Bilingual local term extraction and alignment combined learning method based on mixed graph neural network

Abstract

The invention discloses a bilingual native term extraction and alignment joint learning method based on a mixed graph neural network. The method comprises the steps of obtaining a bilingual corpus data set; constructing a bilingual local term extraction and alignment joint learning model based on the mixed graph neural network, wherein the bilingual local term extraction and alignment joint learning model comprises an input layer, a semantic coding layer, a multi-channel graph generation layer, a multi-head graph attention network layer, a constraint decoding layer, an alignment joint layer and an output layer; dividing the bilingual corpus data set into a training set, a verification set and a test set according to a preset proportion; obtaining an optimal model through the training set and the verification set; and bilingual local terms of the test set are extracted and aligned through the optimal model. The problem that bilingual local term extraction and alignment efficiency is low and precision is insufficient due to the fact that an existing method lacks a dynamic optimization mechanism, semantic sparseness and ambiguity exist, semantic expression is limited by static expression, semantic hierarchy modeling is insufficient and feedback-driven continuous learning is lacked is solved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a bilingual native term extraction and alignment joint learning method based on hybrid graph neural networks. Background Technology

[0002] In the context of increasingly frequent global cultural exchanges, language, as the core carrier of cultural transmission, makes the accurate transmission and understanding of its native terminology ever more important. Terminology is a key linguistic unit for constructing professional knowledge systems; it is both the core carrier for expressing concepts in specific fields and a tool for conveying disciplinary ideas. Given the importance of terminology in multiple fields such as translation studies, information retrieval, and semantic modeling, terminology extraction is gradually becoming one of the core tasks of linguistic and interdisciplinary research in computational linguistics.

[0003] In the early stages of the research, a rule-based approach was used, involving manually annotating corpora and constructing structured databases. While this approach ensured high quality, it faced challenges in terms of efficiency and coherence when dealing with cross-linguistic, lengthy, and multi-domain topics.

[0004] Since the rise of deep learning around 2010, term extraction research has gradually transitioned from traditional rule-based and statistical methods to automated methods based on deep learning. Neural network models can automatically learn the complex relationships between words and semantics in large-scale corpora, thus achieving significant progress in term recognition and boundary determination. Compared to traditional methods that rely on handcrafted features, deep learning not only improves extraction accuracy but also enhances the model's generalization ability across different domains and corpus types. However, the following problems still exist: (1) Lack of dynamic optimization mechanism: When facing multi-domain or cross-linguistic corpora, the current model is difficult to adaptively adjust parameters according to changes in context.

[0005] (2) Semantic sparsity and ambiguity: Especially in professional domain corpora, terms are often semantically concentrated but have limited context, making it difficult for the model to accurately capture semantic boundaries.

[0006] (3) Static representation restricts semantic expression: Traditional deep models rely on fixed word vectors or embeddings, which makes it difficult to capture the hierarchical and contextual dependencies between terms.

[0007] (4) Insufficient semantic level modeling: Terms usually have a multi-level structure of lexical, phrase and conceptual layers, but most methods fail to fully model this level of information.

[0008] (5) Lack of feedback-driven continuous learning: Current models mostly remain in the offline training stage, making it difficult to continuously optimize the extraction results and adapt to emerging fields through feedback mechanisms. Summary of the Invention

[0009] This invention provides a bilingual native term extraction and alignment joint learning method based on hybrid graph neural networks to overcome the above-mentioned technical problems.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows: A bilingual native term extraction and alignment joint learning method based on hybrid graph neural networks, specifically including the following steps: S1: Obtain a corpus containing bilingual (Chinese and English) local terms, and use the BIO annotation method to perform lexical annotation on the bilingual (Chinese and English) local terms in the corpus to obtain corpus sample sequences and corresponding tag sequences; and the tag sequences include B-Terms for marking the start words of terms, I-Terms for marking the words inside the terms, and O-Terms for marking the non-term parts; And based on the label sequence, we define real B-Term label pairs for bilingual local term type alignment to obtain a bilingual corpus dataset; S2: Construct a bilingual native term extraction and alignment joint learning model based on hybrid graph neural networks, which includes an input layer, a semantic encoding layer, a multi-channel graph generation layer, a multi-head graph attention network layer, a constraint decoding layer, an alignment joint layer, and an output layer; The input layer is used to input the bilingual corpus dataset into the semantic coding layer; The semantic encoding layer is used to perform semantic encoding on the bilingual corpus dataset to obtain initial feature vectors for Chinese and English texts; and the initial feature vectors include Chinese feature vectors and English feature vectors; The graph construction layer is used to construct feature graphs in parallel by using the initial feature vectors as graph nodes. Furthermore, the feature graph includes a syntactic graph constructed based on dependency parsing; a semantic similarity graph constructed based on the cosine similarity between word feature vectors; a sequence connection graph constructed based on the physical positional relationship of words; and a word co-occurrence graph constructed based on the word co-occurrence relationship within a sliding window. The multi-head graph attention network layer is used to adaptively learn and fuse feature information from different graph structures output by the graph construction layer through a self-attention mechanism to obtain a fused feature vector of multi-dimensional structural information; and the fused feature vector includes Chinese fused vector and English fused vector. The constraint decoding layer is used to perform joint decoding of the fused feature vectors based on the Conditional Random Field (CRF) module to obtain the optimal label sequence; and the optimal label sequence includes the optimal Chinese label sequence and the optimal English label sequence. The alignment joint layer is used to obtain the B-Terms of the corresponding optimal Chinese tag sequence and the optimal English tag sequence, so as to obtain the first B-Term set of the corresponding Chinese term type and the second B-Term set of the corresponding English term type; at the same time, it obtains the tag type pairs of the same term type in the first B-Term set and the second B-Term set, thereby realizing bilingual local term extraction and alignment. S3: Divide the bilingual corpus dataset into training set, validation set, and test set according to a preset ratio; The bilingual native term extraction and alignment joint learning model based on hybrid graph neural networks is trained and validated using training and validation sets to obtain the optimal model. S4: The optimal model is used to extract and align bilingual local terms on the test set.

[0011] Furthermore, the method for obtaining the initial feature vectors of the Chinese and English texts in S2 specifically includes the following steps: S21: Input the corpus sample sequences from the bilingual corpus dataset into the preset... Model and location embedding module; S22: Through The model obtains the word embedding matrix of the corresponding corpus sample sequence. and ,in Indicates the length of the corpus sample sequence; Indicates the embedding dimension; The position embedding module obtains the position embedding matrix of the corresponding corpus sample sequence based on the sine position encoding function or the cosine position encoding function. ; S23: For the embedding matrix With position embedding matrix Perform element-wise addition to obtain the initial feature vectors of the corresponding Chinese and English texts.

[0012] Furthermore, the method for constructing the semantic similarity graph described in S2 is as follows: The initial feature vector is normalized to obtain a normalized vector, and the formula for obtaining the normalized vector is:

[0013] In the formula: Represents the words in the initial feature vector The original embedding vector; Indicates word elements Normalized vector; The adjacency matrix is ​​constructed by calculating the cosine similarity between tokens based on the normalized vectors, and the formula for constructing the adjacency matrix is ​​as follows:

[0014]

[0015] In the formula: express and Cosine similarity; Indicates word elements Normalized vector; Represents the adjacency matrix; 1 indicates the set similarity threshold; According to the adjacency matrix Get words Neighbor set The Mean Aggregation layer is used to aggregate neighbor features to obtain local information features, and the formula for obtaining the local information features is as follows:

[0016] In the formula: Indicates the first Layered word elements Feature representation; Indicates the first Layer weight matrix; Indicates the activation function; Indicates the first Layered word elements Feature representation; No. Layered word elements Feature representation; By pre-configuring the self-attention mechanism of the Transformer layer, global dependencies between lexical units are obtained based on local information features to optimize lexical units. The feature representation is used to construct a semantic similarity graph.

[0017] Furthermore, the method for constructing the syntactic graph described in S2 is as follows: The syntactic adjacency matrix is ​​obtained based on expert experience through dependency parsing, and the formula for constructing the syntactic adjacency matrix is ​​as follows:

[0018] In the formula: Represents the syntactic adjacency matrix; Based on the syntactic adjacency matrix Positional encoding is performed on the dependency relationships between lexical units to obtain the encoded positions. and call the function Dependency relationships are mapped to numerical identifiers; based on a pre-configured multi-head attention module, initial attention coefficients are obtained according to the numerical identifiers, and the expression for the initial attention coefficients is as follows:

[0019] In the formula: Indicates word elements with lexical elements The initial attention coefficient; Indicates word elements Relative to lexical elements Dependency weights; Represents the weight matrix; Indicates word elements The corresponding numerical identifier; Indicates word elements The corresponding numerical identifier; Indicates transpose; After normalizing the initial attention coefficients, the outputs of multiple attention heads in the multi-head attention module are concatenated to obtain the final feature representation. The formula for obtaining the final feature representation is as follows:

[0020]

[0021] In the formula: This represents the attention coefficient after normalization. Indicates word elements with lexical elements The initial attention coefficient; Indicates the first The attention coefficient of each attention head and ; Indicates the first The weight matrix of each attention head; Indicates the number of attention heads; Indicates word elements The corresponding final feature representation; Based on the final feature representation combined with the syntactic adjacency matrix Obtain the syntax graph.

[0022] Furthermore, the method for constructing the sequence connection graph described in S2 is as follows: Input the initial feature vector into the preset... The model extracts contextual features from the initial feature vector to obtain the contextual feature vector. and ,in Represents the initial eigenvector; Embedding functions by relative position Obtain the word units in the initial feature vector. with lexical elements Relative position embedding matrix and And embed the matrix according to the relative position. Obtain the location feature representation, and the formula for obtaining the location feature representation is:

[0023] In the formula: Indicates word elements Location feature representation; Through preset Modules are represented based on location features With context feature vector Obtain the sequence connectivity graph, and the formula for obtaining the sequence connectivity graph is as follows:

[0024] In the formula: Indicates word elements The final feature representation in the sequence diagram; Indicates word elements The corresponding context feature vector and .

[0025] Furthermore, the method for constructing the word co-occurrence graph described in S2 is as follows: Based on the sliding window technique, a co-occurrence matrix is ​​obtained according to the co-occurrence relationships of words within the sliding window, and the formula for obtaining the co-occurrence matrix is ​​as follows:

[0026] In the formula: Represents the co-occurrence matrix; The set of sentences represents the initial feature vector, and Represents a set of sentences Elements in; Indicates the width of the sliding window; Indicates the first The word at each position is ; Indicates the first The word at each position is ; The co-occurrence matrix is ​​weighted based on the TF-IDF algorithm to construct a co-occurrence adjacency matrix; And the formula for constructing the co-occurrence adjacency matrix is ​​as follows:

[0027]

[0028] In the formula: Represents the co-occurrence adjacency matrix; 2 indicates based on the weighted processing The value is set as a parameter threshold; For co-occurrence adjacency matrix After symmetric normalization, lexical units in the word co-occurrence graph are obtained through a pre-defined GCN network layer. The final feature representation According to the final feature representation Co-occurrence adjacency matrix Obtain a word co-occurrence graph; And the final feature representation The formula for obtaining it is

[0029]

[0030] In the formula: This represents the co-occurrence adjacency matrix after symmetric normalization. Degree matrix; Represents the identity matrix; Indicates the first The feature matrix of the layer; Indicates the first Layer weight matrix; Indicates the first The feature matrix of the layer.

[0031] Furthermore, the method for obtaining the optimal model in S3 is as follows: S31: Train the bilingual local term extraction and alignment joint learning model based on hybrid graph neural network using the training set to obtain the trained bilingual local term extraction and alignment joint learning model. S32: Based on the model loss function, the trained bilingual local terminology extraction and alignment joint learning model is validated using the validation set; That is, to determine whether the output of the bilingual local term extraction and alignment joint learning model after training has converged; If the output of the bilingual local terminology extraction and alignment joint learning model after training converges, then the bilingual local terminology extraction and alignment joint learning model after training is confirmed to be the optimal model. Otherwise, based on the backpropagation method, the weight parameters of the trained bilingual local terminology extraction and alignment joint learning model are adaptively adjusted, and step S31 is repeated until the weight parameters of the trained bilingual local terminology extraction and alignment joint learning model with converged output are confirmed to be the optimal weight parameters. The bilingual local terminology extraction and alignment joint learning model is then reconstructed to obtain the optimal model.

[0032] Furthermore, the loss function constructed in S32 for:

[0033]

[0034] In the formula: This represents the conditional probability, i.e., the log-likelihood of the CRF; Indicates the predicted label sequence and ; Indicates word elements Assign tags The launch fraction; Indicates from the label To tag The transfer fraction; This indicates that the input sequence is the sequence of language sample data in the bilingual corpus dataset; Represents the input sequence X A possible tag sequence; Represents the input sequence The set of all possible label sequences; Represents the input sequence X Corresponding label sequence The score.

[0035] Beneficial Effects: This invention provides a joint learning method for bilingual native term extraction and alignment based on hybrid graph neural networks. It uses the BIO annotation method to perform lexical annotation on Chinese and English bilingual native terms in a corpus to obtain corpus sample sequences and corresponding label sequences. Then, it constructs a joint learning model for bilingual native term extraction and alignment based on hybrid graph neural networks. A semantic encoding layer performs semantic encoding on the bilingual corpus dataset to obtain initial feature vectors for the Chinese and English texts. Next, a multi-channel graph neural network framework is designed to obtain four types of graph structures: syntactic graph, semantic similarity graph, sequence connection graph, and word co-occurrence graph. A multi-head graph attention network layer aggregates multi-perspective information, and based on a conditional random field (CRF) module, the fused feature vectors are jointly decoded to obtain the optimal label sequence, optimizing sequence annotation consistency and effectively alleviating the resource scarcity problem in linguistics, improving the accuracy and robustness of bilingual term processing. Finally, an alignment joint layer is used to achieve the alignment of bilingual native terms, improving cross-language alignment capabilities. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of the bilingual native term extraction and alignment joint learning method based on hybrid graph neural networks of the present invention; Figure 2 This is a schematic diagram of the structure of the bilingual native term extraction and alignment joint learning model based on hybrid graph neural networks constructed in this invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0039] This embodiment provides a bilingual native terminology extraction and alignment joint learning method based on hybrid graph neural networks, such as... Figure 1 As shown, the specific steps include: S1: Based on known linguistic knowledge, a pre-defined system of native Chinese linguistic terminology categories is established, and a corpus containing native Chinese and English bilingual terms is constructed. Specifically, this embodiment systematically collects and organizes relevant academic works and literature based on knowledge of native Chinese linguistic terminology, and classifies them according to nine pre-defined academic disciplines, including phonetics, philology, syntax, semantics, pragmatics, morphology, lexicology, sociolinguistics, and historical linguistics. Sentence-level alignment is performed on the classified Chinese-English bilingual documents to generate a core dataset composed of parallel sentence pairs. Based on the BIO annotation method, lexical annotation is performed on the native Chinese and English bilingual terms in the corpus, i.e., the core dataset, to obtain corpus sample sequences and corresponding label sequences. The label sequences include B-Terms for marking the start words of terms, I-Terms for marking words within terms, and O-Terms for marking non-term parts. And based on the label sequence, we define real B-Term label pairs for bilingual local term type alignment to obtain a bilingual corpus dataset; S2: Construct a joint learning model for bilingual native terminology extraction and alignment based on hybrid graph neural networks, such as... Figure 2 As shown, it includes an input layer, a semantic encoding layer, a multi-channel graph generation layer, a multi-head graph attention network layer, a constraint decoding layer, an alignment joint layer, and an output layer; The input layer is used to input the bilingual corpus dataset into the semantic coding layer; The semantic encoding layer is used to perform semantic encoding on the bilingual corpus dataset to obtain initial feature vectors for Chinese and English texts; and the initial feature vectors include Chinese feature vectors and English feature vectors; Specifically, in this embodiment, to enable the model to better capture the positional information of words in a sentence, a positional embedding mechanism is introduced for words of length [missing information]. The input sequence is used to embed the position embedding matrix and the input word embedding matrix. Add them together to obtain the initial feature vectors of the corresponding Chinese and English texts. Specifically, it includes: S21: Input the corpus sample sequences from the bilingual corpus dataset into the preset... Model and location embedding module; S22: Through The model obtains the word embedding matrix of the corresponding corpus sample sequence. and ,in Indicates the length of the corpus sample sequence; Indicates the embedding dimension; The position embedding module obtains the position embedding matrix of the corresponding corpus sample sequence based on the sine position encoding function or the cosine position encoding function. ; And based on the sinusoidal position coding function or cosine position encoding function The expression is

[0040] In the formula: Indicates the position index; S23: For the embedding matrix With position embedding matrix Perform element-wise addition to obtain the initial feature vectors for the corresponding Chinese and English texts. This embodiment uses the BIO strategy for annotation, and then utilizes... The model performs deep semantic encoding on the corpus and outputs an initial feature vector. In this embodiment, position embedding is introduced through formulas (1) and (2) to enhance the model's ability to capture word position information.

[0041] The graph construction layer is used to construct feature graphs in parallel using initial feature vectors as graph nodes. The feature graphs include a syntactic graph constructed based on dependency parsing; a semantic similarity graph constructed based on cosine similarity between word feature vectors; a sequence connection graph constructed based on word physical position relationships; and a word co-occurrence graph constructed based on word co-occurrence relationships within a sliding window. Specifically, this embodiment constructs a joint learning model based on a four-way graph neural network through a graph construction layer, deeply fusing multi-dimensional contextual features, including: The method for constructing the semantic similarity graph in this embodiment is as follows: The initial feature vector is normalized to obtain a normalized vector, and the formula for obtaining the normalized vector is:

[0042] In the formula: Represents the words in the initial feature vector The original embedding vector; Indicates word elements Normalized vector; Based on the normalized vectors, cosine similarity between terms is calculated to construct an adjacency matrix to determine the existence of edges in the semantic similarity graph. The formula for constructing the adjacency matrix is ​​as follows:

[0043]

[0044] In the formula: express and Cosine similarity; Indicates word elements Normalized vector; Represents the adjacency matrix; 1 indicates the set similarity threshold; According to the adjacency matrix Get words Neighbor set In the feature update process of the semantic graph, a Mean Aggregation layer is used to aggregate neighbor features to obtain local information features, and the formula for obtaining the local information features is as follows:

[0045] In the formula: Indicates the first Layered word elements Feature representation; Indicates the first Layer weight matrix; Indicates the activation function; Indicates the first Layered word elements Feature representation; No. Layered word elements Feature representation; By pre-configuring the self-attention mechanism of the Transformer layer, global dependencies between lexical units are obtained based on local information features to optimize lexical units. The feature representation is used to construct a semantic similarity graph; the expression for the self-attention mechanism is as follows:

[0046] In the formula: , , These represent the query matrix, key matrix, and value matrix, respectively. Indicates the dimension of the query matrix.

[0047] The method for constructing the syntactic graph in this embodiment is as follows: The syntactic adjacency matrix is ​​obtained based on expert experience through dependency parsing, and the formula for constructing the syntactic adjacency matrix is ​​as follows:

[0048] In the formula: Represents the syntactic adjacency matrix; Based on the syntactic adjacency matrix Positional encoding is performed on the dependency relationships between lexical units to obtain the encoded positions. and call the function Dependency relationships are mapped to numerical identifiers; based on a pre-configured multi-head attention module, initial attention coefficients are obtained according to the numerical identifiers, and the expression for the initial attention coefficients is as follows:

[0049] In the formula: Indicates word elements with lexical elements The initial attention coefficient; Indicates word elements Relative to lexical elements Dependency weights; Represents the weight matrix; Indicates word elements The corresponding numerical identifier; Indicates word elements The corresponding numerical identifier; Indicates transpose; in this embodiment, the numerical identifier is used to calculate the unnormalized attention coefficient. The type of dependency relationship of the edge is encoded, which affects the calculation of the attention score; After normalizing the initial attention coefficients using the soft max function, the outputs of multiple attention heads in the multi-head attention module are concatenated to obtain the final feature representation. The formula for obtaining the final feature representation is as follows:

[0050] The features of a lexical are represented as In this embodiment, multi-head attention is used to enhance the model's learning ability, by concatenating the outputs of multiple attention heads:

[0051] In the formula: This represents the attention coefficient after normalization. Indicates word elements with lexical elements The initial attention coefficient; Indicates the first The attention coefficient of each attention head and ; Indicates the first The weight matrix of each attention head; Indicates the number of attention heads; Indicates word elements The corresponding final feature representation; Based on the final feature representation combined with the syntactic adjacency matrix Obtain the syntax graph.

[0052] The method for constructing the sequence connection graph in this embodiment is as follows: Input the initial feature vector into the preset... The model extracts contextual features from the initial feature vector to obtain the contextual feature vector. and ,in Represents the initial feature vector; embedded through a relative position function Obtain the word units in the initial feature vector. with lexical elements Relative position embedding matrix and And embed the matrix according to the relative position. Obtain the location feature representation, and the formula for obtaining the location feature representation is:

[0053] In the formula: Indicates word elements Location feature representation; Through preset Modules are represented based on location features With context feature vector Obtain the sequence connectivity graph, and the formula for obtaining the sequence connectivity graph is as follows:

[0054] In the formula: Indicates word elements The final feature representation in the sequence diagram; Indicates word elements The corresponding context feature vector and In this embodiment, the sequence connection graph is combined with... Using relative position embeddings to capture sequence dependencies of lexical terms.

[0055] The method for constructing the word co-occurrence graph in this embodiment is as follows: Based on the sliding window technique, a co-occurrence matrix is ​​obtained according to the co-occurrence relationships of words within the sliding window, and the formula for obtaining the co-occurrence matrix is ​​as follows:

[0056] In the formula: Represents the co-occurrence matrix; The set of sentences represents the initial feature vector, and Represents a set of sentences Elements in; Indicates the width of the sliding window; Indicates the first The word at each position is ; Indicates the first The word at each position is ; The co-occurrence matrix is ​​weighted based on the TF-IDF algorithm to construct a co-occurrence adjacency matrix; And the formula for constructing the co-occurrence adjacency matrix is ​​as follows:

[0057]

[0058] In the formula: Represents the co-occurrence adjacency matrix; 2 indicates based on the weighted processing The value is set as a parameter threshold; For co-occurrence adjacency matrix After symmetric normalization, lexical units in the word co-occurrence graph are obtained through a pre-defined GCN network layer. The final feature representation According to the final feature representation Co-occurrence adjacency matrix Obtain a word co-occurrence graph; And the final feature representation The formula for obtaining it is

[0059]

[0060] In the formula: This represents the co-occurrence adjacency matrix after symmetric normalization. Degree matrix; Represents the identity matrix; Indicates the first The feature matrix of the layer; Indicates the first Layer weight matrix; Indicates the first The feature matrix of the layer; In this embodiment, four heterogeneous adjacency matrices are constructed in parallel based on the initial feature vector, each representing one of four text structure information types. These four adjacency matrices include: a syntactic graph adjacency matrix constructed based on dependency parsing results; a semantic graph adjacency matrix constructed based on the cosine similarity between lexical feature vectors; a sequence graph adjacency matrix constructed based on the physical positional relationships of lexical units; and a co-occurrence graph adjacency matrix constructed based on the co-occurrence relationships of lexical units within a sliding window. This embodiment receives the initial feature vector and performs joint learning, which includes a heterogeneous graph construction unit and a graph attention fusion unit. Among them, the heterogeneous graph construction unit is used to construct the syntactic graph, semantic similarity graph, sequence connection graph and word co-occurrence graph in parallel. During the construction process, the semantic similarity graph is constructed and the features are updated according to formulas (3)-(7), the syntactic graph is constructed and the features are updated according to formulas (8)-(11), the sequence connection graph is constructed and the features are updated according to formulas (12)-(16), and the word co-occurrence graph is constructed and the features are updated according to formulas (17)-(21). The graph attention fusion unit is used to adaptively learn and fuse the structural information from the four types of graphs through the self-attention mechanism, and finally output an enhanced feature vector with highly condensed information.

[0061] The multi-head graph attention network layer (GAT) is used to adaptively learn and fuse feature information from different graph structures output by the graph construction layer through a self-attention mechanism to obtain a fused feature vector of multi-dimensional structural information; and the fused feature vector includes Chinese fused vector and English fused vector; in this embodiment, GAT adaptively learns the importance weights of different neighbor nodes, aggregates weighted information from four types of graphs, and calculates an enhanced feature vector that fuses multi-dimensional structural information for each word.

[0062] The constraint decoding layer is used to perform joint decoding of the fused feature vectors based on the Conditional Random Field (CRF) module to obtain the optimal label sequence; and the optimal label sequence includes the optimal Chinese label sequence and the optimal English label sequence. Specifically, the high-dimensional feature representations, namely the Chinese fusion vector and the English fusion vector, are input into a Conditional Random Field (CRF) layer for the final joint decoding of the label sequence. The CRF layer learns the transition patterns and dependencies between labels to globally optimize the entire label sequence, ensuring the structural legality of the extracted terms. In the CRF layer, the score of the label sequence is first calculated:

[0063] In the formula: Represents the input sequence; Indicates the predicted label sequence; Indicates position Assign tags The launch fraction; Indicates from the label arrive The transition score is calculated, and then the conditional probability is used as the log-likelihood of the CRF. This conditional probability is negativeized during training and used as the loss function to update the model parameters.

[0064] In the formula: Represents the input sequence The set of all possible label sequences; Represents the input sequence X A possible tag sequence; Represents the input sequence X Corresponding label sequence The score; During the inference phase, the optimal label sequence is found using the Viterbi algorithm:

[0065] In this embodiment, the constraint decoding layer is used to perform structural optimization and decoding on the fused features. It includes a conditional random field (CRF) unit, which learns the transition probabilities between labels, performs global optimization on the predicted label sequence, corrects illogical label combinations, and completes the accurate identification of terms. During the optimization process, the label sequence score and conditional probability are calculated according to formulas (22)-(24), and the optimal label sequence is found.

[0066] The alignment joint layer is used to obtain the B-Terms of the corresponding optimal Chinese tag sequence and the optimal English tag sequence, so as to obtain the first B-Term set of the corresponding Chinese term type and the second B-Term set of the corresponding English term type; at the same time, it obtains the tag type pairs of the same term type in the first B-Term set and the second B-Term set, thereby realizing bilingual local term extraction and alignment. This embodiment achieves accurate recognition of linguistic terms in Chinese and English texts by constraining the decoding results of the decoding layer. Simultaneously, a pre-built cross-language alignment module is used to automatically align Chinese and English terms. Specifically, after CRF decoding, the B-label sets of the predicted sequences in Chinese and English are directly compared to check if the same term type is identified on both sides, thus determining term alignment. In the term recognition performance evaluation, precision, recall, and F1 score are used as evaluation metrics, expressed as:

[0067]

[0068]

[0069] In the formula: This indicates the number of true positives. Indicates the number of false positives. This indicates the number of false negatives.

[0070] In the terminology alignment performance evaluation, label type alignment accuracy is introduced ( ), tag type alignment accuracy ( ), tag type alignment recall ( ) and label type alignment F1 value ( The calculation formula is as follows:

[0071]

[0072]

[0073]

[0074] In the formula: This represents the set of sentence pairs with truly aligned label types, i.e., the set of label type pairs where the first B-Term set and the second B-Term set have the same term type. Indicates sentence pair The set of predicted aligned label types, Indicates sentence pair The set of true alignment label types, The indicator function is represented. In this embodiment, the cross-language alignment module is used to match the term types (i.e., B-label sets) identified in the optimal Chinese label sequence with the term types (i.e., B-label sets) identified in the optimal English label sequence to obtain label pairs. If they are consistent with the "terms that appear in both Chinese and English" in the manual pre-annotation, the label pair is confirmed to be successfully aligned. Finally, the alignment accuracy is obtained by dividing the successful sentence pair (i.e., label pair) by the total number of sentence pairs that need to be aligned. When evaluating the alignment effect, the relevant evaluation indicators are calculated using formulas (28)-(31). This embodiment also includes a knowledge base generation unit, which is used to structure and store the identified and aligned bilingual term pairs to form a knowledge base. S3: Divide the bilingual corpus dataset into training set, validation set, and test set according to a preset ratio; The bilingual native term extraction and alignment joint learning model based on hybrid graph neural networks is trained and validated using training and validation sets to obtain the optimal model. The method for obtaining the optimal model in this embodiment is as follows: S31: Train the bilingual local term extraction and alignment joint learning model based on hybrid graph neural network using the training set to obtain the trained bilingual local term extraction and alignment joint learning model. S32: Based on the model loss function, the trained bilingual local terminology extraction and alignment joint learning model is validated using the validation set; The loss function The expression is:

[0075]

[0076] In the formula: This represents the conditional probability, i.e., the log-likelihood of the CRF; Indicates the predicted label sequence and ; Indicates word elements Assign tags The launch fraction; Indicates from the label To tag The transfer fraction; This indicates that the input sequence is the sequence of language sample data in the bilingual corpus dataset; Represents the input sequence X A possible tag sequence; Represents the input sequence The set of all possible label sequences; Represents the input sequence X Corresponding label sequence The score; That is, to determine whether the output of the bilingual local term extraction and alignment joint learning model after training has converged; If the output of the bilingual local terminology extraction and alignment joint learning model after training converges, then the bilingual local terminology extraction and alignment joint learning model after training is confirmed to be the optimal model. Otherwise, based on the backpropagation method, the weight parameters of the trained bilingual local terminology extraction and alignment joint learning model are adaptively adjusted, and step S31 is repeated until the weight parameters of the trained bilingual local terminology extraction and alignment joint learning model with converged output are confirmed to be the optimal weight parameters, and the bilingual local terminology extraction and alignment joint learning model is reconstructed to obtain the optimal model. S4: The optimal model is used to extract and align bilingual local terms on the test set.

[0077] The method described in this embodiment preprocesses and semantically encodes a parallel corpus of Chinese and English linguistic terminology: a parallel corpus of Chinese and English linguistic terminology is constructed based on linguistic knowledge, and the corpus is annotated using a BIO annotation strategy; subsequently, the annotated corpus is divided into a training set, a validation set, and a test set; finally, a large-scale multilingual pre-trained model is introduced. Deep semantic encoding is performed on the text to efficiently extract initial feature vectors from both Chinese and English texts. A four-way graph structure is constructed through a graph construction layer to deeply fuse multi-dimensional contextual features: the initial feature vectors are used as graph nodes to construct syntactic graphs, semantic similarity graphs, sequence connectivity graphs, and word co-occurrence graphs in parallel; these four heterogeneous graph information are fed into a graph attention network (GAT), which adaptively learns and fuses effective information from different graph structures through a self-attention mechanism, thereby generating a more information-rich and discriminative unified feature representation. Sequence decoding and constraint optimization are performed on the fused features: the high-dimensional feature representation generated by joint learning, i.e., the unified feature representation, is input into a conditional random field (CRF) layer for final joint decoding of the label sequence; the CRF layer performs global optimization of the entire label sequence by learning the transition rules and dependencies between labels to ensure the structural legality of the extracted terms. Terminology recognition and alignment application: Using the results of joint decoding of tag sequences, accurate recognition of linguistic terms in Chinese and English texts is achieved; at the same time, using a pre-built cross-language alignment module, the Chinese and English terms are automatically aligned based on the recognized optimal tag sequences, and the results are used to build a bilingual terminology knowledge base.

[0078] The beneficial effects of the method described in this embodiment are as follows: First, a bilingual corpus covering nine subfields of linguistics (containing 22,828 samples and 136,872 terms) is constructed, and 19 types of term labels are annotated using the BIO annotation scheme; then, a multi-channel graph neural network framework is designed, integrating four types of graph structures: sequence graph, semantic graph, syntactic graph, and co-occurrence graph, through an encoder layer ( Semantic embedding + positional embedding, positional embedding is calculated according to formula (1)-(2)) to capture text features, graph fusion layer (each graph feature is updated according to the corresponding formula (3)-(21)) aggregates multi-perspective information, CRF layer (according to formula (22)-(24)) optimizes sequence label consistency; finally, in the Chinese-English term extraction task, the F1 values ​​of 92.54% and 83.11% respectively are achieved (F1 value is calculated according to formula (27)), and the bilingual label type alignment F1 value reaches 98.39% (alignment F1 value is calculated according to formula (31)), which effectively alleviates the problem of resource scarcity in the field of linguistics and improves the accuracy and robustness of bilingual term processing.

[0079] In this embodiment, multi-channel graph structure collaborative modeling is implemented: four types of heterogeneous graphs are innovatively constructed to capture language features of different dimensions. The semantic graph (Graph SAGE) models deep semantic associations based on cosine similarity (calculated according to formula (4)). The syntactic graph (GAT) captures grammatical structure through dependency analysis (adjacency matrix construction according to formula (8)). The sequence graph ( +Relative position embedding, the relative position embedding is calculated according to formulas (14)-(15) to encode local order dependencies, the co-occurrence graph (GCN) uses a sliding window (the co-occurrence matrix is ​​calculated according to formula (17)) to statistically analyze word co-occurrence patterns and achieve multi-view feature complementarity. Cross-language semantic alignment enhancement: adopts The pre-trained model generates multilingual context embedding vectors, and combines bilingual dependency analysis (Chinese zh_core_web_sm, English en_core_web_sm) to construct a language-independent feature representation space, improving the accuracy of semantic mapping of Chinese and English terms, especially suitable for low-resource or noisy scenarios. Structured annotation and graph fusion mechanism: Based on 9 linguistic classification systems, 19 types of BIO labels are designed to clarify term boundaries and types; the graph fusion layer uses attention weight allocation (such as the attention coefficient calculation in GAT based on formula (9)) and feature aggregation to enable dynamic interaction and reinforcement of four types of graph structures, solving the problem of insufficient capture of complex terms by a single graph structure, and at the same time introducing TF-IDF algorithm weighting (based on formula (18)) to optimize the edge weight of co-occurrence graphs. Global sequence optimization of CRF layer: A CRF layer is superimposed on the graph fusion features to model the transition probability between labels (the transition score is calculated based on formula (30)), and the optimal label sequence is decoded by the Viterbi algorithm (based on formula (24)) to balance the local emission score and the global label consistency, thereby improving the recognition accuracy of long terms and nested terms.

[0080] Comparison with the beneficial effects of existing technologies 1. Superior term extraction performance: Compared to traditional sequence models (such as LSTM, ... The multi-channel graph structure significantly improves feature representation capabilities, with the Chinese F1 score improving by 27.9 percentage points compared to LSTM (64.64%), and the English F1 score improving by [missing information]. (48.86%) improved by 34.25 percentage points; compared with pre-trained models (such as BERT, RoBERTa), the Chinese F1 value still leads by 2.73-0.94 percentage points, proving the complementary advantages of multi-image fusion (F1 value is calculated according to formula (25)).

[0081] 2. Stronger cross-language alignment capability: The model achieves 98.39% by introducing the label type alignment index (Align F1, calculated according to formula (29)), which is 5.26 percentage points higher than RNN (93.13%) and 0.17 percentage points higher than RoBERTa (98.22%). It can effectively handle the structural differences between Chinese and English terms (such as Chinese compound words vs. English descriptive phrases) and reduce cross-language semantic misalignment.

[0082] 3. Higher domain adaptability and robustness: To address the scarcity of resources in the field of linguistics, a large-scale bilingual corpus (covering 9 sub-domains such as phonology and syntax) was built. The model still maintains stable performance in class-imbalanced data (such as 5,852 morphological samples and 253 philological samples). Ablation experiments show that removing any graph structure leads to a decrease in performance. In particular, the absence of semantic graphs causes the English F1 score to drop sharply by 15.41 percentage points, verifying the robustness of multi-component collaboration (the F1 score in the ablation experiment is calculated according to formula (27)).

[0083] 4. Balance between efficiency and interpretability: The Mean Aggregator (feature aggregation based on formula (6)) is used to aggregate semantic graph neighbor features, and GAT multi-head attention (multi-head attention output splicing based on formula (11)) is used to optimize the syntactic graph, taking into account both computational efficiency and model expressive ability; the four types of graph structures correspond to interpretable dimensions such as semantics and syntax, which makes it easier to trace the basis of term recognition (such as checking the frequency of term collocation through co-occurrence graph, and analyzing grammatical dependency relationship through syntactic graph), reducing the debugging difficulty of black box model.

[0084] Specific embodiments of the method described in this example: S100: Corpus Construction and Encoding S1001: Terminology Classification System Design: Systematically review Chinese linguistics textbooks, monographs, and journal articles, establish 9 core areas (phonology, philology, etc.), clarify the representative terms of each area (e.g., phonology includes "initial consonant" and "final consonant," syntax includes "inversion" and "clause"), and form a hierarchical classification framework. Figure 1 ).

[0085] S1002: Corpus Collection and Preprocessing: Collected parallel Chinese and English linguistic texts, covering 22,828 samples in 9 subdomains. Cleaned special symbols, unified capitalization, used jieba for Chinese word segmentation and spaCy for English word segmentation; adopted the BIO annotation scheme, based on 9 domains, expanded to 19 types of tags (such as B-Pho, I-Pho corresponding to the beginning and inside of phonological terms).

[0086] S1003: Corpus Statistics and Classification: Count the number of terms in various categories (64,361 terms in phonology, 1,372 terms in historical linguistics, etc.), and divide them into training set (18,262 terms), validation set (2,283 terms), and test set (2,283 terms) in a ratio of 8:1:1 to ensure that the samples of the same subdomain are evenly distributed in each set.

[0087] S1004: Semantic and positional encoding: Input the labeled corpus The model obtains a 768-dimensional semantic embedding and generates a positional embedding using formulas (1)-(2). The two are then added together to obtain the initial feature vector \(X\). S200: Construction of Multi-channel Graph Neural Networks S2001: Encoder layer feature generation: Semantic embedding: employing The model takes Chinese-English sentence pairs as input (in Chinese). <sep>Separate), output 768-dimensional context embedding and fine-tune the classification layer with fixed pre-trained parameters, dropout rate set to 0.1; position embedding: generate sinusoidal position vector (768 dimensions) according to formula (2), and add it to the semantic embedding according to formula (1) ( Ensure the model captures word order information and adapts to sequences with a maximum length of 256 words; S2002: Construction and Training of Four Types of Graph Structures: Semantic similarity graph (Graph SAGE): Word embeddings are normalized according to formula (3), and cosine similarity is calculated according to formula (4). Formula (5) sets a threshold Construct an adjacency matrix; use a Mean Aggregator (Formula (6)) to aggregate neighbor features, stack a Transformer layer (Formula (7)) to optimize attention weights, and iterate 3 times to update node representations to obtain a semantic similarity graph; Syntactic Graph (GAT): Use spaCy to obtain Chinese-English dependency relations, and use formula (8) to construct the adjacency matrix (if there is a dependency relation) The dependency types (such as subject-verb and verb-object) are encoded as numerical values; the GAT layer calculates the attention coefficient according to formula (9). Formula (10) normalizes and updates node features synchronously, and 8-head attention is used to concatenate (Formula (11)) to obtain the syntactic graph; Sequence connection graph ( +Relative position): Formula (14) will input the encoder output. Obtain a 256-dimensional context representation. Formula (14) generates relative position embeddings. (Maximum relative distance 8), the average position representation is obtained by formula (15). Formula (16) fusion and Then dropout is used to output sequence structure features. To obtain a sequence connectivity graph; Word Co-occurrence Graph (GCN): Formula (17) counts the number of word co-occurrences using a window size of 5, and Formula (18) uses TF-IDF weighted optimization of the co-occurrence matrix. Formula (19) sets a threshold Constructing an adjacency matrix Formula (20) Symmetrical Normalization (Add self-loop), Formula (21) aggregates neighbor features through a 2-layer GCN and uses ReLU as the activation function to obtain the word co-occurrence graph; S2003: Graph Fusion and CRF Optimization Feature Aggregation: Concatenate the 256-dimensional features output by the four types of graph structures, compress them to 256 dimensions through a fully connected layer, introduce attention weights (learn the contribution ratio of each type of graph), and generate a fused feature vector.

[0088] CRF Training: Calculate the label sequence score using formula (22) (including emission scores and transition scores , calculate the conditional probability using formula (23), and maximize the log-likelihood of the correct label sequence during training; during inference, decode the optimal label sequence using the Viterbi algorithm with formula (24).

[0089] S300: Model Training and Evaluation S3001: Training Parameter Settings: Use the Adam optimizer (learning rate 2e-05, weight decay 1e-4), batch size 64, gradient accumulation step 2, train for 50 epochs, and early stop if the validation set performance does not improve for 5 consecutive epochs; set the random seed to 42 to ensure the reproducibility of the experiment (hardware configuration: Windows 10, RTX 4090, CUDA 11.1).

[0090] S3002: Performance Evaluation: Monolingual NER Evaluation: Calculate Precision, Recall, and F1 using formulas (23)-(25). For the Chinese test set, it reaches , , , and for English, it reaches , , . Cross-lingual Alignment Evaluation: Calculate Align Acc (92.24%) and Align F1 (98.39%) using formulas (26)-(29), compare with the baseline model (e.g., RoBERTa Align F1 = 98.22%), and verify the cross-lingual semantic mapping ability.

[0091] S3.3: Ablation Experiment and Error Analysis: Removing any graph structure results in a 7.38 percentage point decrease in Chinese F1 and a 15.41 percentage point decrease in English F1 due to the lack of semantic graphs, which is the most influential component; the errors mainly stem from the structural differences between Chinese and English (e.g., Chinese "etymology" vs English "the origins of words"), and there are some boundary recognition omission problems for some terms (in the ablation experiment, the F1 value is calculated according to formula (27), and the alignment-related metrics are calculated according to formulas (28)-(31)).

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.< / sep>

Claims

1. A bilingual native terminology extraction and alignment joint learning method based on hybrid graph neural networks, characterized in that, The specific steps include: S1: Obtain a corpus containing bilingual (Chinese and English) local terms, and use the BIO annotation method to perform lexical annotation on the bilingual (Chinese and English) local terms in the corpus to obtain corpus sample sequences and corresponding tag sequences; and the tag sequences include B-Terms for marking the start words of terms, I-Terms for marking the words inside the terms, and O-Terms for marking the non-term parts; And based on the label sequence, we define real B-Term label pairs for bilingual local term type alignment to obtain a bilingual corpus dataset; S2: Construct a bilingual native term extraction and alignment joint learning model based on hybrid graph neural networks, which includes an input layer, a semantic encoding layer, a multi-channel graph generation layer, a multi-head graph attention network layer, a constraint decoding layer, an alignment joint layer, and an output layer; The input layer is used to input the bilingual corpus dataset into the semantic coding layer; The semantic encoding layer is used to perform semantic encoding on the bilingual corpus dataset to obtain initial feature vectors for Chinese and English texts; and the initial feature vectors include Chinese feature vectors and English feature vectors; The graph construction layer is used to construct feature graphs in parallel by using the initial feature vectors as graph nodes. Furthermore, the feature graph includes a syntactic graph constructed based on dependency parsing; a semantic similarity graph constructed based on the cosine similarity between word feature vectors; a sequence connection graph constructed based on the physical positional relationship of words; and a word co-occurrence graph constructed based on the word co-occurrence relationship within a sliding window. The multi-head graph attention network layer is used to adaptively learn and fuse feature information from different graph structures output by the graph construction layer through a self-attention mechanism to obtain a fused feature vector of multi-dimensional structural information; and the fused feature vector includes Chinese fused vector and English fused vector. The constraint decoding layer is used to perform joint decoding of the fused feature vectors based on the Conditional Random Field (CRF) module to obtain the optimal label sequence; and the optimal label sequence includes the optimal Chinese label sequence and the optimal English label sequence. The alignment joint layer is used to obtain the B-Terms of the corresponding optimal Chinese tag sequence and the optimal English tag sequence, so as to obtain the first B-Term set of the corresponding Chinese term type and the second B-Term set of the corresponding English term type; at the same time, it obtains the tag type pairs of the same term type in the first B-Term set and the second B-Term set, thereby realizing bilingual local term extraction and alignment. S3: Divide the bilingual corpus dataset into training set, validation set, and test set according to a preset ratio; The bilingual native term extraction and alignment joint learning model based on hybrid graph neural networks is trained and validated using training and validation sets to obtain the optimal model. S4: The optimal model is used to extract and align bilingual local terms on the test set.

2. The bilingual native terminology extraction and alignment joint learning method based on hybrid graph neural networks according to claim 1, characterized in that, The method for obtaining the initial feature vectors of Chinese and English texts in S2 includes the following steps: S21: Input the corpus sample sequences from the bilingual corpus dataset into the preset... Model and location embedding module; S22: Through The model obtains the word embedding matrix of the corresponding corpus sample sequence. and ,in Indicates the length of the corpus sample sequence; Indicates the embedding dimension; The position embedding module obtains the position embedding matrix of the corresponding corpus sample sequence based on the sine position encoding function or the cosine position encoding function. ; S23: For the embedding matrix With position embedding matrix Perform element-wise addition to obtain the initial feature vectors of the corresponding Chinese and English texts.

3. The bilingual native terminology extraction and alignment joint learning method based on hybrid graph neural networks according to claim 2, characterized in that, The method for constructing the semantic similarity graph described in S2 is as follows: The initial feature vector is normalized to obtain a normalized vector, and the formula for obtaining the normalized vector is: In the formula: Represents the words in the initial feature vector The original embedding vector; Indicates word elements Normalized vector; The adjacency matrix is ​​constructed by calculating the cosine similarity between tokens based on the normalized vectors, and the formula for constructing the adjacency matrix is ​​as follows: In the formula: express and Cosine similarity; Indicates word elements Normalized vector; Represents the adjacency matrix; 1 indicates the set similarity threshold; According to the adjacency matrix Get words Neighbor set The Mean Aggregation layer is used to aggregate neighbor features to obtain local information features, and the formula for obtaining the local information features is as follows: In the formula: Indicates the first Layered word elements Feature representation; Indicates the first Layer weight matrix; Indicates the activation function; Indicates the first Layered word elements Feature representation; No. Layered word elements Feature representation; By pre-configuring the self-attention mechanism of the Transformer layer, global dependencies between lexical units are obtained based on local information features to optimize lexical units. The feature representation is used to construct a semantic similarity graph.

4. The bilingual native terminology extraction and alignment joint learning method based on hybrid graph neural networks according to claim 3, characterized in that, The method for constructing the syntactic graph described in S2 is as follows: The syntactic adjacency matrix is ​​obtained based on expert experience through dependency parsing, and the formula for constructing the syntactic adjacency matrix is ​​as follows: In the formula: Represents the syntactic adjacency matrix; Based on the syntactic adjacency matrix Positional encoding is performed on the dependency relationships between lexical units to obtain the encoded positions. and call the function Dependency relationships are mapped to numerical identifiers; based on a pre-configured multi-head attention module, initial attention coefficients are obtained according to the numerical identifiers, and the expression for the initial attention coefficients is as follows: In the formula: Indicates word elements with lexical elements The initial attention coefficient; Indicates word elements Relative to lexical elements Dependency weights; Represents the weight matrix; Indicates word elements The corresponding numerical identifier; Indicates word elements The corresponding numerical identifier; Indicates transpose; After normalizing the initial attention coefficients, the outputs of multiple attention heads in the multi-head attention module are concatenated to obtain the final feature representation. The formula for obtaining the final feature representation is as follows: In the formula: This represents the attention coefficient after normalization. Indicates word elements with lexical elements The initial attention coefficient; Indicates the first The attention coefficient of each attention head and ; Indicates the first The weight matrix of each attention head; Indicates the number of attention heads; Indicates word elements The corresponding final feature representation; Based on the final feature representation combined with the syntactic adjacency matrix Obtain the syntax graph.

5. The bilingual native terminology extraction and alignment joint learning method based on hybrid graph neural networks according to claim 4, characterized in that, The method for constructing the sequence connection graph described in S2 is as follows: Input the initial feature vector into the preset... The model extracts contextual features from the initial feature vector to obtain the contextual feature vector. and ,in Represents the initial eigenvector; Embedding functions by relative position Obtain the word units in the initial feature vector. with lexical elements Relative position embedding matrix and And embed the matrix according to the relative position. Obtain the location feature representation, and the formula for obtaining the location feature representation is: In the formula: Indicates word elements Location feature representation; Through preset Modules are represented based on location features With context feature vector Obtain the sequence connectivity graph, and the formula for obtaining the sequence connectivity graph is as follows: In the formula: Indicates word elements The final feature representation in the sequence diagram; Indicates word elements The corresponding context feature vector and .

6. The bilingual native terminology extraction and alignment joint learning method based on hybrid graph neural networks according to claim 5, characterized in that, The method for constructing the word co-occurrence graph described in S2 is as follows: Based on the sliding window technique, a co-occurrence matrix is ​​obtained according to the co-occurrence relationships of words within the sliding window, and the formula for obtaining the co-occurrence matrix is ​​as follows: In the formula: Represents the co-occurrence matrix; The set of sentences represents the initial feature vector, and Represents a set of sentences Elements in; Indicates the width of the sliding window; Indicates the first The word at each position is ; Indicates the first The word at each position is ; The co-occurrence matrix is ​​weighted based on the TF-IDF algorithm to construct a co-occurrence adjacency matrix; And the formula for constructing the co-occurrence adjacency matrix is ​​as follows: In the formula: Represents the co-occurrence adjacency matrix; 2 indicates based on the weighted processing The value is set as a parameter threshold; For co-occurrence adjacency matrix After symmetric normalization, lexical units in the word co-occurrence graph are obtained through a pre-defined GCN network layer. The final feature representation According to the final feature representation Co-occurrence adjacency matrix Obtain a word co-occurrence graph; And the final feature representation The formula for obtaining it is In the formula: This represents the co-occurrence adjacency matrix after symmetric normalization. Degree matrix; Represents the identity matrix; Indicates the first The feature matrix of the layer; Indicates the first Layer weight matrix; Indicates the first The feature matrix of the layer.

7. The bilingual native terminology extraction and alignment joint learning method based on hybrid graph neural networks according to claim 6, characterized in that, The method for obtaining the optimal model in S3 is as follows: S31: Train the bilingual local term extraction and alignment joint learning model based on hybrid graph neural network using the training set to obtain the trained bilingual local term extraction and alignment joint learning model. S32: Based on the model loss function, the trained bilingual local terminology extraction and alignment joint learning model is validated using the validation set; That is, to determine whether the output of the bilingual local term extraction and alignment joint learning model after training has converged; If the output of the bilingual local terminology extraction and alignment joint learning model after training converges, then the bilingual local terminology extraction and alignment joint learning model after training is confirmed to be the optimal model. Otherwise, based on the backpropagation method, the weight parameters of the trained bilingual local terminology extraction and alignment joint learning model are adaptively adjusted, and step S31 is repeated until the weight parameters of the trained bilingual local terminology extraction and alignment joint learning model with converged output are confirmed to be the optimal weight parameters. The bilingual local terminology extraction and alignment joint learning model is then reconstructed to obtain the optimal model.

8. The bilingual native terminology extraction and alignment joint learning method based on hybrid graph neural networks according to claim 7, characterized in that, The loss function constructed in S32 for: In the formula: This represents the conditional probability, i.e., the log-likelihood of the CRF; Indicates the predicted label sequence and ; Indicates word elements Assign tags The launch fraction; Indicates from the label To tag The transfer fraction; This indicates that the input sequence is the sequence of language sample data in the bilingual corpus dataset; Represents the input sequence X A possible tag sequence; Represents the input sequence The set of all possible label sequences; Represents the input sequence X Corresponding label sequence The score.

Images