in a newer system, and might have even had transitional names like `party_id`. All these evolutionary paths are fully recorded. The database is built from the following types of information:

[0048] Version comparison record: Compare the data dictionary or interface definition file of historical versions to extract the change trajectory of field names;

[0049] System migration logs: Field mapping documents created by developers during system upgrades or refactoring;

[0050] The output of semantic fusion process: manual or automatic pairing results generated during past field alignment or fusion processes;

[0051] Structured field annotation comparison: Based on the semantic matching of field descriptions, it discovers the evolution of fields with inconsistent names but the same meaning.

[0052] When a field in the current system cannot be directly matched in the domain mapping dictionary, the naming history database is searched to see if the field has a historical naming inheritance or substitution relationship with certain standard fields. If such a relationship exists, the field can be determined to be an evolutionary descendant of a certain standard field and merged into a unified field set.

[0053] S13, Calculating Frequency Features and Identifying Drift Patterns: In S12, a set of naming variants were extracted for each field. These variants may be synonyms, abbreviations, historical names, etc. The task in S13 is to determine whether the field has undergone structural drift and identify the specific type of drift by statistically analyzing the frequency of occurrence and evolution trajectory of these naming variants.

[0054] For a given naming variant, we check how many times it appeared in historical data and how many times it appeared in current system data, obtaining two frequency values. These represent the variant's past popularity and current popularity, respectively. If a variant appeared frequently in the past but is rarely used now, it may be an outdated alias. If a variant consistently appears frequently, it may be a mainstream standard usage. In other words, we statistically analyze each naming variant. In historical datasets Compared with current multi-source datasets The frequency of occurrence in the vector is used to construct a frequency feature vector: ;in, Indicates naming variants Frequency of occurrence in historical data Indicates naming variants Frequency of occurrence in the current data;

[0055] Considering the varying importance of different time periods in the judgment, a time decay weight function is introduced. This function assigns a higher weight to recently occurring variants and a gradually decreasing weight to older variants. This allows for a more accurate assessment of which variants remain active and which are merely historical remnants. The time decay weight function is expressed as follows: ;in, For the current time slice, Indicate variant The most recent time of occurrence, The attenuation coefficient represents the time sensitivity.

[0056] based on , The naming evolution behavior of each field is classified by naming trajectory sequences and semantic edit distances between variants, identifying the following structural drift patterns (multiple patterns can be labeled for each field):

[0057] 1. Stable standard type;

[0058] 2. High-frequency synonym drifting type;

[0059] 3. Low-frequency outdated alias type;

[0060] 4. Cross-system semantic forking type.

[0061] Record each field The set of structural drift modes is as follows: It belongs to the stable standard type, the high-frequency synonym drift type, the low-frequency outdated alias type, and the cross-system semantic fork type.

[0062] Specifically:

[0063] 1. Frequency characteristics: This includes historical and current frequencies. First, the frequency of each field's naming variants in historical and current data is calculated separately to determine whether their usage activity has changed. For example, if a variant appears frequently in the past and present, it indicates stable usage; if a variant was commonly used in the past but rarely appears now, it may be an obsolete expression; if multiple variants are used frequently in the same period, there may be synonym drift.

[0064] 2. Time decay weight: To account for the impact of time distance, a decay mechanism is introduced, assigning lower weights to older data. For example, a variant that has appeared frequently in the past may have diminished reference value if it has hardly appeared in recent years; conversely, variants that have appeared frequently recently are more likely to be the current mainstream usage, helping to determine which variants are active and which are merely historical legacies.

[0065] 3. Naming Trajectory Sequence: Review the naming trajectory history to see how the name of a field changes in different versions or platforms of the system. If the name of a field changes frequently in different systems or versions, or if each change points to a different semantic entity, it indicates that the naming of the field is at risk of evolution instability or semantic differentiation.

[0066] 4. Semantic edit distance between naming variants: Analyze the linguistic similarity between naming variants, including spelling differences, semantic similarity, etc. If multiple variants switch frequently but their semantics are basically consistent, it may just be a change in naming style; but if these variants are significantly different in semantics, it means semantic splitting.

[0067] The logic for identifying structural drift patterns is as follows: Based on the combined performance of the above four conditions, determine whether each field falls into one or more of the following four drift patterns:

[0068] 1. Stable standard type: The typical characteristics are that there are very few naming variations, or the same standard field name is used for a long time; the current frequency is high and the historical frequency is also high; there is no semantic divergence and no obvious jumps in the trajectory.

[0069] The judgment logic is as follows: a certain naming variant is dominant in both historical and current frequency; the naming trajectory has basically not evolved or has changed very little; and the semantics between variants are extremely similar. Such fields represent an ideal state of semantic stability and standardized naming.

[0070] 2. High-frequency synonym drifting type: The typical characteristics are that there are multiple frequently used variants of the same semantic field; the variants are typical synonyms or abbreviations; and they often switch between these variants in different scenarios or at different times.

[0071] The judgment logic is as follows: two or more naming variants have medium to high frequency in both the current and historical data; they remain relatively active even after time decay weighting; they have high semantic similarity, small edit distance, or similar word meanings; and the naming trajectory shows multiple high-frequency variants used alternately. This indicates that a unified standard for field naming has not been formed, but the semantics are consistent.

[0072] 3. Low-frequency obsolete aliases: The typical characteristics are that some variants were frequently used in the past but are now almost never seen; they may be field names left over from old systems; they usually only appear in earlier versions of the data.

[0073] The judgment logic is as follows: the variant has a high historical frequency but a low current frequency; the corresponding time point is far from the present, resulting in a lower decay weight; the naming trajectory indicates that the variant only exists in earlier versions; the variant may have a high semantic similarity to the main field. Such variants indicate that the field name was once widely used, but has now been replaced by a more standard name.

[0074] 4. Cross-system semantic bifurcation type: The typical characteristic is that the same field uses names with large semantic differences in different systems; the naming trajectory shows that multiple variants evolve on different platforms; some variants may represent customer numbers in one system and partner numbers in another system, resulting in semantic ambiguity.

[0075] The judgment logic is as follows: naming trajectories show that variants have evolved across platforms and have not been aggregated into a unified field; each variant has a high frequency in its respective system; there are significant semantic differences between different variants, with large edit distances and inconsistent word meanings; and there are differences in the field usage logic between systems. This is the most complex type of structural drift, which may lead to deviations in data alignment and analysis results.

[0076] In summary, the naming evolution behavior of each field is analyzed from multiple dimensions as described above. Based on the frequency change trend, time period, historical evolution path, and semantic distance, the stability and consistency of the field's naming are judged, and finally, one or more structural drift pattern labels are assigned to it.

[0077] S14, Generate a structural drift annotation label set: for each field Generate structural drift annotation labels, including: ;in, For a set of field drift patterns, For a collection of named variants, For each variant, the frequency feature vector, This represents the confidence score for each naming variant, ultimately forming a machine-readable set of structure drift annotation labels, denoted as: .

[0078] Furthermore, the confidence score is calculated using word vector similarity. A word vector maps a word to a real-valued vector in a high-dimensional space; commonly used tools include Word2Vec, GloVe, and FastText. The cosine similarity between two word vectors is calculated to determine their semantic closeness. This mechanism is used to evaluate whether a named variant is semantically consistent with a standard field. The specific process is as follows:

[0079] Vectorization: Convert standard field names and named variants into word vectors, which can be whole word vectors or average vectors composed of multiple sub-words.

[0080] Calculate semantic similarity: Perform cosine similarity calculation on the two vectors to obtain a value between 0 and 1, which represents the degree of semantic similarity: close to 1: semantically very close; close to 0: almost unrelated; use this similarity value as the confidence score of the variant, indicating the degree of confidence that the variant is accepted as the orthodox name of the current field.

[0081] S2 completes the redundant semantic stripping and core semantic unit stability screening of unstructured modalities: the structural drift label set output from S1 is used as a fuzzy interference factor library to guide the semantic flux analysis of content related to structured fields in unstructured modal data, and calculates content consistency and expression compactness to extract core semantic candidate units; at the same time, a pseudo-feedback backtracking mechanism is introduced to inject the core semantic candidate units into the preset insight model to evaluate their response stability, and finally screen out the standardized target semantic unit set.

[0082] S21, Constructing a Fuzzy Interference Factor Library: The aim is to extract field information from structured data that is prone to semantic confusion and establish a dedicated fuzzy interference factor library to guide semantic cleaning and filtering in unstructured text. In actual foreign trade operations, different systems, versions, and teams may have different naming conventions for the same field. Some conventions are standard, while others are outdated, vague, or even have different meanings in different contexts. These are potential fuzzy factors or noise factors.

[0083] S1 has already performed standard naming, name variant identification, frequency analysis, and structural drift pattern classification on each structured field. Two types of field variants with high naming risks have been identified: first, low-frequency outdated aliases, such as a field that was once called `partner_code` but is now almost never used; second, cross-system semantic bifurcation, such as `party_id` representing a customer number in system A and a partner ID in system B, resulting in severe semantic confusion. These variant names are prone to causing ambiguity or misjudgment in unstructured text, and therefore need to be specifically marked as key interference factors for subsequent semantic cleaning. Each structured field is traversed one by one, extracting the following two types of information: the standard name of the field and, among all naming variants of the field, variants belonging to the low-frequency outdated or cross-system semantic bifurcation types. These are considered fuzzy noise. Thus, each field corresponds to a set of [standard name + noise variant], forming a field-level interference factor entry. The entries for "standard field name + noise variant set" corresponding to all fields are integrated together to form a system-level fuzzy interference factor library. This library is a structured knowledge set that describes which words / phrases may be mentioned in a field in the text; which expressions are prone to semantic misunderstanding, structural drift or redundant description. It will play a guiding role in locating interference, removing redundancy and identifying core semantics in subsequent unstructured data processing.

[0084] Specifically, it involves the structural drift annotation label set output from stage S1. As a fuzzy interference factor library The basic input. Each structured field... Include:

[0085] Standard field name ;

[0086] Named variant set ;

[0087] Drift type set If it contains low-frequency outdated aliases or cross-system semantic bifurcation types, then its corresponding variants are considered as potential noise keywords, denoted as the noise set. ;

[0088] The final fuzzy interference factor library is represented as follows: .

[0089] For ease of understanding, an example of a fuzzy interference factor library is provided below.

[0090] Field 1: Customer ID, standard field name: customer_id;

[0091] Standard field name: customer_id;

[0092] Noise naming variants were identified as low-frequency outdated aliases or cross-system semantic bifurcation types:

[0093] partner_code, used in older systems, is now obsolete;

[0094] party_id refers to the customer ID in system A and the partner ID in system B;

[0095] client_no, which occasionally appeared in previous versions, is no longer recommended.

[0096] Field 2: Order creation time, standard field name: order_create_time;

[0097] Standard field name: order_create_time;

[0098] Noise naming variants:

[0099] booking_time is a polysemous term: it can refer to both the time the order is placed and the time the reservation is made.

[0100] created_dt was used in early CRM systems and was infrequent.

[0101] insert_time refers to the time when the data is written to the database in some systems, but this is not strictly equivalent.

[0102] Field 3: Logistics Carrier, Standard Field Name: logistics_carrier_name;

[0103] Standard field name: logistics_carrier_name;

[0104] Noise naming variants:

[0105] Shipper may refer to the shipper rather than the carrier in different systems;

[0106] delivery_party is often confused with customer-side and is not unique;

[0107] express_name, an abbreviation for a courier company, has a severely overgeneralized semantic meaning.

[0108] It is important to note the fuzzy interference factor library:

[0109] One-to-one mapping: Each structure field corresponds to a standard name plus several fuzzy variants;

[0110] These variants were identified in the S1 stage through structural drift analysis, and are based on actual usage frequency and semantic evolution.

[0111] This library is not used for naming conventions, but rather for locating potentially distracting segments in unstructured text;

[0112] S22 performs semantic flux analysis and extracts candidate units. The goal is to automatically identify expression fragments from unstructured text that may correspond to the meanings of structured fields, perform semantic quality assessment on these fragments, and extract semantically clear and content-focused core units as candidate inputs for subsequent alignment. This is for the original unstructured modality dataset. The following processing flow will be executed:

[0113] S221, Text Segmentation and Cleaning: For each piece of unstructured text data A collection of text segments divided into sentences or phrases. Perform cleaning operations such as removing stop words, removing punctuation, and standardizing formatting. Indicates the first An unstructured text record.

[0114] S222, Matching fuzzy interference factors: For each text segment ,exist The matching process is performed to identify segments that mention or are indirectly related to a certain field and its noisy variants, resulting in a set of segments of interest.

[0115] This expression represents a text fragment. The process involves filtering out all fragments containing ambiguous or distracting keywords (such as named variants or noisy field names) for subsequent semantic analysis. Indicates the first The first in the text S222 is a collection of keywords that are part of a larger collection of text fragments. The S222 collection is a collection of keywords that are part of a larger collection of text fragments ...

[0116] S223, Calculate Content Consistency Score: For each text segment marked as relevant, determine whether its semantics are consistent with other segments describing the same field. Compare the semantic vectors of the current segment with other segments mentioning the same field and calculate their average similarity. The higher the score, the more semantically stable and conceptually consistent the content of the segment is with other segments, and the higher its credibility. Specifically, this includes each named entity-related segment... Select descriptions of the corresponding entities in different fragments to construct a comparison pair. Calculate the mean of their vector semantic similarity:

[0117] This formula represents the calculation segment. The average semantic similarity to other descriptive fragments of the same field is used to assess whether its semantics are stable and concentrated. A higher value indicates more consistent content. Representing fragments Word vector representation, Representing fragments Word vector representation, Indicates contrasting segments Word vector representation, Indicates and A set of semantic comparison fragments belonging to the same entity attribute. The cosine similarity function between word vectors A score indicating content consistency.

[0118] S224, Calculating the Expression Compactness Score: Two indicators are introduced: keyword density, which is the proportion of truly core words in the paragraph; and sentence complexity, which refers to the complexity of the grammatical structure and the amount of modifying elements. The score is obtained by dividing the keyword density by the sentence complexity. The higher the score, the more direct, semantically focused, and less redundant the passage. Therefore, the expression compactness score is defined as follows:

[0119] This formula measures whether a text fragment is concise and clear. High keyword density and low sentence complexity indicate focused expression with minimal redundancy, resulting in a higher score. Among these, This indicates keyword density, such as the proportion of core field-related terms. Indicators of sentence complexity, such as the number of nested clauses, average sentence length, etc. The higher the tightness, the more direct and clear the expression, and the less redundancy.

[0120] S225, Set a uniform threshold for filtering candidate semantic units. Filter all that meet the following criteria: The set of fragments serves as the core semantic candidate unit set: This indicates that from all segments that hit the interference word, those with consistent semantics and compact expression are selected to form a core semantic candidate unit set, providing input for the next step of pseudo-feedback backtracking. Among these, The threshold for content consistency scoring is 0.75 to 0.9. A value below 0.70 indicates that the semantics of the text fragment differ significantly from those of other fragments, which may be due to semantic drift or strong context dependence, making it unsuitable as a stable field. A value above 0.95 is too stringent and may discard some sentences or segments with different language styles or expressions but consistent semantics. A value of 0.80 can be fine-tuned in experiments. The selection threshold for expressing the compactness score is set in the range of [0.8, 1.5]. A value less than 0.8 often indicates that the sentence is interfered with by modifiers and parenthetical phrases and lacks core information. A value greater than 2.0 may be a calculation error. It is recommended to normalize the value according to the actual sentence length. Generally, it is set to 1.0, which means that it can be adopted when the keywords are sufficiently concentrated and the language structure is not too complex.

[0121] S23, Pseudo-feedback backtracking and stability assessment: [The following text appears to be incomplete and requires Each semantic unit Pre-set intelligent insight models for foreign trade scenarios Multiple rounds of simulated inference were conducted. During the inference process, the following perturbations were applied: changing the input context and mixing with other candidate units. After each perturbation, the semantic unit along with its changed context was fed into the model. The model's output was observed, the range of change in the model's output was recorded, and the stability score of the semantic response was calculated.

[0122] ;in, This indicates that the model performs well under multiple perturbations. The variance of the output response result, The closer the value is to 1, the more stable the model response and the more reliable the unit semantics.

[0123] The pre-defined intelligent insight model is a multi-task semantic understanding model, possessing capabilities such as named entity recognition, field mapping, context-aware reasoning, and semantic stability assessment. Its core structure can be divided into four layers:

[0124] 1. Input layer: Receives text fragments and contextual information as input, and encodes the input using multi-source embedding methods, including word vectors, field name vectors, and domain tag embedding, preserving word-level, phrase-level, and field-level semantic information.

[0125] 2. Semantic Awareness Layer: Introduces a context window and captures the dependency relationship between semantic units and their context through an attention mechanism. Inputs are the target semantic unit and the perturbed context; outputs a fused semantic representation vector, preserving context-dependent features. This layer can model:

[0126] Semantic alignment: Whether the meaning is consistent between the variant and the standard field;

[0127] Semantic ambiguity: Does the same word represent different fields?

[0128] Contextual conflict: Whether the expression is invalid in the current business context.

[0129] 3. Field Mapping and Label Prediction Layer: This layer performs structural field mapping prediction on keywords or semantic fragments in the input statement, including whether they can be mapped to structural fields, the name of the mapped field, and the mapping probability, which is the confidence level.

[0130] 4. Response Stability Analysis Module: An integrated logical branch outside the core model is used to evaluate the output stability of the same semantic unit under different contextual perturbations. Multiple rounds of testing are performed, with slightly different inputs in each round. The model output is observed to ensure consistency, and the variance stability score is calculated.

[0131] In this invention, the model operates as a validator, not for training or extracting semantic units. Instead, it serves as a judgment and verification system, receiving the core semantic candidate units selected in S22 and evaluating their semantic quality. The operation flow is as follows: a candidate text fragment is extracted from S22, and multiple perturbation versions with different contexts are constructed. Each perturbation text serves as the model input, and the model outputs the field recognition result. Model output metrics include whether it can be recognized as a structural field, which field it maps to, the confidence level of the model's recognition, whether the output before and after perturbation is consistent, and whether polysemous mapping or non-recognition occurs.

[0132] Output response variance Its purpose is to measure the stability of a semantic expression and its susceptibility to contextual perturbations; essentially, it is a semantic robustness metric. The core concept is the perturbation input and the model response, i.e., given an unstructured semantic fragment... We aim to determine whether its semantic expression is stable and reliable. To this end, we employ a pseudo-feedback mechanism, which involves artificially constructing multiple slightly different contexts, referred to as perturbation inputs, and comparing them with... The data is fed into the model for inference, and the model's response is recorded for each iteration. Each response is represented by the identified structural field label and the prediction confidence for that label. The variance calculation target is the same semantic unit. The model's confidence level fluctuates drastically under different perturbation contexts. Large fluctuations indicate that the semantic unit is significantly affected by the context and has poor robustness; small fluctuations indicate clear and stable semantic expression and high reliability. The specific calculation steps are as follows:

[0133] 1. Construction A perturbation input version: around the semantic unit ,structure Disturbance in contextual text: , Indicates the first The input of perturbation versions (including) These inputs differ slightly in terms of grammatical structure, contextual information, modifiers, and word order.

[0134] 2: Input each disturbance into the model and obtain the response confidence: Input each disturbance into the model. Input Model The confidence level of its prediction as the target field is obtained: ;in, For the first Wheel-in-the-Wheel Model The confidence level of a field being identified as the target field, typically ranging from [0,1], with higher values ​​indicating greater confidence.

[0135] 3: All The confidence scores of the rounds of testing are averaged: .

[0136] 4. Calculate the variance of the response fluctuation based on statistical definitions:

[0137] .

[0138] S24, Filtering the standardized target semantic unit set: For each candidate semantic unit... A comprehensive score is generated: ;in, For the weighting coefficients, satisfying The selection of candidate units is determined by parameter tuning in the actual scenario. All candidate units are sorted in descending order of comprehensive score, and the top-scoring units are selected. One is selected as the final retained unit.

[0139] The sorting and filtering process is as follows:

[0140] 1. Overall score ranking: All candidate semantic units are ranked from highest to lowest based on their overall score.

[0141] 2. Set a cutoff for the top k, i.e., set an upper limit for the number to be retained, or set a lower limit for the score.

[0142] 3. Standardization Processing: The final retained fragments undergo semantic standardization at the linguistic level, including eliminating synonyms, simplifying expressions, and removing redundant words, to form a standardized set of target semantic units. Used for semantic alignment and feature fusion in subsequent S3.

[0143] S3, based on the field evolution graph, performs semantic unit alignment and dynamic feature fusion: the standardized target semantic unit set obtained in S2 is aligned with the structured field set constructed in S1 to obtain the field evolution graph, and dynamic adaptation is performed based on the field inheritance relationship and temporal evolution path in the field evolution graph to generate a fusion feature vector set for subsequent intelligent insight tasks.

[0144] S31, Constructing a Field Evolution Graph: This requires modeling the semantic evolution relationships of fields involved in structured data. The goal is to enable subsequent semantic alignment between unstructured semantic information and structured fields, and to achieve cross-temporal information fusion. The structured field set output from S31 is used as an example. Each standardized field in As the master node in the graph, construct a directed field evolution graph: ; is a directed graph where each node is a field name, each edge represents a semantic relationship, and each path in the graph reflects the semantic migration history of a field. This graph will be used in the subsequent semantic alignment process of unstructured semantic fragments. The nodes are expanded to include naming variants. In actual business data, the same semantic field may have different names due to factors such as time, region, language, and system differences. These different naming variants will also be added as other nodes in the graph and establish evolutionary or synonymous relationships with the standard field nodes. Therefore, in the formula, This represents a collection of nodes, containing all standard field nodes and their named variations. This represents a set of edges, where each edge represents a semantic inheritance, substitution, or evolutionary relationship between fields. Each edge... Defined as:

[0145] ;in, Represents the evolution timestamp, representing the field. Replace or evolve into a field The time of occurrence, Semantic relation weights represent the semantic distance and naming similarity between two fields, essentially reflecting the semantic relationship between the fields. to field The semantic conversion reliability or semantic similarity is calculated by comprehensively considering multiple factors, as follows: ;

[0146] in, This represents the edit distance similarity based on the field name string, i.e., the Levenshtein distance. This indicates the frequency ratio of occurrences within the same data sample or document. These are the weighting coefficients.

[0147] S32, Execution Graph Structure Alignment: The goal is to accurately align the extracted target semantic units with the semantic evolution of historical fields, thereby achieving semantic consistency and structural adaptation across versions or sources of data. This involves aligning the standardized target semantic unit set obtained in S2... Each semantic unit in the graph serves as a node to be aligned. Calculate its relationship with the main field node in the graph. Semantic similarity: If the following conditions are met:

[0148] Furthermore, there exists a transition from the field evolution graph. The reachable relationship path from the starting point, i.e. If the main field node is reachable, then establish a bidirectional alignment edge between the semantic unit and the field: ;in, This represents the semantic alignment threshold, ranging from 0.7 to 0.8. If using context-aware models like BERT, a value of 0.7 is recommended; if using a bag-of-words model, it is recommended to increase it to 0.8. This represents the alignment confidence score, taken as 1. Similarity is already a normalized metric between 0 and 1, so it can be directly used as the confidence score without the need for secondary standardization or a scorer. It is a vector representation obtained through word vector encoding, reflecting their relationships in the semantic space. Through this operation, semantic units are attached to the corresponding main field nodes in the field evolution graph, forming an aligned subgraph. Field Evolution Diagram The existing graph only contains fields and their evolutionary relationships. Semantic units are used to match fields, retaining only the successfully aligned parts of the structure, forming a smaller, more focused graph, the alignment subgraph. The field evolution graph is a panoramic map containing all fields and their historical evolution, substitution, and synonym relationships. Semantic units are linguistic expressions extracted from unstructured text, such as "customer number," "order amount," and "arrival time." The alignment operation "attaches" these semantic units to their most matching field nodes. The alignment subgraph focuses on the effective structure from the perspective of the current task, making the computation more focused. Subsequent graph attention and feature aggregation are only performed on relevant fields and are not interfered with by irrelevant fields.

[0149] S33, Dynamic Feature Adaptation Based on Evolutionary Path: Alignment Subgraph Any primary field node traverse its evolution path over time Collect feature representations of historical field nodes and their attached semantic unit nodes. Let each historical node... The basic characteristics are Its time decay weighted expression is defined as: ;in, For the current time, For nodes timestamp, The time decay factor represents the decay of the influence of historical information over time, calculated by the average evolution period of the field nodes. The input is inversely proportional to it: ;in It is the adjustment coefficient.

[0150] Next, a graph attention mechanism is used to perform multi-hop feature aggregation on neighbor nodes: for each main field node Then, based on its neighboring nodes, a graph attention mechanism is used to calculate the importance, i.e., contribution, of different neighbors, and weighted aggregation is used to obtain the field's adaptation features at the current time: Attention weights Represents a node For nodes The semantic contribution satisfies:

[0151] ;in, For nodes The neighborhood group, For nodes The input feature vector, The feature transformation weight matrix, For attention mechanism parameters, It is a non-linear activation function. This represents a vector concatenation operation. Indicates a node with his neighbors Attention scoring after feature splicing.

[0152] The S32 scheme aims to fully utilize historical field evolution information and contextual semantics. By introducing time decay weighting and graph attention mechanisms, it dynamically generates feature representations for each field node that are more semantically stable and historically context-aware. This mechanism solves the problems of unstable feature semantics, complex field evolution paths, and missing contextual information in the context of structural drift and field changes. By controlling the weight of historical field information through time decay, newer fields dominate the expression, while older fields decay in an orderly manner, avoiding semantic confusion. Through graph attention mechanisms, it focuses on historical neighbor nodes that have made significant semantic contributions, dynamically aggregating effective features on the evolution path, improving the semantic representativeness of each field in the aligned subgraph, and ultimately forming context-enhanced node representations. This provides a stable and expressive feature foundation for fusing the mapping between structured fields and unstructured semantic units.

[0153] S34, Generate the fused feature vector set: For each main field node in the aligned subgraph Its final fusion feature representation consists of the following three parts:

[0154] 1. Self-standard eigenvectors: In other words, in the initial structured field extraction stage, the normalized raw numerical values ​​or embedded representations extracted from the source data or field descriptions, such as word vectors of field names, field positions, and frequencies of occurrence.

[0155] 2. Context-enhanced feature vectors: This feature is obtained by aggregating neighbor node information in the field evolution graph through a graph attention mechanism, reflecting the contextual semantic association of the current field on the evolutionary history path.

[0156] 3. Structure drift type encoding vector: ; indicates the evolution category of the current field, such as Stable, Synonym, Obsolete, Split, etc., represented by one-hot encoding or embedding.

[0157] The three elements are combined to form a fused feature representation: ;

[0158] Then, dimensionality reduction transformation is performed through a fully connected MLP projection layer: ;

[0159] The final fused feature vector set is formed: .

[0160] This set serves as the unified input feature space for subsequent intelligent insight models, encompassing multi-dimensional information such as structured field features, unstructured semantic alignment information, and semantic evolution history.

[0161] The final fused feature vector set can be used as the input feature space for any machine learning / deep learning model for:

[0162] Modal fusion analysis, such as text-structure joint modeling;

[0163] Intelligent reports are automatically generated, with fields grouped and sorted according to semantic relevance;

[0164] Field lifecycle evolution prediction;

[0165] Abnormal structural change detection, identifying abnormal variations in fields;

[0166] Intelligent question answering / search enhancement: Better understanding user queries based on field meaning.

[0167] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0168] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multimodal data adaptation and feature association method for intelligent insight, characterized in that, Includes the following steps: S1 standardizes and organizes structured data fields from data sources in different systems in foreign trade business scenarios to form a structured field set; based on a preset domain mapping dictionary and naming trajectory history, it extracts field naming variants, semantic heterogeneous forms and frequency features, identifies field drift patterns, and generates a structure drift annotation tag set. S2 uses the structure drift label set output from S1 as a fuzzy interference factor library to guide the semantic flux analysis of content related to structured fields in unstructured modal data, calculate content consistency and expression compactness, and extract core semantic candidate units. At the same time, a pseudo-feedback backtracking mechanism is introduced to inject the core semantic candidate units into the preset insight model to evaluate their response stability, and finally select a standardized target semantic unit set. S3 aligns the standardized target semantic unit set obtained in S2 with the structured field set execution graph structure constructed in S1 to obtain the field evolution graph, and performs dynamic adaptation based on the field inheritance relationship and temporal evolution path in the field evolution graph.

2. The multimodal data adaptation and feature association method for intelligent insight according to claim 1, characterized in that, The construction of the structured field set includes: cleaning the original data fields from different business systems in the foreign trade business scenario, removing prefixes, suffixes and special delimiters, and normalizing the field format based on a unified data type template; dividing the cleaned fields into multiple data types according to the entity objects and business attributes they describe, forming an initial structured field set.

3. The multimodal data adaptation and feature association method for intelligent insight according to claim 1, characterized in that, Based on a preset domain mapping dictionary, the field names in the initial structured field set are matched. The domain mapping dictionary defines standard field names and their corresponding synonyms, abbreviations, and common spelling variations. At the same time, the naming trajectory history database is invoked. The naming trajectory history database records the historical name change sequence of the same field in each data source as the system version iterates. By comparing the current field name with the historical name sequence, the field naming variations and implicit semantic heterogeneous forms are extracted. Statistically analyze the frequency of each naming variant and semantic heterogeneous form in historical data and current multi-source data, and calculate its frequency distribution and time decay weight; Based on frequency distribution characteristics and the naming trajectory history, field drift patterns are identified and labeled as: stable standard type, high-frequency synonym drift type, low-frequency outdated alias type, and cross-system semantic bifurcation type.

4. The multimodal data adaptation and feature association method for intelligent insight according to claim 1, characterized in that, For each field in the initial structured field set, associate it with the identified drift pattern type, all extracted named variants and their confidence and frequency feature vectors to generate a machine-readable structured drift annotation label set.

5. The multimodal data adaptation and feature association method for intelligent insight according to claim 3, characterized in that, The fuzzy interference factor library includes the standard names of each structured field, their corresponding naming variants, and noise keywords corresponding to the low-frequency outdated alias type and cross-system semantic bifurcation drift pattern. For unstructured modal data, text segmentation and basic cleaning are first performed. Based on the fuzzy interference factor library, text segments are scanned and matched to identify and label all text fragments that mention or are associated with structured fields and their variants. For each labeled text fragment, its content consistency and expression compactness are calculated. The content consistency is obtained by calculating the average vectorized semantic similarity of the same entity or attribute described in different text fragments. The expression compactness is determined by analyzing the keyword density and sentence complexity in the text fragment, and eliminating lengthy expressions containing a large amount of modifying, vague or irrelevant information; Text segments with both content consistency and expression tightness scores above the preset threshold are extracted as core semantic candidate units.

6. The multimodal data adaptation and feature association method for intelligent insight according to claim 5, characterized in that, The pseudo-feedback backtracking mechanism includes taking the core semantic candidate unit as input and injecting it into a preset insight model for foreign trade scenarios for multiple rounds of inference testing; by perturbing the input context or mixing other candidate units, observing and recording the fluctuation of the model's response to the information represented by the semantic unit, and calculating its response stability score.

7. The multimodal data adaptation and feature association method for intelligent insight according to claim 6, characterized in that, Based on the response stability score, and combined with the S content consistency and expression compactness scores, the core semantic candidate units are comprehensively ranked and screened; the screened units are semantically normalized and redundant are eliminated to form a standardized target semantic unit set for feature alignment.

8. The multimodal data adaptation and feature association method for intelligent insight according to claim 1, characterized in that, The construction of the field evolution graph includes taking each standardized field in the structured field set and its structure drift label set as the master node, and establishing directed edges between field nodes with synonym, inheritance or replacement relationships according to the naming trajectory history and drift pattern. The directed edges are accompanied by time stamps and relationship weights to form a field evolution graph that represents the semantic evolution relationship of the fields.

9. A multimodal data adaptation and feature association method for intelligent insight according to claim 8, characterized in that, The execution graph structure alignment includes taking each semantic unit in the standardized target semantic unit set as a node to be aligned, calculating its semantic similarity with each main node in the field evolution graph; based on semantic similarity threshold and relation path analysis, mapping each semantic unit node to the most relevant single or multiple structured field main nodes, and establishing bidirectional alignment edges to complete the expansion and alignment of the graph structure. Based on the structure of the aligned subgraph in the field evolution graph, the temporal evolution path between key field nodes and their associated semantic unit nodes is traversed. For each temporal evolution path, a temporal decay factor is introduced to weight the features of historical nodes. Based on the graph attention mechanism, the feature information of multi-hop neighbor nodes and along the evolution path is aggregated to dynamically generate the context-enhanced feature representation of each node.

10. A multimodal data adaptation and feature association method for intelligent insight according to claim 9, characterized in that, The S3 also includes the generation of a fusion feature vector set, specifically including extracting the standardized features, context-enhanced features, and type encoding of the drift mode of each node in the dynamically adapted alignment subgraph; concatenating and reducing the dimensionality of the three types of features to generate a fusion feature vector that integrates multimodal alignment and evolution information for each original data instance, thus forming a fusion feature vector set for subsequent intelligent insight tasks.

Images