A vietnamese dependency syntax analysis method based on fine-grained multilingual alignment enhancement

Abstract

The application relates to a Vietnamese dependency syntax analysis method based on fine-grained multilingual alignment enhancement, and belongs to the field of natural language processing. The application comprises the following steps: in the pre-training stage, word vector representation is constructed, and Chinese and English word vectors are cached; then, the word vectors are sent into an encoding layer and a decoding layer in batches for preliminary joint training; when the model tends to be stable, the alignment training stage is entered: the cached Chinese and English word vectors are taken out in the Vietnamese round, fine-grained cross-language alignment fusion is carried out by combining learnable dynamic gating parameters, enhanced word vectors are generated, a cross-language alignment loss function is designed, and standard dependency loss is jointly optimized; and finally, the trained model is saved. The method realizes accurate migration of Chinese and English semantic knowledge to Vietnamese through a dynamic gating mechanism, effectively alleviates the problem of Vietnamese dependency syntax data scarcity without generating additional data, and significantly improves the accuracy and generalization ability of Vietnamese dependency syntax analysis.

Description

Technical Field

[0001] This invention relates to a Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement, belonging to the field of natural language processing technology. Background Technology

[0002] Language, as the fundamental medium of cross-cultural communication, plays an increasingly crucial role. Vietnamese and Chinese share deep historical ties, and systematic study and research of Vietnamese will not only deepen cultural exchange between the two countries but also provide strong support for regional economic integration.

[0003] However, cross-lingual natural language processing, especially the core task of dependency parsing, still faces significant technical bottlenecks. This task aims to automatically identify grammatical modification relationships between words in a sentence, and its analysis accuracy directly limits the effectiveness of downstream applications such as machine translation and information extraction. For Chinese-Vietnamese translation, the accuracy of Vietnamese-language syntax parsing is a crucial factor affecting the quality of the translation.

[0004] The main challenge lies in the fact that Vietnamese is a resource-scarce language, lacking large-scale, high-quality annotated corpora. Its unique grammatical characteristics also make it difficult to directly transfer and apply syntactic analysis models trained on high-resource languages ​​(such as English and Chinese), resulting in widespread performance degradation. Therefore, exploring efficient and robust dependency parsing schemes for Vietnamese has become an important research topic.

[0005] Against this backdrop, this invention proposes a Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement. This method aims to effectively transfer knowledge from Chinese and English to Vietnamese by fully utilizing rich linguistic resources and corpora in both languages ​​and combining them with the linguistic characteristics of Vietnamese, thereby improving the syntactic parsing performance of Vietnamese. Although Vietnamese, Chinese, and English belong to different language families, they share some structural similarities. For example, Vietnamese and Chinese are both isolating languages ​​and share similarities in syntactic structure; furthermore, during modernization, Vietnamese has absorbed many Chinese and English words and expressions. These similarities provide strong support for cross-linguistic alignment learning.

[0006] This invention achieves accurate transfer of semantic knowledge from Chinese and English to Vietnamese through a fine-grained multilingual alignment enhancement mechanism. It effectively alleviates the scarcity of Vietnamese dependency syntax data without requiring additional data generation, significantly improving the accuracy and generalization ability of Vietnamese dependency syntax analysis. This fine-grained multilingual alignment enhancement strategy improves the parsing performance of Vietnamese dependency syntax, effectively alleviating the scarcity of Vietnamese dependency syntax data and significantly enhancing the accuracy and generalization ability of Vietnamese dependency syntax analysis.

[0007] Experimental results show that fine-grained multilingual alignment enhancement significantly improves the accuracy and stability of Vietnamese dependency parsing, especially when the model can accurately identify the language, which greatly improves the dependency parsing performance of Vietnamese.

[0008] In summary, fine-grained multilingual alignment enhancement not only provides an innovative approach to constructing Vietnamese dependency parsing but also accumulates valuable experience for grammatical analysis research of other low-resource languages. Through continuous optimization and improvement, this method is expected to be applied and promoted in a wider range of cross-lingual natural language processing tasks, thereby driving the development of multilingual technology and promoting communication and understanding between different cultures. Summary of the Invention

[0009] This invention provides a Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement, which can effectively transfer syntactic knowledge from Chinese and English to Vietnamese, improving the performance of Vietnamese dependency parsing. It is used to solve the problem of insufficient syntactic parsing performance caused by the scarcity of Vietnamese dependency parsing treebank resources, and also provides a reference for natural language processing tasks of other low-resource languages.

[0010] The technical solution of this invention is: a Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement, the method comprising:

[0011] Step 1: Collect aligned datasets of Chinese, English and Vietnamese as experimental data, and collect standard dependency syntax trees of Vietnamese as test data;

[0012] Step 2: Load the Chinese, English, and Vietnamese sentences from the unlabeled aligned data above into the multilingual pre-trained language model XLM-RoBERTa and the character-level bidirectional long short-term memory network Char-BiLSTM, respectively, according to the rounds. Combine the above results to obtain the corresponding output word vectors. , , The word vectors for the Chinese and English rounds need to be cached for later use.

[0013] Step 3: Separate word vectors into batches. The data is fed into subsequent encoding and decoding layers for model training, and finally... Update the model parameters to enable the model to learn the syntactic patterns of each language and establish a preliminary shared semantic space;

[0014] Step 4: Once the model has stabilized, continue to process the word vectors in each round. The data is fed into the encoding and decoding layers for alignment training, and finally, a loss function is applied using dependency parsing. and alignment loss function Update model parameters; significantly improve target language parsing performance through alignment-supervised training;

[0015] Step 5: When the final total loss function... Once the model stabilizes, save the final model and use the UD data to test its performance.

[0016] Further, Step 1 includes:

[0017] Step 1.1: Download unlabeled parallel sentences in Chinese, English, and Vietnamese from the standard multilingual alignment corpus FLORES-200 as experimental data. Perform word segmentation on each sentence and input the segmentation results into a pre-trained dependency parser to generate corresponding dependency parsing tree annotations for subsequent supervised training.

[0018] Step 1.2: Download the tagged Vietnamese general dependency syntax tree library from the Universal Dependencies dataset as experimental test data for subsequent performance testing.

[0019] Furthermore, Step 2 includes:

[0020] Step 2.1: Input the segmented Chinese, English, and Vietnamese unlabeled data into the multilingual pre-trained language model XLM-RoBERTa for feature extraction, and name the obtained feature representations respectively. , as well as Specifically, for each language, the last four layers of features are extracted from the last n layers of XLM-RoBERTa. and use learnable weights The final feature representation is obtained by normalizing the data using the Softmax function, then weighting and summing the results, and finally multiplying by the scaling parameter γ. ;

[0021] Step 2.2: Simultaneously, the unlabeled Chinese, English, and Vietnamese data after word segmentation are input into the character-level bidirectional long short-term memory network Char-BiLSTM to encode the character sequence of each word, thus obtaining the character-level representation. ;

[0022] Step 2.3: Generate a randomly initialized word embedding matrix for each language. ;

[0023] Step 2.4: Represent the features of each language separately. With randomly initialized word embeddings Concatenate and combine with character representations The final word vector representation is formed by combining the elements. ;

[0024] Step 2.5: In the training rounds for Chinese and English, the generated word vectors will be... Stored in the cache and simultaneously fed into subsequent encoding and decoding layers for model training;

[0025] Step 2.6: In the training rounds for Vietnamese, obtain the corresponding word vectors through the aforementioned steps. It will not be cached and will be used in Step 4.2.

[0026] Furthermore, Step 3 includes:

[0027] Step 3.1: Divide the word vectors according to the round. First, the word vector is fed into the encoding layer. It then passes through a language-specific BiLSTM encoder to model the context of each word vector in the sentence, resulting in a language-specific context representation. ;

[0028] Step 3.2: Simultaneously, the word vectors are also fed into a BiLSTM encoder common to all languages ​​to extract cross-linguistic common contextual representations. ;

[0029] Step 3.3, and By directly adding and fusing them, we obtain a context vector representation that has linguistic characteristics and commonalities. ;

[0030] Step 3.4: Analyze the context vectors for each language in each round. The loss function is obtained by inputting into a multilayer perceptron (MLP) and two affine layers (Biaffines) and using dependency parsing. Update the model parameters until the model stabilizes;

[0031] Furthermore, Step 4 includes:

[0032] Step 4.1, when word vectors During each round, the data is cached while the normal encoding and decoding process is initiated for model training.

[0033] Step 4.2, when entering word vectors At the next round, retrieve the previously cached word vectors. By combining the dynamic gating parameter λ, dynamic alignment fusion is performed to obtain the alignment enhancement vector. ;

[0034] Step 4.3, Use Training is performed at the encoding and decoding layers, while incorporating word vectors. and dynamic gating parameters The alignment loss function is designed to update the model parameters, reducing alignment interference while updating them. It affects the fusion ratio.

[0035] The present invention also provides a Vietnamese dependency parsing system based on fine-grained multilingual alignment enhancement, the system comprising: a module for executing the Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement.

[0036] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement.

[0037] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement.

[0038] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement.

[0039] The beneficial effects of this invention are:

[0040] 1. By effectively utilizing the rich resources of Chinese and English and combining the linguistic similarities of Vietnamese, this invention significantly reduces the need for a large amount of manually labeled data, thereby reducing the time and cost of developing a Vietnamese dependency parsing model;

[0041] 2. This invention leverages the similarity between the Chinese and English models and Vietnamese to enhance the model's generalization ability in multilingual environments, enabling it to achieve higher accuracy and stability in Vietnamese dependency parsing, especially when dealing with complex syntactic structures.

[0042] 3. The Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement proposed in this invention not only effectively improves the performance of Vietnamese natural language processing tasks, but also provides a useful reference for the processing of other low-resource languages, opening up a new direction for the development of the field of natural language processing.

[0043] 4. This invention provides technical support for Sino-Vietnamese language and cultural exchange by constructing a Vietnamese dependency syntax treebank, while also promoting the application and development of multilingual technology and helping to improve the use and understanding of Vietnamese in the international market. Attached Figure Description

[0044] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0045] Example 1: As Figure 1 As shown, a Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement is presented, the method comprising:

[0046] Step 1: Collect aligned datasets of Chinese, English and Vietnamese as experimental data, and collect standard dependency syntax trees of Vietnamese as test data;

[0047] Further, Step 1 includes:

[0048] Step 1.1: Download unlabeled parallel sentences in Chinese, English, and Vietnamese from the standard multilingual alignment corpus FLORES-200 as experimental data. Perform word segmentation on each sentence and input the segmentation results into a pre-trained dependency parser to generate corresponding dependency parsing tree annotations for subsequent supervised training.

[0049] Step 1.2: Download the tagged Vietnamese general dependency syntax tree library from the Universal Dependencies dataset as experimental test data for subsequent performance testing.

[0050] Step 2: Load the Chinese, English, and Vietnamese sentences from the unlabeled aligned data above into the multilingual pre-trained language model XLM-RoBERTa and the character-level bidirectional long short-term memory network Char-BiLSTM, respectively, according to the rounds. Combine the above results to obtain the corresponding output word vectors. , , The word vectors for the Chinese and English rounds need to be cached for later use.

[0051] Furthermore, Step 2 includes:

[0052] Step 2.1: Input the segmented Chinese, English, and Vietnamese unlabeled data into the multilingual pre-trained language model XLM-RoBERTa for feature extraction, and name the obtained feature representations respectively. , as well as Specifically, for each language, the last four layers of features are extracted from the last n layers of XLM-RoBERTa. and use learnable weights The final feature representation is obtained by normalizing the data using the Softmax function, then weighting and summing the results, and finally multiplying by the scaling parameter γ. The calculation formula is as follows:

[0053]

[0054]

[0055] Step 2.2: Simultaneously, the unlabeled Chinese, English, and Vietnamese data after word segmentation are input into the character-level bidirectional long short-term memory network Char-BiLSTM to encode the character sequence of each word, thus obtaining the character-level representation. ;

[0056] Step 2.3: Generate a randomly initialized word embedding matrix for each language. ;

[0057] Step 2.4: Represent the features of each language separately. With randomly initialized word embeddings Concatenate and combine with character representations The final word vector representation is formed by combining the elements. The calculation formula is as follows:

[0058]

[0059] Where ⊕ represents a vector concatenation operation, which is then mapped to a unified dimension D through a linear transformation;

[0060] Step 2.5: In the training rounds for Chinese and English, the generated word vectors will be... Stored in the cache and simultaneously fed into subsequent encoding and decoding layers for model training;

[0061] Step 2.6: In the training rounds for Vietnamese, obtain the corresponding word vectors through the aforementioned steps. It will not be cached and will be used in Step 4.2.

[0062] Step 3: Separate word vectors into batches. The data is fed into subsequent encoding and decoding layers for model training, and finally... Update the model parameters to enable the model to learn the syntactic patterns of each language and establish a preliminary shared semantic space;

[0063] Furthermore, Step 3 includes:

[0064] Step 3.1: Divide the word vectors according to the round. First, the word vector is fed into the encoding layer. It then passes through a language-specific BiLSTM encoder to model the context of each word vector in the sentence, resulting in a language-specific context representation. ;

[0065] Step 3.2: Simultaneously, the word vectors are also fed into a BiLSTM encoder common to all languages ​​to extract cross-linguistic common contextual representations. ;

[0066] Step 3.3, and By directly adding and fusing them, we obtain a context vector representation that has linguistic characteristics and commonalities. ;

[0067] Step 3.4: Analyze the context vectors for each language in each round. The loss function is obtained by inputting into a multilayer perceptron (MLP) and two affine layers (Biaffines) and using dependency parsing. Update the model parameters until the model stabilizes;

[0068] Step 4: Once the model has stabilized, continue to process the word vectors in each round. The data is fed into the encoding and decoding layers for alignment training, and finally, a loss function is applied using dependency parsing. and alignment loss function Update model parameters; significantly improve target language parsing performance through alignment-supervised training;

[0069] Furthermore, Step 4 includes:

[0070] Step 4.1, when word vectors During each round, the data is cached while the normal encoding and decoding process is initiated for model training.

[0071] Step 4.2, when entering word vectors At the next round, retrieve the previously cached word vectors. By combining the dynamic gating parameter λ, dynamic alignment fusion is performed to obtain the alignment enhancement vector. ;

[0072] Step 4.3, Use Training is performed at the encoding and decoding layers, while incorporating word vectors. and dynamic gating parameters The alignment loss function is designed to update the model parameters, reducing alignment interference while updating them. It affects the fusion ratio.

[0073] Further, Step 4.2 includes:

[0074] When the model inputs Vietnamese word vectors During the training rounds, we extract the corresponding batch of Chinese word vectors from the cache. English word vectors And combined with globally learnable dynamic gating parameters For each Vietnamese word Perform fine-grained cross-language alignment fusion; specifically, for the i-th word in a Vietnamese sentence... It is related to the j-th word in the Chinese sentence. Cosine similarity is defined as:

[0075] ;

[0076] Similarly, the cosine similarity with the k-th word e_{wk} in the English sentence is:

[0077]

[0078] Therefore, a similarity matrix between Vietnam and China is constructed. Similarity matrix between Vietnam and English Where n is the number of words in the current Vietnamese sentence, and m and h are the number of words in the corresponding Chinese and English sentences, respectively;

[0079] For each Vietnamese word We were respectively in and Selected from The index of the most similar words is denoted as:

[0080]

[0081] Subsequently, these Top-k similarity values ​​are Softmax normalized to generate normalized attention weights:

[0082]

[0083]

[0084] Next, the Chinese and English word vectors are weighted and fused using a dynamic gating parameter λ:

[0085]

[0086] Finally, the fused vector is added to the original Vietnamese word vectors to obtain the alignment-enhanced representation:

[0087]

[0088] This process applies to all Vietnamese words. Execution is performed in parallel, and the final output is an enhanced word vector sequence.

[0089] Step 4.3, Use Training is performed at the encoding and decoding layers, while incorporating word vectors. and dynamic gating parameters The alignment loss function is designed to update the model parameters, reducing alignment interference while updating them. Affects the fusion ratio;

[0090] Furthermore, Step 4.3 includes:

[0091] The aligned and enhanced word vectors output from Step 4.2 The input is fed into the encoding layer; firstly, language-specific features are extracted using a Vietnamese-specific BiLSTM, and cross-lingual common features are extracted using a multilingual BiLSTM with shared parameters. These two features are then concatenated or weighted to generate a context-aware representation. This representation is then fed into the decoding layer, where dependency prediction is performed using an MLP classifier and a Biaffine module. Further improvements in semantic consistency and guidance of dynamic gating parameters are then implemented. For adaptive adjustment, we introduce cross-language alignment loss. Its goal is to constrain the original Vietnamese word vectors. Chinese characters in the cache ,English Overall semantic consistency between them:

[0092]

[0093] in The Frobenius norm, i.e., the sum of squares of all elements; dynamic gating parameter. From learnable parameters Obtained by Sigmoid mapping and scaling:

[0094]

[0095] Should The gating parameters used for fusion in Step 4.2 are fully shared and jointly optimized during training via backpropagation, enabling the model to adaptively determine which source language the Vietnamese semantics are closer to. Finally, the model's total loss function is:

[0096]

[0097] in For standard dependency loss, For alignment loss.

[0098] Step 5, when the last Once the model stabilizes, save the final model and use the UD data to test its performance.

[0099] Furthermore, the present invention also provides a Vietnamese dependency parsing system based on fine-grained multilingual alignment enhancement, the system comprising:

[0100] The collection and preprocessing module is used to collect aligned parallel corpora in Chinese, English and Vietnamese as experimental data, and to perform word segmentation and dependency syntax tree annotation to build a supervised training dataset.

[0101] The multilingual word vector construction module is used to input the segmented Chinese, English, and Vietnamese data into the multilingual pre-trained models XLM-RoBERTa and Character-level Bidirectional LSTM (Char-BiLSTM), respectively, and fuse the outputs of the two to obtain the word vector representations of the corresponding languages. The Chinese and English word vectors are cached during training for use in subsequent Vietnamese rounds.

[0102] The initial joint training module is used to batch train word vectors. The data is fed into the encoding and decoding layers, where language-specific and common features are extracted using private and shared BiLSTM. Dependency relationships are then predicted using MLP and Biaffine modules, and standard dependency loss is applied. Update model parameters and establish a preliminary multilingual semantic space;

[0103] The dynamic alignment enhancement module is used to retrieve cached Chinese and English word vectors from the Vietnamese rounds after the model has stabilized, and combine them with learnable dynamic gating parameters. For each Vietnamese word, fine-grained cross-linguistic similarity calculation, Top-k selection, Softmax normalization, and weighted fusion are performed to generate aligned and enhanced word vectors. ;

[0104] The joint optimization training module is used to align and enhance word vectors. The signal is fed into the encoder-decoder layer for training, while introducing cross-language alignment loss. This loss uses the original Vietnamese word vectors as anchors, constraining their overall semantic consistency with the Chinese and English word vectors, and is consistent with the standard dependency loss. Joint optimization enables dynamic gating parameters The model adapts and adjusts during training to amplify the contribution of source language with higher semantic matching.

[0105] The model saving module is used when the total loss... Once the model stabilizes, save the final trained model and use the Universal Dependency Tree Library (UD) for performance evaluation and practical application.

[0106] The final model was evaluated using a test set, and the LAS (Labeled Attachment Score) and UAS (Unlabeled Attachment Score) were analyzed. The evaluation revealed that the model trained using our method significantly outperformed the original model. The specific formula used is as follows:

[0107] (1) LAS predicts the accuracy of the model in correctly predicting dependencies and dependency labels;

[0108]

[0109] (2) UAS predicts the accuracy of dependencies (without considering dependency labels);

[0110]

[0111] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement.

[0112] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement.

[0113] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement.

[0114] To verify the effectiveness of the method proposed in this invention, the FLORES-200 multilingual alignment corpus was selected as training data for training the model, and finally a systematic evaluation was performed on the Universal Dependencies (UD) standard test set.

[0115] The purpose of this operation is to effectively align the syntactic knowledge of Chinese and English with that of Vietnamese, thereby significantly improving the efficiency, generalization ability, and processing performance of the dependency parsing model. This method can alleviate the problem of scarce Vietnamese dependency parsing data to some extent, thus improving the overall performance of Vietnamese dependency parsing.

[0116] This invention selects two baseline models for comparison: the first is the "Fulsha" model, a traditional dual affine parser structure, whose encoding layer only uses a shared BiLSTM and does not explicitly model language specificity; the second is the "Share_Private" model, which additionally inputs language identifiers into a private BiLSTM encoder to guide the model to learn language-specific contextual representations, thereby improving modeling capabilities in multilingual scenarios. Neither of these models incorporates the dynamic alignment and fusion mechanism proposed in this invention as a baseline reference.

[0117] Subsequently, our method was incorporated into the model to form enhanced versions—namely, "Fulsha+Our" and "Share_Private+Our," where "Our" represents the complete alignment enhancement framework proposed in Step 4 of this invention, including: extracting Chinese and English word vectors from the cache, calculating cross-language similarity, Top-k selection, Softmax weight normalization, dynamic gating fusion, and jointly optimizing the alignment loss. and Dependency Loss The entire process of [the process].

[0118] The experimental results of this invention are shown in Table 1. The table shows that the method proposed in this invention outperforms the benchmark model in all evaluation metrics. Of particular note is the fact that the "Share_Private+Our" model performs best among all variants, indicating that introducing linguistic identifiers at the encoding layer to distinguish linguistic features, combined with the dynamic alignment enhancement mechanism of this invention, can achieve a dual gain of language-specific modeling and cross-linguistic knowledge transfer, thereby maximizing the performance of Vietnamese dependency parsing.

[0119] Finally, by comprehensively applying the series of alignment enhancement strategies proposed in this invention—from multilingual word vector construction to preliminary joint training, and then to dynamic gating alignment enhancement and joint loss optimization—the parsing efficiency and robustness of the model on Vietnamese were significantly improved. This achievement not only promotes the development of low-resource language natural language processing technology but also provides strong support for the construction of Sino-Vietnamese language and cultural exchange and multilingual information processing systems.

[0120] Table 1. Performance comparison of this method with traditional methods on the model.

[0121] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement, characterized in that: The method includes: Step 1: Collect aligned datasets of Chinese, English and Vietnamese as experimental data, and collect standard dependency syntax trees of Vietnamese as test data; Step 2: Load the Chinese, English, and Vietnamese sentences from the unlabeled aligned data above into the multilingual pre-trained language model XLM-RoBERTa and the character-level bidirectional long short-term memory network Char-BiLSTM, respectively, according to the rounds. Combine the above results to obtain the corresponding output word vectors. , , The word vectors for the Chinese and English rounds need to be cached for later use. Step 3: Separate word vectors into batches. The data is fed into subsequent encoding and decoding layers for model training, and finally, a loss function is applied using dependency parsing. Update the model parameters to enable the model to learn the syntactic patterns of each language and establish a preliminary shared semantic space; Step 4: Once the model has stabilized, continue to process the word vectors in each round. The data is fed into the encoding and decoding layers for alignment training, and finally, a loss function is applied using dependency parsing. and alignment loss function Update model parameters; significantly improve target language parsing performance through alignment-supervised training; Step 5: When the final total loss function... Once the model stabilizes, save the final model and use the UD data to test its performance.

2. The Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement according to claim 1, characterized in that: Step 1 includes: Step 1.1: Download unlabeled parallel sentences in Chinese, English and Vietnamese from the standard multilingual alignment corpus FLORES-200 as experimental data. Perform word segmentation on each sentence and input the segmentation results into the pre-trained dependency parser to generate corresponding dependency parsing tree annotations for subsequent supervised training. Step 1.2: Download the tagged Vietnamese general dependency syntax tree library from the Universal Dependencies dataset as experimental test data for subsequent performance testing.

3. The Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement according to claim 1, characterized in that: Step 2 includes: Step 2.1: Input the segmented Chinese, English, and Vietnamese unlabeled data into the multilingual pre-trained language model XLM-RoBERTa for feature extraction, and name the obtained feature representations respectively. , as well as Specifically, for each language, the last four layers of features are extracted from the last n layers of XLM-RoBERTa. and use learnable weights The final feature representation is obtained by normalizing the data using the Softmax function, then weighting and summing the results, and finally multiplying by the scaling parameter γ. ; Step 2.2: Simultaneously, the unlabeled Chinese, English, and Vietnamese data after word segmentation are input into the character-level bidirectional long short-term memory network Char-BiLSTM to encode the character sequence of each word, thus obtaining the character-level representation. ; Step 2.3: Generate a randomly initialized word embedding matrix for each language. ; Step 2.4: Represent the features of each language separately. With randomly initialized word embeddings Concatenate and combine with character representations The final word vector representation is formed by combining the elements. ; Step 2.5: In the training rounds for Chinese and English, the generated word vectors will be... Stored in the cache and simultaneously fed into subsequent encoding and decoding layers for model training; Step 2.6: In the training rounds for Vietnamese, obtain the corresponding word vectors through the aforementioned steps. It will not be cached and will be used in Step 4.

4. The Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement according to claim 1, characterized in that: Step 3 includes: Step 3.1: Represent the word vectors according to the round. First, the word vector is fed into the encoding layer. It then passes through a language-specific BiLSTM encoder to model the context of each word vector in the sentence, resulting in a language-specific context representation. ; Step 3.2: Simultaneously, the word vectors are also fed into a BiLSTM encoder common to all languages ​​to extract cross-linguistic common contextual representations. ; Step 3.3, and By directly adding and fusing them, we obtain a context vector representation that has linguistic characteristics and commonalities. ; Step 3.4: Analyze the context vectors for each language in each round. The loss function is obtained by inputting into a multilayer perceptron (MLP) and two affine layers (Biaffines) and using dependency parsing. Update the model parameters until the model is stable.

5. The Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement according to claim 1, characterized in that: Step 4 includes: Step 4.1: When Chinese and English word vectors are used... During each round, the data is cached while the normal encoding and decoding process is initiated for model training. Step 4.2: When entering the Vietnamese word vector... At the next round, retrieve the previously cached word vectors. Combined with dynamic gating parameters Dynamic alignment fusion is performed to obtain the alignment enhancement vector. ; Step 4.3, Use Training is performed at the encoding and decoding layers, while incorporating word vectors. and dynamic gating parameters The alignment loss function is designed to update the model parameters, reducing alignment interference while updating them. It affects the fusion ratio.

6. A Vietnamese dependency parsing system based on fine-grained multilingual alignment enhancement, characterized in that, The system includes a module for performing a Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement as described in any one of claims 1 to 5.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the program, it implements a Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the Vietnamese dependency parsing method based on fine-grained multilingual alignment enhancement as described in any one of claims 1 to 5.

Images