Translation method and system based on multi-agent cooperation and multi-stage optimization

Abstract

The application discloses a translation method and system based on multi-agent cooperation and multi-stage optimization, and belongs to the technical field of computers. The method obtains complete text and pre-extracts terms through OCR recognition and document structure analysis; generates a translation strategy through input preprocessing and translation task planning; constructs a semantic structure using semantic analysis, and enhances professional terms and domain knowledge in combination with knowledge graph retrieval; dynamically generates multiple translation agents to perform translation in parallel, and generates candidate translations; unifies term expression through term consistency processing, and selects the optimal translation by using multi-dimensional quality evaluation; and finally generates structured translations and updates the knowledge graph. The application significantly improves translation accuracy, term consistency and translation fluency through multi-agent cooperation, knowledge graph enhancement and multi-stage optimization mechanism, and is especially suitable for professional document and publishing level translation scenarios.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and specifically to a translation method and system based on multi-agent cooperation and multi-stage optimization. Background Technology

[0002] With the development of artificial intelligence and natural language processing technologies, machine translation technology has been widely used in cross-language information exchange. Existing machine translation systems typically employ end-to-end neural network models, achieving text translation through training on large-scale corpora. However, in practical applications, existing translation systems still have several shortcomings.

[0003] First, traditional machine translation models typically use sentences as the basic translation unit, making it difficult to maintain overall contextual consistency when processing long documents or complex texts, easily leading to semantic deviations or contextual breaks. Second, in professional domain translations, such as those in medicine, law, or computer technology, texts often contain a large number of technical terms. Existing translation systems lack stable terminology management mechanisms, easily resulting in inconsistent terminology translations or terminology drift. Furthermore, traditional machine translation systems usually rely on a single translation model to generate translations, lacking multi-strategy translation and multi-round optimization mechanisms, resulting in shortcomings in language fluency, grammatical correctness, and stylistic consistency.

[0004] On the other hand, in publishing-level or professional document translation scenarios, the quality of the translation not only needs to ensure semantic accuracy, but also needs to meet requirements such as terminology consistency, language standardization, and stylistic consistency. Existing machine translation systems are unable to simulate the multi-round translation, proofreading, and editing processes in a real translation workflow, and therefore struggle to achieve high-quality publishing-level translation standards.

[0005] Therefore, how to build a translation system that combines semantic understanding, knowledge enhancement, and multi-round translation optimization mechanisms to improve translation accuracy, terminology consistency, and overall translation quality has become a pressing technical problem in the field of machine translation.

[0006] To address the above problems, this invention provides a translation method, system, and approach based on multi-agent collaboration and multi-stage optimization. Summary of the Invention

[0007] To address the shortcomings of existing technologies, a translation system based on multi-agent cooperation and multi-stage optimization is proposed, characterized by the following components:

[0008] The document parsing module is used to receive translation documents input by users and perform content recognition and structure parsing. When the input is a scanned file, image or PDF document, the text content is obtained through OCR recognition and the chapter structure, paragraph level and chart description are identified through document structure parsing. At the same time, possible professional terms are pre-extracted in the whole text to form a term candidate list.

[0009] The input preprocessing module is used to standardize the parsed text, including text segmentation, text cleaning and punctuation standardization, and to identify the source language and target language types, thereby generating standardized translation input data;

[0010] The translation task planning module is used to perform a comprehensive analysis of translation tasks. It generates translation strategies through text type recognition, domain recognition, and text complexity analysis to guide the subsequent translation process.

[0011] The semantic parsing module is used to perform semantic analysis on the source language text, including syntactic dependency analysis, named entity recognition, semantic role labeling, and entity relation extraction, thereby generating a structured semantic representation;

[0012] The knowledge graph retrieval module is used to retrieve translation pairs of technical terms, conceptual relationships, and historical translation records from cross-language knowledge graphs based on semantic structure, thereby constructing a knowledge-enhanced context.

[0013] The dynamic agent generation module is used to dynamically generate different types of translation agents according to the translation task requirements to adapt to different translation scenarios. Based on the translation task planning results and text feature information, this module analyzes the task complexity, domain type and text structure, and instantiates the corresponding functional agents as needed.

[0014] The translation generation module is used to generate candidate translations through multiple translation agents and form a candidate translation set. In this module, different translation agents work together to perform translation tasks under the system's scheduling. This includes a literal translation agent generating a basic translation, a semantic optimization agent adjusting the semantics of the translation, a language polishing agent improving the fluency of expression, and a style optimization agent adjusting the style of the translation according to the text type. The translation results generated by each agent are then collected to form a candidate translation set.

[0015] The terminology consistency processing module is used to identify terms in the translation and match them with standard translations in the terminology database. It also performs term disambiguation based on the context and supports user-defined term locking rules to ensure consistent terminology throughout the text.

[0016] The translation quality assessment module is used to conduct multi-dimensional quality assessments of candidate translations, including semantic consistency, language fluency, terminology accuracy, and grammatical correctness, and selects the best translation based on the comprehensive score.

[0017] The output and knowledge update module is used to generate the final translation and update the knowledge graph, while performing a final quality check on the translation to ensure the accuracy and consistency of the translation results.

[0018] Furthermore, this invention also relates to a translation method based on multi-agent collaboration and multi-stage optimization. This method is applied to the aforementioned translation system based on multi-agent collaboration and multi-stage optimization. It achieves continuous improvement in translation quality by having multiple functional agents collaborate to complete the translation task and by combining a knowledge graph memory mechanism. Specifically, it includes:

[0019] Step S1, Document Acquisition and OCR Text Parsing: The input document is identified and parsed through the OCR recognition module and the document structure parsing module to obtain the complete text content and build a document structure tree. At the same time, core terms are pre-extracted in the whole text to form a task-level term list, thereby establishing a complete text structure context before translation and avoiding the context fragmentation problem caused by traditional sentence-by-sentence translation.

[0020] Step S2, Input Preprocessing and Individual Baseline Establishment: The system receives the normalized text sequence processed in step S1, and sequentially performs text segmentation, text cleaning, punctuation standardization and language detection, and establishes the input baseline for the current translation task;

[0021] Step S3, Translation Task Planning: The task planning agent performs multi-dimensional analysis on the preprocessed text and generates a structured translation strategy;

[0022] Step S4, Semantic parsing and translation task context graph construction: The semantic parsing agent performs deep semantic analysis on the source language text, generates intermediate semantic representations through syntactic dependency analysis, named entity recognition, semantic role labeling and entity relation extraction, and constructs a translation task context graph to characterize entities, events and their semantic relationships in the text;

[0023] Step S5: Knowledge Graph Retrieval and Enhancement. Retrieve translation pairs of professional terms, conceptual relationships, and historical translation records related to the current translation task from the cross-linguistic semantic knowledge graph, and generate knowledge-enhanced context KAC.

[0024] Step S6: Dynamic specialized intelligent agent generation. Based on translation strategies and text features, domain-specific intelligent agents are instantiated as needed. The differentiated responsibilities of each dynamic intelligent agent are focused on domain-specific sentence structure processing and special scenario processing that cannot be covered by existing fixed steps.

[0025] Step S7: Multi-path parallel candidate translation generation. This is achieved through concurrent execution of multiple types of translation agents. Each path generates candidate translations independently, and the translations are pooled together before proceeding to the next step. Each path generates translations in parallel without interdependence, and the coordinator schedules their concurrent execution.

[0026] Step S8: Terminology consistency processing. The terminology alignment agent performs terminology recognition, terminology knowledge base query, and context-aware terminology disambiguation processing on all candidate translations in the candidate pool of step S7, and achieves full-text terminology consistency by combining user-defined terminology locking rules.

[0027] Step S9: Multi-dimensional translation quality assessment and optimal candidate selection. The quality assessment agent evaluates all candidate translations in the pool one by one and selects the one with the highest comprehensive score as the optimal output. The comprehensive score includes semantic consistency, language fluency, terminology accuracy, grammatical correctness and style consistency indicators. If the optimal candidate score is lower than the quality threshold, the process returns to step S6 for iterative optimization.

[0028] Step S10: Continuous updating of the knowledge graph, structured output and publication-level quality control. After the translation is completed, the system sequentially performs three post-processing operations: knowledge graph back-writing, structured output of the translation, and publication-level quality control. The new knowledge from this translation is deposited into the knowledge graph, a structured translation that maintains the original document format is generated, and the final output quality is ensured through a publication-level proofreading mechanism.

[0029] The present invention also discloses a non-volatile storage medium, characterized in that the non-volatile storage medium includes a stored program, wherein the program, when running, controls the device where the non-volatile storage medium is located to execute the above-described method.

[0030] The present invention also discloses an electronic device, characterized in that it comprises a processor and a memory; the memory stores computer-readable instructions, and the processor is used to execute the computer-readable instructions, wherein the computer-readable instructions execute the method described above.

[0031] Beneficial effects

[0032] Compared with the prior art, the present invention has the following beneficial effects.

[0033] First, this invention executes translation tasks through a multi-agent collaborative mechanism, with different functional agents responsible for tasks such as semantic parsing, translation generation, terminology management, and quality assessment, thereby achieving multi-strategy collaborative translation and improving the overall processing capability of the translation system.

[0034] Secondly, this invention processes the input document using OCR recognition and document structure parsing technology, and extracts core terms and document structure information before translation, thereby establishing a complete text context environment and effectively avoiding the context fragmentation problem caused by traditional sentence-by-sentence translation.

[0035] Furthermore, this invention constructs a cross-language knowledge graph and combines it with a terminology database to uniformly manage professional terms during the translation process and supports user-defined term locking functions, thereby significantly improving the consistency and accuracy of professional terminology translation.

[0036] Furthermore, this invention employs a multi-stage translation optimization process, including literal translation generation, semantic restructuring, language polishing, and style calibration, to ensure that the translation retains the semantics of the original text while conforming to the expression habits of the target language, thereby enhancing the naturalness and readability of the translation.

[0037] Furthermore, this invention employs a multi-dimensional translation quality assessment mechanism to evaluate candidate translations based on indicators such as semantic consistency, language fluency, terminology accuracy, and grammatical correctness. The optimal translation is selected based on the comprehensive score, thereby further improving the quality of the translation results.

[0038] Therefore, this invention can significantly improve the translation quality of machine translation systems in complex documents and professional texts, and can meet the application requirements of high-quality translation and publication-level translation. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0040] Example 1

[0041] A translation system based on multi-agent collaboration and multi-stage optimization includes the following:

[0042] The document parsing module receives user-input translation documents and performs content recognition and structural analysis. When the input document is a scanned document, image file, or PDF file, the system uses the OCR recognition unit to recognize the text content and extract text information. Simultaneously, the document structure analysis unit analyzes the document's chapter structure, paragraph levels, headings, and chart descriptions to construct a complete document structure tree. The terminology pre-extraction unit identifies possible technical terms throughout the entire text and generates a candidate term list, providing a foundation for consistent terminology during subsequent translation.

[0043] Input preprocessing module: The input preprocessing module is used to standardize the text data output by the document parsing module. The system first divides the text into sentence-level translation units through the text segmentation unit, then removes irrelevant symbols and noise characters from the text through the text cleaning unit, and unifies the punctuation expression in different language environments through the punctuation standardization unit. Subsequently, the language recognition unit is used to identify the source language type and the target language type, thereby generating standardized translation input data.

[0044] Translation Task Planning Module: This module performs overall analysis and planning for translation tasks. The system first identifies the text type (e.g., technical documents, academic papers, or news text) using a text type recognition unit. Then, it identifies the text's professional field (e.g., medicine, law, or computer technology) using a domain recognition unit. A text complexity analysis unit assesses the text's sentence length, term density, and semantic complexity. Finally, a translation strategy generation unit generates a translation execution strategy based on the analysis results, guiding subsequent translation processes.

[0045] Semantic parsing module: The semantic parsing module is used to perform deep semantic understanding of the source language text; the system analyzes the syntactic dependency relations in the sentence through the syntactic dependency analysis unit, and identifies entity information in the text, such as person names, organization names and place names, through the named entity recognition unit; the semantic role labeling unit is used to identify semantic role relations in the sentence, such as agent, patient and action relations; the entity relation extraction unit is used to identify the semantic relations between entities and generate structured semantic representations to support the subsequent translation generation process;

[0046] Knowledge Graph Retrieval Module: The knowledge graph retrieval module provides knowledge support during the translation process. Based on the semantic structure generated by the semantic parsing module, the system retrieves relevant information from the cross-language knowledge graph. The information in the knowledge graph includes translation pairs of professional terms, conceptual relationships, and historical translation records. The system obtains professional term translations through the terminology knowledge retrieval unit, obtains domain knowledge through the conceptual relationship retrieval unit, and obtains historical translation data through the historical translation record retrieval unit, thereby constructing a knowledge-enhanced context.

[0047] Dynamic Agent Generation Module: The dynamic agent generation module is used to generate different types of specialized agents according to the specific needs of the translation task. The system can dynamically generate domain knowledge agents, long text context agents, chart translation agents, and code translation agents based on the task planning results, thereby providing targeted processing capabilities in different translation scenarios. By dynamically generating agents, the system can automatically expand its processing capabilities according to different translation tasks.

[0048] Translation Generation Module: The translation generation module is used to generate candidate translations. This module contains multiple translation agents, including a literal translation agent, a free translation agent, a language polishing agent, and a style optimization agent. Each translation agent executes translation tasks in parallel under system scheduling and generates multiple candidate translations. The system collects the translations generated by each translation agent to form a candidate translation set for subsequent quality evaluation module screening.

[0049] Terminology Consistency Processing Module: The terminology consistency processing module is used to ensure the consistency of terminology in the translation results. The system first identifies the terminology expressions in the translation through the terminology recognition unit, and then queries the corresponding standard translation in the terminology database through the terminology matching unit. The terminology disambiguation unit determines the specific meaning of the term based on the context semantics and selects the most appropriate translation expression. When the user sets terminology rules in advance, the terminology locking unit can execute fixed translation rules on the specified terms, thereby ensuring the consistency of terminology expressions throughout the text.

[0050] Translation Quality Assessment Module: This module assesses the quality of candidate translations. The system uses a semantic consistency assessment unit to determine if the translation accurately conveys the meaning of the original text, a fluency assessment unit to evaluate the naturalness of the language, a terminology accuracy assessment unit to check the correctness of terminology translation, a grammatical correctness assessment unit to check the correctness of the grammatical structure, and a style consistency assessment unit to evaluate the overall style of the translation. Based on these multiple indicators, the system calculates a comprehensive score and selects the translation with the highest score as the optimal translation. When the optimal translation's score falls below a preset quality threshold, the system can re-trigger the translation generation module for optimization.

[0051] Output and Knowledge Update Module: The output and knowledge update module is used to generate the final translation and update the system knowledge. The system first restores the translation to a document structure consistent with the original text through the translation structure restoration unit. Then, the knowledge graph update unit writes the new terms, new translation relationships, and translation frequency information generated during the translation process into the knowledge graph. Finally, the publication-level quality proofreading unit performs a final quality check on the translation to ensure that the translation meets high-quality translation standards.

[0052] Through the collaborative work of the above modules, the present invention can achieve a high-quality translation process based on multi-agent cooperation and multi-stage optimization, thereby significantly improving translation accuracy, terminology consistency and language expression quality.

[0053] Example 2

[0054] A translation method based on multi-agent cooperation and multi-stage optimization, applied to the aforementioned translation system based on multi-agent cooperation and multi-stage optimization, includes the following steps:

[0055] Step S1, Document Acquisition and OCR Text Parsing: Through OCR recognition and document structure parsing, the system can acquire complete text structure information before translation, construct a document structure tree, and extract key terms in advance, avoiding the context fragmentation problem caused by traditional sentence-by-sentence translation;

[0056] Upon receiving a translation task, the system first performs content acquisition and structure parsing on the input document; this step is executed when the input document is a scanned file, image file, or PDF document; when the input is in structured text format, this step is skipped and the process proceeds directly to step S2; specifically including:

[0057] Step S11: Deep learning text recognition, using the OCR recognition module for recognition, supporting documents in mainstream languages ​​such as Chinese and English, as well as mixed language documents;

[0058] Step S12: Document structure hierarchy recognition. The document chapter hierarchy relationship is automatically identified through a structure parsing algorithm to construct a document structure tree.

[0059] Recognition content and subsequent processing strategies for each structural type: The document chapter hierarchy is automatically identified through a structural parsing algorithm, extracting the following five types of structural information and constructing a document structure tree: Title hierarchy structure, which is constructed after recognition to drive differentiated configuration of chapter-level translation strategies; Paragraph structure, which maintains paragraph boundaries and avoids semantic confusion across paragraphs; Chart and graph captions, which are triggered by the chart and graph translation agent to perform structure-aware translation after recognition; Footers and annotations, which are processed independently to maintain consistency with the citation relationship in the main text; Mathematical formulas, which retain the original symbols and translate only the text descriptions.

[0060] The constructed document structure tree organizes the full text content in a hierarchical nested manner. The top-level node is the document root node Document, which contains Chapter nodes, Section nodes, Paragraph nodes, Table / Figure leaf nodes, and Appendix and Footnote as independent branches outside the main chapters.

[0061] Step S13: Terminology pre-extraction. A terminology pre-extraction agent identifies core terms in advance across the entire text, constructing a task-level terminology list (TL). A hybrid strategy combining TF-IDF weights and domain relevance scoring is employed.

[0062] Term_score(w) = TF-IDF(w) × Domain_relevance(w, d)

[0063] Wherein, Term_score(w) represents the score of w as a term, where w is a candidate word in the document, i.e., the basic processing unit for term scoring; TF-IDF(w) is the term frequency-inverse document frequency score of candidate word w in the current document, which measures the importance and global scarcity of word w in the document. The higher the value, the more important and scarce w is in the document; Domain_relevance(w, d) is the relevance score of word w to document domain d, calculated by a pre-trained domain classification model, with a value range of [0,1], reflecting the degree of matching between the word and the domain to which the current translation task belongs; d is the document domain label, such as medicine, law, etc. The document domain label is given manually or can be automatically determined by the trained domain classification model; when Term_score(w) exceeds the threshold, the term is triggered to be stored in TL, which can be adjusted according to the domain specialization, and is set adaptively by the system according to d by default;

[0064] Step S2, Input Preprocessing and Individual Baseline Establishment: The system receives the normalized text sequence processed in step S1, and sequentially performs text segmentation, text cleaning, punctuation standardization and language detection, and establishes the input baseline for the current translation task;

[0065] Step S21: Text segmentation. A rule-based and statistical hybrid segmentation algorithm is used to handle ambiguity in sentence boundaries caused by periods, ellipses, abbreviations, etc. within quotation marks. Under the constraints of the document structure tree, the semantic integrity of chapter boundaries and paragraph boundaries is maintained, and a standardized sentence sequence is generated.

[0066] Step S22: Text cleaning, removing HTML tags, special control characters, repeated spaces and meaningless noise characters; performing post-processing correction based on language model for possible recognition errors in the OCR recognition results (confusion of similar characters, word segmentation errors, etc.);

[0067] Step S23: Standardize punctuation, unify Chinese and English punctuation (e.g., convert full-width commas to their corresponding language standard forms), handle nested quotation marks and paired parentheses, and standardize list symbols;

[0068] Step S24: Language detection. The source language type is confirmed through an n-gram language model, supporting fragment-level language annotation and segmentation processing for mixed language input.

[0069] Step S25: Input baseline establishment, calculate the text statistical baseline for the current translation task:

[0070] Baseline = { L_avg_base, V_base, D_base}

[0071] Wherein, L_avg_base is the average number of words in sentences in the input baseline, reflecting the baseline level of sentence length complexity in the current document; V_base is the average frequency ranking of words in the input baseline, reflecting the baseline level of word difficulty distribution, with a higher ranking indicating more obscure words; D_base is the density of domain terminology in the input baseline, defined as the proportion of the set of domain terminology in the full-text vocabulary set, with a higher value indicating a more specialized document.

[0072] Step S3, Translation Task Planning: The task planning agent performs multi-dimensional analysis on the preprocessed text and generates a structured translation strategy; specifically including:

[0073] Step S31: Text type recognition, using a multi-label classifier based on a pre-trained language model to classify texts into technical documents, news reports, dialogue texts, academic papers, legal documents, literary works, etc.

[0074] Step S32: Domain identification, combining a keyword dictionary with a pre-trained language model classifier, supports no fewer than 32 fine-grained domains including artificial intelligence, medicine, law, and finance, and records the score f_score;

[0075] Step S33: Text complexity analysis, calculate the overall complexity index C_index:

[0076] C_index = w1 · L_avg_norm + w2 · Clause_density_norm + w3 · T_density_norm

[0077] Wherein, L_avg_norm is the average sentence length after normalization based on the input baseline L_avg_base in step S1, eliminating the sentence length scoring bias caused by different document types; Clause_density_norm is the normalized clause density index, which estimates syntactic complexity by statistically analyzing the ratio of conjunctions in a sentence to its sentence length. A higher value indicates more nested clauses and greater syntactic complexity; T_density_norm is the terminology density after normalization based on D_base in step S1, reflecting the terminology richness of the current document relative to the general vocabulary baseline; w1 is the weight coefficient of the average sentence length dimension, controlling the strength of this dimension's contribution to C_index; w2 is the weight coefficient of the clause density dimension, controlling the strength of the contribution to syntactic structure complexity, which has the greatest impact on translation difficulty; w3 is the weight coefficient of the terminology density dimension, controlling the strength of the contribution to terminology density, and the three satisfy w1+w2+w3=1.

[0078] Step S34: Generate a structured translation strategy. Based on the above analysis results, output a structured translation strategy. The translation strategy shall include at least the following parameters:

[0079] translation_mode: literal / semantic / free, determines the weighting of multiple rounds of translation;

[0080] terminology_mode (terminology processing mode): strict / standard / flexible, determines the strength of terminology locking;

[0081] style (literary style parameter): academic / journalistic / technical / literary, determines the target style for style calibration;

[0082] quality_threshold (quality threshold) θ_q: Adaptively set according to document type. The default value for publication-level documents is θ_q=0.88, and the default value for general documents is θ_q=0.80.

[0083] Step S4, Semantic parsing and translation task context graph construction: The semantic parsing agent performs deep semantic analysis on the source language text to generate a semantic intermediate representation and a translation task context graph (TCG);

[0084] Step S41: Syntactic dependency parsing. The Universal Dependencies framework is used to parse core grammatical relations such as subject, verb, and object, as well as modification relations, and generate a dependency tree structure.

[0085] Step S42: Named entity recognition, using a prefix tree model to identify entity types such as person names, organization names, place names, product names, and professional terms; secondary correction is performed on entity boundary errors that may exist in the OCR-recognized document;

[0086] Step S43: Semantic role labeling, labeling semantic roles such as agent, patient, time, place, and manner with the predicate as the center, and generating an event semantic framework;

[0087] Step S44: Extract entity relationships and identify relationships such as attributes, causality, and comparison between entities;

[0088] Step S45: Construction of the Translation Task Context Graph (TCG), using a Generative Neural Network (GNN) to model the semantic parsing results into a graph structure.

[0089] TCG = (V, E, A)

[0090] Wherein, V is the set of semantic nodes, each node representing an entity (such as a person's name, organization name, product name, or technical term), event, or abstract concept in the text, and terms in the task-level term list TL are attached to the corresponding nodes as prior constraints; E is the set of relation edges, each edge representing a semantic association between two nodes, such as causal relationship, attribute relationship, comparison relationship, or subordinate relationship, and the directionality of the edge reflects the agent and the recipient of the relationship; A is the attribute mapping function between nodes and edges, which assigns semantic attributes to each node or edge, including entity type, relation type, confidence level, etc.

[0091] Step S5, Knowledge Graph Retrieval and Enhancement: The memory and knowledge graph agent retrieves semantically related knowledge from the cross-linguistic semantic knowledge graph and generates a knowledge-enhancing context (KAC).

[0092] The knowledge retrieval content includes:

[0093] Step S51: Retrieve professional terminology translation pairs, query the corresponding translations of source language terms in the target language and their applicable conditions in context; the knowledge graph covers more than 10 professional fields, contains millions of terminology translation pairs, and is continuously expanding;

[0094] Step S52: Retrieve related concepts and hyponyms / hypernyms to assist in processing concept transfer and metaphorical expression;

[0095] Step S53: Retrieve historical translation memory (TM), extract high-confidence historical translation fragments for reuse, and assign a translation memory matching score (TM_score) of:

[0096] TM_score(src, tmEntry) = cos(Enc_cross(src), Enc_cross(tmEntry.src))

[0097] Wherein, src is the source language sentence to be translated; tmEntry is a historical translation record in the translation memory, containing the historical source language sentence tmEntry.src and the corresponding historical translation tmEntry.tgt; Enc_cross(·) is a cross-lingual semantic encoding function that maps any language text to a unified multilingual semantic vector space, making semantically similar content in different languages ​​close in distance in the vector space, based on multilingual pre-trained models such as LaBSE or mBERT; cos(·,·) is a cosine similarity function, and the historical translation is directly reused when TM_score is greater than or equal to a certain threshold, with a default of 0.95;

[0098] Step S54: Organize the output of steps S51-S53 into a knowledge-enhancing context (KAC);

[0099] Step S6: Dynamic specialized agent generation. Based on translation strategies and text features, domain-specific specialized agents are instantiated as needed. The differentiated responsibilities of each dynamic agent focus on domain-specific sentence structure processing and special scenario processing that cannot be covered by existing fixed steps. The specific generation rules are as follows:

[0100] When a document scores f_score in a certain domain and exceeds a certain threshold (0.7 by default), a domain-specific agent is instantiated to handle the domain's expression specifications.

[0101] When the text length exceeds a certain limit (default 1024), a long document entity tracking agent is instantiated to maintain cross-paragraph name, organization name, and pronoun reference chains, preventing entity translation drift in long documents.

[0102] When a document contains charts, instantiate a chart translation agent to perform the translation of the chart content;

[0103] When a document contains code blocks, a code mixing agent is instantiated to handle the translation boundaries between code blocks and natural language in the document, preserving the original code and translating only comments and explanatory text.

[0104] When the source language belongs to the low-resource language set, instantiate a low-resource language agent, activate a dedicated few-shot translation model, and process low-resource languages ​​such as Tibetan and Swahili.

[0105] Step S7: Multi-path parallel candidate translation generation. This is achieved through concurrent execution of multiple types of translation agents. Each path generates candidate translations independently, and the translations are pooled together before proceeding to the next step. Each path generates translations in parallel without interdependence, and the coordinator schedules their concurrent execution.

[0106] Parallel Path 1 Direct Translation Agent

[0107] Using the source language's syntactic structure as a reference, terminology constraints from TL and KAC are injected into the decoding process to generate a basic translation that preserves the original meaning; the decoding strategy is dynamically switched for different node types in the document structure tree.

[0108] Parallel Path P2 Semantic Translation Agent

[0109] Driven by the semantic structure in step S4, it does not adhere to the original syntactic form and pursues the maximum semantic fidelity to the target language; it handles information misalignment and rearrangement caused by word order differences in the source language, semantic equivalent substitution of unnatural word-by-word translation, and cross-language conversion of metaphorical expressions and culturally loaded words.

[0110] Parallel Path 3 Domain Translation Agent

[0111] By combining KAC domain knowledge with specialized intelligent agent resources dynamically generated by step S6, candidate translations with higher accuracy in the professional domain are generated; the domain-specific sentence structure norms provided by the S6 dynamic intelligent agent are utilized to handle domain-specific expressions that cannot be handled correctly by the literal translation path.

[0112] Parallel Path 4: Smoothness Optimization Agent

[0113] With the naturalness of the target language expression as the primary optimization objective, fine-tuning is performed based on a large-scale target language monolingual corpus (with a scale of no less than 10 billion words) to generate the most fluent candidate translations; the perplexity level (PPL) of the target language is used as the optimization metric.

[0114] Parallel Path 5: Style Calibration Agent

[0115] Based on the style parameter in the translation strategy of step S3, candidate translations with clear style targets are generated independently; the following target style types are supported:

[0116] academic writing: objective statement tone, passive voice, conforming to academic writing norms;

[0117] journalistic (news reporting): concise, lively, and highlighting the core information;

[0118] Technical documents: precise and concise, retaining the original text of technical terms;

[0119] legal (legal documents): formal legal language, maintaining the precision of the clause structure;

[0120] literary works: Emphasize the beauty of language and appropriately employ free translation techniques;

[0121] Step S8: Terminology consistency processing. The terminology alignment agent performs systematic terminology consistency processing on all candidate translations in the candidate pool from step S7. System terminology resource scale: covering professional fields, including no less than 1 million high-quality terminology translation pairs that have been manually verified, and maintaining a cross-language semantic terminology knowledge graph.

[0122] Step S81, terminology recognition, adopts a hybrid strategy of CRF and dictionary matching to identify professional terms and multi-word expressions in the candidate translations; at the same time, the TL pre-extracted in step S13 is used as a priori constraint execution priority marker;

[0123] Step S82: Knowledge base query: Search the database of millions of terms and knowledge graphs to obtain a list of candidate translations for the identified terms in the target language; for terms that already have corresponding translations in the TL, directly use the pre-extracted results;

[0124] Step S83: Context-aware term disambiguation. For ambiguous terms with multiple target language equivalents, disambiguation is performed by calculating the cosine similarity between the term candidate and the context vector.

[0125] Step S84: User-defined terminology locking. The system supports user-specified forced translation expressions, which have higher priority than the translations recommended by the knowledge graph system.

[0126] Step S85: Full-text unified replacement and consistency verification: Perform consistency replacement on all candidate translations in the candidate pool;

[0127] Step S9: Multi-dimensional translation quality assessment and optimal candidate selection. The quality assessment agent evaluates all candidate translations in the pool one by one and selects the one with the highest comprehensive score as the optimal output. If the optimal candidate score is lower than the quality threshold θ_q, the process returns to step S6 for iterative optimization.

[0128] Step S91: Calculate the cross-language semantic fidelity score (semantic consistency):

[0129] semantic_score = cos(CLSE(src), CLSE(tgt))

[0130] Here, `semantic_score` is the semantic consistency score of the candidate translation; `src` is the source language text to be translated; `tgt` is the candidate target language translation; `CLSE(·)` is a cross-lingual semantic encoding function that maps source or target language text to the same multilingual semantic vector space, implemented based on LaBSE; `cos(·, ·)` is a cosine similarity function with a value range of [0,1]. The closer it is to 1, the more faithful the translation is to the semantics of the original text; when it is below the fidelity threshold (default 0.85), a semantic alignment refinement operation is triggered.

[0131] Step S92, fluency_score language fluency calculation, based on adaptive normalized perturbation of the input baseline:

[0132] fluency_score = 1 - [PPL(tgt) - PPL_min] / [PPL_max - PPL_min]

[0133] Here, `fluency_score` is the fluency score of the candidate translation; the closer it is to 1, the more natural and fluent the translation. `PPL(tgt)` is the perplexity score calculated by a large-scale language model of the target language for the candidate translation `tgt`; the lower the score, the more the translation conforms to the target language's habits. This model is trained on a single-language corpus of the target language containing at least 10 billion words. `PPL_min` is the lower bound of the perplexity score of high-quality translations in the current domain, dynamically calibrated based on `D_base`. `PPL_max` is the upper bound of the perplexity score of the reference corpus in the current domain, also dynamically calibrated based on `D_base`. Together, they form an adaptively normalized reference interval to ensure a fair and consistent fluency evaluation standard for different types of documents. When `PPL(tgt) ≤ PPL_min`, `fluency_score` = 1; when `PPL(tgt) ≥ PPL_max`, `fluency_score` = 0.

[0134] Step S93: Calculate the terminology_score, combining the system terminology accuracy with the user-locked terminology compliance rate.

[0135] terminology_score = λ·(|T_correct| / |T_total|) + (1-λ) · (|T_locked_complied| / |T_locked_total|)

[0136] Wherein, terminology_score is the comprehensive terminology accuracy score of the candidate translation, with a value range of [0,1]; λ is the weight balance coefficient between the system terminology accuracy and the user-locked term compliance rate, with a default value of 0.6. The more user-locked terms, the lower the value of λ can be to increase the priority of locked terms; |T_correct| is the number of times terms consistent with the million-level terminology database appear in the candidate translation, obtained through cross-validation of the system terminology knowledge base; |T_total| is the total number of times all controlled terms appear in the candidate translation, which, together with |T_correct|, is used to calculate the system terminology accuracy; |T_locked_complied| is the number of times locked terms that comply with user-defined locking rules appear in the candidate translation, with locked terms having a higher priority than system knowledge base recommendations; |T_locked_total| is the total number of times user-defined locked terms appear in the candidate translation, which, together with |T_locked_complied|, is used to calculate the locked term compliance rate;

[0137] Step S94: Calculate grammar_score for grammatical correctness, using a syntax tree parsing tool to evaluate the grammatical compliance of the translation;

[0138] Step S95, style_consistency_score style consistency calculation:

[0139] style_consistency_score = 1 -norm( Style_drift)

[0140] Style_drift = (1 / N)·Σ i ||Style_vec(p i ) - Style_mean‖ 2 ;

[0141] Where style_consistency_score is the global style consistency score of the candidate translations; Style_drift is the average L2 distance squared between the style vectors of each paragraph and the mean style vector, which characterizes the dispersion of the style of the whole text. The larger the value, the more severe the style drift; N is the total number of paragraphs in the candidate translations; p i For the i-th paragraph text in the candidate translation, the unit is the Paragraph node in the document structure tree; Style_vec(p i ) represents the i-th paragraph p i The style feature vector is extracted from the paragraph text by the style classifier, encoding style features such as formality, tone, and style. The style classifier is trained on a publication-grade corpus. Style_mean is the mean vector of style vectors of all paragraphs in the whole text, representing the overall average style of the current candidate translation. The square of the L2 norm of the vector is used to calculate the squared Euclidean distance between the style vector of each paragraph and the mean style vector; norm is used for normalization.

[0142] Step S95, Comprehensive score calculation formula:

[0143] FULLScore = α_sem·semantic + α_flu·fluency + α_term·terminology +α_gram·grammar + α_sty·style_consistency

[0144] Among them, FULLScore is the weighted comprehensive quality score of the candidate translations, with a value range of [0,1]. The highest score in the candidate pool is selected as the optimal candidate. α_sem is the weight coefficient for semantic consistency, with a default value of 0.30; α_flu is the weight coefficient for language fluency, with a default value of 0.20; α_term is the weight coefficient for terminology accuracy, with a default value of 0.25, which can be increased to 0.35 for terminology-intensive documents; α_gram is the weight coefficient for grammatical correctness, with a default value of 0.15; and α_sty is the weight coefficient for style consistency, with a default value of 0.10, which can be dynamically adjusted according to the translation scenario.

[0145] Step S10: Continuous updating of the knowledge graph, structured output, and publication-level quality control.

[0146] After the translation is completed, the system performs three post-processing operations in sequence: knowledge graph back-writing, structured output of the translation, and publication-level quality control. The new knowledge from this translation is stored in the knowledge graph, a structured translation that maintains the original document format is generated, and the final output quality is guaranteed through a publication-level proofreading mechanism.

[0147] Step S101: The knowledge graph is continuously updated. The memory and knowledge graph agent writes the new knowledge generated in this translation back to the cross-language semantic knowledge graph. Specifically, this includes: writing high-confidence term pairs that appear for the first time in this translation as new nodes into the knowledge graph; updating the semantic association edge weights between entities to strengthen high-frequency co-occurrence relationships; accumulating the usage frequency of existing nodes for priority ranking in subsequent searches. Through the above continuous update mechanism, the knowledge graph is continuously enriched as the amount of translation accumulates, and the accuracy of system terms continues to improve as the amount of usage increases.

[0148] Step S102, Translation Backfilling and Structured Output: The optimal translation is backfilled into the corresponding node of the document structure tree constructed in step S1, generating a structured translation document that retains the original document chapter hierarchy, chart position and format; at the same time, the quality score from step S9, the terminology usage record from step S8 and the translation metadata are gathered to generate a structured output package.

[0149] Step S103, Publishing-Level Quality Control: This process references real-world publishing industry editing and proofreading standards, and is trained on a publishing-level parallel corpus of at least 500 million words to simulate the proofreading decision-making logic of professional editors; specifically, it performs the following four checks:

[0150] First, terminology consistency review: Based on terminology processing, a second scan is performed to check the consistency rate of terminology throughout the text and the compliance with user-locked terminology.

[0151] Secondly, grammar and punctuation proofreading, based on the target language grammar rule base and publishing standard templates, detects grammatical errors, punctuation misuse and improper sentence segmentation, automatically corrects them and generates error type statistics;

[0152] Third, style consistency check: perform style uniformity check on the entire text, locate style-drifting paragraphs and perform local corrections;

[0153] Fourth, verify the figures and annotations, checking the consistency of the figures and units in the translated text with the original text, and verifying the numbering sequence of footnotes and endnotes;

[0154] After publication-level quality control is completed, a quality control report is generated, including: the pass and correction status of each check, a comparison of the translation before and after correction, a list of remaining issues, and a confidence score. If the overall score is still lower than the publication-level threshold after quality control, the system marks it as requiring manual review and highlights the paragraphs that need special attention in the quality control report.

[0155] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A translation system based on multi-agent cooperation and multi-stage optimization, characterized by: The document parsing module is used to receive translation documents input by users and perform content recognition and structure parsing. When the input is a scanned file, image or PDF document, the text content is obtained through OCR recognition, and the chapter structure, paragraph level and chart description information are identified through document structure parsing. At the same time, possible professional terms are pre-extracted in the whole text to form a term candidate list. The input preprocessing module is used to standardize the parsed text, including text segmentation, text cleaning and punctuation standardization, and to identify the source language and target language types, thereby generating standardized translation input data; The translation task planning module is used to perform a comprehensive analysis of translation tasks. It generates translation strategies through text type recognition, domain recognition, and text complexity analysis to guide the subsequent translation process. The semantic parsing module is used to perform semantic analysis on the source language text, including syntactic dependency analysis, named entity recognition, semantic role labeling, and entity relation extraction, thereby generating a structured semantic representation; The knowledge graph retrieval module is used to retrieve translation pairs of technical terms, conceptual relationships, and historical translation records from cross-language knowledge graphs based on semantic structure, thereby constructing a knowledge-enhanced context. The dynamic agent generation module is used to dynamically generate different types of translation agents according to the needs of translation tasks, so as to adapt to different translation scenarios; Based on the translation task planning results and text feature information, this module analyzes the task complexity, domain type and text structure, and instantiates the corresponding functional agents as needed. The translation generation module is used to generate candidate translations through multiple translation agents and form a candidate translation set. In this module, different translation agents work together to perform translation tasks under the system's scheduling. This includes a literal translation agent generating a basic translation, a semantic optimization agent adjusting the semantics of the translation, a language polishing agent improving the fluency of expression, and a style optimization agent adjusting the style of the translation according to the text type. The translation results generated by each agent are then collected to form a candidate translation set. The terminology consistency processing module is used to identify terms in the translation and match them with standard translations in the terminology database. It also performs term disambiguation based on the context and supports user-defined term locking rules to ensure consistent terminology throughout the text. The translation quality assessment module is used to conduct multi-dimensional quality assessments of candidate translations, including semantic consistency, language fluency, terminology accuracy, and grammatical correctness, and selects the best translation based on the comprehensive score. The output and knowledge update module is used to generate the final translation and update the knowledge graph, while performing a final quality check on the translation to ensure the accuracy and consistency of the translation results.

2. A translation method based on multi-agent cooperation and multi-stage optimization, wherein the method is applied to the translation system based on multi-agent cooperation and multi-stage optimization as described in claim 1, characterized in that: Step S1, Document Acquisition and OCR Text Parsing: The input document is identified and parsed through the OCR recognition module and the document structure parsing module to obtain the complete text content and build a document structure tree. At the same time, core terms are pre-extracted in the whole text to form a task-level term list, thereby establishing a complete text structure context before translation and avoiding the context fragmentation problem caused by traditional sentence-by-sentence translation. Step S2, Input Preprocessing and Individual Baseline Establishment: The system receives the normalized text sequence processed in step S1, and sequentially performs text segmentation, text cleaning, punctuation standardization and language detection, and establishes the input baseline for the current translation task; Step S3, Translation Task Planning: The task planning agent performs multi-dimensional analysis on the preprocessed text and generates a structured translation strategy; Step S4, Semantic parsing and translation task context graph construction: The semantic parsing agent performs deep semantic analysis on the source language text, generates intermediate semantic representations through syntactic dependency analysis, named entity recognition, semantic role labeling and entity relation extraction, and constructs a translation task context graph to characterize entities, events and their semantic relationships in the text; Step S5: Knowledge Graph Retrieval and Enhancement. Retrieve translation pairs of professional terms, conceptual relationships, and historical translation records related to the current translation task from the cross-linguistic semantic knowledge graph, and generate knowledge-enhanced context KAC. Step S6: Dynamic specialized intelligent agent generation. Based on translation strategies and text features, domain-specific intelligent agents are instantiated as needed. The differentiated responsibilities of each dynamic intelligent agent are focused on domain-specific sentence structure processing and special scenario processing that cannot be covered by existing fixed steps. Step S7: Multi-path parallel candidate translation generation. This is achieved through concurrent execution of multiple types of translation agents. Each path generates candidate translations independently, and the translations are pooled together before proceeding to the next step. Each path generates translations in parallel without interdependence, and the coordinator schedules their concurrent execution. Step S8: Terminology consistency processing. The terminology alignment agent performs terminology recognition, terminology knowledge base query, and context-aware terminology disambiguation processing on all candidate translations in the candidate pool of step S7, and achieves full-text terminology consistency by combining user-defined terminology locking rules. Step S9: Multi-dimensional translation quality assessment and optimal candidate selection. The quality assessment agent evaluates all candidate translations in the pool one by one and selects the one with the highest comprehensive score as the optimal output. The comprehensive score includes semantic consistency, language fluency, terminology accuracy, grammatical correctness and style consistency indicators. If the optimal candidate score is lower than the quality threshold, the process returns to step S6 for iterative optimization. Step S10: Continuous updating of the knowledge graph, structured output and publication-level quality control. After the translation is completed, the system sequentially performs three post-processing operations: knowledge graph back-writing, structured output of the translation, and publication-level quality control. The new knowledge from this translation is deposited into the knowledge graph, a structured translation that maintains the original document format is generated, and the final output quality is ensured through a publication-level proofreading mechanism.

3. The translation method based on multi-agent cooperation and multi-stage optimization according to claim 2, characterized in that: Step S1 further includes the following: Step S11: Deep learning text recognition, using the OCR recognition module for recognition, supporting documents in mainstream languages ​​and mixed languages; Step S12: Document structure hierarchy recognition. The document chapter hierarchy relationship is automatically identified through a structure parsing algorithm to construct a document structure tree. Recognition content and subsequent processing strategies for each structural type: The document chapter hierarchy is automatically identified through the structural parsing algorithm, and the following five types of structural information are extracted and a document structure tree is constructed: Title hierarchy structure, which is identified and a document structure tree is constructed to drive the differentiated configuration of chapter-level translation strategies; Paragraph structure, which maintains paragraph boundaries and avoids semantic confusion across paragraphs; Chart and graph captions, which are identified and the chart and graph translation agent is triggered to perform structure-aware translation; Footers and annotations, which are processed independently to maintain consistency with the citation relationship with the main text. Mathematical formulas, retaining the original symbols, only the textual explanations are translated; The constructed document structure tree organizes the full text content in a hierarchical nested manner. The top-level node is the document root node Document, which contains Chapter nodes, Section nodes, Paragraph nodes, Table / Figure leaf nodes, and Appendix and Footnote as independent branches outside the main chapters. Step S13: Terminology pre-extraction. A terminology pre-extraction agent identifies core terms in advance across the entire text, constructing a task-level terminology list (TL). A hybrid strategy combining TF-IDF weights and domain relevance scoring is employed. Term_score(w) = TF-IDF(w) × Domain_relevance(w, d) Wherein, Term_score(w) represents the score of w as a term, where w is a candidate word in the document, i.e., the basic processing unit for term scoring; TF-IDF(w) is the term frequency-inverse document frequency score of candidate word w in the current document, which measures the importance and global scarcity of word w in the document. The higher the value, the more important and scarce w is in the document; Domain_relevance(w, d) is the relevance score of word w to document domain d, calculated by a pre-trained domain classification model, with a value range of [0,1], reflecting the degree of matching between the word and the domain to which the current translation task belongs; d is the document domain label, which is given manually or can be automatically determined by the trained domain classification model; when Term_score(w) exceeds the threshold, the term is triggered to be stored in TL, which can be adjusted according to the domain expertise, and is set adaptively by the system according to d by default.

4. The translation system based on multi-agent cooperation and multi-stage optimization according to claim 1, characterized in that: step S2 further includes the following: Step S21: Text segmentation. A rule-based and statistical hybrid segmentation algorithm is used to handle the ambiguity of sentence boundaries caused by periods, ellipses, and abbreviations within quotation marks. Under the constraints of the document structure tree, the semantic integrity of chapter boundaries and paragraph boundaries is maintained, and a standardized sentence sequence is generated. Step S22: Text cleaning, removing HTML tags, special control characters, repeated spaces, and meaningless noise characters; performing post-processing correction based on language model for possible recognition errors in the OCR recognition results; Step S23: Standardize punctuation, unify Chinese and English punctuation, handle nested quotation marks and paired parentheses, and standardize list symbols; Step S24: Language detection. The source language type is confirmed through an n-gram language model, supporting fragment-level language annotation and segmentation processing for mixed language input. Step S25: Input baseline establishment, calculate the text statistical baseline for the current translation task: Baseline = { L_avg_base, V_base, D_base}; in, L_avg_base is the average number of words in the sentences in the input baseline, reflecting the baseline level of sentence length complexity in the current document; V_base is the average frequency ranking of words in the input baseline, reflecting the baseline level of word difficulty distribution, with a higher ranking indicating more obscure words; D_base is the density of domain terminology in the input baseline, defined as the proportion of the set of domain terminology in the full-text vocabulary, with a higher value indicating a more specialized document.

5. The translation method based on multi-agent cooperation and multi-stage optimization according to claim 2, characterized in that: Step S3 further includes the following: Step S31: Text type recognition, using a multi-label classifier based on a pre-trained language model for classification; Step S32: Domain identification, combining a keyword dictionary with a pre-trained language model classifier, supports fine-grained domains, and records the score f_score; Step S33: Text complexity analysis, calculate the overall complexity index C_index: C_index = w1·L_avg_norm + w2·Clause_density_norm +w3·T_density_norm Wherein, L_avg_norm is the average sentence length after normalization based on the input baseline L_avg_base in step S1, eliminating the sentence length scoring bias caused by different document types; Clause_density_norm is the normalized clause density index, which estimates syntactic complexity by statistically analyzing the ratio of conjunctions in a sentence to its sentence length. A higher value indicates more nested clauses and greater syntactic complexity; T_density_norm is the terminology density after normalization based on D_base in step S1, reflecting the terminology richness of the current document relative to the general vocabulary baseline; w1 is the weight coefficient of the average sentence length dimension, controlling the strength of this dimension's contribution to C_index; w2 is the weight coefficient of the clause density dimension, controlling the strength of the contribution to syntactic structure complexity, which has the greatest impact on translation difficulty; w3 is the weight coefficient of the terminology density dimension, controlling the strength of the contribution to terminology density, and the three satisfy w1+w2+w3=1. Step S34: Generate a structured translation strategy. Based on the above analysis results, output a structured translation strategy. The translation strategy shall include at least the following parameters: translation_mode: literal / semantic / free, determines the weighting of multiple rounds of translation; terminology_mode: strict / standard / flexible, determines the strength of terminology locking; The style parameters are: academic / journalistic / technical / literary, which determine the target style for style calibration. quality_threshold θ_q: The quality threshold is set adaptively based on the document type. The default value for publication-level documents is θ_q=0.88, and the default value for general documents is θ_q=0.

80.

6. The translation method based on multi-agent cooperation and multi-stage optimization according to claim 2, characterized in that: step S4 further includes the following: Step S41: Syntactic dependency parsing. The Universal Dependencies framework is used to parse the core grammatical relations of subject, verb, and object, as well as the modification relations, and generate a dependency tree structure. Step S42: Named entity recognition, using a prefix tree model to identify entity types such as person names, organization names, place names, product names, and professional terms; perform secondary correction for potential entity boundary errors in the OCR-recognized document; Step S43: Semantic role labeling, labeling the agent, patient, time, place and manner semantic roles with the predicate as the center, and generating an event semantic framework; Step S44: Extract entity relationships and identify the attributes, causal relationships, and comparison relationships between entities; Step S45: Construction of the Translation Task Context Graph (TCG), using a Generative Neural Network (GNN) to model the semantic parsing results into a graph structure. TCG = (V, E, A); in, V is a set of semantic nodes, where each node represents an entity, event, or abstract concept in the text. Terms in the task-level term list TL are attached to the corresponding nodes as prior constraints. E is a set of relation edges, where each edge represents a semantic association between two nodes. The directionality of the edge reflects the agent and the recipient of the relationship. A is an attribute mapping function between nodes and edges, which assigns semantic attributes to each node or edge, including entity type, relation type, and confidence level.

7. The translation method based on multi-agent cooperation and multi-stage optimization according to claim 2, characterized in that: step S5 further includes the following: Step S51: Retrieve professional terminology translation pairs, query the corresponding translations of source language terms in the target language and their applicable conditions in context; the knowledge graph covers more than 10 professional fields, contains millions of terminology translation pairs, and is continuously expanding; Step S52: Retrieve related concepts and hyponyms / hypernyms to assist in processing concept transfer and metaphorical expression; Step S53: Historical Translation Records (TM) retrieval; extract high-confidence historical translation fragments for reuse; Translation Memory Matching Score (TM_score) is: TM_score(src, tmEntry) = cos(Enc_cross(src), Enc_cross(tmEntry.src)); in, `src` is the source language sentence to be translated; `tmEntry` is a historical translation record in the translation memory, containing the historical source language sentence `tmEntry.src` and the corresponding historical translation `tmEntry.tgt`; `Enc_cross(·)` is a cross-lingual semantic encoding function that maps any language text to a unified multilingual semantic vector space, making semantically similar content in different languages ​​close in distance in the vector space, based on a multilingual pre-trained model; `cos(·, ·)` is a cosine similarity function, where the historical translation is directly reused when `TM_score` is greater than or equal to a certain threshold, with a default of 0.

95. Step S54: Organize the output of steps S51-S53 into a knowledge enhancement context (KAC).

8. The translation method based on multi-agent cooperation and multi-stage optimization according to claim 2, characterized in that: step S6 further includes the following: When a document scores f_score in a certain domain and exceeds a certain threshold (0.7 by default), a domain-specific agent is instantiated to handle the domain's expression specifications. When the text length exceeds a certain limit (default 1024), a long document entity tracking agent is instantiated to maintain cross-paragraph name, organization name, and pronoun reference chains, preventing entity translation drift in long documents. When a document contains charts, instantiate a chart translation agent to perform the translation of the chart content; When a document contains code blocks, a code mixing agent is instantiated to handle the translation boundaries between code blocks and natural language in the document, preserving the original code and translating only comments and explanatory text. When the source language belongs to the low-resource language set, instantiate a low-resource language agent and activate a dedicated few-shot translation model.

9. The translation method based on multi-agent cooperation and multi-stage optimization according to claim 2, characterized in that: Step S7 further includes the following: Parallel Path 1: Direct Translation Agent Using the source language's syntactic structure as a reference, terminology constraints from TL and KAC are injected into the decoding process to generate a basic translation that preserves the original meaning; the decoding strategy is dynamically switched for different node types in the document structure tree. Parallel path P2 semantic translation agent: Driven by the semantic structure in step S4, it does not adhere to the original syntactic form and pursues the maximum semantic fidelity to the target language; it handles information misalignment and rearrangement caused by word order differences in the source language, semantic equivalent substitution of unnatural word-by-word translation, and cross-language conversion of metaphorical expressions and culturally loaded words. Parallel Path 3 Domain Translation Agent: By combining KAC domain knowledge with specialized intelligent agent resources dynamically generated by step S6, candidate translations with higher accuracy in the professional domain are generated; the domain-specific sentence structure norms provided by the S6 dynamic intelligent agent are utilized to handle domain-specific expressions that cannot be handled correctly by the literal translation path. Parallel Path 4: Smoothness Optimization Agent: With the naturalness of the target language expression as the primary optimization goal, fine-tuning is performed based on a large-scale target language monolingual corpus to generate the most fluent candidate translations; The target language perplexity level (PPL) was used as the optimization metric. Parallel Path 5: Style Calibration Agent Based on the style parameter in the translation strategy of step S3, candidate translations with clear style objectives are generated independently; The following target style types are supported: Academic writing: objective statement tone, passive voice, and conformity to academic writing norms; Journalistic news reporting: concise, lively, and highlighting the core information; Technical documentation: precise and concise, retaining the original text of technical terms; Legal documents: formal legal language, maintaining precise clause structure; Literary works: Emphasize the beauty of language and appropriately employ paraphrasing techniques.

10. The translation method based on multi-agent cooperation and multi-stage optimization according to claim 2, characterized in that: Step S9 further includes the following: Step S91: Calculate the cross-language semantic fidelity score (semantic_score) for semantic consistency. semantic_score = cos(CLSE(src), CLSE(tgt)); Wherein, `semantic_score` is the semantic consistency score of the candidate translation; `src` is the source language text to be translated; `tgt` is the candidate target language translation; `CLSE(·)` is a cross-lingual semantic encoding function that maps source or target language text to the same multilingual semantic vector space, implemented based on LaBSE; `cos(·, ·)` is a cosine similarity function with a value range of [0,1]. The closer it is to 1, the more faithful the translation is to the semantics of the original text; when it is below the fidelity threshold, the semantic alignment refinement operation is triggered by a default value of 0.

85. Step S92, fluency_score language fluency calculation, based on adaptive normalized perturbation of the input baseline: fluency_score = 1 - [PPL(tgt) - PPL_min] / [PPL_max - PPL_min]; Here, `fluency_score` is the fluency score of the candidate translation; the closer it is to 1, the more natural and fluent the translation. `PPL(tgt)` is the perplexity score calculated by a large-scale language model of the target language for the candidate translation `tgt`; the lower the score, the more the translation conforms to the target language's habits. This model is trained on a single-language corpus of the target language containing at least 10 billion words. `PPL_min` is the lower bound of the perplexity score of high-quality translations in the current domain, dynamically calibrated based on `D_base`. `PPL_max` is the upper bound of the perplexity score of the reference corpus in the current domain, also dynamically calibrated based on `D_base`. Together, they form an adaptively normalized reference interval to ensure a fair and consistent fluency evaluation standard for different types of documents. When `PPL(tgt) ≤ PPL_min`, `fluency_score` = 1; when `PPL(tgt) ≥ PPL_max`, `fluency_score` = 0. Step S93: Calculate the terminology_score, combining the system terminology accuracy with the user-locked terminology compliance rate. terminology_score = λ · (|T_correct| / |T_total|) + (1-λ) · (|T_locked_complied| / |T_locked_total|); Wherein, terminology_score is the comprehensive terminology accuracy score of the candidate translation, with a value range of [0,1]; λ is the weight balance coefficient between the system terminology accuracy and the user-locked term compliance rate, with a default value of 0.

6. The more user-locked terms, the lower the value of λ can be to increase the priority of locked terms; |T_correct| is the number of times terms consistent with the million-level terminology database appear in the candidate translation, obtained through cross-validation of the system terminology knowledge base; |T_total| is the total number of times all controlled terms appear in the candidate translation, which, together with |T_correct|, is used to calculate the system terminology accuracy; |T_locked_complied| is the number of times locked terms that comply with user-defined locking rules appear in the candidate translation, with locked terms having a higher priority than system knowledge base recommendations; |T_locked_total| is the total number of times user-defined locked terms appear in the candidate translation, which, together with |T_locked_complied|, is used to calculate the locked term compliance rate; Step S94: Calculate grammar_score for grammatical correctness, using a syntax tree parsing tool to evaluate the grammatical compliance of the translation; Step S95, style_consistency_score style consistency calculation: style_consistency_score = 1 -norm( Style_drift); Style_drift = (1 / N)·Σᵢ‖Style_vec(pᵢ) - Style_mean‖ 2 ; Where style_consistency_score is the global style consistency score of the candidate translations; Style_drift is the average L2 distance squared between the style vectors of each paragraph and the mean style vector, which characterizes the dispersion of the style of the whole text. The larger the value, the more severe the style drift; N is the total number of paragraphs in the candidate translations; p i For the i-th paragraph text in the candidate translation, the unit is the Paragraph node in the document structure tree; Style_vec(p i ) represents the i-th paragraph p i The style feature vector is extracted from the paragraph text by a style classifier, encoding formality, tone, and stylistic features. The style classifier is trained on a publication-grade corpus. Style_mean is the mean vector of style vectors across all paragraphs in the entire text, representing the overall average style of the current candidate translation. 2 The square of the L2 norm of the vector is used to calculate the squared Euclidean distance between the style vector of each paragraph and the mean style vector; norm is used for normalization. Step S96, Comprehensive score calculation formula: FULLScore = α_sem·semantic + α_flu·fluency + α_term·terminology + α_gram·grammar + α_sty·style_consistency; Among them, FULLScore is the weighted comprehensive quality score of the candidate translations, with a value range of [0,1]. The highest score in the candidate pool is selected as the best candidate. α_sem is the weight coefficient for semantic consistency, with a default value of 0.30; α_flu is the weight coefficient for language fluency, with a default value of 0.20; α_term is the weight coefficient for terminology accuracy, with a default value of 0.25, which can be increased to 0.35 for terminology-intensive documents; α_gram is the weight coefficient for grammatical correctness, with a default value of 0.15; and α_sty is the weight coefficient for style consistency, with a default value of 0.10, which can be dynamically adjusted according to the translation scenario.

Images