Text classification mapping method based on hierarchical semantic matching and fact data fusion

Abstract

The invention provides a text classification mapping method based on hierarchical semantic matching and fact data fusion, and relates to the technical field of text classification. The problems that in an existing text classification mapping technology, cross-category logic errors exist, the matching accuracy rate and the coverage rate are difficult to balance, and noise-containing external fact data cannot be effectively purified and fused are solved. According to the method, text classification mapping is realized through a two-stage process: firstly, data is preprocessed and vectorized, then full-text and high-level semantic similarity is subjected to weighted fusion, and prediction mapping is generated by using double thresholds; and then purifying external fact data, finally fusing the two types of mappings and labeling sources, and outputting a high-quality mapping table. According to the method, the mapping accuracy and logic consistency are effectively improved, noise data are filtered to enhance the reliability of a result, the accuracy and the coverage rate are balanced through a double-threshold strategy, high automation of the process is realized to improve the efficiency, and the technical framework is high in universality and can adapt to multi-field text classification mapping requirements.

Description

Technical Field

[0001] This invention relates to the field of text classification technology, and in particular to a text classification mapping method based on hierarchical semantic matching and factual data fusion. Background Technology

[0002] In numerous fields such as enterprise data management, industry analysis, and knowledge engineering, it is often necessary to integrate data from different sources. This data frequently follows its own independent classification standards, such as the nationally published "National Industrial Classification of Economic Activities" (GB / T4754—2017) and the proprietary industry chain labeling system used internally by enterprises. These two systems differ significantly in textual description, hierarchical structure, and granularity of classification, making data interoperability difficult. Therefore, automatically and accurately establishing the mapping relationship between them has become a key technological bottleneck in achieving data fusion and value-added applications.

[0003] Currently, the mainstream technical solutions for solving this type of text mapping problem and their shortcomings mainly include: 1. Keyword-based or bag-of-words model-based matching methods: This is the most traditional technique. This method extracts keywords, N-gram fragments, or constructs bag-of-words vectors from the text, then calculates word overlap or similarity between texts for matching. However, these methods have significant limitations. They cannot understand the deep semantic meaning of words and are highly sensitive to synonyms, polysemous words, and different expressions, leading to low matching accuracy. For example, in practical applications, a system relying solely on the keyword "wheat" might incorrectly match "wheat cultivation," belonging to the agricultural category, to "wheat processing machinery manufacturing," belonging to the manufacturing category, because they share the core keyword but their industries are completely different. This error, arising from a lack of understanding of context and classification logic, is prevalent in traditional methods and severely affects the usability of the mapping results.

[0004] 2. Matching method based on general semantic similarity: With the development of deep learning technology, using pre-trained language models (such as BERT) to convert text into high-dimensional semantic vectors, and then judging text relevance by calculating the cosine similarity between vectors, has become a more advanced approach. This approach is the closest existing technology to this invention. Compared with keyword matching, it can better capture the deep semantics of text and solves the problems of synonyms and sentence structure variations to some extent. However, this approach still has the following core shortcomings: A lack of awareness of hierarchical structure easily leads to logical errors: General semantic models, during training, primarily learn the co-occurrence relationships of words in large corpora, forming a kind of "semantic association" ability. This ability can cause serious cross-category logical errors when processing classification system mappings. The text labels in a classification system are not isolated strings; they contain a strict hierarchical structure (e.g., category-major category-intermediate category-minor category). When performing matching, general semantic models treat all texts as flattened data at the same level, completely ignoring this inherent structural information. A typical failure is that the model might give a high similarity score because "wheat planting" and "grain combine harvester" are highly related in context, but ignore the fundamental classification difference that the former belongs to the primary industry (agriculture) while the latter belongs to the secondary industry (manufacturing). This kind of "semantically related but logically flawed" matching is a fundamental problem that current technology cannot solve.

[0005] The inability to systematically balance matching accuracy and coverage: In practical applications, it is necessary to cover as many entries as possible while ensuring the accuracy of the mapping relationship. Existing technologies typically use a single similarity threshold for judgment. If the threshold is set too high, a large number of potential correct matches will be missed (low coverage); if the threshold is set too low, a large number of incorrect matches will be introduced (low accuracy). Existing technologies lack a systematic mechanism to dynamically balance this contradiction.

[0006] The lack of an effective mechanism for integrating and verifying external factual data is a significant issue. To improve the accuracy of mapping, third-party data containing objective correlations (e.g., a directory of companies labeled with two different categories) can be introduced as "factual" evidence. However, this factual data often contains a lot of "noise." For example, a company whose main business is "canned fruit and vegetable manufacturing" may also include "intellectual property services" in its business scope. Directly associating the two labels of this company would result in a semantically unrelated "pseudo-fact" mapping from "canned fruit and vegetable manufacturing" to "intellectual property services." Existing technologies typically only passively integrate this data, lacking an intelligent means to internally verify and "purify" the factual data source, thus facing the risk of "garbage in, garbage out," and even contaminating the final mapping result due to the introduction of noisy data.

[0007] In summary, existing technologies, whether traditional keyword matching or more advanced general semantic matching, have failed to effectively solve the complex technical challenges of ensuring logical consistency, balancing accuracy and coverage, and intelligently verifying and integrating external factual data in text classification mapping. Summary of the Invention

[0008] This invention proposes a text classification mapping method based on hierarchical semantic matching and factual data fusion, implemented through an innovative two-stage process. The first stage is "hierarchical semantic prediction mapping generation," which generates a set of prediction mappings that take into account both macro-level categories and micro-level semantics using a weighted fusion algorithm. The second stage is "factual mapping purification and fusion," which constructs factual mappings by introducing external data and purifies them using semantic consistency verification. Finally, the results of the two stages are integrated to generate the final mapping table, thus solving the problems of ensuring logical consistency, balancing accuracy and coverage, and intelligently verifying and fusing external factual data in text classification mapping.

[0009] A text classification mapping method based on hierarchical semantic matching and factual data fusion, the method comprising the following steps: S1. Obtain source classification system data, target classification system data, and external bridging data containing the objective relationship between the two, and perform filtering and preprocessing on the external bridging data, retaining the records of classification labels in the source classification system with the finest granularity. S2. Select a pre-trained text vectorization model suitable for the target language, input the relevant text descriptions from the source classification system data, the target classification system data, and the external bridging data into the model, and generate a fixed-dimensional high-dimensional semantic vector representing the semantic information of each text. S3. Calculate the full-text semantic similarity matrix between each item in the source classification system and the target classification system, as well as the high-level semantic similarity matrix based only on the highest-level category text of the two systems. The two matrices are weighted and fused by preset weights to obtain the final similarity matrix. Then, a dual-threshold downgrading matching strategy is used to match the items and generate a set of prediction mappings with high / low confidence labels. S4. Using entities as intermediate keys, construct a preliminary fact mapping between the source classification system and the target classification system based on the preprocessed external bridging data. Construct context-enhanced text for the triples corresponding to the preliminary fact mapping and calculate their semantic consistency score with the target classification text. Filter according to a preset threshold to obtain a high-purity fact mapping set. S5. Merge the predicted mapping set with the high-purity fact mapping set, label the source of each mapping relationship, and generate and output the final text classification mapping table.

[0010] Furthermore, S1 includes the following steps: S11. Obtain source text classification system data: Load the first text classification system data as source text classification system data, wherein the source text classification system data has a preset hierarchical structure and corresponding hierarchical text descriptions; S12. Obtain target classification system data: Load the second text classification system data as the target classification system data, wherein the target classification system data contains text descriptions corresponding to multi-level classifications; S13. Obtain external bridging data: Load a third-party dataset as external bridging data. Each record in the external bridging data contains at least one entity, the classification result of the source text classification system corresponding to the entity, and the classification result of the target classification system corresponding to the entity. S14. Bridging data preprocessing: Perform a filtering preprocessing operation on the external bridging data, retaining only the records of the source text classification system corresponding to the entities in the external bridging data that are accurate to the finest granularity of the preset hierarchical structure.

[0011] Furthermore, S2 includes the following steps: S21. Model selection: Select a pre-trained text vectorization model suitable for the target language; S22, Vectorization Execution: Input all relevant text descriptions from the source text classification system data, the target classification system data, and the external bridging data into the pre-trained text vectorization model. The pre-trained text vectorization model generates a fixed-dimensional high-dimensional semantic vector for each text. The high-dimensional semantic vector represents the semantic information contained in the corresponding text through its position and direction in the multi-dimensional space.

[0012] Furthermore, S3 includes the following steps: S31. Calculate the full-text semantic similarity matrix. Calculate the cosine similarity between the semantic vector of each category entry in the source text classification system data and the semantic vector of each category entry in the target classification system data, forming an m×n full-text semantic similarity matrix. Where m is the total number of entries in the source text classification system data, and n is the total number of entries in the target classification system data, the... It reflects the micro-semantic correlation between all entries in the complete text description of the two classification systems; S32. Calculate the high-level semantic similarity matrix. Extract the highest-level category text for each entry in the source and target text classification systems, calculate the cosine similarity between the semantic vectors of the highest-level category texts, and form an m×n high-level semantic similarity matrix. The high-level semantic similarity matrix It reflects the similarity of the two classification systems' entries in terms of macro-industry categories; S33. Weighted fusion calculation of the final similarity matrix A hierarchical weighted guidance mechanism is adopted, through formulas.

[0013] Will and Weighted fusion yields the final similarity matrix. ;in, This is a preset hyperparameter with a value range of [0,1], used to adjust the importance of high-level category matching. A higher value emphasizes the consistency of macro-categories, and the weighted fusion calculation yields the final similarity. The calculation formula combines macroscopic structural constraints. With micro-semantic content This guides the generation of semantically relevant and logically correct judgments. S34, Dual-threshold degradation matching strategy: Balance accuracy and coverage through two rounds of matching.

[0014] Furthermore, S34 includes the following steps: S341, Round 1 – High-Confidence Matching: Setting a High Similarity Threshold Traverse the final similarity matrix All similarity scores greater than or equal to The source text classification system data entries and the target classification system data entries are paired and marked as high-confidence prediction maps; S342, Second Round – Low-Confidence Matching: For the source text classification system data entries that failed to match in the first round, a low similarity threshold is set. Among the unmatched entries, find all similarity scores between [ , Pairing source text classification system data entries with target text classification system data entries between them, and marking them as low-confidence prediction mappings; S343. Output a set of predicted maps with confidence labels, where the confidence labels are high or low.

[0015] Furthermore, S4 includes the following steps: S41. Constructing preliminary factual relationships: Using the external bridging data preprocessed in S1, with entities as the intermediate key, construct preliminary factual mapping relationships between the source text classification system data classification and the target text classification system data classification. S42. Fact purification: The initial fact mapping relationship is self-verified through semantic matching technology to filter out noise.

[0016] Furthermore, S42 includes the following steps: S421. Construct context-enhanced text: For each triple consisting of external bridging data, concatenate the entity name with the source text classification system data classification text to form context-enhanced text; S422, Calculate the semantic vector of the context-enhanced text: Input the context-enhanced text into the pre-trained text vectorization model selected in S2 to generate the semantic vector corresponding to the context-enhanced text; S423. Calculate the semantic consistency score: Calculate the cosine similarity between the semantic vector of the context-enhanced text and the semantic vector of the corresponding target classification system data classification text, and define the cosine similarity as the semantic consistency score of the corresponding fact association. S424. Screening for high-purity fact maps: Setting a fact purification threshold. The semantic consistency score is greater than or equal to Fact association, as a high-purity fact mapping, filters out semantic consistency scores below [a certain threshold]. Noise correlation; S425. Output a set of verified, high-purity fact maps.

[0017] Furthermore, S5 includes the following steps: S51. Dataset merging: Merge the set of prediction maps generated in S3 with the set of high-purity fact maps generated in step 4. During the merging process, prioritize retaining the effective association between high-confidence prediction maps and high-purity fact maps. S52. Source Labeling and Deduplication: For each mapping relationship in the merged dataset, source labeling is performed and duplicate associations are removed. The sources of the labels are divided into three categories: Intersection mapping: A mapping relationship that exists simultaneously in both the prediction mapping set and the high-purity fact mapping set, possessing the highest reliability; AI-predicted mappings only: mappings predicted solely by the AI ​​model in step three and not verified in the high-purity fact mapping set; Fact-only mappings: mapping relationships that exist only in the set of high-purity fact mappings and have not been recognized by the S3 AI model (or have not reached the AI ​​model matching threshold); S53. Generate the final mapping table: Output the final text classification mapping table containing the above three types of mapping relationships and corresponding source annotations.

[0018] A storage medium storing a computer program that, when executed by a processor, implements the aforementioned text classification mapping method based on hierarchical semantic matching and factual data fusion.

[0019] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described text classification mapping method based on hierarchical semantic matching and factual data fusion.

[0020] Compared with the prior art, the present invention achieves the following significant beneficial effects through the above technical solution: This invention's unique hierarchical weighting guidance mechanism quantifies the macroscopic structural information of the classification system and integrates it into similarity calculations, fundamentally solving the cross-category logical errors caused by the lack of hierarchical awareness in general semantic models. For example, it effectively avoids mismatching entries for "agriculture" with entries for "manufacturing," ensuring the correctness of the mapping results in terms of industry logic.

[0021] The "fact purification" step is another core innovation of this invention. Instead of passively accepting external data, it actively "cleans" the fact data source using semantic consistency checks, effectively eliminating a large amount of "pseudo-fact" noise caused by factors such as the overly broad scope of a company's operations. This significantly improves the quality of the final fused data, making the mapping results more reliable and trustworthy.

[0022] The dual-threshold degradation matching strategy provides a systematic solution to the trade-off between accuracy and coverage. It first ensures the quality of the core mapping set through high standards, then cautiously expands the coverage using slightly lower standards, and explicitly labels results from different sources. In a simulation test, the fused mapping scheme achieved 99.09% coverage of enterprise data, demonstrating its efficiency and comprehensiveness in practical applications.

[0023] The entire process of this invention, from data preprocessing and vectorization to two-stage matching and fusion, can be automated, greatly reducing the time and cost required for manual establishment and verification of mapping relationships. With appropriate hardware configuration, the core matching algorithm can complete calculations on thousands of entries within seconds, demonstrating extremely high processing efficiency.

[0024] The "hierarchical prediction + fact purification fusion" technical framework proposed in this invention has good versatility. It is not limited to a specific classification system, and its core ideas and methodology can be widely applied between any two hierarchical text classification systems that need to be mapped. By simply adjusting the corresponding parameters, it can be adapted to application scenarios in different fields. Attached Figure Description

[0025] Figure 1 This is a schematic diagram illustrating the hierarchical weighted matching process of categorized text directly through an embedding model; Figure 2 A schematic diagram illustrating the process of supplementing, filtering, and labeling the results of the previous stage by introducing relevant external data. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] A text classification mapping method based on hierarchical semantic matching and factual data fusion, the method comprising the following steps: S1. Obtain source classification system data, target classification system data, and external bridging data containing the objective relationship between the two, and perform filtering and preprocessing on the external bridging data, retaining the records of classification labels in the source classification system with the finest granularity. S2. Select a pre-trained text vectorization model suitable for the target language, input the relevant text descriptions from the source classification system data, the target classification system data, and the external bridging data into the model, and generate a fixed-dimensional high-dimensional semantic vector representing the semantic information of each text. S3. Calculate the full-text semantic similarity matrix between each item in the source classification system and the target classification system, as well as the high-level semantic similarity matrix based only on the highest-level category text of the two systems. The two matrices are weighted and fused by preset weights to obtain the final similarity matrix. Then, a dual-threshold downgrading matching strategy is used to match the items and generate a set of prediction mappings with high / low confidence labels. S4. Using entities as intermediate keys, construct a preliminary fact mapping between the source classification system and the target classification system based on the preprocessed external bridging data. Construct context-enhanced text for the triples corresponding to the preliminary fact mapping and calculate their semantic consistency score with the target classification text. Filter according to a preset threshold to obtain a high-purity fact mapping set. S5. Merge the predicted mapping set with the high-purity fact mapping set, label the source of each mapping relationship, and generate and output the final text classification mapping table.

[0028] Furthermore, S1 includes the following steps: S11. Obtain source classification system data (System A): Load the data for the first text classification system, such as the "National Economic Industry Classification". This data contains a clear hierarchical structure, such as "category", "major category", "intermediate category", "minor category" and their corresponding text descriptions; S12. Obtain target classification system data (system B): Load the data for the second text classification system. This data also contains text descriptions of multi-level classifications; S13. Obtain external bridging data: Load a third-party dataset containing objective relationships between System A and System B, such as a directory of enterprise data. Each record in this data must contain at least one entity (such as an enterprise name) and its corresponding System A and System B categories; S14. Bridging Data Preprocessing: Perform critical filtering on the external bridging data. To ensure the baseline quality of the factual data source, only records with classification labels in System A accurate to the finest granularity (e.g., "minor category") are retained. This aims to eliminate the significant uncertainty and potential noise introduced by coarse classification granularity (e.g., only labeled to "category" or "major category"). In one specific embodiment, this step refines the original 15,301 enterprise records to 12,371 high-precision records.

[0029] This step aims to prepare all the data sources required for subsequent processes and to perform the necessary cleaning and formatting.

[0030] Furthermore, S2 includes the following steps: S21. Model Selection: Select a pre-trained text vectorization model suitable for the target language. In a preferred embodiment, the shibing624 / text2vec-base-chinese model is used. S22. Vectorized Execution: Input all relevant text descriptions (such as national standard classification names, industry chain link names, company names, etc.) from system A, system B, and external bridging data into the pre-trained model to generate a fixed-dimensional semantic vector (e.g., 768-dimensional) for each text. The position and orientation of these vectors in multidimensional space represent the semantic information they contain.

[0031] This step utilizes a pre-trained deep learning language model to convert all involved text descriptions into high-dimensional mathematical vectors that can be computed by machines.

[0032] Furthermore, S3 includes the following steps: S31. Calculate the full-text semantic similarity matrix. The cosine similarity between the semantic vector of each category entry in system A and the semantic vector of each category entry in system B is calculated to form an m×n similarity matrix. , where m and n are the total number of entries in system A and system B, respectively. This matrix reflects the micro-semantic correlation between all entries in the complete text description of the two systems; S32. Calculate the high-level semantic similarity matrix. Extract only the highest-level category text (e.g., the "category" name in the national standard classification) for each entry in system A and system B, and calculate the cosine similarity of the semantic vectors of these high-level texts to form another m×n similarity matrix. This matrix reflects the similarity of entries in terms of their affiliation within macro-industry categories; S33. Weighted fusion calculation of the final similarity. To address the cross-category logical error problem mentioned in the background section, this invention proposes a hierarchical weighted guidance mechanism. The two similarity matrices are weighted and fused using the following formula to obtain the final similarity matrix. :

[0033] in, This is a preset hyperparameter with a value range of [value range missing], used to adjust the importance of high-level category matching. The larger the value, the more emphasis is placed on the consistency of macro-categories. In a preferred embodiment, the weights are experimentally adjusted. The value is set to 0.4. The essence of this formula is to apply macroscopic structural constraints... With micro-level semantic content By combining these approaches, the model can be guided to make judgments that are both semantically relevant and logically correct. S34, Dual-threshold degradation matching strategy: Balance accuracy and coverage through two rounds of matching.

[0034] This stage does not rely on external bridging data; it generates a set of "predictive maps" solely through intelligent analysis of the texts of system A and system B.

[0035] Furthermore, S34 includes the following steps: S341, First Round (High-Confidence Matching): Set a high similarity threshold. Traversal A matrix that includes all similarity scores greater than or equal to The (system A entry, system B entry) pairings are labeled as "high-confidence prediction mappings". In a preferred embodiment, Set to 0.75; S342, Second Round (Low-Confidence Matching): For System A items that failed to match successfully in the first round, a lower similarity threshold is set. Among these unmatched entries, find all those with similarity scores between [ ]. , Pairing source text classification system data entries with target text classification system data entries between them, and marking them as low-confidence prediction mappings; S343. The output of this stage is a set of prediction maps with confidence labels (high / low).

[0036] Furthermore, S4 includes the following steps: S41. Constructing preliminary factual relationships: Using preprocessed external bridging data, construct preliminary factual mapping pairs (Classification A, Classification B) with entities (such as company names) as intermediate keys. S42. Fact Purification (Core Innovation): To solve the noise problem of fact data sources themselves, this invention proposes a "fact purification" method, which essentially uses semantic matching technology to perform self-verification of fact data.

[0037] The core of this phase is to introduce and "cleanse" external bridging data to build a highly reliable "fact map" and merge it with the predictive map from the first phase.

[0038] Furthermore, S42 includes the following steps: S421. For each triple (entity name, category A, category B) consisting of bridging data, construct a context-enhanced text. For example, concatenate the "entity name" with the "category A text" to form a descriptive text such as "XX Company, whose main business is:...". S422. Calculate the semantic vector of the context-enhanced text; S423. Calculate the cosine similarity between the vector and the semantic vector of the corresponding “System B Classification” text. This score is defined as the “semantic consistency score” of this fact association. S424. Set a fact purification threshold. Only those with a semantic consistency score greater than or equal to Only factual correlations that meet this threshold are considered valid and high-purity, and are retained; those below this threshold are considered noise and filtered out. In a preferred embodiment, The value is set to 0.6. This step effectively filters out semantically irrelevant pseudo-fact associations such as "fruit canning" and "intellectual property." In one specific embodiment, this step refines 121,186 preliminary factual associations into 40,309 high-purity factual associations, significantly improving data quality. S425. The output of this stage is a set of verified and purified "high-purity fact maps".

[0039] Furthermore, S5 includes the following steps: S51. Dataset merging: Merge the "prediction mapping" (especially the high-confidence part) generated in the first stage with the "high-purity fact mapping" generated in the second stage; S52. Source Labeling and Deduplication: In the final mapping table, the source of each mapping relationship is labeled, mainly in three categories: Intersection mapping: Intersection mapping: A relationship that exists simultaneously in both prediction mapping and fact mapping, and has the highest credibility; AI-predicted mappings only: Relationships predicted solely by AI models but not corroborated by factual data; Fact-only mapping: Relationships that exist in high-purity fact data but have not been recognized by AI models (or have not reached a threshold); S53. Generate the final mapping table: Output a final mapping table containing the above three types of mapping relationships and annotated with source tags. This table has a clear structure and high transparency, allowing users to evaluate the credibility of each mapping based on the source tags.

[0040] To ensure that those skilled in the art can reproduce this invention, the following table lists the key parameter configurations used in a preferred embodiment:

[0041] A storage medium storing a computer program that, when executed by a processor, implements the aforementioned text classification mapping method based on hierarchical semantic matching and factual data fusion.

[0042] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the text classification mapping method based on hierarchical semantic matching and factual data fusion as described in any one of claims 1-8.

[0043] When this invention maps and annotates enterprise directories with a data volume of hundreds of thousands of records, the hardware platform used has up to 16GB of video memory and 20GB of RAM, running in a Python 3.9 software environment. This patented achievement has relatively low platform memory and video memory usage. During annotation testing, it achieved an overall precision of 9.37%, a recall of 16.51%, and an F1 score of 11.96%. For this application example, the above test data is sufficient to demonstrate that the solution effectively annotates the core industry chain; and manual evaluation of the annotation results indicates that the annotation results are relatively accurate and reach a practical level.

[0044] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A text classification mapping method based on hierarchical semantic matching and fact data fusion, characterized in that, The method comprises the following steps: S1, obtaining source classification system data, target classification system data and external bridging data containing objective correlation of both, and performing screening and preprocessing on the external bridging data to retain records of classification labels in the source classification system accurate to the finest granularity; S2, selecting a pre-trained text vectorization model suitable for the target language, inputting related text descriptions in the source classification system data, target classification system data and external bridging data into the model, and generating a fixed-dimension high-dimensional semantic vector representing the semantic information of each text; S3, calculating a full-text semantic similarity matrix between entries of the source classification system and the target classification system and a high-level semantic similarity matrix based only on the text of the highest level categories of both, weighting and fusing the two matrices by a preset weight to obtain a final similarity matrix, and using a double-threshold degradation matching strategy to match the entries to generate a predicted mapping set with high / low confidence labels; S4, constructing a preliminary fact mapping between the source classification system and the target classification system based on the preprocessed external bridging data with entities as intermediate keys, constructing context-enhanced text for the triple corresponding to the preliminary fact mapping, calculating the semantic consistency score of the context-enhanced text with the target classification text, and screening to obtain a high-purity fact mapping set according to a preset threshold; S5, merging the predicted mapping set and the high-purity fact mapping set, labeling the source of each mapping relationship, generating and outputting a final text classification mapping table. 2.The text classification mapping method based on hierarchical semantic matching and fact data fusion according to claim 1, characterized in that, S1 comprises the following steps: S11, obtaining source text classification system data: loading first text classification system data as source text classification system data, the source text classification system data having a preset hierarchical structure and text descriptions corresponding to the levels; S12, obtaining target classification system data: loading second text classification system data as target classification system data, the target classification system data containing text descriptions corresponding to multiple levels of classification; S13, obtaining external bridging data: loading a third-party data set as external bridging data, each record in the external bridging data containing at least one entity, a source text classification system classification result corresponding to the entity, and a target classification system classification result corresponding to the entity; S14, bridging data preprocessing: performing screening and preprocessing operations on the external bridging data, and retaining only records in the external bridging data in which the source text classification system classification result corresponding to the entity is accurate to the finest granularity of the preset hierarchical structure. 3.The text classification mapping method based on hierarchical semantic matching and fact data fusion according to claim 1, characterized in that, S2 comprises the following steps: S21, model selection: selecting a pre-trained text vectorization model suitable for the target language; S22, vectorization execution: inputting all related text descriptions in the source text classification system data, target classification system data and external bridging data into the pre-trained text vectorization model, generating a fixed-dimension high-dimensional semantic vector for each text by the pre-trained text vectorization model, and the high-dimensional semantic vector representing the semantic information contained in the corresponding text through its position and direction in a multi-dimensional space. 4.The text classification mapping method based on hierarchical semantic matching and fact data fusion according to claim 1, characterized in that, S3 comprises the following steps: S31、calculating the full-text semantic similarity matrix : calculating the cosine similarity between the semantic vector of each classification item in the source text classification system data and the semantic vector of each classification item in the target classification system data, forming an m x n full-text semantic similarity matrix , where m is the total number of items in the source text classification system data, n is the total number of items in the target classification system data, and the reflects the microscopic semantic correlation of all items in the two classification systems in the complete text description; S32, calculating a high-level semantic similarity matrix : extracting the highest level category text of each entry in the source text classification system data and the target classification system data, calculating the cosine similarity between the semantic vectors of the highest level category text, forming a high-level semantic similarity matrix of m x n , the high-level semantic similarity matrix reflects the similarity of the two classification system entries in the macro industry category; S33, calculating a final similarity matrix by weighted fusion : using a hierarchical weight guiding mechanism, through formula Will With Weighted fusion, get the final similarity matrix ; Wherein, is a preset hyperparameter, the value range is [0, 1], used to adjust the importance of high-level category matching, The greater the value, the greater the emphasis on the consistency of macro category, the final similarity is calculated by weighted fusion The calculation formula combines macro structure constraint And micro semantic content , guide to generate semantic related and logically correct judgment; S34, double-threshold degradation matching strategy: balancing accuracy and coverage through two rounds of matching.

5. The text classification mapping method based on hierarchical semantic matching and fact data fusion according to claim 4, characterized in that, S34 comprises the following steps: S341、first round - high confidence matching: set a high similarity threshold , traverse the final similarity matrix , pair all source text classification system data entries and target classification system data entries with similarity scores greater than or equal to , mark as high confidence prediction mapping; S342、second round - low confidence matching: for the source taxonomy data entries that are not successfully matched in the first round, set a low similarity threshold Among the unmatched entries, find all source taxonomy data entries and target taxonomy data entries pairs with similarity scores between [ , ) and mark them as low confidence predicted mappings; S343, output the predicted mapping set with the credibility label, the credibility label is high or low.

6. The text classification mapping method based on hierarchical semantic matching and fact data fusion according to claim 1, characterized in that, S4 includes the following steps: S41, construct a preliminary fact association: use the externally bridged data preprocessed in S1 to construct a preliminary fact mapping relationship pair between the source text classification system data classification and the target classification system data classification, taking the entity as the intermediate key; S42, fact purification: self-check the preliminary fact mapping relationship through semantic matching technology to filter noise.

7. The text classification mapping method based on hierarchical semantic matching and fact data fusion according to claim 6, characterized in that, S42 includes the following steps: S421, construct context-enhanced text: for each triple composed of external bridged data, concatenate the entity name with the source text classification system data classification text to form context-enhanced text; S422, calculate the semantic vector of the context-enhanced text: input the context-enhanced text into the pre-trained text vectorization model selected in S2 to generate the semantic vector corresponding to the context-enhanced text; S423, calculate the semantic consistency score: calculate the cosine similarity between the semantic vector of the context-enhanced text and the semantic vector of the corresponding target classification system data classification text, and define the cosine similarity as the semantic consistency score of the corresponding fact association; S424, screening high-purity fact mappings: setting a fact purification threshold , retaining fact associations with semantic consistency scores greater than or equal to as high-purity fact mappings, and filtering noise associations with semantic consistency scores less than ​ S425, output the verified high-purity fact mapping set. 8.The text classification mapping method based on hierarchical semantic matching and fact data fusion according to claim 1, characterized in that, S5 includes the following steps: S51, data merging: merge the predicted mapping set generated in S3 with the high-purity fact mapping set generated in step four, and preferentially retain the effective associations of high-credibility predicted mapping and high-purity fact mapping during the merging process; S52, source labeling and deduplication: source labeling and deduplication are performed on each mapping relationship in the merged data set, and the labeled sources are divided into three categories: Intersection mapping: the mapping relationship existing in both the predicted mapping set and the high-purity fact mapping set has the highest credibility; Only AI predicted mapping: a mapping relationship predicted only by the AI model in step three and not confirmed in the high-purity fact mapping set; Only fact mapping: a mapping relationship existing only in the high-purity fact mapping set and not identified by the AI model in S3 (or not reaching the AI model matching threshold); S53, generate the final mapping table: output the final text classification mapping table containing the above three types of mapping relationships and corresponding source labels.

9. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to realize the text classification mapping method based on hierarchical semantic matching and fact data fusion according to any one of claims 1-8.

10. A computer device, comprising: It includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to realize the text classification mapping method based on hierarchical semantic matching and fact data fusion according to any one of claims 1-8.

Images