A scientific-technology knowledge entity disambiguation and cross-carrier alignment method based on asymmetric semantic manifold mapping

Abstract

The application discloses a scientific-technology knowledge entity disambiguation and cross-carrier alignment method based on asymmetric semantic manifold mapping, atomized concept representation is constructed by carrying out morphological decoupling and normalization processing on scientific paper and technical patent keywords; semantic anchor points are screened based on rank-frequency distribution and curvature mutation detection mechanism; semantic vectors are generated by using a double-tower type Transformer model and are constrained on a unit hypersphere manifold; semantic aggregation is realized by combining a near neighbor topological structure and a high similarity threshold value, and a stable concept unit is formed; a cross-language unified latent space is established through an asymmetric encoding and contrast learning mechanism; finally, one-to-one alignment of English and Chinese concepts is realized based on near neighbor search and an adaptive threshold value determination; the application can effectively solve the alignment difficulty problem caused by the difference in cross-carrier semantic expression, and improve the knowledge correlation precision and system robustness.

Description

Technical Field

[0001] This invention relates to the fields of knowledge engineering and natural language processing, specifically to a method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping. Background Technology

[0002] Because scientific papers and technological patents differ significantly in writing paradigms, linguistic environments, and target audiences, the keyword generation and annotation process is highly dependent on the author's expression habits. This objectively leads to severe fragmentation in the representation of knowledge entities. Specifically, the same technical concept or scientific entity commonly exhibits diverse synonyms, coexistence of abbreviations and full names, inconsistent granularity of expression, and cross-linguistic semantic asymmetry at the textual level. These phenomena cause knowledge units with the same connotation to be split into multiple isolated terms at the string level. Especially when jointly analyzing heterogeneous Chinese and English data, this knowledge representation bias caused by language barriers directly leads to distorted measurement of the scientific-technological evolution path, making it difficult to reconstruct the true knowledge flow map.

[0003] Given the limitations of existing technologies, how to penetrate the surface of strings and delve into their semantic core to establish a high-fidelity framework for the standardization and association of knowledge entities has become a critical issue that urgently needs to be addressed in the field of scientific and technological intelligence analysis. Traditional cleaning methods based on dictionaries or manual rules often face challenges such as low coverage, poor scalability, and inability to capture implicit semantic relationships when dealing with massive amounts of dynamic and highly segmented scientific and technological texts. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping, which solves the problems of heterogeneity of expression, fragmentation of concept representation, and inconsistency of cross-language semantic distribution between scientific papers and technical patents in the context of big data in science and technology.

[0005] Technical Solution: The present invention provides a method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping, comprising the following steps:

[0006] Step 1: Perform heuristic multidimensional morphological decoupling and normalization on keywords derived from scientific papers and technical patents, construct a many-to-one mapping relationship from keywords to atomized concepts, and realize the atomized representation of knowledge entities;

[0007] Step 2: Perform frequency statistics and rank-frequency distribution modeling on the normalized entity set, adaptively determine the frequency threshold based on the curvature change detection mechanism of the cumulative distribution function, and select the core semantic anchor points according to the threshold;

[0008] Step 3: Construct a dual-tower coding model, map the selected semantic anchors into high-dimensional semantic vectors, and normalize the semantic vectors, constraining them to a unit hyperspherical manifold;

[0009] Step 4: Construct a topological structure based on a nearest neighbor graph within the monolingual manifold space. Based on a preset high similarity threshold, perform neighborhood aggregation on semantic vectors and merge vectors that are highly similar in meaning into a unified conceptual unit.

[0010] Step 5: Construct an asymmetric encoding architecture, mapping Chinese and English concept units to a unified latent space through different encoding functions, and using a contrastive learning strategy to optimize the cross-linguistic semantic representation within this space;

[0011] Step 6: In the unified latent space, perform cross-language nearest neighbor search for each Chinese concept unit, and at the same time, make judgments based on the maximum similarity criterion and adaptive similarity threshold to establish a one-to-one mapping relationship between Chinese concept units and English concept units, and perform isolation processing on concept units that do not meet the threshold conditions.

[0012] Furthermore, in step 1, the heuristic multidimensional morphological decoupling and normalization process is as follows: for English keywords, case unification, separator normalization, word form restoration and abbreviation expansion are performed; for Chinese keywords, semantic boundary detection, functional suffix stripping and synonym mapping are performed.

[0013] Furthermore, in step 2, the curvature abrupt change detection mechanism based on the cumulative distribution function adaptively determines the frequency threshold by: calculating the second difference of the function and taking the frequency corresponding to the position with the largest curvature change as the adaptive frequency threshold.

[0014] Furthermore, in step 3, all semantic vectors are constrained to a unit hyperspherical manifold, and the semantic similarity between any two vectors is obtained by calculating their inner product.

[0015] Furthermore, in step 4, the topological structure based on the nearest neighbor graph is a hierarchical navigation small-world graph; the high similarity threshold is adaptively determined in the high similarity region on the right based on the probability density function of the monolingual similarity distribution; by identifying the connected components in the nearest neighbor graph structure, the semantic vectors are aggregated to form a unified conceptual unit.

[0016] Furthermore, in step 5, the asymmetric encoding architecture consists of two encoding functions with independent parameters: a query-side encoding function and an anchor-side encoding function. The query-side encoding function is used to encode Chinese concept units, and the anchor-side encoding function is used to encode English concept units.

[0017] Furthermore, in step 5, the contrastive learning strategy employs an optimization objective function with a temperature parameter to adjust the smoothness of the cross-language matching probability distribution.

[0018] Furthermore, in step 6, the adaptive similarity threshold is dynamically determined by scanning a preset similarity interval and identifying the inflection point of the hit rate change rate; the one-to-one mapping relationship forces each Chinese concept unit to match only one English concept unit with the highest similarity that exceeds the threshold, thereby ensuring the uniqueness of the alignment result.

[0019] The present invention discloses a system for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping, comprising:

[0020] The heuristic morphological decoupling and atomization normalization module is used to perform morphological decoupling and normalization on multi-source heterogeneous keywords and construct atomized concept representations.

[0021] An adaptive semantic density estimation and noise reduction module is used to filter core semantic anchors and noise by modeling rank-frequency distribution and detecting curvature abrupt changes;

[0022] The semantic vector manifold representation module is used to map entities to semantic vectors and constrain them to a unit hyperspherical manifold;

[0023] The semantic consensus aggregation module is used to aggregate neighborhoods on the nearest neighbor graph topology using a high similarity threshold to form stable and compact conceptual units.

[0024] The asymmetric latent space mapping and alignment determination module is used to construct a unified latent space through asymmetric encoding and contrastive learning, and to perform one-to-one strong association mapping of cross-linguistic concept units based on the maximum similarity criterion and adaptive threshold.

[0025] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0026] (1) Reduce the impact of cross-domain semantic differences and achieve high-precision knowledge association. Unlike traditional methods based on string matching or shallow dictionary mapping, this invention constructs a unit hyperspherical manifold space, which to some extent reduces the impact of differences in expression between scientific papers and technical patents on semantic matching. At the same time, it introduces an asymmetric coding paradigm to model the problem of inconsistent expression granularity across contexts, which helps to improve the accuracy of cross-carrier semantic alignment and reduce semantic offset.

[0027] (2) It has adaptive noise reduction capabilities, improving the stability of massive data processing. This invention introduces an adaptive anchor point recognition mechanism based on curvature change detection, which can dynamically determine the screening threshold according to the distribution characteristics of data in different fields, thereby identifying core semantic units and filtering long-tail noise. Compared with the fixed threshold method, this mechanism reduces the deviation caused by manually set parameters, which helps to maintain the stability and consistency of entity representation in large-scale heterogeneous data environments.

[0028] (3) Achieving semantic aggregation and improving the compactness of knowledge representation. By implementing a neighborhood aggregation strategy under similarity threshold constraints, this invention can merge synonyms, abbreviations, and morphological variations into unified conceptual units, thereby alleviating the fragmentation problem of knowledge representation to a certain extent. Under the premise of maintaining semantic consistency, the compression of the concept level helps to improve the efficiency of subsequent knowledge graph construction and analysis processes.

[0029] (4) Ensure alignment uniqueness and improve the interpretability of results. This invention achieves a one-to-one matching relationship between cross-language entities by introducing the dual constraints of the maximum similarity criterion and the similarity threshold, thereby reducing the ambiguity caused by multiple matching. At the same time, by constructing a multi-level mapping relationship structure, the mapping path from the normalized concept to the original term is preserved, which helps to improve the traceability and interpretability of the results. Attached Figure Description

[0030] Figure 1 A flowchart of the present invention. Detailed Implementation

[0031] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0032] like Figure 1 As shown, this embodiment of the invention provides a method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping, including the following steps:

[0033] S1. Heuristic Multidimensional Morphological Decoupling and Feature Atomization Standardization: For keywords from heterogeneous scientific papers and technical patents, a multilingual morphological filtering engine is constructed. Morphological decoupling and standardization operations are performed on the English and Chinese sides respectively. Finally, a many-to-one association mapping relationship of "concept-atomic item" is established to achieve preliminary dimensionality reduction and semantic focus at the character level.

[0034] In this embodiment, step S1 specifically includes the following sub-steps:

[0035] Step S11: Construct a multi-source keyword input set. Specifically, this involves extracting keyword sets from scientific paper databases and technology patent databases as the original input corpus. Considering the differences in writing paradigms between papers and patents, the same technical concepts may have diverse expressions.

[0036] Step S12: English side morphological decoupling and normalization. Perform morphological decoupling and suffix stripping operations on English keywords based on regular semantic operators, specifically including:

[0037] S121: Case-sensitive processing, converting all English keywords to lowercase;

[0038] S122: Delimiter normalization processing, uniformly replacing hyphens, underscores and forward slashes with spaces;

[0039] S123: Lexical restoration and suffix stripping: a rule-based stemming algorithm is used to restore plural forms, gerunds and past tenses, and a domain stop word list is combined to avoid misprocessing of proper nouns.

[0040] S124: Abbreviation extension processing, based on the domain abbreviation dictionary, performs full name mapping on the abbreviations in the keywords, and establishes a bidirectional index relationship between abbreviations and full names.

[0041] Step S13: Perform semantic boundary detection (based on word segmentation model) and redundant information removal on Chinese keywords, specifically including:

[0042] S131: Functional suffix stripping, based on a predefined suffix list and the longest matching strategy, removes redundant suffixes such as "technology", "method", "system", and "model", and uses word frequency to perform semantic integrity verification to avoid over-pruning.

[0043] S132: Synonym standardization processing, using a domain synonym mapping table to unify different expression forms;

[0044] S133: Format cleaning process, removing extra spaces and mixed full-width / half-width formatting noise from keywords.

[0045] Step S14: Construct the "original keyword - normalized atom item" association mapping relationship. Based on the keyword sets processed by the English and Chinese sides, establish a many-to-one mapping relationship. This method maps different morphological variants to normalized atomic terms. The mapping relationship is stored using a hash index structure, where the key is the normalized feature representation of the keyword, and the value is the identifier of the corresponding atomic term, thus supporting a backtracking mapping from the normalized result to the original term.

[0046] Step S2, Adaptive core anchor point extraction based on semantic space density: This step aims to use probability distribution modeling to perform sparsity compression on the normalized entity set. By performing rank-frequency distribution feature extraction on keywords, core semantic anchor points are adaptively identified and random noise in the long-tail distribution is removed.

[0047] In this embodiment, step S2 specifically includes the following sub-steps:

[0048] Step S21: Frequency Statistics and Rank-Frequency Distribution Modeling. Perform word frequency statistics on the keyword set after normalization in Step S1. Let the normalized keyword set be... ,in This represents the total number of different keywords. For any given keyword... Its frequency of occurrence is recorded as The keywords are sorted in descending order of frequency to obtain the sequence: ,in This indicates the number after sorting by frequency. Using keywords, a rank-frequency distribution relationship can be constructed.

[0049] Step S22: Constructing the cumulative distribution function. Based on the sorted frequency sequence, construct the normalized cumulative distribution function. Defined as , This represents the sequence number of the keywords after being sorted in descending order of frequency.

[0050] Step S23: Curvature abrupt change detection. For discrete functions... Curvature analysis is performed to identify key inflection point locations. First, a first-order difference is constructed. Further construct the second-order difference The curvature abrupt change point is determined by calculating the absolute value of the second-order difference. , .in, This indicates the location where the curvature change is most significant, corresponding to the inflection point of the cumulative distribution curve. Used to characterize the local curvature of the cumulative distribution curve.

[0051] Step S24: Adaptive threshold determination and semantic anchor point selection. Based on the inflection point position obtained in step S23... The adaptive frequency threshold is defined as: For any keyword Its screening and judgment rules are as follows: .in, This represents the set of core semantic anchor points obtained through filtering, while other keywords are treated as long-tail noise and filtered out.

[0052] S3. Domain-Aware Multi-Dimensional Semantic Vector Manifold Representation: The goal of this step is to construct a dual-tower feature extraction network based on the Transformer architecture, mapping the anchor-filtered atomic keywords to a high-dimensional continuous vector space, and then using global... Norm normalization constrains knowledge entities to a unit hyperspherical manifold.

[0053] In this embodiment, step S3 specifically includes the following sub-steps:

[0054] Step S31: Construction of the Transformer Dual-Tower Encoding Network. A dual-tower feature extraction network based on a pre-trained language model is constructed. The pre-trained language model is obtained through pre-training on a large-scale corpus. Let the English encoder be a function... The Chinese encoder is a function. Both consist of multi-layer Transformer encoders, with keyword representations as input and corresponding semantic vector representations as outputs. The dual-tower structure is used to model the semantic spaces of Chinese and English respectively, providing a basic representation for subsequent cross-language semantic alignment.

[0055] Step S32: Keyword embedding vector generation. The semantic anchor set obtained in step S2 is used... Input into the encoder for the corresponding language. For any keyword Its embedding vector representation is defined as .in, Keywords Language type For the embedding vector dimension. Further, the embedding vector... It can be obtained from the sequence-level semantic representation output by the Transformer, including but not limited to the hidden state corresponding to the [CLS] tag or the pooled representation result.

[0056] Step S33: Global Norm normalization. For embedded vectors implement Norm normalization process yields a normalized vector. Among them, the vector Norm is defined as After normalization, it satisfies: , Indicates the first The first keyword embedding vector of the first keyword Each dimension component.

[0057] Step S34: Representation of the unit hyperspherical manifold constraint. The set of all normalized embedding vectors is represented as: Then we have: That is, all semantic vectors are constrained to On a unit hyperspherical manifold. express In 3D space A 1 / 2-dimensional hypersphere. Through the above processing, any two keywords... and The semantic similarity between them can be measured by cosine similarity: This eliminates the influence of differences in vector magnitude, making semantic similarity primarily determined by vector direction.

[0058] Step S4: High-threshold semantic consensus aggregation based on topological graph constraints. This step constructs a hierarchical navigation small-world (HNSW) topological index structure within the monolingual manifold space, captures the local semantic topological relationships of entities using near nearest neighbor search, and performs neighborhood aggregation based on similarity threshold constraints under high-confidence consensus threshold constraints, merging the sets of entity vectors with extremely close semantic distances within the manifold space into unified conceptual units.

[0059] Step S41: HNSW Topology Index Construction. This involves constructing the unit hypersphere embedding vector set obtained in Step S3. Constructing a hierarchical navigation small world graph structure , where the set of nodes Corresponding embedding vector set edge set This represents the nearest neighbor connection relationship. During the construction process, for each node... Establish multiple adjacent edges that satisfy: .in, The maximum number of nearest neighbors allowed for each node. Additionally, a candidate queue size parameter is introduced during the approximate nearest neighbor search process. It is used to control the trade-off between search accuracy and efficiency.

[0060] Step S42: Monolingual semantic similarity measurement. For any two normalized vectors within the same language manifold space... and Its semantic similarity is defined as: ,because ,therefore: .

[0061] Step S43: High-confidence consensus threshold modeling. Based on the similarity set of all candidate nearest neighbor pairs: Construct its probability density function estimate The kernel density estimation method is used to express it as follows: ,in, For kernel function, This is the bandwidth parameter. A conservative threshold is selected in the high similarity region to the right of the probability density function. This is used to constrain aggregation to be performed only under conditions of high-confidence semantic consistency. The threshold for monolingual aggregation is adaptively determined based on the similarity distribution.

[0062] Step S44: Semantic aggregation determination. Define the aggregation determination function as follows: ,when At that time, the node and They are merged into the same conceptual unit.

[0063] Step S45: Constructing conceptual units based on transitive closures. Constructing an undirected graph: Then each connected component Defined as a conceptual unit: .in, This is a set of edges that satisfy the high similarity threshold condition. The aggregation process is constrained by the HNSW topology and is only performed within the constraints of the local nearest neighbor graph, thereby avoiding semantic mis-aggregation caused by global dense connections.

[0064] Step S46: Determine the representative items of the conceptual unit. For each conceptual unit... The element with the highest keyword frequency within it is selected as the representative tag, defined as: .

[0065] Step S5: Cross-carrier knowledge alignment based on asymmetric latent space projection. This step constructs a heterogeneous latent space mapping bridge, adopts an asymmetric coding paradigm, defines Chinese concept units as query-end manifolds and English concept units as anchor space channel ends, establishes a unidirectional projection path from the Chinese concept space to the English concept space, and uses it to support semantic alignment between different knowledge carriers such as scientific papers and technical patents.

[0066] In this embodiment, step S5 specifically includes the following sub-steps:

[0067] Step S51: Construction of Asymmetric Encoding Architecture. A dual encoder structure based on the asymmetric paradigm is constructed to map Chinese and English concept units to a unified latent space. Let the set of Chinese concept units be... The English concept unit set is Define the query-side encoding function and the anchor-side encoding function as follows: For any Chinese concept unit With English concept unit Their vector representations are as follows: The two encoding functions are built on the Transformer architecture, but their parameters are independent of each other.

[0068] Step S52: Cross-lingual contrastive learning embedding space construction. The encoder parameters are initialized using a multilingual pre-trained model, enabling semantic representations from different languages ​​to map to a shared latent space. Furthermore, domain-specific Chinese and English terminology is used to construct the set. The encoder is fine-tuned and optimized to enhance cross-language semantic alignment capabilities within a specific domain. This indicates the number of known pairs of Chinese and English terms or concepts.

[0069] Step S53: Definition of Asymmetric Mapping Function. Define the cross-language matching score function as an inner product: Based on this, temperature parameters are introduced. Construct a normalized matching probability distribution: .in, Indicates the current batch or candidate English concept unit index. This is a temperature parameter used to adjust the smoothness of the distribution.

[0070] Step S54: Construct the contrastive learning objective function. For positive sample pairs... Construct an optimization objective based on maximum likelihood estimation: By minimizing the loss function This maximizes the similarity of correctly aligned cross-linguistic concept units in the latent space, while simultaneously increasing the distance between them and negative samples. Negative samples are sampled from misaligned Chinese or English concept units in the same batch.

[0071] Step S55: Determine the cross-carrier semantic alignment mapping. For any Chinese concept unit... Its optimal alignment result in the English concept space is defined as: This establishes a one-way mapping relationship from the Chinese knowledge domain to the English knowledge domain: .

[0072] Step S56: Explanation of the Asymmetric Mapping Mechanism. Through the aforementioned asymmetric encoding and contrastive learning mechanism, the query end (Chinese context) and the anchor end (English context) achieve semantic alignment in a unified latent space while maintaining their respective semantic distribution characteristics. This one-way mapping mechanism can effectively characterize the knowledge transfer path from scientific knowledge to technological applications, improving the accuracy and stability of cross-carrier semantic matching.

[0073] Step S6: Alignment determination and verification under global-local consistency constraints. This step uses the English concept manifold space as the global benchmark anchor point, performs cross-space nearest neighbor optimization for each Chinese concept, and combines a threshold-sensitive feedback mechanism to make the final alignment determination.

[0074] In this embodiment, step S6 specifically includes the following sub-steps:

[0075] Step S61: Cross-spatial nearest neighbor search. For each Chinese concept unit... Its query vector is represented as In the English concept vector set Execute Top- Nearest neighbor search yields a candidate set: The similarity function is defined as cosine similarity: In this embodiment, since the vector has been normalized, we have: ,and In this embodiment, the number of candidate nearest neighbors is set to 10.

[0076] Step S62: Define the cross-language alignment decision operator. Define the alignment decision operator: ,in When the judgment value is 1, a mapping relationship is established: .

[0077] Step S63: Threshold scanning and performance metric definition. Within the preset threshold range... Step size is For each candidate threshold The following evaluation metrics are defined:

[0078] (1) Hit Rate:

[0079] (2) Average Matches: ,in, This is an indicator function.

[0080] Step S64: Adaptive threshold determination. Based on hit rate function. Calculate its discrete first-order difference based on the changing trend of . Select the threshold corresponding to the inflection point where the rate of change is significant: At the same time, combined with the average number of matches Constraints should be imposed to avoid multiple matching. Used for monolingual concept aggregation Used for final alignment determination across languages.

[0081] Step S65: Conservative isolation processing. For any Chinese concept unit... ,like: If multiple English concept units have the same highest similarity score as the Chinese concept unit, or if the difference between the highest and second-highest similarity scores is less than a preset interval parameter, then... If it does not meet the uniqueness condition, then the following judgment is made: At this point, no cross-language mapping relationship is established, and the conceptual unit is marked as unmatched and retained in the original set.

[0082] Step S66: Maximum Similarity Uniqueness Constraint. By introducing the maximum similarity criterion, the cross-language mapping relationship is constrained into a functional form: .in: Through the aforementioned constraint mechanism, a balance between global consistency and local optimality is achieved in cross-language knowledge alignment, ensuring that the mapping results possess high precision, uniqueness, and interpretability. Among these, This indicates no matching results or an isolated state.

[0083] This invention constructs a multi-layered semantic processing framework, including heuristic morphological decoupling, adaptive semantic density estimation, and asymmetric latent space mapping, to standardize the representation and structure the modeling of multi-source knowledge entities. By mapping knowledge entities from different carriers to a unified semantic representation space and combining semantic similarity constraints to align conceptual units, this invention effectively reduces semantic noise interference, improves the accuracy and stability of the association between scientific and technological knowledge, and thus achieves accurate matching and fusion of cross-carrier knowledge.

Claims

1. A method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping, characterized in that, Includes the following steps: Step 1: Perform heuristic multidimensional morphological decoupling and normalization on keywords derived from scientific papers and technical patents, construct a many-to-one mapping relationship from keywords to atomized concepts, and realize the atomized representation of knowledge entities; Step 2: Perform frequency statistics and rank-frequency distribution modeling on the normalized entity set, adaptively determine the frequency threshold based on the curvature change detection mechanism of the cumulative distribution function, and select the core semantic anchor points according to the threshold; Step 3: Construct a dual-tower coding model, map the selected semantic anchors into high-dimensional semantic vectors, and normalize the semantic vectors, constraining them to a unit hyperspherical manifold; Step 4: Construct a topological structure based on a nearest neighbor graph within the monolingual manifold space. Based on a preset high similarity threshold, perform neighborhood aggregation on semantic vectors and merge vectors that are highly similar in meaning into a unified conceptual unit. Step 5: Construct an asymmetric encoding architecture, mapping Chinese and English concept units to a unified latent space through different encoding functions, and using a contrastive learning strategy to optimize the cross-linguistic semantic representation within this space; Step 6: In the unified latent space, perform cross-language nearest neighbor search for each Chinese concept unit, and at the same time, make judgments based on the maximum similarity criterion and adaptive similarity threshold to establish a one-to-one mapping relationship between Chinese concept units and English concept units, and perform isolation processing on concept units that do not meet the threshold conditions.

2. The method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping according to claim 1, characterized in that, In step 1, the heuristic multidimensional morphological decoupling and normalization process is as follows: for English keywords, case unification, delimiter normalization, word form restoration and abbreviation expansion are performed; for Chinese keywords, semantic boundary detection, functional suffix stripping and synonym mapping are performed.

3. The method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping according to claim 1, characterized in that, In step 2, the curvature abrupt change detection mechanism based on the cumulative distribution function adaptively determines the frequency threshold by calculating the second difference of the function and taking the frequency corresponding to the position with the largest curvature change as the adaptive frequency threshold.

4. The method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping according to claim 1, characterized in that, In step 3, all semantic vectors are constrained to a unit hyperspherical manifold, and the semantic similarity between any two vectors is obtained by calculating their dot product.

5. The method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping according to claim 1, characterized in that, In step 4, the topological structure based on the nearest neighbor graph is a hierarchical navigation small-world graph; the high similarity threshold is adaptively determined in the high similarity region on the right based on the probability density function of the monolingual similarity distribution; by identifying the connected components in the nearest neighbor graph structure, the semantic vectors are aggregated to form a unified conceptual unit.

6. The method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping according to claim 1, characterized in that, In step 5, the asymmetric encoding architecture consists of two encoding functions with independent parameters: a query-side encoding function and an anchor-side encoding function. The query-side encoding function is used to encode Chinese concept units, and the anchor-side encoding function is used to encode English concept units.

7. A method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping according to claim 6, characterized in that, In step 5, the contrastive learning strategy employs an optimization objective function with a temperature parameter to adjust the smoothness of the cross-language matching probability distribution.

8. The method for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping according to claim 1, characterized in that, In step 6, the adaptive similarity threshold is dynamically determined by scanning the preset similarity interval and identifying the inflection point of the hit rate change rate; the one-to-one mapping relationship forces each Chinese concept unit to match only one English concept unit with the highest similarity that exceeds the threshold, thereby ensuring the uniqueness of the alignment result.

9. A system for disambiguation and cross-carrier alignment of scientific and technological knowledge entities based on asymmetric semantic manifold mapping, characterized in that, include: The heuristic morphological decoupling and atomization normalization module is used to perform morphological decoupling and normalization on multi-source heterogeneous keywords and construct atomized concept representations. An adaptive semantic density estimation and noise reduction module is used to filter core semantic anchors and noise by modeling rank-frequency distribution and detecting curvature abrupt changes; The semantic vector manifold representation module is used to map entities to semantic vectors and constrain them to a unit hyperspherical manifold; The semantic consensus aggregation module is used to aggregate neighborhoods on the nearest neighbor graph topology using a high similarity threshold to form stable and compact conceptual units. The asymmetric latent space mapping and alignment determination module is used to construct a unified latent space through asymmetric encoding and contrastive learning, and to perform one-to-one strong association mapping of cross-linguistic concept units based on the maximum similarity criterion and adaptive threshold.

Images