Data analysis method and system based on artificial intelligence

Abstract

The invention discloses a data analysis method and system based on artificial intelligence, and relates to the technical field of artificial intelligence data process.The method comprises the steps that an input text is split, and semantic representation is output through a multi-scale pyramid; generating semantic anchor points through a dynamic time warping mechanism, aligning input semantic representations, constructing a domain knowledge structure, and converting the domain knowledge structure into a semantic prior tensor; inputting the semantic priori tensor into a priori injection channel of a text encoder, and encoding the semantic representation to form an encoded representation; constructing a semantic graph according to the coded representation and the semantic representation, inputting the semantic graph into a graph neural network, and outputting a graph enhanced representation; according to the coding representation and the graph enhancement representation, constructing a joint representation, training an analysis model, and outputting a cross-language representation; clustering the cross-language representation to generate an adversarial sample, and constructing a semantic causal path diagram with an original sample; the problems of semantic segmentation, alignment deviation and weak cross-language migration are solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence data processing technology, and more specifically, to a data analysis method and system based on artificial intelligence. Background Technology

[0002] With the rapid development of artificial intelligence technology, data analysis is playing an increasingly prominent role in fields such as financial risk control, public opinion monitoring, intelligent customer service, intelligent search, and cross-language retrieval. Modern application scenarios exhibit multi-source, multimodal, and multilingual data formats, with text, voice, images, and other information types collectively forming a complex data ecosystem. Especially in cross-language and cross-modal tasks, accurately understanding semantics, achieving effective alignment, and constructing robust semantic structures have become core challenges for intelligent analysis systems. However, existing data analysis methods still have significant shortcomings in semantic granularity, structural expression, cross-language alignment, and knowledge utilization, making it difficult to support the demands for high-precision and highly robust intelligent analysis.

[0003] For example, the invention patent with publication number CN120448563A discloses a semantic understanding-driven cross-modal information fusion and retrieval method and system. This method acquires raw text, image, and audio data, and extracts initial feature sets for each modality using deep neural networks. It dynamically allocates modal weight coefficients based on an attention mechanism, weighting and fusing the initial feature sets to obtain a cross-modal fused feature representation. Through a cross-modal semantic association analysis model, it extracts high-dimensional semantic association features from the fused feature representation to generate a semantically enhanced feature vector. Based on this vector, it constructs a cross-modal semantic graph network to complete missing modal features, generating an optimized multimodal feature set. Finally, it inputs the optimized feature set and query samples into a comparative learning model to calculate semantic similarity scores, and generates a ranked list of cross-modal retrieval results based on the scores.

[0004] For example, the semantic sentiment analysis method that integrates deep features and temporal models disclosed in the invention patent with publication number CN112560503A includes: text preprocessing, converting the text into a word vector matrix with a uniform format; extracting local semantic sentiment text features and extracting contextual semantic sentiment text features using the word vector matrix as input; generating fused semantic sentiment text features by weighting the local semantic sentiment text features and contextual semantic sentiment text features using an attention mechanism; concatenating the local semantic sentiment text features, the contextual semantic sentiment text features, and the fused semantic sentiment text features to generate global semantic sentiment text features; and performing the final text sentiment semantic analysis and recognition using a softmax classifier with the global semantic sentiment text features as input.

[0005] The above-disclosed technical solutions have at least the following technical problems: Traditional models have limited ability to align semantic sequences of varying lengths, especially in cross-linguistic and cross-modal tasks, often resulting in large alignment deviations and semantic misalignments. At the same time, sequence-based models have inherent limitations in handling long-distance dependencies, cross-sentence relationships, and semantically structured information.

[0006] To address the above problems, this invention proposes a solution. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a data analysis method and system based on artificial intelligence, which solves the problems of semantic fragmentation, alignment deviation, insufficient knowledge utilization, and weak cross-language transfer by multi-scale semantic fusion, dynamic alignment, knowledge injection, and graph augmentation modeling.

[0008] To achieve the above objectives, the present invention provides the following technical solution: An AI-based data analysis method includes: semantically decomposing the input text to obtain multi-scale semantic units, mapping features at each level to a unified semantic space through a multi-scale pyramid, and outputting a semantic representation; generating semantic anchors through a dynamic time warping mechanism to align the input semantic representation, constructing a domain knowledge structure, and converting the output domain knowledge tensor into a semantic prior tensor; inputting the semantic prior tensor into the prior injection channel of a text encoder, fusing the semantic prior tensor with the bias term of an attention structure, and jointly encoding the semantic representation to form an encoded representation; constructing a semantic graph based on the encoded representation and semantic representation, and inputting it into a graph neural network for node information transfer and aggregation, outputting a graph-enhanced representation; constructing a speech-text joint representation based on the encoded representation and graph-enhanced representation, training a transferable semantic kernel to form an analysis model, and outputting a cross-linguistic representation; clustering the cross-linguistic representation to form error categories, generating adversarial examples based on the error categories, forming a self-evolving data pool with the original samples, and constructing a semantic causal path graph.

[0009] In a preferred embodiment, the semantic segmentation of the input text to obtain multi-scale semantic units, and the mapping of features at each level to a unified semantic space through a multi-scale pyramid, outputting a semantic representation, is as follows: The input text is preprocessed by dividing it into sentence sequences based on syntactic rules and generating sentence boundary markers to obtain an initial text structure; based on the initial text structure, character segmentation, word segmentation, phrase extraction, and paragraph aggregation are performed sequentially to form a set of multi-scale semantic units; the multi-scale semantic features in the set of multi-scale semantic units are input to the multi-scale pyramid according to semantic granularity from low to high; through cross-layer feature fusion and inter-layer mapping mechanisms, intermediate features associated with the unified semantic space are generated at each layer; based on the intermediate features generated by the multi-scale pyramid, the sequence length of the low-granularity layer features is mapped to the high-granularity feature sequence through upsampling; for each scale, the upsampled features are projected to the unified semantic dimension through a shared projection matrix, and the features of all scales are fused in hierarchical order to form a fused semantic vector; the fused semantic vector is normalized, and through a structure rearrangement module, the fused vector is reorganized according to the initial text structure order to obtain a semantic representation.

[0010] In a preferred embodiment, the multi-scale semantic features from the multi-scale semantic unit set are input into a multi-scale pyramid according to semantic granularity from low to high. Through cross-layer feature fusion and inter-layer mapping mechanisms, intermediate features associated with a unified semantic space are generated at each layer. Specifically, the process is as follows: character, word, phrase, paragraph, and topic features from the multi-scale semantic unit set are input into the multi-scale pyramid according to semantic granularity from low to high; the pyramid performs cross-layer attention alignment using high-level features as queries and low-level features as keys and values, extracting fine-grained semantic information from the low-level layers; a shared projection matrix is ​​obtained through multi-scale contrastive learning pre-training during the pyramid initialization phase; the aligned low-level features are mapped to a unified semantic dimension using the shared projection matrix, and the low-level mapped features are fused with the current layer's high-level features through a gated weighted fusion mechanism to form fused features; the fused features are input into the Transformer within the layer for semantic modeling, serving as the intermediate semantic state of that layer. This process is repeated layer by layer to generate intermediate features, ultimately forming a complete intermediate semantic representation corresponding to the unified semantic space at the top layer.

[0011] In a preferred embodiment, the step of generating semantic anchors through a dynamic time warping mechanism, aligning the input semantic representations, constructing a domain knowledge structure, and converting the output domain knowledge tensor into a semantic prior tensor is as follows: The semantic representation is used as the input sequence, and dynamic time warping is performed in conjunction with a reference semantic sequence to calculate the optimal alignment path between sequences. The semantic position corresponding to each input semantic representation and the reference sequence is determined, and the key positions of the alignment path are marked as semantic anchors. Based on the semantic anchors, the input sequence is aggregated through an attention mechanism to generate the aligned semantic representation. The aligned semantic representation is then used as... The process begins by extracting entity, relation, or concept-level information to construct a domain knowledge structure graph. This graph is then represented as a set of nodes and edges. Semantic feature vectors are extracted from each node, and fixed-dimensional vectors are generated using embedding methods. All node vectors are then stacked to form a node feature matrix. An adjacency matrix is ​​constructed based on the edge set and subjected to symmetric normalization. The node feature matrix is ​​combined with the normalized adjacency matrix to form a three-dimensional tensor, which is then input into a graph neural network to generate a domain knowledge tensor. Finally, the domain knowledge tensor is fused with the aligned semantic representation through a linear mapping to generate a semantic prior tensor.

[0012] In a preferred embodiment, the step of inputting the semantic prior tensor into the prior injection channel of the text encoder, fusing the semantic prior tensor with the bias term of the attention structure, and jointly encoding the semantic representation to form an encoded representation is as follows: The semantic prior tensor is input into a linear mapping layer, and its dimension is transformed using a trained shared projection matrix to obtain dimensionally aligned semantic prior features; the aligned semantic prior features are then input into two different linear transformation layers for transformation to obtain prior key vectors and prior value vectors; the aligned semantic prior features are input into a bias generation network, and a bias matrix with the same size as the attention score matrix is ​​calculated through feedforward; the input text is processed by the encoder to generate text key vectors and text value vectors; the text key vectors and prior key vectors are then processed along the sequence... Dimensional concatenation yields an extended key vector set; text value vectors and prior value vectors are concatenated in the same way to form an extended value vector set; the query vector is sequentially multiplied by each key vector in the extended key vector set, and then scaled by the square root of the feature dimension to obtain the basic attention score matrix; the basic attention score matrix is ​​element-wise added to the bias matrix to obtain the bias-enhanced attention score matrix, i.e., the bias-enhanced attention matrix; the bias-enhanced attention matrix is ​​subjected to softmax normalization to obtain the final attention weight matrix; the attention weight matrix is ​​weighted and summed with the extended value vector set to obtain the fused output; the fused output is input into the feedforward network inside the encoder, and then processed by residual connections and layer normalization to finally generate a joint encoded representation.

[0013] In a preferred embodiment, the step of constructing a semantic graph based on the encoded representation and semantic representation, and inputting it into a graph neural network for node information transfer and aggregation, outputting a graph-enhanced representation, is as follows: Using the encoded representation as the basic semantic vector and the semantic representation as the supplementary semantic vector, nodes are constructed one-to-one; the correlation strength between any two nodes is output using the cosine similarity method; when the correlation strength exceeds a preset threshold, an edge connection is established between the nodes, and a relation type label is attached to the edge, forming a semantic graph structure; the node content in the semantic graph is converted into node feature vectors, and the connection relationships between nodes are converted into adjacency structures; the node feature vectors and adjacency structures are input into the graph neural network, and the features of adjacent nodes are weighted and combined using the neighborhood aggregation mechanism of the graph neural network to obtain the updated representation of each node; the node updated representation of the previous layer is used as input in the graph neural network, and neighborhood information aggregation is performed layer by layer to form multi-layer node representations; in the output layer of the graph neural network, the multi-layer node representations are collected and organized to become a high-level semantic feature expression of the entire graph, forming a graph-enhanced representation.

[0014] In a preferred embodiment, the step of inputting node feature vectors and adjacency structures into a graph neural network, and weighting and combining the features of adjacent nodes through the neighborhood aggregation mechanism of the graph neural network to obtain the updated representation of each node, is as follows: The feature vector corresponding to each node and the adjacency structure of the entire semantic graph are input into the graph convolutional layer of the graph neural network; In the graph convolutional layer, the graph neural network determines the direct neighbor range of each node based on the pre-encoded node connection information in the adjacency structure, and retrieves the feature vectors of all neighboring nodes connected to the node according to the adjacency relationship; A set of independent weight parameters is assigned to the neighbor features of each node, and the neighbor feature vectors are multiplied by the corresponding weights during calculation, while a trainable bias is added so that the features of each neighboring node undergo a structured transformation before fusion; After the transformation is completed, the contribution ratio of the neighboring nodes to the target node is determined based on the edge type attached to the semantic graph; The transformed features of each neighboring node are multiplied by its corresponding contribution weight, and the weighted results of all neighboring nodes are summed to obtain the updated representation of each node.

[0015] In a preferred embodiment, the construction of a joint speech-text representation based on the encoded representation and graph augmentation representation, and the training of a transferable semantic kernel forming analysis model to output a cross-linguistic representation, specifically involves: dividing the speech signal into several frames with a fixed frame length; extracting frequency domain features and acoustic features from each frame using short-time Fourier transform to form a speech feature sequence; inputting the speech feature sequence into a speech encoder, extracting high-level acoustic representations through a Transformer structure to obtain a speech encoded vector; calculating the minimum cost alignment path between the two sequences using a dynamic time warping algorithm based on the speech encoded vector and the encoded representation to form an alignment mapping table; adjusting the dimensions of the encoded representation and the graph augmentation representation through independent linear transformation layers, and then concatenating the adjusted vectors to generate a text augmentation table. The speech encoding vector is first identified; then, according to the alignment mapping table, the speech encoding vector is matched with the corresponding text enhancement vector to generate a paired vector sequence; the paired vector sequence is input into the joint encoder, and bidirectional interactive attention is calculated through a cross-modal attention mechanism to generate a multimodal interactive representation for each unit; the multimodal interactive representation is processed by a feedforward network, layer normalization, and residual connections to obtain a joint representation sequence; using the joint representation as input, a trainable semantic kernel network is built to output semantic kernel representations between different samples; according to task requirements, the semantic kernel module and the joint encoder are jointly trained through cross-entropy; after joint training, the joint encoder and semantic kernel module are fixed to build a cross-language analysis model and use it for inference, thus forming the analysis model; new speech-text input pairs are fed into the analysis model to obtain cross-language representations.

[0016] In a preferred embodiment, the process of clustering cross-linguistic representations to form error categories, generating adversarial examples based on these error categories, forming a self-evolving data pool with the original samples, and constructing a semantic causal path graph is as follows: The original data samples corresponding to the target task are obtained and preprocessed to form original samples; the original samples are input into an analysis model for processing to obtain the cross-linguistic representation of each sample, forming a cross-linguistic semantic representation set; the cross-linguistic representation set is analyzed using a density clustering algorithm, and samples with low semantic similarity are classified into different error categories, with each sample's error category label output, forming error samples; based on each error category, the original samples are perturbed using an adversarial example generator to generate corresponding adversarial examples; the original samples, adversarial examples, and error samples are combined to form a self-evolving data pool; based on the self-evolving data pool and the internal reasoning path of the analysis model, nodes and edges are extracted, and a semantic causal path graph is formed through structured modeling.

[0017] A system for data analysis based on artificial intelligence includes a semantic segmentation module, an anchor module, a joint encoding module, a semantic graph module, an analysis model module, and a clustering output module, with connections between the modules. The semantic segmentation module performs semantic segmentation on the input text to obtain multi-scale semantic units, and maps features at each level to a unified semantic space using a multi-scale pyramid to output a semantic representation. The anchor module generates semantic anchors through a dynamic time warping mechanism to align the input semantic representation and construct a domain knowledge structure, converting the output domain knowledge tensor into a semantic prior tensor. The joint encoding module inputs the semantic prior tensor into the text encoder. The system comprises the following modules: a priori injection module, which fuses semantic prior tensors with bias terms of the attention structure and jointly encodes semantic representations to form encoded representations; a semantic graph module, which constructs semantic graphs based on encoded and semantic representations and inputs them into a graph neural network for node information transfer and aggregation, outputting graph-enhanced representations; an analysis model module, which constructs speech-text joint representations based on encoded and graph-enhanced representations and trains transferable semantic kernels to form an analysis model, outputting cross-linguistic representations; and a clustering output module, which clusters cross-linguistic representations to form error categories, generates adversarial examples based on error categories, forms an evolutionary data pool with the original samples, and constructs a semantic causal path graph.

[0018] The technical effects and advantages of the data analysis method and system based on artificial intelligence of this invention are as follows: 1. This invention establishes a unified semantic space through multi-scale semantic decomposition and a multi-scale pyramid structure, achieving full-granularity semantic feature fusion from characters, words, phrases to paragraphs and topics. The multi-scale pyramid fully absorbs fine-grained and high-level semantic information during cross-layer feature fusion and inter-layer mapping, resulting in a final semantic representation with a consistent semantic scale and higher expression density, solving the problems of single-granularity and fragmented cross-layer information in traditional text representations. With the support of normalization and structural rearrangement, the semantic representation can strictly maintain the sequence consistency of the input text structure, thereby improving the stability of subsequent alignment, injection, and graph modeling.

[0019] 2. This invention generates semantic anchors through a dynamic temporal warping mechanism, which dynamically seeks the optimal alignment path for temporal differences between different semantic sequences, enabling precise semantic position mapping of text semantic representations across sequences of unequal length. Simultaneously, the domain knowledge structure construction driven by semantic anchors allows for the structured expression of knowledge at the entity, relation, and concept levels, which is then encoded into a semantic prior tensor via a graph neural network, giving the model explicit knowledge alignment and enhancement capabilities. This introduction of semantic priors addresses the problem of insufficient utilization of domain knowledge in end-to-end models. Attached Figure Description

[0020] Figure 1This is a flowchart illustrating a data analysis method based on artificial intelligence according to the present invention.

[0021] Figure 2 This is a schematic diagram of the system structure of a data analysis method based on artificial intelligence according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1, Figure 1 This invention provides a data analysis method based on artificial intelligence, comprising: S1 performs semantic segmentation on the input text to obtain multi-scale semantic units, and maps the features of each level to a unified semantic space through a multi-scale pyramid to output semantic representation; In this embodiment, the input text is semantically segmented to obtain multi-scale semantic units, and the features at each level are mapped to a unified semantic space through a multi-scale pyramid to output a semantic representation, as follows: The input text is preprocessed, including character filtering, punctuation normalization and special symbol cleanup, and the text is divided into sentence sequences based on syntactic rules, and sentence boundary markers are generated to obtain the initial text structure; Based on the initial text structure, character segmentation, word segmentation, phrase extraction, and paragraph aggregation are performed sequentially to generate semantic units of characters, words, phrases, and paragraphs, forming a multi-scale semantic unit set; Multi-scale semantic features from the set of multi-scale semantic units are input into the multi-scale pyramid according to semantic granularity from low to high. Through cross-layer feature fusion and inter-layer mapping mechanism, intermediate features associated with the unified semantic space are generated at each layer. Based on the intermediate features generated by the multi-scale pyramid, the low-granularity layer features are mapped to the sequence length of high-granularity features through upsampling methods (such as linear interpolation or transposed convolution); For each scale, the upsampled features are projected onto a unified semantic dimension through a shared projection matrix to ensure that features at different scales can be fused in the same semantic space. Features at all scales are fused in a hierarchical order to form a cross-scale consistent fused semantic vector. The fused semantic vectors are normalized (e.g., LayerNorm or L2 normalization) to eliminate differences in feature magnitude at different scales. The fused vectors are then reorganized according to the initial text structure order through the structure rearrangement module, so that the vector sequence is consistent with the semantic level of the original text, and finally a unified semantic representation is obtained.

[0024] In this embodiment, based on the initial text structure, character segmentation, word segmentation, phrase extraction, and paragraph aggregation are performed sequentially to generate character-level, word-level, phrase-level, and paragraph-level semantic units, forming a multi-scale semantic unit set, as detailed below: Each sentence is split into a character sequence, and the character sequence is vectorized using a one-hot encoding method to obtain character features; Using sentences as input, a word segmenter (statistical word segmentation) is used to generate word sequences, and the word sequences are encoded using a word vector model (Word2Vec or GloVe) to form word features; Using word sequences as input, a syntactic parser (dependency parsing or constituent parsing) identifies phrase units such as noun phrases and verb phrases, and performs phrase embedding (such as phrase pooling or phrase encoder) on the identified phrase units to generate phrase features; Sentences within the same paragraph are aggregated according to paragraph boundaries, and their corresponding character features, word features, and phrase features are collected. These feature sequences are then input into a paragraph encoder (such as an RNN) to generate paragraph features through positional encoding and attention mechanisms. Using paragraph features as input, topic semantic vectors are extracted and generated through text clustering methods (such as LDA and neural topic models). The character features, word features, phrase features, paragraph features, and topic-level semantic vectors are mapped according to granularity and organized into a hierarchical structure to form the final multi-scale semantic unit set.

[0025] In this embodiment, the multi-scale semantic features in the multi-scale semantic unit set are input into the multi-scale pyramid according to the semantic granularity from low to high. Through cross-layer feature fusion and inter-layer mapping mechanisms, intermediate features associated with the unified semantic space are generated at each layer, as follows: Characters, words, phrases, paragraphs, and topic features from the multi-scale semantic unit set are input into the multi-scale pyramid in order of semantic granularity from low to high. The pyramid uses high-level features as queries and low-level features as keys and values ​​to perform cross-layer attention alignment, extracting fine-grained semantic information from the low layer. Character and word features belong to low-level features, while phrase, paragraph, and topic features belong to high-level features. The fine-grained semantic information includes character combination patterns, word boundaries, and basic word meanings. During the pyramid initialization phase, a shared projection matrix is ​​obtained through multi-scale contrastive learning pre-training. By sharing the projection matrix, the aligned low-level features are mapped to a unified semantic dimension, and the low-level mapped features are fused with the current high-level features through a gated weighted fusion mechanism to form a hierarchical fused feature that combines details and abstract semantics. The Transformer within the fused feature input layer is used for semantic modeling, serving as the intermediate semantic state of that layer. This process is repeated layer by layer to generate intermediate features for each layer, and a complete intermediate semantic representation corresponding to a unified semantic space is formed at the top layer.

[0026] In this embodiment, aligned low-level features are mapped to a unified semantic dimension using a shared projection matrix, and the low-level mapped features are fused with the current high-level features using a gated weighted fusion mechanism to form a hierarchical fused feature that combines both detailed and abstract semantics, as detailed below: Dimensional mapping is performed using a unified projection matrix, which means performing a linear transformation on the low-level feature vectors to project them from the original dimension to a semantic dimension consistent with the current high-level features. The mapped low-level features are concatenated with the current high-level features and input into a gating network. The gating network consists of linear transformation and Sigmoid activation and is used to calculate the gating weights that represent the relative importance of the two types of features. Based on the gating weights, weights are applied to the low-level mapping features and the high-level features respectively, and element-wise weighted fusion is performed so that the low-level features contribute fine-grained semantics and the high-level features contribute abstract semantics, ultimately resulting in a hierarchical fusion feature vector that combines low-level details with high-level semantic abstraction capabilities.

[0027] In this embodiment, semantic modeling is performed on the Transformer within the fused feature input layer, serving as the intermediate semantic state of that layer. This process is repeated layer by layer to generate intermediate features for each layer, culminating in a unified intermediate semantic representation at the top layer that corresponds to a unified semantic space. Specifically: The hierarchical fusion features obtained by gating fusion are fed into the Transformer encoder of this layer. The semantic correlation between each vector within the feature is calculated through the self-attention mechanism, and then weighted convergence is performed to form the context-enhanced representation of this layer. By performing nonlinear transformation on the context representation through a feedforward network, the features are made capable of local semantic abstraction, and the output is stabilized through residual connection and normalization operations to obtain the intermediate semantic state of this layer. By using intermediate semantic states as input to the next layer of Transformer, it can perform self-attention modeling, context aggregation, and semantic abstraction again at a higher semantic granularity, thereby obtaining intermediate features at different semantic scales layer by layer upwards; When the pyramid reaches the top, the intermediate semantic states of each layer are spliced ​​or weighted and integrated in a preset order, and then dimension reorganization and semantic alignment are performed through a unified mapping network to finally form a complete intermediate semantic representation corresponding to the unified semantic space.

[0028] S2 generates semantic anchors through a dynamic time warping mechanism, aligns the input semantic representation, constructs a domain knowledge structure, and converts the output domain knowledge tensor into a semantic prior tensor. In this embodiment, a dynamic time warping mechanism is used to generate semantic anchors, align the input semantic representation, construct a domain knowledge structure, and convert the output domain knowledge tensor into a semantic prior tensor, as detailed below: The semantic representation is used as the input sequence. It is combined with the reference semantic sequence (such as standard text in the same domain or semantic template sequence extracted from the knowledge base) for dynamic time warping, calculate the optimal alignment path between the sequences, determine the semantic position of each input semantic representation and the reference sequence, and mark the key positions of the alignment path as semantic anchors. Based on semantic anchors, the input sequence is aggregated through an attention mechanism to align the sequence with the reference sequence in both time and semantic dimensions, thereby generating an aligned semantic representation. Based on the aligned semantic representation, entity, relation or concept hierarchical information is extracted to construct a domain knowledge structure graph, where nodes represent entities or concepts and edges represent causal or hierarchical relationships. The constructed domain knowledge structure is represented as a set of nodes and a set of edges, where nodes represent entities or concepts and edges represent the relationships between nodes; Extract the semantic feature vector of each node, generate a fixed-dimensional vector through an embedding method, and stack all node vectors to form a node feature matrix; Construct an adjacency matrix based on the edge set, and then perform symmetric normalization on the adjacency matrix, where: This indicates that there is a relationship between nodes i and j. This indicates that there is no relationship. The node feature matrix is ​​combined with the normalized adjacency matrix to form a three-dimensional tensor, which is then input into the graph neural network to generate a domain knowledge tensor. A semantic prior tensor is generated by fusing the domain knowledge tensor with the aligned semantic representation through a linear mapping.

[0029] In this embodiment, semantic representations are used as input sequences, and dynamic time warping is performed in conjunction with reference semantic sequences (such as standard text from the same domain or semantic template sequences extracted from a knowledge base). The optimal alignment path between sequences is calculated, the semantic position corresponding to each input semantic representation and the reference sequence is determined, and the key positions of the alignment path are marked as semantic anchors, as follows: The input semantic representation sequence and the reference semantic sequence are used as inputs to Dynamic Time Warping (DTW). The similarity of each pair of vectors is calculated using cosine similarity to form a distance matrix. The optimal cumulative path from the starting point to the ending point is searched on the distance matrix by dynamic programming. At each step of the path, the path can move to the right, down, or diagonally to minimize the cumulative distance, thereby determining the semantic position in the input sequence that best matches the reference sequence. After obtaining the optimal cumulative path, key points on the optimal cumulative path (such as points where the cumulative distance changes significantly at fixed step intervals) are marked as semantic anchors.

[0030] In this embodiment, based on semantic anchors, the input sequence is aggregated through an attention mechanism to align the sequence with the reference sequence in both time and semantic dimensions, generating an aligned semantic representation, as follows: Based on the semantic anchors generated by dynamic time warping, each anchor vector in the reference sequence is used as the query vector, and the vectors related to the semantic anchors in the input sequence are used as the key vector and the corresponding value vector. The weight of the input vector to the anchor is obtained by calculating the scaled dot product attention. The method for obtaining the weight of the input vector to the anchor point by calculating the scaled dot product attention is as follows: first, the query vector is multiplied by each key vector, then the vector is divided by the square root of the vector dimension for scaling, and then the results of all relevant key vectors under the same anchor point are normalized by the softmax function to obtain the weight coefficient. The corresponding value vectors are weighted and summed according to the weight coefficients to generate the alignment vector for each anchor point. Then, all alignment vectors are arranged in the order of the reference sequence to form a semantic representation that is aligned with the reference sequence in both time and semantic dimensions.

[0031] S3, input the semantic prior tensor into the prior injection channel of the text encoder, fuse the semantic prior tensor with the bias term of the attention structure, and jointly encode the semantic representation to form the encoded representation; In this embodiment, the semantic prior tensor is input into the prior injection channel of the text encoder, the semantic prior tensor is fused with the bias term of the attention structure, and the semantic representation is jointly encoded to form the encoded representation, as follows: The semantic prior tensor is input into the linear mapping layer, and its dimension is transformed by the trained shared projection matrix so that the dimension of the output semantic vector is consistent with the feature dimension inside the text encoder, thus obtaining the dimension-aligned semantic prior features. The aligned semantic prior features are input into two different linear transformation layers for transformation to obtain the prior key vector and the prior value vector. The aligned semantic prior features are input into the bias generation network, and a bias matrix with the same size as the attention score matrix is ​​obtained through feedforward calculation. The input text is processed by an encoder to generate text key vectors and text value vectors, which serve as the basic feature source for the attention mechanism. Concatenate the text key vector with the prior key vector along the sequence dimension to obtain the extended key vector set; The text value vector and the prior value vector are concatenated in the same way to form an extended value vector set; The query vector is sequentially multiplied by each key vector in the extended key vector set, and then scaled by the square root of the feature dimension to obtain the basic attention score matrix, which represents the similarity between the text query vector and all key vectors. The basic attention score matrix and the bias matrix are added element by element to obtain the bias-enhanced attention score matrix, i.e., the bias-enhanced attention matrix, which allows domain knowledge to directly influence attention weights. The bias-enhanced attention matrix is ​​normalized using softmax to obtain the final attention weight matrix, which represents the weighted ratio of the query vector to the text information and prior knowledge. Perform a weighted summation operation on the attention weight matrix and the extended value vector set to obtain the fused output; The feedforward network inside the fusion input and output encoders is then processed through residual connections and layer normalization to finally generate a joint encoded representation that simultaneously incorporates textual semantics and domain knowledge priors.

[0032] In this embodiment, the aligned semantic prior features are input into two different linear transformation layers for transformation to obtain prior key vectors and prior value vectors, as follows: The dimension-aligned semantic prior features are sequentially input into two structurally independent linear transformation layers. Each linear transformation layer contains a set of trainable weight parameters and bias parameters. During forward computation, the system performs a vector multiplication operation between the semantic prior features and the weights of one of the linear transformation layers, and adds the corresponding bias to obtain the first set of output features. Since this linear layer is specifically designed to generate features that express similarity, the output is defined as the prior key vector; Subsequently, the same set of semantic prior features is input into another linear transformation layer. The weights and bias parameters of this linear layer are trained independently and are not shared with the previous linear layer. The output of the second set of features is used to represent content information and is defined as a prior value vector.

[0033] The input text is processed by an encoder to generate text key vectors and text value vectors, as follows: First, the input text is converted into a word segmentation sequence and mapped into word vectors. Then, these word vectors are input into the text encoder in sequence. Text encoders typically employ a Transformer architecture, which includes multiple layers of self-attention modules and a feedforward network. When processing text sequences, the system sets up two independent linear transformation units within each layer of the self-attention module: one set is used to generate the key vector of the text, and the other set is used to generate the value vector of the text. The two linear transformation units each hold independently trained weights and bias parameters. When the encoder runs, it inputs each word vector into the linear transformation unit used to generate key vectors. Through vector multiplication and bias adjustment, it obtains key vectors that represent semantic matching relationships. At the same time, the system inputs the same word vector into the linear transformation unit used to generate the value vector, and obtains the value vector used to express the text content in the same way; As the text is processed layer by layer in the multi-layer encoder, the key vector and value vector are continuously updated at different levels, eventually generating a complete sequence of text key vectors and text value vectors at the encoder output.

[0034] S4. Construct a semantic graph based on the encoding representation and semantic representation, and input it into the graph neural network for node information transmission and aggregation, outputting an enhanced graph representation; In this embodiment, a semantic graph is constructed based on the encoded representation and the semantic representation, and then input into a graph neural network for node information transmission and aggregation, outputting an enhanced graph representation, as follows: Using encoded representation as the basic semantic vector and semantic representation as the supplementary semantic vector, the two are constructed into nodes one by one. Each node contains semantic content, syntactic features and contextual representation, which are used as the node set of the semantic graph. The cosine similarity method is used to output the association strength between any two nodes. When the association strength exceeds a preset threshold, an edge is established between the nodes, and a relation type label is attached to the edge, such as modification relation, subject-predicate relation, semantic consistency relation, etc., thus forming a semantic graph structure composed of nodes and edges. The content of nodes in the semantic graph is converted into node feature vectors, and the connection relationship between nodes is converted into an adjacency structure description, including node index, edge connection information and relationship type. The adjacency structure tells the graph neural network the reference path and information transmission direction between nodes, ensuring that the subsequent graph convolution can be executed effectively. The node feature vectors and adjacency structure are input into the graph neural network. The features of adjacent nodes are weighted and combined through the neighborhood aggregation mechanism of the graph neural network to complete the information transfer process between nodes and obtain the updated representation of each node. In graph neural networks, the node update representation of the previous layer is used as input, and neighborhood information aggregation is performed layer by layer. This allows the nodes to gradually absorb semantic information from a wider range of neighborhoods during the propagation of multiple layers, forming a multi-layer node representation that includes local and global semantic structures. In the output layer of the graph neural network, the multi-layer node representations obtained from the final layer are collected and organized to become a high-level semantic feature representation of the entire graph, thus forming a graph-enhanced representation.

[0035] In this embodiment, the node feature vectors and adjacency structure are input into the graph neural network. The features of adjacent nodes are weighted and combined through the neighborhood aggregation mechanism of the graph neural network to complete the information transfer process between nodes and obtain the updated representation of each node, as follows: The feature vector corresponding to each node, along with the adjacency structure of the entire semantic graph, is input into the graph convolutional layer of the graph neural network. In the graph convolutional layer, the graph neural network determines the range of direct neighbors of each node based on the node connection information pre-encoded in the adjacency structure, and retrieves the feature vectors of all neighboring nodes connected to the node according to the adjacency relationship. Each node's neighbor features are assigned a set of independent weight parameters, and the neighbor feature vectors are multiplied by the corresponding weights during computation. At the same time, a trainable bias is added so that the features of each neighbor node undergo a structured transformation before fusion, thereby enhancing the ability to distinguish between different neighbors. After the transformation is completed, the contribution ratio of neighboring nodes to the target node is determined based on the edge type attached to the semantic graph. For example, if the relationship between two nodes comes from syntactic subject-predicate connection, the system will assign a higher aggregation weight to the neighboring node. If the relationship is weak, a lower weight will be assigned, thus realizing differentiated information fusion based on structural relationship. Multiply the transformed features of each neighboring node by its corresponding contribution weight, and then sum the weighted results of all neighboring nodes to obtain the neighborhood feature aggregation result of that node, which is the updated representation of each node.

[0036] S5. Based on the encoding representation and graph augmentation representation, construct the speech-text joint representation, train the transferable semantic kernel to form an analysis model, and output cross-linguistic representation; In this embodiment, a joint speech-text representation is constructed based on the encoded representation and the graph-enhanced representation, and a transferable semantic kernel forming analysis model is trained to output a cross-linguistic representation, as follows: The speech signal is divided into several frames with a fixed frame length. Frequency domain features are extracted from each frame by short-time Fourier transform, and acoustic features (such as log-Mel spectrum or Mel-Cepstral coefficients) are also extracted. At the same time, time position markers are added to each frame to form a speech feature sequence. The speech feature sequence is input into the speech encoder, and the high-level acoustic representation is extracted through the Transformer structure to obtain the speech coding vector; Based on the speech coding vector and the coding representation, the speech coding vector is used as the first sequence and the coding representation is used as the second sequence. The minimum cost alignment path between the two sequences is calculated by the dynamic time warping algorithm, and the text unit corresponding to each speech frame in the path is recorded as a mapping relationship to form a speech-text alignment mapping table. The dimensionality of the encoded representation and the graph augmentation representation is adjusted by an independent linear transformation layer to make them have a consistent vector dimension. Then, the adjusted vectors are concatenated to generate a text augmentation representation vector that simultaneously contains local text semantic features and cross-node structural semantic features. According to the alignment mapping table, the speech coding vector is matched with the corresponding text enhancement vector. If some speech frames fail to align to the text units, they are filled in by interpolation to generate a complete pairing vector sequence. The paired vector sequence is input into the joint encoder, and bidirectional interactive attention is calculated through a cross-modal attention mechanism to generate a multimodal interactive representation for each unit; The multimodal interaction representation is processed by a feedforward network, layer normalization, and residual connections to obtain a speech-text joint representation sequence with cross-modal semantic fusion capability. The multimodal interaction representation first enters the feedforward network for nonlinear transformation, then undergoes layer normalization to improve training stability, and then the input features are added to the feedforward output through residual connections to generate a joint representation sequence that has expressive capabilities in both speech and text semantic channels. Using the joint representation as input, a trainable semantic kernel network is constructed, including a trainable projection layer, a similarity calculation layer, and a normalized output layer, which are used to output semantic kernel representations between different samples. The trainable projection layer is used to map the joint representation to the semantic kernel space. The similarity calculation layer measures the semantic closeness of different samples based on cosine similarity, dot product, or a trainable kernel function. The normalized output layer performs scaling on the similarity vector to make the output semantic kernel representation stable and comparable. Based on the task requirements, the semantic kernel module and the joint encoder are jointly trained by cross-entropy, enabling the model to distinguish between cross-linguistic sample pairs that are semantically similar and semantically dissimilar. After joint training is completed, the joint encoder and semantic kernel module are fixed to build a cross-language analysis model for inference. Specifically, the trained parameters are solidified into an inference mode, the joint encoder is responsible for generating speech-text fusion representations, and the semantic kernel module performs semantic measurement, thus forming a complete cross-language analysis model. By feeding new speech-text input pairs into the analysis model, cross-linguistic representations in a unified semantic space are obtained.

[0037] S6 clusters cross-linguistic representations to form error categories, generates adversarial examples based on the error categories, forms a self-evolving data pool with the original samples, and constructs a semantic causal path graph; In this embodiment, cross-language representations are clustered to form error categories. Adversarial examples are generated based on these error categories and combined with the original samples to form a self-evolving data pool. A semantic causal path graph is then constructed, as detailed below: Obtain the original data samples corresponding to the target task, including text and speech, and perform word segmentation or encoding on the text, preprocessing and feature extraction on the speech to form the original samples; The original samples are input into the analysis model (including the joint encoder and the transferable semantic kernel) for processing, and cross-linguistic representations of each sample are obtained, forming a set of cross-linguistic semantic representations; The cross-linguistic representation set is analyzed using a density clustering algorithm. Samples with low semantic similarity are divided into different error categories, and the error category label of each sample is output to form error samples. Based on each error category, the original sample is perturbed by an adversarial sample generator, including text perturbation (synonym replacement, spelling transformation), speech perturbation (adding noise, fine-tuning acoustic features), and feature space perturbation, to generate corresponding adversarial samples; The original samples, adversarial samples, and error samples are combined to form a self-evolving data pool; Based on the self-evolving data pool and the internal reasoning path of the analysis model, nodes (sample features, error categories, perturbation features) and edges (causal relationships or dependencies) are extracted. A semantic causal path graph is formed through structured modeling. The internal reasoning path of the analysis model refers to the intermediate feature transmission and decision information generated by the analysis model in the process of reasoning on the input samples. Specifically, it includes: the encoder's hidden vector, joint representation vector, semantic kernel output vector, and the feature dependencies formed in the forward propagation of the model.

[0038] In this embodiment, a density clustering algorithm is used to analyze the cross-lingual representation set, classifying samples with low semantic similarity into different error categories, and outputting the error category label for each sample to form error samples, as detailed below: Collect the cross-linguistic representations of all original samples after processing by the analysis model, use them as input for cluster analysis, and output the Euclidean distance of each pair of vectors to form a semantic similarity matrix; By using the DBSCAN density clustering algorithm, core point radius and minimum neighbor number parameters are set to identify dense clusters and sparse samples. According to the density clustering rules, the vectors are divided into high-density clusters, and samples that do not meet the core point conditions are marked as noise. Noisy samples that are not assigned to any cluster or samples that are far from the cluster center are marked as samples of the wrong category. Generate an error category label (category number or error category identifier) ​​for each sample, and collect all samples labeled as error categories to form an error sample set.

[0039] In this embodiment, based on the self-evolving data pool and the internal reasoning path of the analysis model, nodes (sample features, error categories, perturbation features) and edges (causal relationships or dependency relationships) are extracted, and a semantic causal path graph is formed through structured modeling, as follows: First, the features of the original sample, the speech-text joint representation generated by the model, the error category labels obtained by clustering, and the perturbation features in the adversarial sample are respectively used as nodes, and their corresponding feature vectors are recorded. Subsequently, three types of traceable dependency edges were extracted from the model: one type comes from the cross-modal attention mechanism, which uses the higher-weighted connections in the attention matrix to represent the information transmission relationship between nodes; One type of gradient comes from the backpropagation stage, which analyzes the sensitivity of features to the output and uses the features with significant gradients as causal dependency edges. Another type comes from the model's forward propagation process, which forms the feature flow edge by recording the path of generating representations of sample features in each layer and passing them to the output; Finally, nodes and the three types of edges are added to the same directed graph, so that the graph can simultaneously reflect the causal links between feature input, semantic fusion, perturbation effects and error categories, thus obtaining a semantic causal path graph that can be used to explain the model decision-making process and error causes.

[0040] Example 2, Figure 2 The present invention provides a system for a data analysis method based on artificial intelligence, comprising a semantic segmentation module, an anchor module, a joint encoding module, a semantic graph module, an analysis model module, and a clustering output module, wherein the modules are interconnected. The semantic segmentation module is used to perform semantic segmentation on the input text, obtain multi-scale semantic units, and map the features of each level to a unified semantic space through a multi-scale pyramid to output semantic representations. The anchor module is used to generate semantic anchors through a dynamic time warping mechanism, align the input semantic representation, construct the domain knowledge structure, and convert the output domain knowledge tensor into a semantic prior tensor. The joint encoding module is used to input the semantic prior tensor into the prior injection channel of the text encoder, fuse the semantic prior tensor with the bias terms of the attention structure, and jointly encode the semantic representation to form the encoded representation; The semantic graph module is used to construct a semantic graph based on the encoded representation and the semantic representation, and input it into the graph neural network for node information transmission and aggregation, and output the graph augmented representation; The analysis model module is used to construct a joint speech-text representation based on the encoded representation and graph augmentation representation, and train a transferable semantic kernel to form an analysis model, outputting a cross-linguistic representation; The clustering output module is used to cluster cross-language representations to form error categories, generate adversarial examples based on the error categories, form a self-evolving data pool with the original samples, and construct a semantic causal path graph.

[0041] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0042] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0043] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0044] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0045] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0046] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data analysis method based on artificial intelligence, characterized in that, include: The input text is semantically segmented to obtain multi-scale semantic units, and the features at each level are mapped to a unified semantic space through a multi-scale pyramid to output a semantic representation. Semantic anchors are generated through a dynamic time warping mechanism to align the input semantic representation and construct a domain knowledge structure, converting the output domain knowledge tensor into a semantic prior tensor. The semantic prior tensor is input into the prior injection channel of the text encoder, and the semantic prior tensor is fused with the bias term of the attention structure. The semantic representation is then jointly encoded to form the encoded representation. A semantic graph is constructed based on the encoding and semantic representations, and then input into a graph neural network for node information transfer and aggregation, outputting an enhanced graph representation. Based on the encoding representation and graph augmentation representation, a joint speech-text representation is constructed, and a transferable semantic kernel is trained to form an analysis model, outputting a cross-linguistic representation; Cross-linguistic representations are clustered to form error categories. Adversarial examples are generated based on the error categories and combined with the original examples to form a self-evolving data pool. A semantic causal path graph is then constructed.

2. The data analysis method based on artificial intelligence according to claim 1, characterized in that, The input text is semantically segmented to obtain multi-scale semantic units, and the features at each level are mapped to a unified semantic space through a multi-scale pyramid to output a semantic representation, as follows: The input text is preprocessed by dividing it into sentence sequences based on syntactic rules and generating sentence boundary markers to obtain the initial text structure. Based on the initial text structure, character segmentation, word segmentation, phrase extraction and paragraph aggregation are performed sequentially to form a multi-scale semantic unit set; Multi-scale semantic features from the set of multi-scale semantic units are input into the multi-scale pyramid according to semantic granularity from low to high. Through cross-layer feature fusion and inter-layer mapping mechanism, intermediate features associated with the unified semantic space are generated at each layer. Based on the intermediate features generated by the multi-scale pyramid, the sequence length of mapping low-granularity layer features to high-granularity features is obtained by upsampling. For each scale, the upsampled features are projected onto a unified semantic dimension through a shared projection matrix, and the features from all scales are fused in hierarchical order to form a fused semantic vector. The fused semantic vector is normalized, and then the fused vector is reorganized according to the initial text structure order through the structure rearrangement module to obtain the semantic representation.

3. The data analysis method based on artificial intelligence according to claim 2, characterized in that, The process involves inputting multi-scale semantic features from the multi-scale semantic unit set into a multi-scale pyramid according to semantic granularity from low to high. Through cross-layer feature fusion and inter-layer mapping mechanisms, intermediate features associated with a unified semantic space are generated at each layer, as detailed below: Characters, words, phrases, paragraphs, and topic features from the multi-scale semantic unit set are input into the multi-scale pyramid in order of semantic granularity from low to high. The pyramid uses high-level features as queries and low-level features as keys and values ​​to perform cross-layer attention alignment, extracting fine-grained semantic information from the low layer. During the pyramid initialization phase, a shared projection matrix is ​​obtained through multi-scale contrastive learning pre-training. Aligned low-level features are mapped to a unified semantic dimension by sharing a projection matrix, and the low-level mapped features are fused with the current high-level features by a gated weighted fusion mechanism to form fused features. The Transformer within the fused feature input layer is used for semantic modeling, serving as the intermediate semantic state of that layer. This process is repeated layer by layer to generate intermediate features for each layer, and a complete intermediate semantic representation corresponding to a unified semantic space is formed at the top layer.

4. The data analysis method based on artificial intelligence according to claim 1, characterized in that, The process involves generating semantic anchors through a dynamic time warping mechanism, aligning the input semantic representation, constructing a domain knowledge structure, and converting the output domain knowledge tensor into a semantic prior tensor, as detailed below: The semantic representation is used as the input sequence, and dynamic time warping is performed in combination with the reference semantic sequence. The optimal alignment path between the sequences is calculated, the semantic position of each input semantic representation and the reference sequence is determined, and the key positions of the alignment path are marked as semantic anchors. Based on semantic anchors, the input sequence is aggregated through an attention mechanism to generate an aligned semantic representation; Based on the aligned semantic representation, extract entity, relation or concept hierarchical information to construct a domain knowledge structure graph; The constructed domain knowledge structure is represented as a set of nodes and a set of edges. The semantic feature vector of each node is extracted, a fixed-dimensional vector is generated by embedding method, and all node vectors are stacked to form a node feature matrix. Construct an adjacency matrix based on the edge set, and then perform symmetric normalization on the adjacency matrix; The node feature matrix is ​​combined with the normalized adjacency matrix to form a three-dimensional tensor, which is then input into the graph neural network to generate a domain knowledge tensor. A semantic prior tensor is generated by fusing the domain knowledge tensor with the aligned semantic representation through a linear mapping.

5. The data analysis method based on artificial intelligence according to claim 1, characterized in that, The semantic prior tensor is input into the prior injection channel of the text encoder, the semantic prior tensor is fused with the bias term of the attention structure, and the semantic representation is jointly encoded to form the encoded representation, as follows: The semantic prior tensor is input into the linear mapping layer, and the dimension of it is transformed by the trained shared projection matrix to obtain the dimension-aligned semantic prior features. The aligned semantic prior features are input into two different linear transformation layers for transformation to obtain the prior key vector and the prior value vector. The aligned semantic prior features are input into the bias generation network, and a bias matrix with the same size as the attention score matrix is ​​obtained through feedforward calculation. The input text is processed by an encoder to generate text key vectors and text value vectors; Concatenate the text key vector with the prior key vector along the sequence dimension to obtain the extended key vector set; The text value vector and the prior value vector are concatenated in the same way to form an extended value vector set; The query vector is sequentially multiplied by each key vector in the extended key vector set, and then scaled by the square root of the feature dimension to obtain the basic attention score matrix. The basic attention score matrix and the bias matrix are added element by element to obtain the bias-enhanced attention score matrix, i.e., the bias-enhanced attention matrix. The bias-enhanced attention matrix is ​​normalized using a softmax method to obtain the final attention weight matrix. Perform a weighted summation operation on the attention weight matrix and the extended value vector set to obtain the fused output; The feedforward network inside the fused input / output encoder is then processed through residual connections and layer normalization to finally generate a joint coded representation.

6. The data analysis method based on artificial intelligence according to claim 1, characterized in that, The process involves constructing a semantic graph based on the encoded and semantic representations, inputting it into a graph neural network for node information transfer and aggregation, and outputting an enhanced graph representation, as detailed below: Using the encoded representation as the basic semantic vector and the semantic representation as the supplementary semantic vector, nodes are constructed by corresponding the two one by one. The cosine similarity method is used to output the association strength between any two nodes. When the association strength exceeds a preset threshold, an edge is established between the nodes, and a relationship type label is attached to the edge to form a semantic graph structure. The content of nodes in the semantic graph is converted into node feature vectors, and the connection relationships between nodes are converted into adjacency structures. The node feature vectors and adjacency structure are input into the graph neural network. The features of adjacent nodes are weighted and combined through the neighborhood aggregation mechanism of the graph neural network to obtain the updated representation of each node. In graph neural networks, the node update representation of the previous layer is used as input, and neighborhood information aggregation is performed layer by layer to form multi-layer node representations; In the output layer of the graph neural network, the multi-layer node representations are collected and organized to become a high-level semantic feature representation of the whole graph, forming a graph-enhanced representation.

7. The data analysis method based on artificial intelligence according to claim 1, characterized in that, The node feature vectors and adjacency structure are input into the graph neural network. The features of adjacent nodes are weighted and combined through the neighborhood aggregation mechanism of the graph neural network to obtain the updated representation of each node, as follows: The feature vector corresponding to each node, along with the adjacency structure of the entire semantic graph, is input into the graph convolutional layer of the graph neural network. In the graph convolutional layer, the graph neural network determines the range of direct neighbors of each node based on the node connection information pre-encoded in the adjacency structure, and retrieves the feature vectors of all neighboring nodes connected to the node according to the adjacency relationship. Each node is assigned a set of independent weight parameters for its neighbor features, and the neighbor feature vectors are multiplied by their corresponding weights during computation. At the same time, a trainable bias is added so that the features of each neighbor node undergo a structured transformation before fusion. After the transformation is completed, the contribution ratio of neighboring nodes to the target node is determined based on the edge types attached to the semantic graph. Multiply the transformed features of each neighboring node by its corresponding contribution weight, and then sum the weighted results of all neighboring nodes to obtain the updated representation of each node.

8. The data analysis method based on artificial intelligence according to claim 7, characterized in that, The process involves constructing a joint speech-text representation based on the encoded representation and graph-enhanced representation, training a transferable semantic kernel to form an analysis model, and outputting a cross-linguistic representation, as detailed below: The speech signal is divided into several frames with a fixed frame length. Frequency domain features are extracted from each frame by short-time Fourier transform, and acoustic features are also extracted to form a speech feature sequence. The speech feature sequence is input into the speech encoder, and the high-level acoustic representation is extracted through the Transformer structure to obtain the speech coding vector; Based on the speech coding vector and coding representation, the minimum cost alignment path between two sequences is calculated using the dynamic time warping algorithm to form an alignment mapping table; The dimensions of the encoded representation and the graph augmentation representation are adjusted by an independent linear transformation layer, and then the adjusted vectors are concatenated to generate a text augmentation representation vector. Based on the alignment mapping table, the speech coding vector is matched with the corresponding text enhancement vector to generate a pairing vector sequence; The paired vector sequence is input into the joint encoder, and bidirectional interactive attention is calculated through a cross-modal attention mechanism to generate a multimodal interactive representation for each unit; The multimodal interaction representation is processed by a feedforward network, layer normalization, and residual connections to obtain a joint representation sequence; Using joint representation as input, a trainable semantic kernel network is constructed to output semantic kernel representations between different samples; According to the task requirements, joint training of the semantic kernel module and the joint encoder is performed using cross-entropy; After joint training is completed, the joint encoder and semantic kernel module are fixed, a cross-language analysis model is constructed and used for reasoning, thus forming the analysis model; The new speech-text input pairs are fed into the analysis model to obtain cross-linguistic representations.

9. The data analysis method based on artificial intelligence according to claim 1, characterized in that, The process involves clustering cross-linguistic representations to form error categories, generating adversarial examples based on these error categories, combining them with the original samples to form a self-evolving data pool, and constructing a semantic causal path graph, as detailed below: Obtain the original data samples corresponding to the target task and preprocess them to form the original samples; The original samples are input into the analysis model for processing, resulting in a cross-linguistic representation of each sample, forming a set of cross-linguistic semantic representations; The cross-linguistic representation set is analyzed using a density clustering algorithm. Samples with low semantic similarity are divided into different error categories, and the error category label of each sample is output to form error samples. Based on each error category, the original sample is perturbed by an adversarial sample generator to generate the corresponding adversarial sample; The original samples, adversarial samples, and error samples are combined to form a self-evolving data pool; Based on the self-evolving data pool and the internal reasoning path of the analysis model, nodes and edges are extracted, and a semantic causal path graph is formed through structured modeling.

10. A system using an artificial intelligence-based data analysis method as described in any one of claims 1-9, characterized in that, It includes a semantic segmentation module, an anchor point module, a joint encoding module, a semantic graph module, an analysis model module, and a clustering output module, with connections between the modules; The semantic segmentation module is used to perform semantic segmentation on the input text, obtain multi-scale semantic units, and map the features of each level to a unified semantic space through a multi-scale pyramid to output semantic representations. The anchor module is used to generate semantic anchors through a dynamic time warping mechanism, align the input semantic representation, construct the domain knowledge structure, and convert the output domain knowledge tensor into a semantic prior tensor. The joint encoding module is used to input the semantic prior tensor into the prior injection channel of the text encoder, fuse the semantic prior tensor with the bias terms of the attention structure, and jointly encode the semantic representation to form the encoded representation; The semantic graph module is used to construct a semantic graph based on the encoded representation and the semantic representation, and input it into the graph neural network for node information transmission and aggregation, and output the graph augmented representation; The analysis model module is used to construct a joint speech-text representation based on the encoded representation and graph augmentation representation, and train a transferable semantic kernel to form an analysis model, outputting a cross-linguistic representation; The clustering output module is used to cluster cross-language representations to form error categories, generate adversarial examples based on the error categories, form a self-evolving data pool with the original samples, and construct a semantic causal path graph.

Images