A cross-language text clustering method and system based on high-dimensional vector space manifold alignment

Abstract

The application provides a cross-language text clustering method and system based on high-dimensional vector space manifold alignment, and relates to the technical field of natural language processing.The method comprises the following steps: constructing a source language and target language feature matrix through feature extraction, generating a pseudo-anchor point matrix by unsupervised topological matching, performing space centering processing, and executing orthogonal Procrustes analysis; based on the orthogonal transformation matrix, scaling factor and translation vector, performing rigid body transformation on the source language feature matrix, fusing the target language feature matrix, constructing a unified manifold space, performing clustering processing, and obtaining the semantic cluster division result of the cross-language text. Through the application, the technical problem of the prior art that the local topological structure of the source language semantic space is destroyed due to the dependence on large-scale parallel corpus for forced space mapping, which affects the cross-language text clustering accuracy and causes clustering drift can be solved, the topological coincidence of different language texts in a unified space is realized, and the cross-language text clustering accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, specifically to a cross-lingual text clustering method and system based on high-dimensional vector space manifold alignment. Background Technology

[0002] In the fields of cross-lingual text clustering and natural language processing, to enable texts from different languages ​​to be compared and clustered in the same vector space, existing mainstream technologies mainly fall into two categories. One category is joint training and forced mapping techniques based on parallel corpora. This relies on large-scale bilingual parallel corpora and utilizes known parallel anchors to calculate a mapping matrix through supervised learning, mapping the source language space to the target language space. The other category is pipelined clustering techniques based on machine translation. For low-resource languages ​​lacking parallel corpora, this approach calls an external machine translation system to uniformly translate all non-central language texts into a high-resource-central language, then extracts feature vectors in a monolingual environment and executes standard clustering algorithms.

[0003] However, supervised mapping methods heavily rely on massive amounts of high-quality parallel corpora, making their construction cost prohibitively high and virtually unavailable for less common languages ​​or specific vertical domains. Furthermore, traditional unconstrained linear mapping matrices contain affine transformations such as scaling, stretching, and shearing, which, when forcibly aligned, disrupt the local topology of the original source language's vector space, leading to spatial distortion and causing texts belonging to the same semantic cluster to be scattered, resulting in a precipitous drop in clustering accuracy. Simultaneously, multilingual joint training and machine translation pipelines fail to effectively separate linguistic and semantic features. The extracted vectors are often linearly coupled between the two, with excessively high weights for linguistic features. Consequently, clustering algorithms prioritize capturing linguistic features such as grammatical structure and translation artifacts, causing cluster drift—incorrectly grouping texts of the same language into language clusters instead of performing cross-language fusion based on objective semantics.

[0004] In summary, existing technologies suffer from the technical problem that relying on large-scale parallel corpora for forced spatial mapping disrupts the local topological structure of the source language semantic space, further affecting the accuracy of cross-language clustering and causing cluster drift. Summary of the Invention

[0005] This application provides a cross-lingual text clustering method and system based on high-dimensional vector space manifold alignment, which addresses the technical problem in existing technologies where forced spatial mapping based on large-scale parallel corpora disrupts the local topological structure of the source language semantic space, further affecting the accuracy of cross-lingual clustering and causing cluster drift.

[0006] In view of the above problems, this application provides a cross-language text clustering method and system based on high-dimensional vector space manifold alignment.

[0007] Firstly, this application provides a cross-lingual text clustering method based on high-dimensional vector space manifold alignment. This method is implemented through a cross-lingual text clustering system based on high-dimensional vector space manifold alignment. The method includes: obtaining a source language text set and a target language text set; extracting features from the source and target language text sets respectively; constructing mutually independent source language feature matrices and target language feature matrices; and performing... Unsupervised topological matching is used to generate pseudo-anchor matrixes for the source language and target language. These pseudo-anchor matrices are then spatially centered to obtain corresponding centered matrices. Orthogonal Prouker analysis of the centered matrices is performed to solve for the globally optimal orthogonal transformation matrix and scaling factor. Based on the orthogonal transformation matrix, scaling factor, and translation vector, a global rigid transformation is performed on the source language feature matrix to obtain an aligned source language feature matrix. The aligned source language feature matrix is ​​then fused with the target language feature matrix to construct a unified manifold space. Clustering is then performed within this unified manifold space to obtain the semantic clustering results for the cross-language texts.

[0008] Optionally, local topological signature vectors are constructed for the corresponding source language nodes and target language nodes in the source language feature matrix and target language feature matrix, respectively. These local topological signature vectors are used to search for the k nearest neighbors within the same language space using Euclidean distance, and the distance values ​​of these k nearest neighbors are extracted to form an ordered distance vector, which serves as the topological signature vector for the corresponding node, as follows: S(x i )=[d1,d2,...,d k ]; where S(x i ) represents the topological signature vector of the source language node, d k The distance value representing the k-th nearest neighbor node, where d1≤d2≤...≤d k Unsupervised topology matching is performed based on the topology signature vector to generate a source language pseudo-anchor matrix and a target language pseudo-anchor matrix.

[0009] Optionally, the topology matching confidence function is defined using the topology signature vector, as follows: ;in, The topological signature vector representing the target language node. To control the bandwidth parameter of the distribution smoothness, σ∈[0.1,1.0]; confidence is calculated according to the topology matching confidence function, and matching is performed by bidirectional nearest neighbor filtering to generate source language pseudo-anchor matrix and target language pseudo-anchor matrix.

[0010] Optionally, the centroid vectors of the source language pseudo-anchor matrix and the target language pseudo-anchor matrix are calculated separately, as follows: ; ; where μ X Let μ be the centroid vector of the pseudo-anchor matrix of the source language. Y Let N be the centroid vector of the pseudo-anchor matrix of the target language, and N be the number of pseudo-anchors. The X-dimensional feature components representing the i-th source language pseudo-anchor point in the embedding space. The Y-dimensional feature components of the i-th target language pseudo-anchor point in the embedding space are represented; spatial centering is performed using the centroid vector to obtain the corresponding centering matrix.

[0011] Alternatively, the space centralization process can be performed as follows: ; ;in, It is a column vector of all 1s. Transpose processing of the centroid vector representing the pseudo-anchor matrix of the source language. Transpose of the centroid vector representing the pseudo-anchor matrix of the target language.

[0012] Alternatively, by performing singular value decomposition of the cross covariance matrix, the solution satisfying R can be obtained. T R=I d The optimal orthogonal rotation matrix R*, where det(R) = 1; calculate the scaling factor and translation vector as follows: ; Where s represents the scaling factor, t is the translation vector, and Tr( The trace represents the matrix.

[0013] Optionally, high-dimensional spatial density clustering is performed based on the aligned unified manifold space matrix to generate a cross-lingual semantic cluster set; the cross-lingual cluster intra-cluster entropy is calculated for each valid cluster in the cross-lingual semantic cluster set, as follows: ; where H(C i Characterizing the effective cluster C i The cross-language cluster intra-entropy, where L is the set of languages ​​involved in the system, p(l|C i Characterizing the effective cluster C i The proportion of text belonging to language l in the total number of texts in the effective cluster is used to evaluate the clustering drift index based on the cross-language cluster entropy. By eliminating residual drift, the semantic clustering results of the cross-language texts are generated.

[0014] Optionally, the cluster drift index is calculated as follows: Where CDI is the cluster drift index, |C| is the total number of effective clusters, and |L| is the number of language types.

[0015] Optionally, when new text data is added, the monolingual pre-trained model is invoked to extract the feature matrix, and the current rigid body transformation operator is reused to project the feature matrix onto a unified manifold space, as follows: ; where △V src Characteristic matrix, 1 △M Let t be a column vector of all 1s. T Let be the transpose of the translation vector; calculate the manifold projection error rate of the current transformation matrix on the new data, as follows: Where MPER is the manifold projection error rate, ||·||² is the L2 norm of the vector, and ΔX i Let ΔY be the difference vector of the i-th pseudo-anchor point in the source language. i Characterization and △X i The corresponding i-th difference vector of the target language; establish cross-language clustering labels for incremental text based on manifold projection error rate.

[0016] Secondly, this application also provides a cross-lingual text clustering system based on high-dimensional vector space manifold alignment, used to execute a cross-lingual text clustering method based on high-dimensional vector space manifold alignment as described in the first aspect, wherein the cross-lingual text clustering system based on high-dimensional vector space manifold alignment includes: a feature extraction module, used to obtain a source language text set and a target language text set, and to extract features from the source language text set and the target language text set respectively, constructing mutually independent source language feature matrices and target language feature matrices; and an unsupervised topology matching module, used to perform unsupervised topology matching based on the local topological structure of the source language feature matrices and the target language feature matrices, generating a source language pseudo-anchor matrix. The system comprises: a source language pseudo-anchor matrix and a target language pseudo-anchor matrix; a spatial centering module for spatially centering the source language pseudo-anchor matrix and the target language pseudo-anchor matrix to obtain the corresponding centered matrix; an orthogonal solution module for performing orthogonal Prouke analysis on the centered matrix to solve for the globally optimal orthogonal transformation matrix and scaling factor; a global rigid transformation module for performing a global rigid transformation on the source language feature matrix based on the orthogonal transformation matrix, scaling factor, and translation vector to obtain the aligned source language feature matrix; and a clustering module for fusing the aligned source language feature matrix and the target language feature matrix to construct a unified manifold space, and performing clustering in the unified manifold space to obtain the semantic clustering results of the cross-language text.

[0017] One or more technical solutions provided in this application have at least the following beneficial effects: By acquiring a source language text set and a target language text set, feature extraction is performed on the source language text set and the target language text set respectively to construct mutually independent source language feature matrices and target language feature matrices; based on the local topological structure of the source language feature matrices and target language feature matrices, unsupervised topological matching is performed to generate source language pseudo-anchor matrix and target language pseudo-anchor matrix; spatial centering processing is performed on the source language pseudo-anchor matrix and target language pseudo-anchor matrix to obtain corresponding centered matrices; orthogonal Prouke analysis is performed on the centered matrices to solve for the globally optimal orthogonal transformation matrix and scaling factor; based on the orthogonal transformation matrix, scaling factor, and translation vector, a global rigid body transformation is performed on the source language feature matrix to obtain an aligned source language feature matrix; the aligned source language feature matrix is ​​fused with the target language feature matrix to construct a unified manifold space, and clustering processing is performed in the unified manifold space to obtain the semantic cluster partitioning results of cross-language texts. In other words, by treating the semantic spaces of different languages ​​as high-dimensional manifolds with similar geometric topologies, a small number of pseudo-anchor points are obtained through unsupervised topological matching. Then, orthogonal Proukt analysis is used to calculate the optimal global rotation, translation, and scaling matrices in the ultra-high-dimensional space. Non-intrusive geometric rigid body transformations are used to eliminate physical displacements caused by language differences, enabling semantic clusters of different languages ​​to achieve topological overlap in a unified space, thereby improving the accuracy of cross-language text clustering. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application 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 merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a cross-language text clustering method based on high-dimensional vector space manifold alignment, as proposed in this application.

[0020] Figure 2 This is a schematic diagram of the structure of a cross-language text clustering system based on high-dimensional vector space manifold alignment according to this application.

[0021] Figure labeling: Feature extraction module 11, Unsupervised topology matching module 12, Spatial centering module 13, Orthogonal solution module 14, Global rigid body transformation module 15, Clustering module 16. Detailed Implementation

[0022] This application provides a cross-lingual text clustering method and system based on high-dimensional vector space manifold alignment. It addresses the technical problem in existing technologies where forced spatial mapping based on large-scale parallel corpora disrupts the local topological structure of the source language's semantic space, further affecting cross-lingual clustering accuracy and causing cluster drift. By treating the semantic spaces of different languages ​​as high-dimensional manifolds with similar geometric topologies, a small number of pseudo-anchor points are obtained through unsupervised topological matching. Then, orthogonal Prouker analysis is used to calculate the optimal global rotation, translation, and scaling matrices in the ultra-high-dimensional space. Non-intrusive geometric rigid body transformations eliminate physical displacements caused by language differences, enabling semantic clusters from different languages ​​to achieve topological overlap in a unified space, thus improving the accuracy of cross-lingual text clustering.

[0023] Example 1, as Figure 1 As shown, this application provides a cross-lingual text clustering method based on high-dimensional vector space manifold alignment. This method is applied to a cross-lingual text clustering system based on high-dimensional vector space manifold alignment. The specific steps of this method are as follows: Obtain a source language text set and a target language text set, extract features from the source language text set and the target language text set respectively, and construct mutually independent source language feature matrices and target language feature matrices.

[0024] Specifically, the process involves acquiring the source language text set and the target language text set input by the user, and then cleaning and segmenting them. For each language, a corresponding monolingual pre-trained language model is invoked to extract deep semantic features of the text, constructing independent high-dimensional vector spaces that do not overlap.

[0025] Each text in both the source and target language text sets is stored as a string with its language tag. Each text is then cleaned and normalized, including removing HTML tags, special characters, and extra spaces; performing word segmentation based on the language; converting all text to lowercase; and removing stop words. The preprocessed text is stored as a token sequence.

[0026] Based on the language type of the text, the corresponding monolingual pre-trained language model is invoked. For Chinese, a model pre-trained specifically on Chinese corpora is used; for English, the corresponding English pre-trained model is used. Each model is a Transformer architecture with approximately 110 million parameters and an output vector dimension of 768. After loading the model, the pre-processed text token sequence is input into the model, and the vectors of all tokens in the last layer are average-pooled to obtain the semantic feature vector of the text. Each dimension of this vector is a floating-point number, ranging approximately from -1 to 1.

[0027] The feature vectors of all source language texts are stacked row-wise to form a two-dimensional matrix with a size equal to the number of source language texts multiplied by the vector dimension of 768; this is called the source language feature matrix. Similarly, the feature vectors of all target language texts are stacked row-wise to form the target language feature matrix. Since the two models have different parameters and different training corpora, the vector spaces of these two matrices are independent of each other; therefore, it is not possible to directly calculate the similarity between a Chinese vector and an English vector. L2 norm normalization is applied to all row vectors within each feature matrix, making the magnitude of each vector equal to 1. Using high-quality models pre-trained independently for each language accurately captures the semantic features within that language, avoiding the semantic dilution problem caused by shared parameters in multilingual joint training.

[0028] Based on the local topological structure of the source language feature matrix and the target language feature matrix, unsupervised topological matching is performed to generate the source language pseudo-anchor matrix and the target language pseudo-anchor matrix.

[0029] Furthermore, this application also includes the following steps: constructing local topological signature vectors for the source language nodes and target language nodes corresponding to the source language feature matrix and target language feature matrix, respectively; the local topological signature vectors search for the k nearest neighbor nodes in the same language space using Euclidean distance, and extracting the distance values ​​of the k nearest neighbor nodes to form an ordered distance vector, which serves as the topological signature vector of the corresponding node, as follows: S(x i )=[d1,d2,...,d k ]; where S(x i ) represents the topological signature vector of the source language node, d k The distance value representing the k-th nearest neighbor node, where d1≤d2≤...≤d k Unsupervised topology matching is performed based on the topology signature vector to generate a source language pseudo-anchor matrix and a target language pseudo-anchor matrix.

[0030] Furthermore, this application also includes the following step: defining a topology matching confidence function using the topology signature vector, as follows: ;in, The topological signature vector representing the target language node. To control the bandwidth parameter of the distribution smoothness, σ∈[0.1,1.0]; confidence is calculated according to the topology matching confidence function, and matching is performed by bidirectional nearest neighbor filtering to generate source language pseudo-anchor matrix and target language pseudo-anchor matrix.

[0031] Specifically, for each text vector in the source language feature matrix, calculate its Euclidean distance to all other source language vectors in the same space. Similarly, for each text vector in the target language feature matrix, calculate its Euclidean distance to all other target language vectors in the same space. For each vector in the source language, sort its Euclidean distances to all other vectors in the same space in ascending order, and take the k smallest distance values ​​to form an ordered distance vector, which serves as the topological signature vector of that node, such as S(x i )=[d1,d2,...,dk], where d1≤d2≤...≤dk, S(x i ) represents the topological signature vector of the source language node; k is the nearest neighbor number, usually set to the range [10, 50], such as k=15; i represents the i-th text vector in the source language feature matrix, 1≤i≤k. Similarly, a topological signature S(y) of the same dimension is also constructed for each vector in the target language feature matrix. j ).

[0032] To quantify the topological similarity between a source language node and a target language node, a Gaussian kernel function is used to calculate the matching score between their topological signatures. ,in, Let S(x) be the squared Euclidean distance between two topological signature vectors, and σ be the bandwidth parameter controlling the smoothness of the distribution, with empirical values ​​σ∈[0.1,1.0]. Specifically, let S(x) be the topological signature vector of the source language node x. i The topological signature vector S(y) of the target language node y and the target language node y j Subtract each element one by one, calculate the sum of squares of the differences, divide by the square of twice the bandwidth parameter σ, and finally take the negative of the exponential function. For example, let σ = 0.5 and k = 5. Source language node x i The topological signature S(x) i = [0.10, 0.20, 0.25, 0.30, 0.40], target language node y j The topological signature S(y) j Given the expression = [0.12, 0.18, 0.26, 0.31, 0.38], calculate the sum of squared differences between the two expressions. =0.0014. Substitute into the formula. The confidence score of 0.9972 indicates that the two nodes have highly consistent local geometric environments in their respective spaces, and are highly likely to be mutually translated semantic concepts.

[0033] The algorithm iterates through all nodes in both the source and target languages, calculating a confidence score for each pair using the formula described above, thus forming a confidence matrix. To filter out high-confidence pseudo-anchors, a strict bidirectional maximum confidence strategy is employed. First, for each source language node, the target node with the highest confidence among all target language nodes is identified; this is the target node corresponding to the maximum confidence. Conversely, for each target language node, the source node with the highest confidence among all source language nodes is identified. Only when a source node and a target node are mutually matched at their maximum confidence are the node pairs retained as candidate pseudo-anchors, effectively avoiding one-to-many fuzzy matching and ensuring that each anchor pair is unique and mutually exclusive.

[0034] Sort all candidate node pairs filtered by bidirectional nearest neighbor from highest to lowest confidence score, and set a target number of pseudo-anchors N, such as N=500. Take the top N pairs after sorting as the final pseudo-anchors. If the total number of candidate pairs that pass the filter is less than N, accept all candidate pairs. Extract the row vectors corresponding to these N source language nodes from the source language feature matrix, and stack them in matching order to form a source language pseudo-anchor matrix X with size N×d and vector dimension d=768. Extract the corresponding N target language node vectors from the target language feature matrix to form a target language pseudo-anchor matrix Y, also with size N×768. The first row of X and the first row of Y are semantically considered to correspond, and so on. By quantizing the topological signature of the local geometric space, relying solely on the inherent structural features of the data manifold itself, high-precision alignment benchmarks can be automatically mined in non-overlapping high-dimensional spaces, reducing the application threshold of cross-language clustering systems in low-resource language scenarios.

[0035] The source language pseudo-anchor matrix and the target language pseudo-anchor matrix are spatially centered to obtain the corresponding centered matrix.

[0036] Furthermore, this application also includes the following steps: calculating the centroid vectors of the source language pseudo-anchor matrix and the target language pseudo-anchor matrix respectively, as follows: ; ; where μ X Let μ be the centroid vector of the pseudo-anchor matrix of the source language. Y Let X be the centroid vector of the pseudo-anchor matrix of the target language, N be the number of pseudo-anchors, and X be the centroid vector of the pseudo-anchor matrix. i Y represents the X-dimensional feature components of the i-th source language pseudo-anchor point in the embedding space. i The Y-dimensional feature components of the i-th target language pseudo-anchor point in the embedding space are represented; spatial centering is performed using the centroid vector to obtain the corresponding centering matrix.

[0037] Furthermore, this application also includes the following steps: spatial centralization processing is as follows: ; ; among which, 1 N It is a column vector of all 1s. Transpose processing of the centroid vector representing the pseudo-anchor matrix of the source language. Transpose of the centroid vector representing the pseudo-anchor matrix of the target language.

[0038] Specifically, the centroid vectors of the source language pseudo-anchor matrix and the target language pseudo-anchor matrix are calculated separately. This involves iterating through all N rows of the source language pseudo-anchor matrix and calculating the arithmetic mean of the N values ​​in each column. Specifically, for the first column, the first components from row 1 to row N are summed and then divided by N to obtain the first component of the centroid vector. The same process is repeated for the second column until the end. The final centroid vector of the source language pseudo-anchor matrix is ​​denoted as μ. X Let represent the geometric center positions of the N pseudo-anchor points of the source language in d-dimensional space. Similarly, the centroid vector μ is obtained by averaging all N rows and d columns of the target language pseudo-anchor point matrix. Y .

[0039] The formula is: ; ;in, Let the centroid vector of the pseudo-anchor matrix of the source language be . Let N be the centroid vector of the pseudo-anchor matrix of the target language, and N be the number of pseudo-anchors. The X-dimensional feature components representing the i-th source language pseudo-anchor point in the embedding space. The Y-dimensional feature component representing the i-th target language pseudo-anchor point in the embedding space.

[0040] To facilitate matrix subtraction, a column vector of length N consisting entirely of 1s is generated, denoted as 1. N Each element of this vector is the value 1, and it is used to broadcast the centroid vector to each row in matrix operations.

[0041] Subtracting the centroid vector from each row of the source language pseudo-anchor matrix yields the source language centered matrix, which satisfies the condition that the sum of each column is zero, meaning the mean of all rows across all dimensions is zero. The formula is as follows: , of which 1 N It is a column vector of all 1s. The transpose of the centroid vector representing the source language pseudo-anchor matrix transforms the row vectors into column vectors. The same method is used to center the target language pseudo-anchor matrix, resulting in a centered target language matrix that also satisfies the condition that the sum of all columns is zero. The formula is as follows: ; among which, 1 N It is a column vector of all 1s. Transpose of the centroid vector representing the pseudo-anchor matrix of the target language.

[0042] Perform orthogonal Pluke analysis on the centered matrix to find the globally optimal orthogonal transformation matrix and scaling factor.

[0043] Furthermore, this application also includes the following steps: solving for the solution that satisfies R by performing singular value decomposition of the cross covariance matrix. T R=I d The optimal orthogonal rotation matrix R, where det(R) = 1. * Calculate the scaling factor and translation vector as follows: ; t=μ Y -s·μ X ·R * Where s represents the scaling factor, t is the translation vector, and Tr( The trace represents the matrix.

[0044] Specifically, the transpose of the source language centering matrix is ​​calculated and multiplied by the target language centering matrix to obtain a 768-row × 768-column square matrix C. Each element of this matrix reflects the covariance between a certain dimension of the source language and a certain dimension of the target language. The singular value decomposition function from the linear algebra library is then used to decompose the square matrix C into the product of three matrices: C = UΣV. T In this matrix, U is an orthogonal matrix, called the left singular vector matrix; ∑ is a diagonal matrix, and the elements on the diagonal are called singular values, arranged in descending order; V is an orthogonal matrix, called the right singular vector matrix. T It is the transpose of V. (This refers to the relationship between U and V.) T Multiplying them yields a candidate rotation matrix R, which automatically satisfies orthogonality, i.e., R T R=I d However, it may involve reflection, and geometrically it means rotating the source language center set to the direction that is closest to the target language center set.

[0045] Calculate the determinant of the candidate rotation matrix. If the determinant is equal to 1, i.e., det(R) = 1, then the matrix is ​​already a pure rotation matrix and is directly accepted as the optimal rotation matrix R. * If the determinant equals -1, it indicates that the transformation involves a reflection, equivalent to mirroring. This is undesirable for semantic space alignment because reflection alters the chirality of vectors, potentially disrupting certain semantic relations. In this case, multiplying the last column of matrix U by -1 changes the sign of the last left singular vector, and then recalculating R... * =UV T The resulting matrix will have a determinant of 1, and it will still be the optimal pure rotation matrix.

[0046] Further calculations are performed to determine the globally optimal scaling factor *s* and translation vector *t*. The scaling factor *s* reflects the overall scale difference between the source language space and the target language space. Where s represents the scaling factor, Tr( The trace of the matrix represents the alignment between the source and target language point sets after rotation, i.e., the trace of the covariance. The numerator is the energy of the source language point set itself, i.e., the sum of the squares of the magnitudes of all vectors. Since the feature extraction models for the source and target languages ​​typically have similar scales, s is generally between 0.9 and 1.1. The translation vector t = μ Y -s·μ X ·R * For example, in a d=768 dimensional space, the trace of the cross covariance is calculated using matrix multiplication. =152.4, Trace of source language-centered autocovariance =148.2. Substituting this into the formula, the scaling factor s is 1.028. The source language centroid μ X After rotation, it becomes [0.5, [0.2,...,0.1], target language centroid μ Y For [0.6, If [0.1,...,0.2], then the first component of the translation vector t is t1 = 0.6. 1.028 × 0.5 = 0.086. This result shows that the spatial scales of the two language models are extremely close (s is close to 1), requiring only minor scaling and physical displacement compensation.

[0047] Based on the orthogonal transformation matrix, scaling factor, and translation vector, a global rigid body transformation is performed on the source language feature matrix to obtain the aligned source language feature matrix.

[0048] Specifically, the orthogonal transformation matrix R obtained by solving * The scaling factor s and the translation vector t are treated as a global rigid body transformation operator and applied to the complete source language feature matrix V. srcThe aligned source language feature matrix is ​​obtained. Multiplying the source language feature matrix by a scaling factor and then right-multiplying it by the orthogonal transformation matrix applies a rotation operation to each source language vector (row vector), causing it to rotate as a whole in high-dimensional space. After rotation and scaling, the orientation of the source language data cloud is consistent with that of the target language data cloud, and the overall scale is adjusted to a similar level, but the origin of the coordinate system is not yet aligned. Based on the rotation and scaling, a translation vector t is added to each transformed vector. Since t is a row vector (length d), it needs to be broadcast to all M rows. An M x d matrix is ​​generated, with each row being t, and then element-wise added to the result matrix from the previous step. The translation operation moves the centroid of the source language data cloud to a position coinciding with the centroid of the target language data cloud, completing the rigid body transformation. The aligned source language feature matrix and the target language feature matrix are located in the same high-dimensional coordinate system, and the semantically corresponding points in the two spaces are geometrically as close as possible. At this point, the Euclidean distance or cosine similarity between the source language vectors and the target language vectors can be directly calculated for subsequent joint clustering. By introducing Pluck analysis with orthogonal constraints, traditional lossy affine mappings are replaced with pure geometric rigid body transformations (rotation, translation, and proportional scaling). This non-invasive manifold alignment perfectly preserves the pure semantic distribution extracted by the monolingual model, eliminating physical displacements caused by language differences. The aligned source language feature matrix is ​​fused with the target language feature matrix to construct a unified manifold space, and clustering is performed in this unified manifold space to obtain semantic clustering results for cross-lingual texts.

[0049] Furthermore, this application also includes the following steps: performing high-dimensional spatial density clustering based on the aligned unified manifold space matrix to generate a cross-lingual semantic cluster set; calculating the cross-lingual cluster intra-cluster entropy for each valid cluster in the cross-lingual semantic cluster set, as follows: ;wherein, H(C i Characterizing the effective cluster C i The cross-language cluster intra-entropy, where L is the set of languages ​​involved in the system, p(l|C i Characterizing the effective cluster C i The proportion of text belonging to language l in the total number of texts in the effective cluster is used to evaluate the clustering drift index based on the cross-language cluster entropy. By eliminating residual drift, the semantic clustering results of the cross-language texts are generated.

[0050] Furthermore, this application also includes the following steps: the cluster drift index is calculated as follows: Where CDI is the cluster drift index, |C| is the total number of effective clusters, and |L| is the number of language types.

[0051] Specifically, the aligned source language feature matrix and target language feature matrix are concatenated along the sample dimension to form a unified manifold space matrix. On this unified space, a density-based clustering algorithm is applied, using cosine distance as the metric, to output the cross-lingual semantic cluster label for each text. The unified manifold space matrix consists of the first M rows of source language text and the last K rows of target language text.

[0052] Density-based adaptive clustering is performed in the aligned unified manifold space. Information entropy theory is introduced to construct a cluster drift index. By quantifying the purity of language distribution within each cluster, residual drift is detected and adaptively fine-tuned. The aligned unified manifold space matrix V... union ∈R (M+K)×d It contains all text vectors of the source and target languages. The aligned unified manifold space matrix V... union The algorithm calculates the pairwise cosine distance matrix for all text vectors. Cosine distance is defined as 1 minus cosine similarity, ranging from 0 to 2; smaller values ​​indicate greater similarity. The pairwise cosine distance matrix is ​​then input into the HDBSCAN algorithm. A weighted nearest neighbor graph is constructed based on the distance matrix. A hierarchical cluster structure is built by continuously removing low-density edges. Finally, stability analysis is used to extract the most significant clusters. Two key parameters need to be set: minimum cluster size (min_cluster_size=15) and minimum number of samples (min_samples=5). This adaptively discovers semantic clusters of arbitrary shapes, marks sparse outliers as noise, and outputs a preliminary cross-linguistic semantic cluster set C={C1,C2,...,C...}. m To quantitatively assess whether language clusters still exist in the clustering results, i.e., erroneous clusters of single languages, the cross-language cluster intra-cluster entropy is calculated for each valid cluster, and the calculation formula is defined as follows: Where L is the set of languages ​​involved in the system, i.e., L = {source language, target language}, H(C i Characterizing the effective cluster C i cross-language cluster intra-entropy, p(l|C i Characterizing the effective cluster C i The proportion of text belonging to language l in the total number of texts in the effective cluster.

[0053] Since density clustering may produce some clusters dominated by a single language, a clustering drift index (CDI) based on information entropy is introduced for evaluation. The proportion of source and target languages ​​in each cluster is calculated, and then the entropy of that cluster is calculated. If the entropy of a cluster is too low, and the cluster size is greater than the minimum cluster size, the system lowers the local distance threshold of that cluster region, forcing it to split, and reassigning the sub-points to neighboring mixed clusters. This fine-tuning process can be iteratively performed until the entropy of all clusters is above a preset threshold.

[0054] The cluster drift index is calculated as follows: Where CDI is the clustering drift index, |C| is the total number of effective clusters, and |L| is the number of language types. In the bilingual scenario, |L|=2, and the ideal maximum entropy log2(2)=1. For example, the cross-language clustering system outputs 3 clusters. Cluster C1 contains 80 Chinese characters and 20 English characters, so p(zh)=0.8, p(en)=0.2, and H(C1) is 0.7219. Cluster C2 contains 50 Chinese characters and 50 English characters, and H(C2) is 1.0. Cluster C3 contains 90 Chinese characters and 10 English characters, and H(C3) is 0.4690. The global average entropy is (0.7219+1.0+0.4690) / 3≈0.7303. Substituting into the formula, we get CDI as 0.2697. The value range of CDI is [0,1]. The closer the CDI is to 0, the more uniform the language mixing within each cluster and the more thorough the cross-language semantic fusion; the closer the CDI is to 1, the more dominant a single language is within the cluster and the more severe the cluster drift.

[0055] Set the drift tolerance threshold τ cdi =0.35. Local cluster 1 H(C i ) / log2(|L|)>τ cdi This will extract the core points within the cluster and locally reduce the distance threshold of HDBSCAN. The parameters cause the suspected language cluster to split and its connectivity with neighboring clusters to be recalculated, forcing it to reassign based on pure semantic distance. This ultimately eliminates residual drift, resulting in the final cross-lingual semantic cluster label and a system-level cluster drift index (CDI) evaluation report.

[0056] Clustering algorithms can only perceive pure semantic distance in a unified space, effectively breaking down language barriers and improving the purity and normalized mutual information (NMI) of cross-linguistic semantic cluster partitioning. By introducing a clustering drift index based on information entropy, the previously imperceptible phenomenon of language dominance is precisely quantified mathematically. Combined with the adaptive fine-tuning mechanism of density clustering, residual drift caused by local density unevenness is further suppressed at the end of the algorithm, ensuring that the final output clusters truly reflect the semantic clusters of the objective physical world, rather than language clusters interfered with by linguistic features.

[0057] Furthermore, this application also includes the following steps: when new text data is added, the monolingual pre-trained model is invoked to extract the feature matrix, and the current rigid body transformation operator is reused to project the feature matrix onto a unified manifold space, as follows: ; where △V src Characteristic matrix, 1 △M Let t be a column vector of all 1s. TLet be the transpose of the translation vector; calculate the manifold projection error rate of the current transformation matrix on the new data, as follows: Where MPER is the manifold projection error rate, ||·||² is the L2 norm of the vector, and ΔX i Let ΔY be the difference vector of the i-th pseudo-anchor point in the source language. i Characterization and △X i The corresponding i-th difference vector of the target language; establish cross-language clustering labels for incremental text based on manifold projection error rate.

[0058] Specifically, in real-world industrial applications, text data is continuously generated. Re-performing full unsupervised anchor point generation and Pruk analysis for each new batch of text would consume enormous computational resources, resulting in extremely high system latency and failing to meet the real-time processing requirements of streaming data. Over time, the semantic distribution of language may shift conceptually, causing historically calculated manifold transformation matrices to gradually become invalid. A lightweight online projection mechanism is designed to achieve millisecond-level access to incremental text by reusing already solved rigid body transformation operators. Simultaneously, a manifold projection error rate monitoring mechanism is introduced to dynamically trigger background matrix updates when errors exceed the limit, ensuring the system's long-term robustness.

[0059] When new text data arrives, the new source language text data stream already has the globally optimal rotation matrix, translation vector, scaling factor, and the current unified manifold space and its cluster center state stored in memory. When the new source language text arrives, the monolingual pre-trained model is invoked to extract its feature matrix. Skipping the complex topology matching step, the current rigid body transformation operator is directly reused to project it onto the unified manifold space. ; where △V src Characteristic matrix, 1 △M Let t be a column vector of all 1s. T This is the transpose of the translation vector. It only involves basic matrix multiplication and addition, resulting in extremely low time complexity. The minimum distance is the cosine distance to the center of each semantic cluster. If the minimum distance is less than a set assignment threshold, such as 0.25, the text is assigned to the corresponding cluster; if it is greater than the threshold, it is temporarily stored in an outlier buffer. When the buffer reaches min_samples, local density clustering is triggered to form a new cluster.

[0060] Set a sliding time window, and at fixed intervals, extract a small number of the latest pseudo-anchor pairs ΔX and ΔY from recently added source and target language texts to calculate the manifold projection error rate (MPER) of the current transformation matrix on the new data. Where MPER is the manifold projection error rate, ||·||² is the L2 norm of the vector, and ΔX i Let ΔY be the difference vector of the i-th pseudo-anchor point in the source language. iCharacterization and △X i The corresponding i-th difference vector of the target language; cross-lingual clustering labels for incremental text are established based on the manifold projection error rate. The numerator represents the physical distance deviation between the new anchor point in the source language and the real new anchor point in the target language after projection through the history matrix; the denominator is the modulus of the target language anchor point, used for normalization. For example, if 50 pairs of new pseudo-anchor points are extracted, for the first pair of anchor points, the Euclidean distance deviation between the projected source language vector and the target language vector is 0.45, and the norm of the target language vector is 15.0, then the single-point error rate is 0.45 / 15.0=0.03, or 3%. Combining the 50 pairs of anchor points, the average MPER is calculated to be 0.042.

[0061] Set the manifold aging tolerance threshold τ err =0.12, meaning a maximum allowable projection relative error of 12%. When the calculated MPER > τ err This indicates that a significant conceptual shift has occurred in the underlying semantic space of the language, as shown by the history matrix R. * Expired. Automatically triggers an asynchronous background task to re-execute the Pruk analysis using all data within the latest time window, smoothly replacing the orthogonal transformation matrix, scaling factor, and translation vector parameters in memory. Outputs cross-language clustering labels for incremental text in real time, and outputs dynamic update instructions for the system transformation matrix when necessary.

[0062] To objectively evaluate the technical effectiveness of this application, a multilingual customer service ticket dynamic analysis scenario is used as an example. The dataset contains 25,000 Chinese tickets and 25,000 English tickets, with 120 true semantic categories. The system of this application is quantitatively compared with a traditional baseline system using lossy MLP mapping based on parallel corpora and KMeans clustering. In the cross-language text clustering task, a pairwise text-based evaluation criterion is used to calculate the clustering quality. For any two text pairs in the dataset, the following basic variables are defined: TP refers to the number of pairs of texts belonging to the same true semantic category and assigned to the same cluster by the system; TP refers to the number of pairs of texts belonging to different true semantic categories but incorrectly assigned to the same cluster by the system; FN refers to the number of pairs of texts belonging to the same true semantic category but incorrectly assigned to different clusters by the system. Based on the above variables, the core evaluation formula for clustering quality is defined as follows: Precision measures the proportion of texts that are truly classified into the same semantic category, where the system assigns them to the same cluster; Precision = TP / (TP+FP); Recall measures the proportion of all text pairs that truly belong to the same semantic category that are successfully classified into the same cluster; Recall = TP / (TP+FN); The F1 score is the harmonic mean of precision and recall, used to comprehensively evaluate clustering quality; F1 score = 2 × (Precision × Recall) / (Precision + Recall). To comprehensively measure the system's overall performance in cross-language fusion, accuracy, and engineering real-time performance, a system-level comprehensive evaluation index is further proposed, namely the cross-language clustering comprehensive performance score, calculated as follows: Where F1 is the F1 score, ranging from 0 to 1; CDI is the clustering drift index, ranging from 0 to 1, with a lower CDI indicating better cross-language integration; T resp It is the average response time of the system for processing a single batch of incremental text; T maxα represents the maximum acceptable response time threshold in industrial scenarios, such as 1000ms; α, β, and γ are weighting coefficients, satisfying α+β+γ=1, such as α=0.5, β=0.3, and γ=0.2. For example, the test yielded TP=45000, TP=43420, and FN=52200, with a precision of 91.2%, recall of 89.6%, and an F1 score of 0.9039. The current system's clustering drift index (CDI) was measured to be -0.15, and the single-batch incremental text response time was 12ms. Substituting these values ​​into the CCES formula yielded a comprehensive score of 0.90455, indicating that the system has reached an industrially usable state in terms of semantic accuracy, language fusion, and streaming processing efficiency. Compared to the cross-lingual clustering accuracy of the traditional baseline system, the precision improved from 72.4% to 91.2%, the recall from 68.5% to 89.6%, and the F1 score from 70.39% to 90.39%. Compared to the cluster drift elimination effect of the traditional baseline system, the cluster drift index was reduced from 0.68 to 0.15. Compared to the incremental text streaming response of the traditional baseline system, the average response time was shortened from 1987.24ms to 130.77ms.

[0063] Orthogonal Prouke analysis is used to mitigate linguistic artifacts commonly found in cross-linguistic text clustering. A non-intrusive rigid body transformation is employed, using geometric methods such as rotation, scaling, and translation to map the feature spaces of different languages ​​to a unified physical-logical space without disrupting the original topology of the monolingual model. This zero-distortion characteristic ensures that the original semantic relationships between texts remain absolutely consistent after the transformation, preventing the clustering of identical languages ​​caused by spatial stretching. After achieving initial spatial overlap, an information entropy-based clustering drift quantification index is introduced to address the potential residual local density unevenness in complex corpora. By monitoring the purity of language distribution within each cluster in real time, pseudo-clusters dominated by language features can be accurately identified. Once a preset drift threshold is triggered, adaptive fine-tuning is automatically initiated, locally adjusting the sensitivity of the density clustering algorithm, forcibly splitting and reclassifying suspected error regions, transforming qualitative distribution perception into quantitative mathematical monitoring, and improving the accuracy of cross-linguistic semantic segmentation. For industrial-grade real-time data stream scenarios, an efficient online projection and error diagnosis system is designed. By reusing trained geometric transformation operators, newly added text is instantaneously projected onto a unified space with extremely low latency, meeting real-time processing requirements. Simultaneously, by using a sliding time window to calculate the manifold projection error rate in real time and dynamically monitoring the evolution of the underlying semantic space, the system can automatically trigger asynchronous updates when facing concept drift, ensuring high availability and robustness over long periods of operation.

[0064] In summary, the cross-lingual text clustering method based on high-dimensional vector space manifold alignment provided in this application has the following beneficial effects: By obtaining a source language text set and a target language text set, features are extracted from both sets to construct independent source language feature matrices and target language feature matrices; based on the local topological structure of the source language feature matrices and target language feature matrices, unsupervised topological matching is performed to generate source language pseudo-anchor matrix and target language pseudo-anchor matrix; spatial centering is performed on the source language pseudo-anchor matrix and target language pseudo-anchor matrix to obtain corresponding centered matrices; orthogonal Prouke analysis is performed on the centered matrices to solve for the globally optimal orthogonal transformation matrix and scaling factor; based on the orthogonal transformation matrix, scaling factor, and translation vector, a global rigid transformation is performed on the source language feature matrix to obtain the aligned source language feature matrix; the aligned source language feature matrix and target language feature matrix are fused to construct a unified manifold space, and clustering is performed in the unified manifold space to obtain the semantic clustering results of the cross-lingual texts. In other words, by treating the semantic spaces of different languages ​​as high-dimensional manifolds with similar geometric topologies, a small number of pseudo-anchor points are obtained through unsupervised topological matching. Then, orthogonal Proukt analysis is used to calculate the optimal global rotation, translation, and scaling matrices in the ultra-high-dimensional space. Non-intrusive geometric rigid body transformations are used to eliminate physical displacements caused by language differences, enabling semantic clusters of different languages ​​to achieve topological overlap in a unified space, thereby improving the accuracy of cross-language text clustering.

[0065] Example 2: Based on the same inventive concept as the cross-lingual text clustering method based on high-dimensional vector space manifold alignment in Example 1, this application also provides a cross-lingual text clustering system based on high-dimensional vector space manifold alignment. Please refer to the appendix. Figure 2The cross-lingual text clustering system based on high-dimensional vector space manifold alignment includes: a feature extraction module 11, used to acquire a source language text set and a target language text set, extract features from the source language text set and the target language text set respectively, and construct mutually independent source language feature matrices and target language feature matrices; an unsupervised topology matching module 12, used to perform unsupervised topology matching based on the local topological structure of the source language feature matrix and the target language feature matrix, and generate a source language pseudo-anchor matrix and a target language pseudo-anchor matrix; and a space centering processing module 13, used to process the source language pseudo-anchor matrix and the target language pseudo-anchor matrix. The pseudo-anchor matrix is ​​spatially centered to obtain the corresponding centered matrix; the orthogonal solution module 14 is used to perform orthogonal Proukt analysis on the centered matrix to solve for the globally optimal orthogonal transformation matrix and scaling factor; the global rigid transformation module 15 is used to perform a global rigid transformation on the source language feature matrix based on the orthogonal transformation matrix, scaling factor and translation vector to obtain the aligned source language feature matrix; the clustering module 16 is used to fuse the aligned source language feature matrix with the target language feature matrix to construct a unified manifold space, and perform clustering in the unified manifold space to obtain the semantic clustering results of cross-language texts.

[0066] Furthermore, the unsupervised topological matching module 12 in the cross-language text clustering system based on high-dimensional vector space manifold alignment is also used to: construct local topological signature vectors for the source language nodes and target language nodes corresponding to the source language feature matrix and target language feature matrix, respectively. The local topological signature vector searches for the k nearest neighbor nodes in the same language space using Euclidean distance, and extracts the distance values ​​of the k nearest neighbor nodes to form an ordered distance vector, which serves as the topological signature vector of the corresponding node, as follows: S(x i )=[d1,d2,...,d k ]; where S(x i ) represents the topological signature vector of the source language node, d k The distance value representing the k-th nearest neighbor node, where d1≤d2≤...≤d k Unsupervised topology matching is performed based on the topology signature vector to generate a source language pseudo-anchor matrix and a target language pseudo-anchor matrix.

[0067] Furthermore, the unsupervised topology matching module 12 in the cross-language text clustering system based on high-dimensional vector space manifold alignment is also used to: define a topology matching confidence function using the topology signature vector, as follows: ;in, The topological signature vector representing the target language node. To control the bandwidth parameter of the distribution smoothness, σ∈[0.1,1.0]; confidence is calculated according to the topology matching confidence function, and matching is performed by bidirectional nearest neighbor filtering to generate source language pseudo-anchor matrix and target language pseudo-anchor matrix.

[0068] Furthermore, the space centering module 13 in the cross-lingual text clustering system based on high-dimensional vector space manifold alignment is also used to: calculate the centroid vectors of the source language pseudo-anchor matrix and the target language pseudo-anchor matrix respectively, as follows: ; ; where μ X Let μ be the centroid vector of the pseudo-anchor matrix of the source language. Y Let X be the centroid vector of the pseudo-anchor matrix of the target language, N be the number of pseudo-anchors, and X be the centroid vector of the pseudo-anchor matrix. i Y represents the X-dimensional feature components of the i-th source language pseudo-anchor point in the embedding space. i The Y-dimensional feature components of the i-th target language pseudo-anchor point in the embedding space are represented; spatial centering is performed using the centroid vector to obtain the corresponding centering matrix.

[0069] Furthermore, the space centering module 13 in the cross-language text clustering system based on high-dimensional vector space manifold alignment is also used for: space centering processing as follows: ; ; among which, 1 N It is a column vector of all 1s. Transpose processing of the centroid vector representing the pseudo-anchor matrix of the source language. Transpose of the centroid vector representing the pseudo-anchor matrix of the target language.

[0070] Furthermore, the orthogonal solution module 14 in the cross-language text clustering system based on high-dimensional vector space manifold alignment is also used to: solve for the solution satisfying R by performing singular value decomposition of the cross covariance matrix. T R=I d The optimal orthogonal rotation matrix R, where det(R) = 1. * Calculate the scaling factor and translation vector as follows: ; Where s represents the scaling factor, t is the translation vector, and Tr( The trace represents the matrix.

[0071] Furthermore, the clustering processing module 16 in the cross-language text clustering system based on high-dimensional vector space manifold alignment is also used to: perform high-dimensional spatial density clustering based on the aligned unified manifold space matrix to generate a cross-language semantic cluster set; and calculate the cross-language cluster intra-entropy for each valid cluster in the cross-language semantic cluster set, as follows: ;wherein, H(C i Characterizing the effective cluster C i The cross-language cluster intra-entropy, where L is the set of languages ​​involved in the system, p(l|C i Characterizing the effective cluster C i The proportion of text belonging to language l in the total number of texts in the effective cluster is used to determine the number of texts in the effective cluster. Based on the entropy within the cross-language cluster, the clustering drift index is evaluated, and the semantic clustering result of the cross-language text is generated by eliminating residual drift.

[0072] Furthermore, the clustering processing module 16 in the cross-lingual text clustering system based on high-dimensional vector space manifold alignment is also used for: calculating the clustering drift index as follows: Where CDI is the cluster drift index, |C| is the total number of effective clusters, and |L| is the number of language types.

[0073] Furthermore, the cross-lingual text clustering system based on high-dimensional vector space manifold alignment further includes: when new text data is added, a monolingual pre-trained model is invoked to extract the feature matrix, and the current rigid body transformation operator is reused to project the feature matrix onto a unified manifold space, as follows: ; where △V src Characteristic matrix, 1 △M Let t be a column vector of all 1s. T Let be the transpose of the translation vector; calculate the manifold projection error rate of the current transformation matrix on the new data, as follows: Where MPER is the manifold projection error rate, ||·||² is the L2 norm of the vector, and ΔX i Let ΔY be the difference vector of the i-th pseudo-anchor point in the source language. i Characterization and △X i The corresponding i-th difference vector of the target language; establish cross-language clustering labels for incremental text based on manifold projection error rate.

[0074] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The cross-lingual text clustering method and specific examples based on high-dimensional vector space manifold alignment in the foregoing embodiment one are also applicable to the cross-lingual text clustering system based on high-dimensional vector space manifold alignment in this embodiment. Through the foregoing detailed description of the cross-lingual text clustering method based on high-dimensional vector space manifold alignment, those skilled in the art can clearly understand the cross-lingual text clustering system based on high-dimensional vector space manifold alignment in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0075] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0076] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A cross-lingual text clustering method based on high-dimensional vector space manifold alignment, characterized in that, include: Obtain a source language text set and a target language text set, extract features from the source language text set and the target language text set respectively, and construct mutually independent source language feature matrices and target language feature matrices; Based on the local topological structure of the source language feature matrix and the target language feature matrix, unsupervised topological matching is performed to generate the source language pseudo-anchor matrix and the target language pseudo-anchor matrix. The source language pseudo-anchor matrix and the target language pseudo-anchor matrix are spatially centered to obtain the corresponding centered matrix; Perform orthogonal Pluke analysis on the centered matrix to find the globally optimal orthogonal transformation matrix and scaling factor; Based on the orthogonal transformation matrix, scaling factor and translation vector, a global rigid body transformation is performed on the source language feature matrix to obtain the aligned source language feature matrix. The aligned source language feature matrix and target language feature matrix are fused to construct a unified manifold space, and clustering is performed in the unified manifold space to obtain the semantic clustering results of cross-language texts.

2. The cross-lingual text clustering method based on high-dimensional vector space manifold alignment as described in claim 1, characterized in that, Generate the source language pseudo-anchor matrix and the target language pseudo-anchor matrix, including: For the source language node and target language node in the source language feature matrix and target language feature matrix respectively, a local topological signature vector is constructed. The local topological signature vector searches for the k nearest neighbor nodes in the same language space using Euclidean distance, and extracts the distance values ​​of the k nearest neighbor nodes to form an ordered distance vector, which serves as the topological signature vector of the corresponding node, as follows: ； in, The topological signature vector representing the source language node. The distance value representing the k-th nearest neighbor node, and ; Unsupervised topological matching is performed based on the topological signature vector to generate a source language pseudo-anchor matrix and a target language pseudo-anchor matrix.

3. The cross-lingual text clustering method based on high-dimensional vector space manifold alignment as described in claim 2, characterized in that, Performing unsupervised topology matching based on the topology signature vector includes: The topology matching confidence function is defined using the aforementioned topology signature vector, as follows: ； in, The topological signature vector representing the target language node. The bandwidth parameter is used to control the smoothness of the distribution. ; Confidence is calculated based on the topology matching confidence function, and matching is performed through bidirectional nearest neighbor filtering to generate source language pseudo-anchor matrix and target language pseudo-anchor matrix.

4. The cross-lingual text clustering method based on high-dimensional vector space manifold alignment as described in claim 1, characterized in that, The source language pseudo-anchor matrix and the target language pseudo-anchor matrix are spatially centered to obtain the corresponding centered matrices, including: The centroid vectors of the source language pseudo-anchor matrix and the target language pseudo-anchor matrix are calculated as follows: ； ； in, Let the centroid vector of the pseudo-anchor matrix of the source language be . Let N be the centroid vector of the pseudo-anchor matrix of the target language, and N be the number of pseudo-anchors. The X-dimensional feature components representing the i-th source language pseudo-anchor point in the embedding space. The Y-dimensional feature component representing the i-th target language pseudo-anchor point in the embedding space; The centroid vector is used to perform spatial centering to obtain the corresponding center matrix.

5. The cross-lingual text clustering method based on high-dimensional vector space manifold alignment as described in claim 4, characterized in that, The space centralization process is as follows: ； ； in, It is a column vector of all 1s. Transpose processing of the centroid vector representing the pseudo-anchor matrix of the source language. Transpose of the centroid vector representing the pseudo-anchor matrix of the target language.

6. The cross-lingual text clustering method based on high-dimensional vector space manifold alignment as described in claim 1, characterized in that, Perform orthogonal Pluke analysis on the centered matrix to solve for the globally optimal orthogonal transformation matrix and scaling factor, including: By performing singular value decomposition of the cross covariance matrix, we can solve for the solution that satisfies R. T R=I d The optimal orthogonal rotation matrix R, where det(R) = 1. * ; The scaling factor and translation vector are calculated as follows: ； ； Where s represents the scaling factor, t is the translation vector, and Tr( The trace represents the matrix.

7. The cross-lingual text clustering method based on high-dimensional vector space manifold alignment as described in claim 1, characterized in that, Clustering is performed in the unified manifold space to obtain semantic clustering results for cross-linguistic texts, including: High-dimensional spatial density clustering is performed based on the aligned unified manifold space matrix to generate a cross-language semantic cluster set; For each valid cluster in the cross-lingual semantic cluster set, the cross-lingual cluster intra-cluster entropy is calculated as follows: ； in, Characterizing effective clusters The cross-language cluster intra-entropy, where L is the set of languages ​​involved in the system. Characterizing effective clusters The proportion of text belonging to language l in the total number of texts in the effective cluster; Cluster drift index evaluation is performed based on the cross-language cluster entropy, and semantic clustering results of cross-language text are generated by eliminating residual drift.

8. The cross-lingual text clustering method based on high-dimensional vector space manifold alignment as described in claim 7, characterized in that, The cluster drift index is calculated as follows: ； Where CDI is the cluster drift index, |C| is the total number of effective clusters, and |L| is the number of language types.

9. The cross-lingual text clustering method based on high-dimensional vector space manifold alignment as described in claim 1, characterized in that, When new text data is added, the monolingual pre-trained model is called to extract the feature matrix, and the current rigid body transformation operator is reused to project the feature matrix onto the unified manifold space, as follows: ； in, Characteristic matrix, It is a column vector of all 1s. This is the transpose of the translation vector; The manifold projection error rate of the current transformation matrix on the new data is calculated as follows: ； Where MPER is the manifold projection error rate. Let L2 norm be the vector. Let i be the difference vector of the i-th pseudo-anchor point in the source language. Characterization and The i-th difference vector of the corresponding target language; Cross-linguistic clustering labels for incremental texts are established based on manifold projection error rates.

10. A cross-lingual text clustering system based on high-dimensional vector space manifold alignment, characterized in that, The steps for implementing the cross-lingual text clustering method based on high-dimensional vector space manifold alignment as described in any one of claims 1 to 9, wherein the cross-lingual text clustering system based on high-dimensional vector space manifold alignment comprises: The feature extraction module is used to obtain a source language text set and a target language text set, extract features from the source language text set and the target language text set respectively, and construct mutually independent source language feature matrices and target language feature matrices; An unsupervised topology matching module is used to perform unsupervised topology matching based on the local topological structure of the source language feature matrix and the target language feature matrix, and generate a source language pseudo-anchor matrix and a target language pseudo-anchor matrix. The spatial centralization processing module is used to perform spatial centralization processing on the source language pseudo-anchor matrix and the target language pseudo-anchor matrix to obtain the corresponding centralized matrix; The orthogonal solution module is used to perform orthogonal Proukt analysis on the centered matrix to solve for the globally optimal orthogonal transformation matrix and scaling factor; The overall rigid body transformation module is used to perform an overall rigid body transformation on the source language feature matrix based on the orthogonal transformation matrix, scaling factor and translation vector to obtain the aligned source language feature matrix. The clustering module is used to fuse the aligned source language feature matrix with the target language feature matrix to construct a unified manifold space, and perform clustering processing in the unified manifold space to obtain the semantic clustering results of cross-language texts.

Images