Vietnamese dependency syntax data enhancement method based on big language model collaboration

Abstract

The invention provides a Vietnamese dependency syntax data enhancement method based on large language model collaboration, and belongs to the field of natural language processing. The method comprises the following steps of: generating an initial dependency syntax tree of Vietnamese sentences by using a plurality of current mainstream large language models as a multi-label model; by introducing an expert large language model GPT-4o-mini with a better analysis effect, scoring and feeding back a dependency tree output by each round, and driving a multi-label model to iteratively optimize until a set threshold value or a maximum round number is reached; and automatically selecting the dependency syntax tree with the highest score from the output of the three annotation models of each sentence to form a high-quality synthesized Vietnamese dependency syntax tree. And the syntactic analysis model is fused with an original UD Vietnamese training set, and training of the syntactic analysis model is completed. According to the method, high-quality automatic expansion of low-resource language Vietnamese dependency syntax data is realized, and the dependency analysis performance is effectively improved.

Description

[0001] Technical Language

[0002] This invention relates to a Vietnamese dependency syntax data augmentation method based on large language model collaboration, belonging to the field of natural language processing technology. Background Technology

[0003] Against the backdrop of the booming development of multilingual NLP, how to effectively construct high-quality dependency syntax treebanks for low-resource languages, including Vietnamese, has become a hot topic of research both domestically and internationally. As a fundamental resource for syntactic analysis, dependency syntax treebanks are crucial for downstream applications such as machine translation, semantic understanding, and information extraction. Traditional dependency parsing methods heavily rely on large amounts of meticulously annotated, high-quality corpora, which is feasible for high-resource languages ​​like English and Chinese, but faces numerous limitations for Vietnamese, a language with extremely scarce resources. First, there is a lack of publicly available large-scale dependency syntax annotation datasets for Vietnamese; second, existing dependency syntax models generally perform poorly on Vietnamese, exhibiting limited generalization ability and low accuracy, severely hindering research progress and industrial applications of Vietnamese in cross-linguistic language processing.

[0004] Meanwhile, Large Language Models (LLMs), such as GPT, LLaMA, Qwen, DeepSeek, and ChatGLM, have made groundbreaking progress in various fields, including natural language generation, question answering systems, dialogue generation, text summarization, and machine translation, thanks to their powerful language understanding and generation capabilities on large-scale corpora. In particular, LLMs have demonstrated good robustness and transferability in multilingual understanding and cross-lingual transfer. Nevertheless, the potential of current large language models in structured grammar tasks has not yet been fully realized, especially in dependency syntax prediction and syntax tree construction; their collaborative corpus building and knowledge integration capabilities still require further exploration.

[0005] This study argues that effectively combining the language generalization advantages of LLMs with the structural modeling capabilities of traditional parsing models (such as the BiAffine Parser), and leveraging multi-model collaboration and expert evaluation mechanisms, can significantly improve the quality of dependency tree construction for low-resource languages ​​such as Vietnamese. By introducing multiple large language models to annotate and parse the same sentence, and utilizing expert large language models to score, filter, and provide feedback on the parsing results, a high-quality dependency treebase that is structurally complete, semantically sound, and grammatically accurate can be automatically constructed without the need for large-scale manual annotation.

[0006] Experimental results show that the strategy of combining collaborative parsing with large language models with an expert scoring feedback mechanism significantly improves the accuracy and robustness of the model in Vietnamese dependency parsing tasks. Compared with traditional single-parsing model approaches, this method effectively improves the quality and consistency of pseudo-labels and reduces the risk of grammatical error propagation in unlabeled corpora by integrating the parsing capabilities of multiple large language models (such as Qwen, ChatGLM, and DeepSeek) and combining the GPT-4o-mini expert large language model to perform multiple rounds of scoring and optimization on the generated syntax tree.

[0007] By constructing a closed-loop process of multi-round collaborative parsing—expert feedback—structural rewriting, this invention not only significantly improves the efficiency of automated construction of Vietnamese dependency syntax treebanks, but also achieves considerable improvements in two metrics: LAS (Labeled Dependency Accuracy) and UAS (Unlabeled Dependency Accuracy). Experimental data show that LAS increased from 63.180 in the original model to 64.882, and UAS increased from 77.025 to 79.807, with an overall performance improvement of approximately 2 percentage points, verifying the effectiveness and generalizability of this method in improving syntactic analysis performance.

[0008] In summary, the Vietnamese dependency syntax data augmentation method based on large language model collaboration proposed in this invention not only provides an innovative approach to alleviating the difficulties of low-resource language annotation but also opens up new paths for structured language processing tasks in other resource-scarce languages. This method shows promising application prospects in multilingual transfer learning, automatic corpus augmentation, and intelligent annotation assistance. In the future, it can be further extended to the construction of structured corpora for Southeast Asian languages ​​such as Burmese, Thai, and Lao, providing strong support for the development of global multilingual NLP technology and the deepening of multilingual communication. This construction method based on the LLM collaborative mechanism provides a new technical path and theoretical support for the automatic generation of low-resource language syntax treebanks, and also lays a solid foundation for promoting the expansion of multilingual NLP to more languages ​​and more tasks. Summary of the Invention

[0009] The technical problem this invention aims to solve is to provide a Vietnamese dependency syntax data augmentation method based on collaborative large language models. This method utilizes multiple mainstream large language models to parse unlabeled Vietnamese corpora, and then uses expert large language models to uniformly evaluate and select high-quality dependency tree results, achieving a closed loop of automatic pseudo-annotation and quality control. Compared with traditional methods, this method eliminates the need for extensive manual annotation, effectively improving corpus construction efficiency while ensuring the grammatical rationality and structural integrity of dependency relations.

[0010] This invention innovatively introduces a "label-expert" collaborative optimization mechanism, combining multi-model competition and feedback iteration, significantly improving the performance of Vietnamese dependency parsing. This method is particularly suitable for addressing the problems of scarce labeled data and insufficient accuracy of parsing models in low-resource languages ​​such as Vietnamese, while also providing a general and scalable solution for syntactic tree construction in other resource-constrained languages.

[0011] The technical solution of this invention is: a Vietnamese dependency syntax data augmentation method based on large language model collaboration, the method comprising:

[0012] Step 1: Obtain labeled Vietnamese dependency syntax data from the Universal Dependencies (UD) dataset and collect unlabeled Vietnamese monolingual corpora, and perform relevant preprocessing operations on both.

[0013] Step 2: Construct a dependency parsing system environment, select a mainstream large language model as the multi-label model, parse the preprocessed unlabeled data, and output a five-column CoNLL-U format dependency parsing tree;

[0014] Step 3: Construct an expert large language model, and have the expert large language model score and provide feedback on the output of each labeled model; evaluate its grammatical rationality and dependency structure accuracy;

[0015] Step 4: The expert large language model performs multiple rounds of optimization on each labeled model based on expert feedback, and scores and provides feedback on the output of each round, sets an early stopping mechanism, and selects the dependency syntax tree with the highest score as the optimal result;

[0016] Step 5: For each sentence, select the optimal dependency syntax tree from the parsing results of each model, convert the optimal dependency tree into a standard 10-column CoNLL-U format for training, and save it to a separate .conllu file, retaining the complete dependency information fields;

[0017] Step 6: Perform structural consistency checks and label filtering on the output dependency syntax tree to remove sentences with abnormal structures or illegal labels. At the same time, combine manual evaluation to select high-quality data.

[0018] Step 7: Merge the selected high-quality dependency parsing trees with the original UD training set to construct a high-quality synthetic Vietnamese dependency parsing tree, and use it to retrain the BiLSTM+MLP+BiAffine Parser parser model.

[0019] Further, Step 1 includes:

[0020] Step 1.1: Obtain the Vietnamese training set from the general dataset UD, and collect unlabeled Vietnamese monolingual corpora from datasets such as ALT and FLORES-200;

[0021] Step 1.2: Use the Underthesea tool to segment the Vietnamese sentences;

[0022] Step 1.3: Preprocess the data, including filtering out duplicates and sentences with abnormal punctuation.

[0023] Furthermore, Step 2 includes:

[0024] Step 2.1: Call multiple mainstream large language models together as a multi-label model, including Qwen, ChatGLM and DeepSeek, to perform dependency parsing on the preprocessed Vietnamese sentences;

[0025] Step 2.2: Generate a preliminary CoNLL-U format dependency syntax tree (five-column simplified version) and output the independent parsing results for each model.

[0026] Furthermore, Step 3 includes:

[0027] Step 3.1: Select GPT-4o-mini as the unified scoring expert big language model, conduct quality assessment and feedback on the output of each labeled model, and provide guidance for the subsequent screening of analysis results and model iteration;

[0028] Step 3.2: The expert big language model outputs scores and natural language feedback information for syntactic structure, dependency relation accuracy, and semantic consistency. The model parses sentences according to a unified prompt to generate dependency syntax trees.

[0029] Furthermore, Step 4 includes:

[0030] Step 4.1: Each labeled model undergoes multiple rounds of parsing, with the expert large language model GPT-4o-mini scoring the output of each round and providing structural improvement suggestions;

[0031] Step 4.2: Record the scoring results in each iteration, find the dependency syntax tree with the highest score from all output results as the optimal result, and stop early and save the tree structure if the score is higher than the set threshold (e.g., 0.95).

[0032] Furthermore, Step 5 includes:

[0033] Step 5.1: Convert the selected optimal results into the standard CoNLL-U ten-column format, including: completing the optimal dependency syntax tree output by each model into the standard CoNLL-U ten-column format, retaining five core fields (word index, vocabulary form, part of speech, headword index, dependency relation), and uniformly filling the remaining columns (such as lexical, XPOS, features, augmented dependencies, annotations) with underscores (_) to ensure that the data format is standardized and consistent;

[0034] Step 5.2: Organize the preliminary results into the manual screening process, manually check the rationality of the grammatical structure and the standardization of tags, and remove low-quality output;

[0035] Step 5.3: Retain fully annotated and structurally correct sentences for subsequent training. The transformation results can be saved to separate .conllu files for each model, ensuring consistent data structure and usability for later training.

[0036] Furthermore, Step 6 includes:

[0037] Step 6.1: Perform multiple consistency checks on the output dependency syntax tree set; including: manually filtering the generated dependency trees by labels, retaining only dependency relation types (such as root, nsubj, obj, case, advmod, etc.) within the set of valid high-frequency labels;

[0038] Step 6.2: Building upon automatic structure verification, a manual review mechanism is introduced. Linguistic annotation experts review the dependency trees selected by the machine for grammatical correctness, semantic rationality, and label consistency. This includes deleting anomalous sentences with invalid labels, incomplete structures, or missing root nodes, and constructing a pseudo-annotated dataset with controllable quality.

[0039] Furthermore, Step 7 includes:

[0040] Step 7.1: Merge all the selected high-quality dependency parsing trees with the original UD training data;

[0041] Step 7.2: Perform consistency checks on the merged dataset, including root checks, label validity checks, and duplicate sequence number checks; construct a high-quality synthetic Vietnamese dependency syntax tree;

[0042] Step 7.3: Use the augmented corpus as input to train the parser model of BiLSTM+MLP+BiAffine Parser.

[0043] The beneficial effects of this invention are:

[0044] 1. This invention fully leverages the language understanding and generation capabilities of Large Language Models (LLMs), and introduces models such as ChatGLM, Qwen, and DeepSeek as multi-annotation models to perform multi-round collaborative parsing of Vietnamese syntax structures. It achieves automatic construction of high-quality dependency syntax trees without requiring extensive manual annotation, effectively reducing annotation costs and improving database construction efficiency.

[0045] 2. This invention introduces expert large language models (such as GPT-4o-mini) for scoring and feedback, which promotes iterative optimization of multi-labeled models, selects the optimal structural results, improves the structural consistency and language accuracy of treebank construction, and effectively avoids structural conflicts and label confusion.

[0046] 3. This invention designs a complete data filtering and quality control process, including format standardization, label validity filtering, multi-round refinement, score threshold limitation and other measures to ensure that the final generated tree library has high structural stability and semantic rationality.

[0047] 4. After the dependency syntax tree library constructed in this invention is fused with the original UD training set, it shows a significant improvement in testing on the BiAffine Parser, verifying the feasibility and practicality of large language model-assisted library construction.

[0048] 5. This invention provides a general, scalable, and low-cost high-quality syntactic treebank construction scheme for low-resource languages ​​such as Vietnamese. It also has good transferability and can be applied to natural language processing tasks in other languages ​​(such as Burmese, Thai, Lao and other Southeast Asian languages), promoting the practical application of multilingual NLP research. Attached Figure Description

[0049] Figure 1 This is a flowchart from the present invention. Detailed Implementation

[0050] Example 1: As Figure 1 As shown, a Vietnamese dependency syntax data augmentation method based on large language model collaboration is proposed. The method first obtains unlabeled Vietnamese corpora from datasets such as ALT and FLORES-200 and preprocesses them. Then, multiple mainstream large language models (such as Qwen, ChatGLM, and DeepSeek) are used as multi-labeled model parsers to generate a five-column CoNLL-U format dependency syntax tree. Finally, an expert large language model (such as GPT-4o-mini) scores and provides feedback on the generated results, driving the model to undergo multiple rounds of optimization.

[0051] The optimized optimal syntax tree was converted to the standard ten-column CoNLL-U format and fused with the original Vietnamese dependency parsing dataset to construct a high-quality training set. Finally, the BiAffine Parser was used for model training, significantly improving the performance of Vietnamese dependency parsing, especially demonstrating better structural accuracy and generalization ability under low-resource conditions. The specific steps of the method are as follows:

[0052] Step 1: Obtain Vietnamese labeled dependency syntax data from the general dataset UD, and collect unlabeled Vietnamese monolingual corpus, and perform relevant preprocessing operations on both.

[0053] Further, Step 1 includes:

[0054] Step 1.1: Obtain manually labeled Vietnamese dependency syntax data from the Universal Dependencies (UD) dataset as high-quality basic training corpus for this project, and collect unlabeled Vietnamese monolingual corpus from multiple public datasets (such as AsianLanguage Treebank (ALT), FLORES-200, etc.) for subsequent automatic construction of dependency structures and model enhancement.

[0055] Step 1.2: The Underthesea Vietnamese natural language processing toolkit was used to perform standardized preprocessing on the labeled and unlabeled corpora. This included sentence-level text segmentation, accurate word segmentation using a language model, deduplication of repeated sentences, and character regularization, symbol standardization, and punctuation unification. The result was a standardized and high-quality Vietnamese input corpus, which served as the foundation for subsequent dependency tree construction.

[0056] Step 2: Construct a dependency parsing system environment, select a mainstream large language model as the multi-label model, parse the preprocessed unlabeled data, and output a five-column CoNLL-U format dependency parsing tree;

[0057] Furthermore, Step 2 includes:

[0058] Step 2.1: Input the cleaned and standardized Vietnamese sentences one by one into several mainstream large language models (LLMs), including Qwen, ChatGLM and DeepSeek, and perform dependency parsing on the preprocessed Vietnamese sentences;

[0059] In this process, each large language model participates as an independent "dependency syntax tree multi-annotation model". Based on its powerful semantic understanding and language generation capabilities, it independently performs dependency parsing on each input sentence and automatically generates preliminary dependency structures to simulate the effect of high-quality manual annotation, providing diverse structural candidates for subsequent screening and training.

[0060] To improve the standardization and accuracy of the parsing structure of large language models, a unified prompt design was designed and applied, employing a dual-guidance structure of "System Role + User Task". The system prompt explicitly identifies the model as a "Vietnamese dependency parsing and annotation expert," while the user prompt defines a specific task: completing the input partial CoNLL-U structure into a complete five-column format, including word index, form, part-of-speech tag, dependency head, and dependency relation tag, thereby ensuring high consistency and parsability of the output structure.

[0061] Step 2.2: All dependency syntax trees output by large language models are uniformly organized into standard five-column CoNLL-U format text files and numbered. This format strictly follows dependency syntax annotation specifications to ensure consistency and traceability in subsequent expert evaluation and comparison of large language models, providing a structured and standardized data foundation for subsequent optimization, selection, and training.

[0062] Step 3: Construct an expert large language model, and use the expert large language model to score and provide feedback on the output of each labeled model;

[0063] Furthermore, Step 3 includes:

[0064] Step 3.1: GPT-4o-mini was selected as the unified expert scoring language model. The output of each labeled model was evaluated and feedback was provided to guide the subsequent selection of analysis results and model iteration. Among these steps,

[0065] To achieve high-quality structural evaluation of dependency syntax trees, an expert large language model (ExpertEvaluator) is constructed, and a guiding and constraining system prompt is designed for it. This prompt clearly defines the expert large language model as a "dependency syntax structure analysis expert," whose task is to comprehensively evaluate the syntactic structural integrity, dependency relation rationality, and label usage standardization of a given five-column CoNLL-U format dependency syntax tree, outputting a structural accuracy score between 0.0 and 1.0, and providing detailed textual evaluation feedback, including potential error types, optimization suggestions, and label replacement suggestions, thus providing guidance for subsequent multi-annotated model optimization.

[0066] Step 3.2: The expert large language model outputs scores and natural language feedback information for syntactic structure, dependency relation accuracy, and semantic consistency. The model parses sentences according to a unified prompt to generate dependency syntax trees; among which:

[0067] The dependency syntax structures generated by multi-annotation models (such as Qwen, ChatGLM, and DeepSeek) are input one by one into the expert large language model, triggering independent evaluation of each dependency tree by experts. The system automatically records the scoring results and comments given by the expert large language model, serving as the core basis for quantifying annotation quality and diagnosing structural defects. These scores and feedback not only provide quantitative and qualitative references for the next round of parsing optimization (refinement), but can also be used to compare the performance of different models, select the optimal dependency structure results, and ultimately drive continuous iterative improvement in syntax tree quality.

[0068] Step 4: The expert large language model performs multiple rounds of optimization on each labeled model based on expert feedback, and scores and provides feedback on the output of each round, sets an early stopping mechanism, and selects the dependency syntax tree with the highest score as the optimal result;

[0069] Furthermore, Step 4 includes:

[0070] Step 4.1: Each labeled model undergoes multiple rounds of parsing, with the expert large language model GPT-4o-mini scoring the output of each round and providing structural improvement suggestions; including:

[0071] The scoring results and structured feedback output by the expert large language model are returned to each annotation model (such as Qwen, ChatGLM, and DeepSeek) as the basis for iterative optimization. Each annotation model modifies and adjusts its original dependency syntax tree structure in a targeted manner based on the feedback (i.e., refines the process), including but not limited to correcting dependency centers, correcting incorrect dependency labels, adjusting dependency arc directions, or supplementing missing structures, thereby generating an updated dependency tree version.

[0072] The updated dependency structure is then re-input into the expert large language model for re-evaluation to obtain new scores and improvement suggestions.

[0073] Step 4.2: Record the scoring results in each iteration, find the dependency syntax tree with the highest score from all output results as the optimal result, and stop early and save the tree structure if the score is higher than the set threshold.

[0074] The process iterates until a certain round of scoring reaches or exceeds a preset score threshold (e.g., 0.95) or reaches the set maximum number of iteration rounds (e.g., 3 rounds). Finally, the dependency tree with the highest score from all rounds is selected as the optimal parsing result of the current multi-label model on that sentence, and used for subsequent data generation and training set expansion.

[0075] Step 5: For each sentence, select the optimal dependency syntax tree from the parsing results of each model, and convert the optimal dependency tree into a standard 10-column CoNLL-U format for training;

[0076] Furthermore, Step 5 includes:

[0077] Step 5.1: Convert the filtered best results into a standard CoLL-U ten-column format; including:

[0078] For the optimal parsing results output by various language models (such as Qwen, ChatGLM, and DeepSeek) on each sentence, structural completion is performed according to the CoNLL-U standard format, expanding the original five-column structure to a ten-column structure to ensure compliance with the common format specifications of dependency parsing treebanks.

[0079] The ten-column CoNLL-U format dependency tree generated and completed for each model is written into its own independent output file, such as Qwen_best.conllu, ChatGLM_best.conllu, and DeepSeek_best.conllu. These files will serve as pseudo-annotated dependency data sources for subsequent data integration, quality screening, and training set enhancement, laying the foundation for building a high-quality Vietnamese dependency syntax tree library.

[0080] Step 5.2: Organize the preliminary results into the manual screening process, manually check the rationality of the grammatical structure and the standardization of tags, and remove low-quality output;

[0081] Step 5.3: Retain sentences with complete annotations and correct structure for use in subsequent training.

[0082] Step 6: Perform structural consistency checks and label filtering on the output dependency syntax tree to remove sentences with abnormal structures or illegal labels. At the same time, combine manual evaluation to select high-quality data.

[0083] Furthermore, Step 6 includes:

[0084] Step 6.1: Perform multiple consistency checks on the output dependency syntax tree set; including:

[0085] Construct a whitelist containing legal dependency relationship tags (such as root, nsubj, obj, advmod, and other common dependency relationships), and perform comprehensive statistics and checks on the tags in each dependency syntax tree to ensure that each dependency relationship conforms to the predetermined standard format in order to maintain data consistency and accuracy.

[0086] Step 6.2: Building upon automatic structure verification, a manual review mechanism is introduced, where linguistic annotation specialists review the dependency trees that have passed machine screening for grammatical correctness, semantic rationality, and label consistency; including:

[0087] Sentences containing illegal labels or non-compliant dependency relations are removed, and these deleted sentences and their corresponding illegal labels are recorded and saved in a log file. The purpose of this operation is to ensure the stability of the dataset quality, avoid the negative impact of spurious data on subsequent training and model optimization, and provide detailed traceability information for subsequent error analysis and correction.

[0088] The output results are further evaluated and screened manually by experts and students in relevant fields to select high-quality data.

[0089] Step 7: Merge the selected high-quality dependency parsing trees with the original UD training set to construct a high-quality synthetic Vietnamese dependency parsing tree, and use it to retrain the BiLSTM+MLP+BiAffine Parser parser model.

[0090] Furthermore, Step 7 includes:

[0091] Step 7.1: Merge all the selected high-quality dependency parsing trees with the original UD training data; including:

[0092] The cleaned and filtered pseudo-data was integrated with the original UD Vietnamese training set to form an enhanced training dataset. During this process, all data was ensured to conform to the format requirements of dependency parsing. Invalid data was removed, and annotation errors were corrected to improve the quality and diversity of the dataset, providing richer samples for model training. Furthermore, the enhanced training dataset can introduce more syntactic variations, improving the model's adaptability and generalization ability to complex Vietnamese sentence structures.

[0093] Step 7.2: Perform consistency checks on the merged dataset, including root checks, label validity checks, and duplicate sequence number checks; construct a high-quality synthetic Vietnamese dependency syntax tree;

[0094] Step 7.3: Train the parser model using the augmented corpus as input. This includes:

[0095] The BiAffine Parser model was trained using the aforementioned mixed dataset. Specifically, a BiLSTM+MLP+BiAffine architecture was employed to predict dependency arcs and label annotations. During training, BiLSTM, acting as a bidirectional LSTM model, was used to extract fine-grained information from the vocabulary; the MLP layer further fused features to optimize the prediction accuracy of dependency relations; and the BiAffine module jointly modeled dependency arcs and labels using a bilinear mapping approach, thereby improving the accuracy and robustness of parsing. After training, the model outputs a high-performance Vietnamese dependency parsing model, providing accurate structural predictions for subsequent syntactic analysis tasks.

[0096] The present invention also includes:

[0097] Step 8: Evaluate the enhanced model using a standard Vietnamese test set and compare the changes in LAS (Labeled Dependency Accuracy) and UAS (Unlabeled Dependency Accuracy) metrics.

[0098] Furthermore, the specific steps of Step 8 are as follows:

[0099] Step 8.1: Apply the enhanced new model to a standard Vietnamese test set for comprehensive evaluation. The test set is used to test the model's performance in real-world language environments, including the accuracy of syntactic parsing and the model's generalization ability to new samples.

[0100] Step 8.2: Quantitatively evaluate the model's performance by calculating commonly used dependency syntax evaluation metrics—LAS (Labeled Attachment Score) and UAS (Unlabeled Attachment Score). Specific metrics are as follows:

[0101] LAS (Labelled Attachment Score): Measures the correctness of the model's predicted dependency arcs and their corresponding labels. The calculation formula is as follows:

[0102]

[0103] UAS (Unlabeled Attachment Score): Measures the accuracy of the dependency arcs predicted by the model, without considering label correctness. The calculation formula is as follows:

[0104]

[0105] Step 8.3: Compare the performance of the new model with the original model to analyze the effect of data augmentation on model performance. By comparing evaluation metrics such as LAS and UAS, the specific effects of data augmentation are quantified, further verifying the effectiveness of the method of collaborative data augmentation using a large language model in Vietnamese dependency syntax construction. Experimental results show that the strategy of combining collaborative parsing using a large language model with an expert scoring feedback mechanism significantly improves the accuracy and robustness of the model in Vietnamese dependency syntax analysis tasks.

[0106] To validate the effectiveness of our research method, we selected the widely used and classic dataset, Universal Dependencies (UD), as our experimental foundation. We trained a preliminary model using this dataset and evaluated it as a benchmark model. Based on this, we used the benchmark model to analyze unlabeled Vietnamese text, generating preliminary Vietnamese syntax trees. During this process, we developed a set of selection rules to choose high-quality results from the generated syntax trees, using these as the basis for constructing a high-quality Vietnamese pseudo-corpus treebase.

[0107] The purpose of constructing this pseudo-corpus treebank is to enhance the model's training effect by training it with the original experimental data, thereby improving the efficiency, generalization ability, and processing accuracy of dependency parsing for Vietnamese. This method effectively alleviates the challenge posed by the scarcity of Vietnamese monolingual corpora and significantly improves the overall performance of the Vietnamese dependency parsing model with limited data support. This approach not only optimizes data utilization efficiency but also provides a novel solution for dependency parsing of low-resource languages.

[0108] The experimental results of this invention are shown in Table 1. The results demonstrate that the proposed method outperforms the benchmark model on multiple evaluation metrics. This improvement validates the effectiveness of joint optimization training, indicating that the accuracy and robustness of Vietnamese dependency parsing are significantly enhanced through these techniques.

[0109] By comprehensively applying these strategies, significant breakthroughs have been achieved in Vietnamese natural language processing tasks. This research not only provides innovative ideas for constructing Vietnamese dependency syntax treebanks but also strongly supports the further development of multilingual technologies (such as Burmese, Thai, Lao, and other Southeast Asian languages) and Sino-Vietnamese language and cultural exchange. Through cross-linguistic collaborative optimization, this method provides a solid foundation for the further application of Vietnamese in the field of natural language processing, promoting the advancement of multilingual technologies and deepening their practical applications.

[0110] Table 1 Comparison of the original treebank and the hybrid syntactic treebank constructed by this method

[0111]

[0112] The specific embodiments of the present invention have been described in detail with reference to the accompanying drawings. However, it should be understood that the present invention is not limited to the embodiments described above. Those skilled in the art, having fully understood the principles and technical framework of the present invention, can make appropriate changes, adjustments, or extensions to the present invention according to specific application needs, provided that such changes and improvements do not deviate from the basic spirit and core ideas of the present invention.

Claims

1. A Vietnamese dependency syntax data augmentation method based on large language model collaboration, characterized in that: The method includes: Step 1: Obtain Vietnamese labeled dependency syntax data from the general dataset UD, and collect unlabeled Vietnamese monolingual corpus, and perform relevant preprocessing operations on both. Step 2: Construct a dependency parsing system environment, select a mainstream large language model as the multi-label model, parse the preprocessed unlabeled data, and output a five-column CoNLL-U format dependency parsing tree; Step 3: Construct an expert large language model, and use the expert large language model to score and provide feedback on the output of each labeled model; Step 4: The expert large language model performs multiple rounds of optimization on each labeled model based on expert feedback, and scores and provides feedback on the output of each round, sets an early stopping mechanism, and selects the dependency syntax tree with the highest score as the optimal result; Step 5: For each sentence, select the optimal dependency syntax tree from the parsing results of each model, and convert the optimal dependency tree into a standard 10-column CoNLL-U format for training; Step 6: Perform structural consistency checks and label filtering on the output dependency syntax tree to remove sentences with abnormal structures or illegal labels. At the same time, combine manual evaluation to select high-quality data. Step 7: Merge the selected high-quality dependency parsing trees with the original UD training set to construct a high-quality synthetic Vietnamese dependency parsing tree, and use it to retrain the BiLSTM+MLP+BiAffine Parser parser model.

2. The Vietnamese dependency syntax data augmentation method based on large language model collaboration according to claim 1, characterized in that: Step 1 includes: Step 1.1: Obtain the Vietnamese training set from the general dataset UD and collect unlabeled Vietnamese monolingual corpus; Step 1.2: Use the Underthesea tool to segment the Vietnamese sentences; Step 1.3: Preprocess the data, including filtering out duplicates and sentences with abnormal punctuation.

3. The Vietnamese dependency syntax data augmentation method based on large language model collaboration according to claim 1, characterized in that: Step 2 includes: Step 2.1: Call multiple mainstream large language models together as a multi-label model, including Qwen, ChatGLM and DeepSeek, to perform dependency parsing on the preprocessed Vietnamese sentences; Step 2.2: Generate a preliminary CoNLL-U format dependency syntax tree and output the independent parsing results for each model.

4. The Vietnamese dependency syntax data augmentation method based on large language model collaboration according to claim 1, characterized in that: Step 3 includes: Step 3.1: Select GPT-4o-mini as the unified scoring expert big language model, conduct quality assessment and feedback on the output of each labeled model, and provide guidance for the subsequent screening of analysis results and model iteration; Step 3.2: The expert big language model outputs scores and natural language feedback information for syntactic structure, dependency relation accuracy, and semantic consistency. The model parses sentences according to a unified prompt to generate dependency syntax trees.

5. The Vietnamese dependency syntax data augmentation method based on large language model collaboration according to claim 1, characterized in that: Step 4 includes: Step 4.1: Each labeled model undergoes multiple rounds of parsing, with the expert large language model GPT-4o-mini scoring the output of each round and providing structural improvement suggestions; Step 4.2: Record the scoring results in each iteration, find the dependency syntax tree with the highest score from all output results as the optimal result, and stop early and save the tree structure if the score is higher than the set threshold.

6. The Vietnamese dependency syntax data augmentation method based on large language model collaboration according to claim 1, characterized in that: Step 5 includes: Step 5.1: Convert the filtered best results into the standard CoLL-U ten-column format; Step 5.2: Organize the preliminary results into the manual screening process, manually check the rationality of the grammatical structure and the standardization of tags, and remove low-quality output; Step 5.3: Retain sentences with complete annotations and correct structure for use in subsequent training.

7. The Vietnamese dependency syntax data augmentation method based on large language model collaboration according to claim 1, characterized in that: Step 6 includes: Step 6.1: Perform multiple consistency checks on the output dependency syntax tree set; Step 6.2: Based on automatic structure verification, a manual review mechanism is introduced, in which linguistic annotation personnel review the dependency trees that have passed the machine screening in terms of grammatical correctness, semantic rationality, and label consistency.

8. The Vietnamese dependency syntax data augmentation method based on large language model collaboration according to claim 1, characterized in that: Step 7 includes: Step 7.1: Merge all the selected high-quality dependency parsing trees with the original UD training data; Step 7.2: Perform consistency checks on the merged dataset, including root checks, label validity checks, and duplicate sequence number checks; construct a high-quality synthetic Vietnamese dependency syntax tree; Step 7.3: Use the augmented corpus as input to train the parser model of BiLSTM+MLP+BiAffine Parser.

Images