Multi-round accurate translation method and system based on large model in self-training financial supervision field

Abstract

The invention provides a multi-round accurate translation method and system based on a large model in the field of self-training financial supervision. The method comprises the following steps: S1, constructing a training data set of the large model in the field of financial supervision; s2, constructing a large model in the financial supervision field, and performing fine adjustment on the general large model; s3, performing multi-round translation on a plurality of to-be-translated texts by utilizing the fine-tuned financial supervision field large model, and constructing a context information base; s4, translating a target text by using the fine-tuned large model in the financial supervision field, and adjusting and optimizing a translation result by using a context information base; the large model in the financial supervision field is a fine-tuning translation model which is built based on a Gamma 227b large language model and aims at the financial supervision field. The translation process is highly automatic, a user only needs to input a text or a document needing to be translated, a high-quality translation result can be obtained, and complex operation is not needed.

Description

Technical Field

[0001] The present invention relates to the field of large model technology, and in particular to a multi-round precise translation method and system based on a self-training large model in the field of financial supervision. Background Art

[0002] Translation in the field of financial supervision requires high professionalism, rigor and accuracy, and its texts usually contain a large number of professional terms, abbreviations, laws and regulations, and complex sentence structures. Traditional translation methods mainly rely on ordinary machine translation + manual translation, which has problems such as low efficiency, lack of professionalism, and difficulty in ensuring consistency, making it difficult to meet the growing needs of cross-border financial supervision cooperation and information exchange.

[0003] In the existing technology, both ordinary translation and AI large model translation have their own limitations, including:

[0004] Although ordinary machine translation systems based on statistical models or rules can handle some simple translation tasks, they have the following problems when dealing with professional texts in the field of financial supervision:

[0005] Inaccurate translation of terms: Lack of understanding of professional terms in the field of financial regulation can easily lead to mistranslation or omission;

[0006] Semantic understanding bias: It is difficult to accurately understand the deep meaning of financial regulatory texts with strong logic and complex semantics, resulting in the translation results not being consistent with the original text;

[0007] Lack of contextual information: Translation is usually done sentence by sentence, lacking understanding of contextual information, making it difficult to ensure translation coherence and consistency.

[0008] AI big model translation using deep learning technology can learn a large amount of text data and solve the problems of ordinary machine translation systems to a certain extent; however, general AI big model translation still has the following limitations in the field of financial supervision:

[0009] Lack of domain expertise: Training data usually comes from the Internet and lacks professional texts in the field of financial regulation, resulting in the model's insufficient understanding of professional terminology, regulations, and writing styles in the field of financial regulation;

[0010] Poor consistency in multiple rounds of translation: It is difficult to maintain consistency of information in multiple rounds of translation, and inconsistencies or inconsistent translation styles are prone to occur. Summary of the invention

[0011] In view of the defects in the prior art, the purpose of the present invention is to provide a multi-round precise translation method and system based on a self-training large model in the field of financial supervision.

[0012] According to the present invention, a multi-round accurate translation method based on a self-training large model in the financial regulatory field is provided, comprising:

[0013] Step S1: Construct a large model training dataset in the field of financial supervision;

[0014] Step S2: construct a large model in the field of financial supervision and fine-tune the large model in the field of financial supervision;

[0015] Step S3: Use the fine-tuned large model in the financial regulatory field to perform multiple rounds of translation on multiple texts to be translated and build a context information database;

[0016] Step S4: Use the fine-tuned large model in the financial regulatory field to translate the target text, and use the context information database to adjust and optimize the translation results;

[0017] The financial regulation domain large model is a fine-tuned translation model for the financial regulation domain built based on the gemma2 27b large language model.

[0018] Preferably, the step S1 comprises:

[0019] Step S1.1: Construct a bilingual corpus and pseudo-parallel corpus in the field of financial regulation, including laws and regulations, regulatory policies, industry reports, and academic papers;

[0020] Step S1.2: preprocessing the bilingual corpus and pseudo-parallel corpus, including: removing duplicate data, correcting errors and unifying the format;

[0021] Step S1.3: Perform word segmentation, part-of-speech tagging, and named entity recognition on the preprocessed bilingual corpus and pseudo-parallel corpus to construct a structured training data set.

[0022] Preferably, step S2 comprises:

[0023] Step S2.1: Use the training data set prepared in step S1 to fine-tune the general large model to obtain a fine-tuned large model in the financial regulatory field;

[0024] Step S2.2: Use self-training technology to fine-tune the fine-tuned large model of the financial regulatory field to obtain a fine-tuned large model of the financial regulatory field.

[0025] Preferably, the step S2.2 comprises:

[0026] Step S2.2.1: Obtain bilingual text data in the field of financial regulation;

[0027] Step S2.2.2: Use the fine-tuned large model in the financial regulation field to translate the source language text in the bilingual text data in the financial regulation field to generate the target language text;

[0028] Step S2.2.3: Evaluate the translation quality of the target language text, and screen high-quality target language texts that meet preset requirements, and form new bilingual text data based on the screened high-quality target language texts and corresponding source language texts as pseudo bilingual text data;

[0029] Step S2.2.4: Use pseudo-bilingual text data to fine-tune the fine-tuned large model in the financial regulatory field to obtain a fine-tuned large model in the financial regulatory field.

[0030] Preferably, the method also includes: evaluating the fine-tuned big model in the financial regulatory field using the BLEU indicator and the ROUGE indicator; when the evaluation shows that the current fine-tuned big model in the financial regulatory field does not meet the preset requirements, re-fine-tuning the big model in the financial regulatory field.

[0031] Preferably, step S3 comprises:

[0032] Step S3.1: Use the fine-tuned large model in the financial regulatory field to translate and obtain historical translation records;

[0033] Step S3.2: Obtain the meaning of specific words in the target text input by the user, and translate them using the fine-tuned large model in the financial regulatory field to obtain translation information of the specific words;

[0034] Step S3.3: Construct a context information database based on historical translation records and specific vocabulary translation information in the target text.

[0035] Preferably, the step S4 adopts: in the process of translating the target text using the fine-tuned large model in the financial regulatory field, the attention mechanism is used to adjust and optimize the translation result in combination with the context information library.

[0036] Preferably, the method further comprises:

[0037] Use document parsing tools to analyze the format and structure of the target text;

[0038] Extract the target text content and split it into multiple translation units;

[0039] Translate multiple translation units separately using the fine-tuned large model in the financial regulatory field;

[0040] The translated content is reassembled according to the format and structure of the analyzed target text to generate a translated document while retaining the format of the original document.

[0041] According to the present invention, a multi-round accurate translation system based on a self-training large model in the financial regulatory field is provided, comprising:

[0042] Module M1: Constructing a large model training dataset in the field of financial regulation;

[0043] Module M2: Build a large model in the field of financial supervision and fine-tune the large model in the field of financial supervision;

[0044] Module M3: Use the fine-tuned large model in the financial regulatory field to perform multiple rounds of translation on multiple texts to build a context information database;

[0045] Module M4: Use the fine-tuned financial regulatory domain model to translate the target text, and use the context information database to adjust and optimize the translation results;

[0046] The financial regulation domain large model is a fine-tuned translation model for the financial regulation domain built based on the gemma2 27b large language model.

[0047] Preferably, the module M1 comprises:

[0048] Module M1.1: Construct a bilingual corpus and pseudo-parallel corpus in the field of financial regulation, including laws and regulations, regulatory policies, industry reports and academic papers;

[0049] Module M1.2: Preprocessing of bilingual corpora and pseudo-parallel corpora, including: removing duplicate data, correcting errors and unifying formats;

[0050] Module M1.3: Perform word segmentation, part-of-speech tagging, and named entity recognition on the preprocessed bilingual corpus and pseudo-parallel corpus to construct a structured training dataset;

[0051] The module M2 comprises:

[0052] Module M2.1: Use the training data set of module M1 to fine-tune the general large model to obtain a fine-tuned large model in the financial regulatory field;

[0053] Module M2.2: Use self-training technology to fine-tune the fine-tuned large model in the financial regulatory field to obtain a fine-tuned large model in the financial regulatory field;

[0054] The module M2.2 includes:

[0055] Module M2.2.1: Obtain bilingual text data in the field of financial regulation;

[0056] Module M2.2.2: Use the fine-tuned large model in the financial regulation field to translate the source language text in the bilingual text data in the financial regulation field to generate the target language text;

[0057] Module M2.2.3: Evaluate the translation quality of the target language text and screen high-quality target language texts that meet the preset requirements. Based on the screened high-quality target language texts and the corresponding source language texts, form new bilingual text data and use them as pseudo-bilingual text data;

[0058] Module M2.2.4: Use pseudo-bilingual text data to fine-tune the fine-tuned financial regulatory model to obtain a fine-tuned financial regulatory model;

[0059] The module M3 comprises:

[0060] Module M3.1: Use the fine-tuned large model in the financial regulatory field to translate and obtain historical translation records;

[0061] Module M3.2: Obtain the meaning of specific words in the target text input by the user, and translate them using the fine-tuned large model in the financial regulatory field to obtain translation information of specific words;

[0062] Module M3.3: Build a contextual information base based on historical translation records and specific vocabulary translation information in the target text;

[0063] The module M4 adopts: in the process of translating the target text using the fine-tuned large model in the financial regulatory field, the attention mechanism is used to adjust and optimize the translation results in combination with the context information library.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1. The present invention constructs a large model in the field of financial supervision and uses professional data sets in the field of financial supervision for fine-tuning, so that it has a stronger ability to understand domain expertise and can more accurately translate professional terms, regulations and writing styles in the field of financial supervision, effectively solving the problem of insufficient translation accuracy of ordinary machine translation and general AI large model translation in the field of financial supervision in the prior art.

[0066] 2. The present invention effectively utilizes context information, such as the content, keywords, entities, etc. of previous rounds of conversations, in the process of multiple rounds of translation, which can better understand the semantics of the current conversation, improve the accuracy and consistency of translation, and solve the problem of poor consistency of multiple rounds of translation in the prior art;

[0067] 3. The present invention has a high degree of automation, can significantly improve translation efficiency, reduce manual translation costs, and solve the problem of low manual translation efficiency in the prior art;

[0068] 4. The present invention has powerful parallel processing capabilities and can process multiple translation tasks at the same time, further improving translation efficiency and meeting large-scale translation needs;

[0069] 5. In the process of multiple rounds of translation, the present invention can ensure the consistency of translation between different rounds of dialogues by recording and analyzing context information, avoid inconsistencies or inconsistent translation styles, and solve the problem of poor consistency of multiple rounds of translation in the prior art;

[0070] 6. The present invention supports the translation of documents in PDF, DOC and other formats, and can retain the format and structure of the original document, such as paragraphs, titles, tables, pictures, etc., thereby improving the readability and practicality of the translation results and solving the problem of format loss in document translation in the prior art; the translation results retain the original document format, which is convenient for users to perform subsequent editing and use, such as directly modifying or annotating in the translated document without reformatting or adjusting the format. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0072] Figure 1 Flowchart of the multi-round precise translation method based on self-training large model in the field of financial supervision. DETAILED DESCRIPTION

[0073] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0074] Example 1

[0075] According to the invention, a multi-round accurate translation method based on a self-training large model in the financial regulatory field is provided. Figure 1 As shown, the following steps are included:

[0076] Step 1: Build a large model training dataset in the field of financial regulation;

[0077] Specifically, the step 1 includes:

[0078] Step 1.1: Collect bilingual corpora in the field of financial regulation, including laws and regulations, regulatory policies, industry reports, academic papers, etc. Step 1.2: Clean and preprocess the corpus, for example, remove duplicate data, correct errors, unify formats, etc.

[0079] Step 1.3: Perform word segmentation, part-of-speech tagging, named entity recognition and other processing on the corpus to build a structured training dataset.

[0080] Step 1.4 manually proofreads and reviews the training data set to ensure the accuracy and reliability of the data.

[0081] In the present embodiment, the purpose of constructing a structured training data set is to improve the learning efficiency of the model, enhance the generalization ability of the model, and enhance the translation quality of a specific field; specifically, structured data organizes different information into clear features (fields), such as titles, publishing agencies, keywords, etc. This makes it easier for the model to understand the contextual information of the text, thereby learning a more accurate translation model. At the same time, structured data is usually preprocessed and cleaned, such as removing irrelevant information, unifying the format, etc. This reduces the interference of noise data on model training, allowing the model to focus on learning translation knowledge. Structured data facilitates feature engineering, such as extracting keyword features, theme features, etc. More effective feature representations can be constructed through feature engineering, thereby enhancing the generalization ability of the model, enabling it to better handle texts that have not been encountered. Through term annotation, the model can be helped to learn professional vocabulary in a specific field and its corresponding translation, thereby improving the translation quality in a specific field.

[0082] Step 2: Self-training large models in the field of financial regulation;

[0083] Step 2.1: Select gemma2 27b as the pre-trained base model.

[0084] Step 2.2: Use the constructed financial regulatory field big model training dataset to fine-tune the basic big model.

[0085] Step 2.3: Use self-training techniques, such as Translation Language Model (TLM), to further fine-tune the fine-tuned model to give it stronger domain expertise understanding capabilities.

[0086] Step 2.4: Evaluate the self-trained model, for example, using indicators such as BLEU and ROUGE to ensure that the translation quality of the model meets the expected requirements.

[0087] In this embodiment, gemma2 27b is selected as the basic large model for pre-training; wherein, gemma2 27b is an open source commercial large language model with 27 billion parameters, which has strong text generation and understanding capabilities. Gemma 2 27B has excellent performance in multiple natural language processing tasks; at the same time, Gemma 2 27B is trained on text data in the legal and regulatory fields, which gives it an advantage in processing professional texts such as regulatory documents. It can better understand the professional terms, expressions and logical relationships in the legal and regulatory fields, thereby providing more accurate and fluent translations; finally, Gemma 2 27B is pre-trained on large-scale text data, which gives it strong knowledge transfer capabilities. Even when training data is limited, it can quickly learn new translation knowledge and achieve good translation results.

[0088] Specifically, the step S2.3 includes:

[0089] Step S2.3.1: Prepare domain-specific data and collect a large amount of domain-specific parallel corpora, such as bilingual text data in the field of financial regulation. These data can come from public datasets, web crawlers, manual translation, etc.

[0090] Step S2.3.2: Use the fine-tuned model for translation. Use the previously fine-tuned translation model to translate the source language text in the domain-specific parallel corpus to generate the target language text.

[0091] Step S2.3.3: Screen high-quality translation results, use some indicators, such as BLEU scores to evaluate the quality of the translation results, and screen out high-quality translation results. The screening criteria can be adjusted according to actual needs, for example, translation results with BLEU scores higher than a certain threshold can be selected.

[0092] Step S2.3.4: Construct a manually corrected pseudo-parallel corpus. The screened high-quality translation results and the corresponding source language texts form a new parallel corpus, called a pseudo-parallel corpus. If the translation results are adjusted, the adjusted corpus becomes a manually corrected pseudo-parallel corpus.

[0093] Step S2.3.5: Self-training using pseudo-parallel corpus: Further train the fine-tuned model using pseudo-parallel corpus, for example, using maximum likelihood estimation or other training methods. Steps S2.3.2 to S2.3.5 can be triggered repeatedly to perform multiple rounds of self-training to continuously improve the model's ability to understand domain expertise.

[0094] Specifically, the step S2.4 includes:

[0095] BLEU is a commonly used machine translation evaluation indicator that evaluates translation quality by calculating the n-gram overlap rate between the machine translation result and the reference translation.

[0096] Calculation formula:

[0097] BLEU=BP*exp(sum(w_n*log p_n))

[0098] Among them, BP: length penalty factor, used to penalize the case where the machine translation result is too short. w_n: n-gram weight, usually set to a uniform weight, such as w_1 = w_2 = w_3 = w_4 = 0.25. p_n: n-gram precision, that is, the proportion of n-grams appearing in the machine translation result in the reference translation. Exp: exponential function.

[0099] ROUGE is a set of metrics for evaluating automatic summarization and machine translation, which focuses on the recall rate between machine translation results and reference translations.

[0100] ROUGE-N=sum(min(count_match(n-gram),count_ref(n-gram))) / sum(count_ref(n-gram))

[0101] Among them, count_match(n-gram): the number of n-grams that appear in the machine translation result and the reference translation; count_ref(n-gram): the number of n-grams that appear in the reference translation.

[0102] When the evaluation does not meet the preset requirements, the model is retrained;

[0103] Specifically, it includes: increasing training data: collecting more domain-specific parallel corpora and using these data to further train the model.

[0104] Adjust model parameters: Try adjusting the model’s hyperparameters, such as learning rate, batch size, hidden layer size, etc., to find a better model configuration.

[0105] Improve training methods: Try using different training methods, such as curriculum learning, adversarial training, etc., to improve the training effect of the model.

[0106] Use a more powerful model: Try using a pre-trained language model with larger parameter size and better performance as the base model.

[0107] Combine with other technologies: Try to combine other technologies to improve the performance of the model, such as knowledge distillation, multi-model fusion, etc.

[0108] Step 3: Multi-round translation context information processing

[0109] Step 3.1: When performing multiple rounds of translation, record the original text and translation of each round of dialogue and build a context information database.

[0110] Among them, the context information base includes historical translation records and session information;

[0111] Historical translation records: refers to previously translated text content, including original text and translation. These historical records can help the model learn translation patterns, professional terms, expression habits, etc., thereby improving the accuracy and consistency of subsequent translations.

[0112] Conversation information: refers to the conversation content before the current translation task, such as questions raised by the user and answers given by the model. This conversation information can help the model understand the user's intentions and needs, thereby providing a translation that better meets the user's expectations.

[0113] This information can help the model better understand user needs and translation scenarios, thereby providing better translation services. Through multiple rounds of translation context information, the model can better understand the background knowledge and contextual relationships of the text, thereby improving the accuracy and fluency of the translation.

[0114] For example, the common translation of YiBiao Tong in the "One-meter Communication for Banking Regulatory Data Collection Interface Standard" is "onemeter communication" or "YiBiao Tong", but its actual meaning is "data sharing platform". When you explain the meaning of YiBiao Tong in the translation context, the subsequent translation will be based on the content related to the data sharing platform.

[0115] Step 3.2: When translating the next round of dialogue, relevant information in the context information library is used as input, such as the content, keywords, entities, etc. of the previous rounds of dialogue.

[0116] Step 3.3: Use attention mechanisms or other techniques to enable the model to effectively utilize contextual information to improve translation accuracy and consistency.

[0117] In this embodiment, the attention mechanism can dynamically adjust the translation context according to the current translation target, focusing more on the parts of the source text related to the current target and ignoring irrelevant parts.

[0118] Solve the problem of long-distance dependency: The attention mechanism can capture long-distance dependencies in the source text, for example, a word may have a semantic association with another word that is far away from it. This is very important for translating long and complex sentences.

[0119] Attentional focus mechanism and paraphrase tendency of contextual information:

[0120] The attention mechanism not only focuses on the current translation target, but also considers the contextual information of the source language, such as the theme of the text. In the process of English-Chinese translation, the abbreviations in the regulatory information field are different from those in other fields. Through the theme attention focus mechanism, the translation is more inclined to the abbreviations in the regulatory field.

[0121] By integrating contextual information, the model can understand the meaning of the source language more accurately and select more appropriate expressions for translation, making the translation results more inclined to free translation.

[0122] Step 4: Document translation format preservation

[0123] Step 4.1: The user inputs the document to be translated (eg, PDF, DOC format).

[0124] Step 4.2: Systematically analyze the format and structure of the document, such as paragraphs, headings, tables, pictures, etc.

[0125] Step 4.3: The system extracts the document content and segments it into multiple translation units, such as sentences or paragraphs.

[0126] Step 4.4: The system inputs each translation unit into the self-trained financial regulatory domain model for translation.

[0127] Step 4.5: The system adjusts and optimizes the translation results according to the context information provided by the context information processing module.

[0128] Step 4.6: The system reassembles the translated content according to the format and structure of the original document to generate a translated document, while retaining the format of the original document.

[0129] Multiple rounds of precise translation

[0130] The user inputs the text or document to be translated; the system selects the corresponding translation process according to the input type (text or document); the system outputs the final translation result to the user.

[0131] The present invention also provides a multi-round precise translation system based on a self-training large model in the field of financial regulation. The multi-round precise translation system based on a self-training large model in the field of financial regulation can be realized by executing the process steps of the multi-round precise translation method based on the self-training large model in the field of financial regulation, that is, those skilled in the art can understand the multi-round precise translation method based on the self-training large model in the field of financial regulation as a preferred implementation of the multi-round precise translation system based on the self-training large model in the field of financial regulation.

[0132] In summary, the present invention provides a multi-round precise translation method and system based on a self-training large model in the financial regulatory field, which can effectively improve the accuracy, consistency and efficiency of translation in the financial regulatory field, and support the retention of document translation format. It has important practical value and promotion significance, and can better meet the needs of cross-border financial regulatory cooperation and information exchange.

[0133] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, the system and its various devices, modules, and units provided by the present invention can be completely realized in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps to realize the same functions. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0134] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A multi-round accurate translation method based on a self-training large model in the financial regulatory field, characterized in that: include: Step S1: Construct a large model training dataset in the field of financial supervision; Step S2: construct a large model in the field of financial supervision and fine-tune the large model in the field of financial supervision; Step S3: Use the fine-tuned large model in the financial regulatory field to perform multiple rounds of translation on multiple texts to be translated and build a context information database; Step S4: Use the fine-tuned large model in the financial regulatory field to translate the target text, and use the context information database to adjust and optimize the translation results; The financial regulation domain large model is a fine-tuned translation model for the financial regulation domain built based on the gemma2 27b large language model.

2. The multi-round accurate translation method based on the self-training financial regulatory domain large model according to claim 1 is characterized in that: The step S1 comprises: Step S1.1: Construct a bilingual corpus and pseudo-parallel corpus in the field of financial regulation, including laws and regulations, regulatory policies, industry reports, and academic papers; Step S1.2: preprocessing the bilingual corpus and pseudo-parallel corpus, including: removing duplicate data, correcting errors and unifying the format; Step S1.3: Perform word segmentation, part-of-speech tagging, and named entity recognition on the preprocessed bilingual corpus and pseudo-parallel corpus to construct a structured training data set.

3. The multi-round accurate translation method based on the self-training financial regulatory domain large model according to claim 1 is characterized in that: The step S2 comprises: Step S2.1: Use the training data set in step S1 to fine-tune the general large model to obtain a fine-tuned large model in the financial regulatory field; Step S2.2: Use self-training technology to fine-tune the fine-tuned large model of the financial regulatory field to obtain a fine-tuned large model of the financial regulatory field.

4. The multi-round accurate translation method based on the self-training financial regulatory domain large model according to claim 3 is characterized in that: The step S2.2 comprises: Step S2.2.1: Obtain bilingual text data in the field of financial regulation; Step S2.2.2: Use the fine-tuned large model in the financial regulation field to translate the source language text in the bilingual text data in the financial regulation field to generate the target language text; Step S2.2.3: Evaluate the translation quality of the target language text, and screen high-quality target language texts that meet preset requirements, and form new bilingual text data based on the screened high-quality target language texts and corresponding source language texts as pseudo bilingual text data; Step S2.2.4: Use pseudo-bilingual text data to fine-tune the fine-tuned large model in the financial regulatory field to obtain a fine-tuned large model in the financial regulatory field.

5. The multi-round accurate translation method based on the self-training financial regulatory domain large model according to claim 1 is characterized in that: The method also includes: evaluating the fine-tuned large model in the financial regulatory field using the BLEU indicator and the ROUGE indicator; when the current fine-tuned large model in the financial regulatory field does not meet the preset requirements, re-fine-tuning the large model in the financial regulatory field.

6. The multi-round accurate translation method based on the self-training financial regulatory domain large model according to claim 1 is characterized in that: The step S3 comprises: Step S3.1: Use the fine-tuned large model in the financial regulatory field to translate and obtain historical translation records; Step S3.2: Obtain the meaning of specific words in the target text input by the user, and translate them using the fine-tuned large model in the financial regulatory field to obtain translation information of the specific words; Step S3.3: Construct a context information database based on historical translation records and specific vocabulary translation information in the target text.

7. The multi-round accurate translation method based on self-training financial regulatory domain large model according to claim 1 is characterized in that: The step S4 adopts: in the process of translating the target text using the fine-tuned large model in the financial regulatory field, the attention mechanism is used to adjust and optimize the translation results in combination with the context information library.

8. The multi-round accurate translation method based on the self-training financial regulatory domain large model according to claim 1 is characterized in that: The method further comprises: Use document parsing tools to analyze the format and structure of the target text; Extract the target text content and split it into multiple translation units; Translate multiple translation units separately using the fine-tuned large model in the financial regulatory field; The translated content is reassembled according to the format and structure of the analyzed target text to generate a translated document while retaining the format of the original document.

9. A multi-round accurate translation system based on a self-training large model in the field of financial supervision, characterized in that: include: Module M1: Constructing a large model training dataset in the field of financial regulation; Module M2: Build a large model in the field of financial supervision and fine-tune the large model in the field of financial supervision; Module M3: Use the fine-tuned large model in the financial regulatory field to perform multiple rounds of translation on multiple texts to build a context information database; Module M4: Use the fine-tuned financial regulatory domain model to translate the target text, and use the context information database to adjust and optimize the translation results; The financial regulation domain large model is a fine-tuned translation model for the financial regulation domain built based on the gemma2 27b large language model.

10. The multi-round accurate translation system based on the self-training financial regulatory domain large model according to claim 9 is characterized in that: The module M1 comprises: Module M1.1: Construct a bilingual corpus and pseudo-parallel corpus in the field of financial regulation, including laws and regulations, regulatory policies, industry reports and academic papers; Module M1.2: Preprocessing of bilingual corpora and pseudo-parallel corpora, including: removing duplicate data, correcting errors and unifying formats; Module M1.3: Perform word segmentation, part-of-speech tagging, and named entity recognition on the preprocessed bilingual corpus and pseudo-parallel corpus to construct a structured training dataset; The module M2 comprises: Module M2.1: Use the training data set of module M1 to fine-tune the general large model to obtain a fine-tuned large model in the financial regulatory field; Module M2.2: Use self-training technology to fine-tune the fine-tuned large model in the financial regulatory field to obtain a fine-tuned large model in the financial regulatory field; The module M2.2 includes: Module M2.2.1: Obtain bilingual text data in the field of financial regulation; Module M2.2.2: Use the fine-tuned large model in the field of financial regulation to translate the source language text in the bilingual text data in the field of financial regulation to generate the target language text; Module M2.2.3: Evaluate the translation quality of the target language text and screen high-quality target language texts that meet the preset requirements. Based on the screened high-quality target language texts and the corresponding source language texts, form new bilingual text data and use them as pseudo-bilingual text data; Module M2.2.4: Use pseudo-bilingual text data to fine-tune the fine-tuned financial regulatory model to obtain a fine-tuned financial regulatory model; The module M3 comprises: Module M3.1: Use the fine-tuned large model in the financial regulatory field to translate and obtain historical translation records; Module M3.2: Obtain the meaning of specific words in the target text input by the user, and translate them using the fine-tuned large model in the financial regulatory field to obtain translation information of specific words; Module M3.3: Build a contextual information base based on historical translation records and specific vocabulary translation information in the target text; The module M4 adopts: in the process of translating the target text using the fine-tuned large model in the financial regulatory field, the attention mechanism is used to adjust and optimize the translation results in combination with the context information library.

Images