Translation method and device, storage medium and program product

Abstract

The invention discloses a translation method and device, a storage medium and a program product, and relates to the technical field of artificial intelligence. According to the method, firstly, a to-be-translated text is translated through different translation models, candidate translations obtained through translation of all the translation models are obtained, and semantic graphs of all the candidate translations are constructed; then, determining a node set of semantic matching among the semantic graphs, and determining a target node in the node set; meanwhile, in the semantic relation edge of each semantic graph, determining a target edge between the target nodes; and finally, constructing a fusion graph according to the target node and the target edge, and generating a target translation of the to-be-translated text according to the fusion graph. The technical problem of low translation accuracy depending on a single translation engine at present is solved, and the translation accuracy of the to-be-translated text is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to translation methods, translation devices, storage media, and computer program products. Background Technology

[0002] With increasingly frequent global exchanges, the demand for professional document translation is growing rapidly, especially in fields such as business and law, where translation quality directly impacts the effectiveness of collaborations. However, current mainstream translation solutions still have significant limitations. Specifically, both commercial translation systems and open-source translation models largely rely on a single translation engine. This single-model architecture has inherent limitations in its capabilities, and its performance bottlenecks are difficult to overcome, making it impossible to simultaneously meet multiple requirements such as terminology accuracy, language fluency, and style adaptability. For example, general-purpose models perform poorly in translating specialized domain terminology, while specialized domain models often lack in terms of language naturalness.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this application is to provide a translation method, translation device, storage medium, and computer program product, which aims to solve the technical problem of low translation accuracy when relying on a single translation engine.

[0005] To achieve the above objectives, this application proposes a translation method, which includes:

[0006] Get the text to be translated; The text to be translated is translated using different translation models to obtain candidate translations from each model, and a semantic graph of each candidate translation is constructed. Determine the set of nodes that are semantically matched among the semantic graphs, and identify the target node in the set of nodes; In the semantic relation edges of each of the semantic graphs, determine the target edges between the target nodes; A fusion graph is constructed based on the target node and the target edge, and the target translation of the text to be translated is generated based on the fusion graph.

[0007] In one embodiment, the method further includes: Obtain capability description data for each translation model in the heterogeneous model pool, and determine the richness of the capability description data; Determine the routing strategy corresponding to the richness level, and select different translation models from the heterogeneous model pool to translate the text to be translated based on the routing strategy.

[0008] In one embodiment, the step of determining the target node in the node set includes: Determine the node information of each node in the node set, wherein the node information includes the authority of the node source translation model, the local consistency of the semantic graph in which the node is located, and the matching degree between the node content and the terminology database; The confidence level of each node is determined based on the node information, and the target node is determined from each node in the node set based on the confidence level.

[0009] In one embodiment, the step of determining the target edges between the target nodes in the semantic relationship edges of each of the semantic graphs includes: In the semantic relation edges of each of the semantic graphs, determine the candidate edges that the target node has been connected to; Based on the number and / or confidence level of the candidate edges, the target edges between the target nodes are determined from the candidate edges.

[0010] In one embodiment, the method further includes: Establish a special evaluation model committee, wherein the special evaluation model committee includes multiple special evaluators; The target translation is evaluated by each of the specialized evaluators in the specialized evaluation model committee, and the preliminary evaluation results of each specialized evaluator are obtained. By integrating the preliminary evaluation results of each of the aforementioned specialized evaluators, the comprehensive evaluation result of the specialized evaluation model committee on the target translation is obtained.

[0011] In one embodiment, after the step of obtaining the comprehensive evaluation result of the target translation by the special evaluation model committee by fusing the preliminary evaluation results of each of the special evaluation devices, the method further includes: Based on the comprehensive evaluation results of the special evaluation model committee, translation suggestions are generated for each translation model to translate the text to be translated. Referring to the translation prompts, the text to be translated is retranslated using each translation model, and a new fusion graph is constructed based on the retranslated text to generate a new target translation.

[0012] In one embodiment, after the step of obtaining the comprehensive evaluation result of the target translation by the special evaluation model committee by fusing the preliminary evaluation results of each of the special evaluation devices, the method further includes: Based on the comprehensive evaluation results of the special evaluation model committee, the translation performance analysis results are determined; Based on the translation performance analysis results, the routing strategy for selecting different translation models in the heterogeneous model pool is dynamically adjusted.

[0013] Furthermore, to achieve the above objectives, this application also proposes a translation method, which includes: The first module is used to obtain the text to be translated; The second module is used to translate the text to be translated using different translation models, obtain candidate translations from each translation model, and construct semantic graphs for each candidate translation. The third module is used to determine the set of nodes that are semantically matched among the semantic graphs, and to determine the target node in the set of nodes; The fourth module is used to determine the target edges between the target nodes in the semantic relationship edges of each of the semantic graphs; The fifth module is used to construct a fusion graph based on the target node and the target edge, and to generate the target translation of the text to be translated based on the fusion graph.

[0014] In addition, to achieve the above objectives, this application also proposes a translation device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the translation method as described above.

[0015] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the translation method described above.

[0016] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the translation method described above.

[0017] One or more technical solutions proposed in this application have at least the following technical effects: In this application, firstly, the text to be translated is translated using different translation models to obtain candidate translations from each model, and a semantic graph of each candidate translation is constructed. Then, a set of nodes that semantically match between the semantic graphs is determined, and target nodes are identified in the node set. At the same time, target edges between target nodes are identified in the semantic relationship edges of each semantic graph. Finally, a fusion graph is constructed based on the target nodes and target edges, and the target translation of the text to be translated is generated based on the fusion graph.

[0018] In other words, by fusing candidate translations obtained from different translation models based on a semantic graph, and generating the target translation of the text to be translated based on the fused graph, the translation capabilities of different translation models can be fully utilized, taking into account their strengths and advantages. This solves the technical problem of low translation accuracy when relying on a single translation engine, and improves the translation accuracy of the text to be translated. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a first flowchart illustrating the first embodiment of the translation method of this application. Figure 2 This is a second flowchart illustrating the first embodiment of the translation method of this application. Figure 3 This is a third flowchart illustrating the first embodiment of the translation method of this application; Figure 4 This is a schematic diagram of the fourth process provided in the first embodiment of the translation method of this application; Figure 5 This is a fifth flowchart illustrating the second embodiment of the translation method of this application; Figure 6 A sixth flowchart provided for the second embodiment of the translation method of this application; Figure 7 This is a schematic diagram of the seventh process provided for the second embodiment of the translation method of this application; Figure 8 This is a schematic diagram of the module structure of the translation device according to an embodiment of this application; Figure 9 This is a schematic diagram of the device structure of the hardware operating environment involved in the translation method in the embodiments of this application.

[0022] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0024] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or translation device capable of performing the above functions. The following description uses a translation device as an example to illustrate this embodiment and the subsequent embodiments.

[0025] Based on this, the embodiments of this application provide a translation method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the translation method of this application.

[0026] In this embodiment, the translation method includes steps S10 to S40: Step S10: Obtain the text to be translated; After receiving input documents in various formats such as PDF, Word, audio, and images, text extraction and standardization can be achieved by integrating technologies such as Optical Character Recognition (OCR) and Automatic Speech Recognition (ASR). Specifically, for scanned documents, the OCR module can not only recognize text but also parse the layout structure and distinguish elements such as titles, body text, and chart titles. For audio input, the ASR module can recognize paralinguistic information such as speaker, speech rate, and pauses, which helps in subsequent tone and style translation. The output after the aforementioned preprocessing is a structured, unified text data stream, which can be used as the content to be translated.

[0027] Furthermore, a pre-trained language model can be used to intelligently segment the content to be translated. This pre-trained language model has been trained on a large number of documents and can recognize semantic boundaries rather than simple line breaks. For example, when encountering transition words such as "However" or "on the other hand," it will consider semantic segmentation even within the same paragraph to ensure translation coherence.

[0028] Furthermore, it can retrieve relevant domain knowledge in real time from a customized terminology database based on RAG technology. The terminology database employs a multi-layered index structure, supporting multi-dimensional retrieval of synonyms, near-synonyms, and related concepts. For example, when translating "transformer," it can automatically distinguish between "transformer" in the power industry and "Transformer architecture" in the AI ​​field based on the context. The terminology database supports professional terminology from multiple fields, allowing users to customize and dynamically update it. Update mechanisms include manual addition, batch import, and automatic extraction of new terms from high-quality translation results.

[0029] Step S20: Translate the text to be translated using different translation models to obtain candidate translations from each model, and construct a semantic graph of each candidate translation. After obtaining the text to be translated through document input, intelligent preprocessing, paragraph segmentation, and knowledge enhancement in step S10, the text is translated using different translation models, such as general large language model, domain expert model, and language pair optimization model, to obtain candidate translations from each model.

[0030] In one feasible implementation, refer to Figure 2 The method further includes steps A11-A12: Step A11: Obtain the capability description data of each translation model in the heterogeneous model pool, and determine the richness of the capability description data; Step A12: Determine the routing strategy corresponding to the richness level, and select different translation models from the heterogeneous model pool to translate the text to be translated based on the routing strategy.

[0031] In this embodiment, a heterogeneous model pool is constructed, comprising diverse models such as general-purpose large language models, domain expert models, and language pair optimization models. Each translation model in the heterogeneous model pool has its unique advantages and applicable scenarios: general-purpose large language models, such as GPT-4 and Claude, excel at understanding complex contexts and generating fluent language, but may lack precision in specific domain terminology; domain expert models are specifically fine-tuned for specific domains (such as medicine, law, and finance), providing accurate terminology but exhibiting less linguistic flexibility; and language pair optimization models are specifically optimized for specific language pairs (such as Chinese-English and Japanese-English), capable of handling unique language expression habits.

[0032] Simultaneously, a dynamically updated multi-dimensional capability profile is established for each translation model in the heterogeneous model pool. The capability description data used to construct the capability profile includes: 1. Domain adaptability: quantified domain professional capability score (e.g., finance 0.95, medical 0.65); 2. Historical translation quality: average score based on past quality arbitration; 3. Response speed: average response time and standard deviation; 4. Cost-effectiveness: computational or economic cost per call; 5. Style matching: style tendency of the generated text (formal, colloquial, academic, etc.).

[0033] Furthermore, based on the richness of the aforementioned capability description data, a routing strategy for selecting translation models from the heterogeneous model pool is determined. When there is a lack of sufficiently rich capability description data, making it difficult to construct a multi-dimensional capability profile of the translation model, an online learning algorithm is used for exploratory routing selection. The following algorithms can be employed: using an ε-greedy strategy and the UCB algorithm to balance the opportunity to utilize known high-performance models with the opportunity to explore potentially better models; using the LinUCB algorithm to perform intelligent predictions based on contextual information such as text features and user needs; the EXP3 algorithm, which is robust in adversarial environments where model performance may change over time; and Thompson sampling, which establishes a probability distribution for each model's performance based on Bayesian methods, supporting uncertainty quantification. In other words, the routing strategy in this case is to select a translation model from the heterogeneous model pool using the aforementioned online learning algorithm.

[0034] Once sufficient capability description data has been accumulated to construct a multi-dimensional capability profile of the translation model, a specialized routing prediction model can be trained. This can be achieved by using features from historical translation tasks and final quality score labels to train a BERT-based sequence classification model or a tree model like XGBoost, or by training a model using reinforcement learning algorithms such as DQN or PPO based on the long-term cumulative reward of the quality arbitration score. In other words, the routing strategy at this point is to select a translation model from a heterogeneous model pool using the aforementioned routing prediction model.

[0035] In one embodiment, a mechanism is provided for switching routing strategies between the aforementioned online learning algorithm and routing prediction model. Specifically, the accumulation status of capability description data is monitored in real time, and the routing strategy is automatically switched. For example, when the capability description data recorded in historical translation tasks exceeds a threshold, the training of the routing prediction model is started. The performance of the online learning algorithm and the routing prediction model are compared periodically, and the routing strategy is automatically switched based on the performance. The routing strategy of the online learning algorithm is gradually transitioned to a hybrid routing strategy that uses both the online learning algorithm and the routing prediction model, and finally the routing strategy of the routing prediction model is adopted.

[0036] In step S20, the candidate translations obtained by each translation model are converted into structured graph representations, i.e., semantic graphs, to explicitly capture the semantic and grammatical relationships between words. Specifically, semantic dependency analysis is performed on the candidate translations to construct their semantic graphs. A semantic dependency analyzer based on deep neural networks, such as a BERT-based biaffine attention parser, can be used to analyze each candidate translation, identifying each word or phrase in the candidate translation and the dependency relationships between them. This outputs a directed graph, or semantic graph, where nodes are words or basic phrases, and edges represent semantic dependency relationships between nodes, such as subject, object, core, and modifier.

[0037] In one application scenario, for the candidate translation "Scientists carefully isolated the compound", semantic dependency analysis may generate the following relations: isolation -- core -- ROOT; scientists -- subject -- isolation; carefully -- modifier -- isolation; compound -- object -- isolation. Each word ("scientists", "isolation", etc.) will be converted into a high-dimensional semantic vector.

[0038] In one embodiment, the semantic graph can also be enriched. Specifically, when performing node enhancement, a semantic embedding vector is calculated for each node, using models such as BERT or Word2Vec, which encodes the semantic information of words. When performing edge enhancement, each dependency edge is assigned a predefined type label and weight, the weights of which can be determined based on the confidence score of the semantic dependency parser.

[0039] Step S30: Determine the set of nodes that are semantically matched between the semantic graphs, and determine the target node in the set of nodes; In one embodiment, by aligning multiple semantic graphs from different candidate translations, the semantically corresponding parts between the different candidate translations are found, that is, the set of semantically matched nodes between the semantic graphs, laying the foundation for subsequent optimal fusion.

[0040] First, node similarity is calculated using either a Graph Attention Network (GAT) or a Graph Convolutional Network (GCN). Each semantic graph is input into GAT / GCN. Taking GAT as an example, its attention mechanism calculates the association strength between each node in the semantic graph and its neighboring nodes, thereby generating a richer node representation that incorporates local graph structure information. For node u from semantic graph A and node v from semantic graph B, their similarity is obtained by calculating the cosine similarity of their representation vectors after GAT enhancement: Sim(u,v) = cosine(GAT(u), GAT(v)).

[0041] Next, cross-graph semantic alignment is performed. Based on the node similarity calculated in the previous step, a greedy matching or maximum weight matching algorithm, such as the Hungarian algorithm, is used to find the optimal node correspondence between the two semantic graphs, that is, the set of semantically matched nodes between the semantic graphs. Specifically, the process is to find a set of node pairings that maximizes the sum of the semantic similarity of the paired nodes. In this way, a semantic bridge is established between words in different candidate translations.

[0042] In one application scenario, the semantic graph of candidate translation A contains the node "scientist," and the semantic graph of candidate translation B contains the node "researcher." After GAT enhancement and similarity calculation, Sim(scientist, researcher) scores very high. The cross-graph semantic alignment algorithm then matches these two nodes as semantically corresponding nodes and adds them to the set of semantically matched nodes. It can be understood that a semantically corresponding node can be a node in the semantic graph of candidate translation A and a node in the semantic graph of candidate translation B, or it can be a node in the semantic graphs of candidate translation A, candidate translation B, candidate translation C, or even nodes from multiple candidate translation semantic graphs.

[0043] In one feasible implementation, refer to Figure 3 Step S30 may include steps B11-B12: Step B11: Determine the node information of each node in the node set. The node information includes the authority of the node source translation model, the local consistency of the semantic graph where the node is located, and the matching degree between the node content and the terminology database. Step B12: Determine the confidence level of each node based on the node information, and determine the target node from among the nodes in the node set based on the confidence level.

[0044] Based on the aforementioned alignment results, i.e. the set of nodes that semantically match between semantic graphs, all candidate graphs are examined from a unified and integrated perspective, and the optimal combination of semantic units is selected, i.e., the target node and target edge are determined, and the final translation is generated.

[0045] In this embodiment, for each pair of semantically matched nodes in the node set, a representative node is selected as the target node and added to the fusion graph according to an optimal selection strategy. The optimal selection strategy may be to select the node with the highest confidence level. This confidence level is determined by various factors, such as the node's information. This information may include: the authority of the node's source translation model, i.e., the suitability of the translation model from which the node originates in this domain; the local consistency of the semantic graph in which the node resides, i.e., the stability of the relationship between the node and its neighboring nodes, which can be characterized by the aforementioned similarity; and the matching degree between the node's content and the terminology database, i.e., if the node is a term, its degree of matching with the standard translations in the terminology database.

[0046] Step S40: In the semantic relation edges of each semantic graph, determine the target edges between target nodes; When constructing the fusion graph, firstly, a new, empty semantic graph is created as the fusion graph; then, all aligned node pairs, i.e., the set of semantically matched nodes, are traversed. For each pair of aligned nodes, a target node is selected and added to the fusion graph according to the aforementioned selection strategy; finally, the selection of edges (dependencies) used to construct the fusion graph is similar to the selection of target nodes. Dependencies shared by most high-quality candidate graphs or with the highest confidence are selected as target edges.

[0047] In one feasible implementation, refer to Figure 4 Step S40 includes steps C11 to C12: Step C11: In the semantic relation edges of each semantic graph, determine the candidate edges that the target node has been connected to. Step C12: Determine the target edges between the target nodes from the candidate edges based on the number of candidate edges and / or the confidence level.

[0048] In this embodiment, a method for determining target edges between target nodes in the semantic relation edges of each semantic graph is proposed. First, candidate edges that have connected the target nodes are determined in the semantic relation edges of each semantic graph. For example, the nodes in the semantic graph of candidate translation A include A1, A2, A3, A4, and A5; the nodes in the semantic graph of candidate translation B include B1, B2, B3, B4, and B5; and the nodes in the semantic graph of candidate translation C include C1, C2, C3, C4, and C5. Referring to Table 1 below, the set of semantically matched nodes includes A1-B1-C2, A2-B2-C3, A3-B3-C4, and A4-B4. That is, assuming that the semantic relations of the nodes in candidate translations A, B, and C are sequentially connected, then some statements A1-A2 in candidate translation A correspond to some statements B1-A2 in candidate translation B. B2 also corresponds to some statements C2-C3 in candidate translation C; similarly, some statements A2-A3 in candidate translation A correspond to some statements B2-B3 in candidate translation B, and similarly correspond to some statements C3-C4 in candidate translation C; similarly, some statements A3-A4 in candidate translation A correspond to some statements B3-B4 in candidate translation B. In the above, A1-A2, B1-B2, and C2-C3 are candidate edges that have been connected to target nodes A1, B1, or C2; A1-A2, B1-B2, C2-C3, A2-A3, B2-B3, and C3-C4 are candidate edges that have been connected to target nodes A2, B2, or C3; and A2-A3, B2-B3, C3-C4, A3-A4, and B3-B4 are candidate edges that have been connected to target nodes A3, B3, or C4. Then, based on the number and / or confidence of each candidate edge, the target edge between the target nodes can be determined from the candidate edges. The candidate edge with the largest number or the highest confidence can be taken as the target edge between the target nodes.

[0049]

[0050] Table 1 Step S50: Construct a fusion graph based on the target node and the target edge, and generate a target translation of the text to be translated according to the fusion graph.

[0051] After constructing the fusion graph based on the target node and the target edge, the fusion graph can be linearized and surfaceized to generate a target translation of the text to be translated.

[0052] Linearizing and surfaceizing the fusion graph, as a generation task from structure to sequence, can adopt a constraint-based search algorithm or a lightweight sequence-to-sequence model to convert the fusion graph back into a grammatically correct and fluent sentence such as a Chinese sentence. It can be understood that this process needs to ensure that the order of words conforms to the grammar rules of the target language, and function words such as auxiliary words and prepositions need to be added correctly. The final target translation obtained needs to be faithful to the semantic structure of the fusion graph.

[0053] In an application scenario, the source text is "The scientist isolated the compound carefully.", and the candidate translations include Translation A from a general model: "The scientist isolated the compound carefully.", and Translation B from a domain expert model: "The researcher isolated the compound carefully." The following is the fusion process: 1. Graph construction: Construct semantic dependency graphs for A and B respectively. 2. Graph alignment: Establish corresponding relationships through GAT and node matching: scientist <-> researcher, isolated <-> isolated out, carefully <-> carefully, compound <-> compound. 3. Optimal fusion: "scientist" vs "researcher": The domain expert model is more reliable in academic texts, and "researcher" is more formal, so "researcher" is selected; "isolated" vs "isolated out": "isolated out" more accurately conveys the perfect tense of "isolate out", so "isolated out" is selected; "carefully" vs "carefully": In a scientific context, "carefully" better reflects the rigor of the operation, so "carefully" is selected; "compound": They are the same, so it is retained. 4. Linearization: According to the semantic structure of the fusion graph (the researcher as the subject, isolated out as the core predicate, carefully as the adverbial modifier, and compound as the object), generate the final translation, that is, the fusion result: "The researcher isolated the compound carefully." Furthermore, this embodiment also provides an intelligent conflict resolution mechanism to detect and resolve various types of conflicts during the fusion process of candidate translations. It is understood that this fusion process includes, but is not limited to, terminological conflicts: different translations of the same concept, such as translating "AI" as "artificial intelligence" or directly using the original "AI"; grammatical conflicts: different syntactic structures leading to subtle differences in meaning; and style conflicts: differences in formality and tone. Therefore, in specific conflict resolution strategies, different conflicting translations can be selected based on multi-dimensional confidence scores. Confidence scores can be calculated based on factors such as the historical accuracy of the translation model in the domain, the authoritative support of the terminology database, contextual consistency checks, and user preference settings. The translation with the highest confidence score is then selected from the conflicting translations.

[0054] In summary, this embodiment transforms the translation fusion problem into a structured semantic graph optimization problem through the collaborative work of semantic dependency analysis, graph attention networks, and graph matching algorithms. It performs deep fusion at the semantic level rather than the surface string level, thereby generating results superior to any single candidate translation. This enables the generation of high-quality translations optimized across multiple dimensions such as terminology, style, and syntax, achieving a true "1+1>2" fusion effect.

[0055] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 5 The translation method further includes steps T10 to T30: Step T10: Construct a special evaluation model committee, which includes multiple special evaluators; Step T20: The target translation is evaluated by each specialized evaluator in the specialized evaluation model committee to obtain the preliminary evaluation results of each specialized evaluator; Step T30: By integrating the preliminary evaluation results of each specialized evaluator, the comprehensive evaluation results of the specialized evaluation model committee on the target translation are obtained.

[0056] Currently, the quality assessment of existing translation systems typically relies on a single judging model. This single model has inherent knowledge gaps and capability limitations, failing to simultaneously possess the professional judgment capabilities to assess multiple dimensions such as terminology accuracy, grammatical fluency, logical consistency, and style suitability. It cannot provide comprehensive evaluation opinions from different professional perspectives, meaning it cannot offer comprehensive and professional improvement suggestions from multiple dimensions such as terminology, grammar, logic, and style. Consequently, the improvement suggestions it provides are often one-sided or inaccurate, ultimately severely limiting the improvement of translation quality and failing to meet the translation needs of high-end professional scenarios. For example, in the translation of professional documents, a single judging model may fail to identify terminology errors in specific fields or provide revision suggestions that conform to professional standards. In other words, the professionalism and comprehensiveness of the suggestions provided by current single judging models are insufficient, lacking a comprehensive evaluation based on multi-dimensional professional advice.

[0057] To address this, this embodiment first proposes a specialized evaluation model committee with a modular and scalable architecture, comprising multiple specialized evaluators. For example, the terminology accuracy evaluator, driven by a domain knowledge graph and terminology database, specifically checks the accuracy of translated professional terms; the grammatical fluency evaluator, trained on large-scale grammatical norms, assesses language naturalness and grammatical correctness; the logical consistency evaluator utilizes a logical reasoning model to ensure the translated text's logic is consistent with the source text; the style fit evaluator, based on a style classification model, checks whether it conforms to the required stylistic requirements; and the cultural appropriateness evaluator, by embedding cultural knowledge, assesses the appropriate handling of cultural elements. Then, each specialized evaluator works independently and provides a score and comments for evaluating the target translation, serving as the preliminary evaluation results for each specialized evaluator. Finally, through weighted voting or a meta-evaluation model, the preliminary evaluation results of each specialized evaluator are merged to generate a comprehensive evaluation result of the specialized evaluation model committee for the target translation. This design ensures diversity in evaluation perspectives, avoids the knowledge blind spots and biases of a single model, and fundamentally overcomes the limitations of a single judge model.

[0058] In one feasible implementation, refer to Figure 6 Step T30 may be followed by steps D11-D12: Step D11: Based on the comprehensive evaluation results of the special evaluation model committee, generate translation suggestions for each translation model for translating the text to be translated. Step D12: Referring to the translation prompts, the text to be translated is retranslated using each translation model. Based on the retranslated text, a new fusion graph is constructed, and a new target translation is generated.

[0059] Most existing translation systems adopt the "one-time translation and direct output" mode, lacking a deep quality assessment of translation results and an iterative optimization mechanism based on the assessment results. Even if some translation systems introduce simple quality assessment, it is only limited to giving scores or simple evaluations and cannot provide specific and operable improvement guidance.

[0060] It can be understood that the comprehensive assessment result of the special assessment model committee is a structured detailed report, including the overall quality score, dimension sub-item scores, specific problem list, improvement suggestions, confidence level, and executable feedback.

[0061] In an application scenario, for the sentence "Our product achieves state-of-the-art performance" in a technical document, the initial translation is "我们的产品实现了最先进的性能". The special assessment model committee conducts a joint review from different dimensions. Among them, the term evaluator points out that "state-of-the-art" should be translated as "顶尖" or retained in English in technical documents, and it is rated as "term inaccuracy"; the style evaluator believes that "最先进的性能" has a relatively strong colloquial color, and it is recommended to use a more formal "实现了顶尖 / state-of-the-art的性能表现" in technical documents. Finally, the detailed report of the comprehensive assessment result is: summarizing the opinions of each expert and generating detailed guidance including the specific modification positions, recommended replacement words, and reasons for modification. Subsequently, the comprehensive improvement suggestions generated by the special assessment model committee are automatically converted into precise translation tips. The following is an example of tip conversion: "Please re-translate the following sentence to ensure: 1. The professional term'state-of-the-art' is retained in English or translated as '顶尖'; 2. The overall language style maintains the formality and rigor of technical documents. Original text: Our product achieves state-of-the-art performance". Finally, the new target translation is: "我们的产品实现了state-of-the-art性能".

[0062] In an embodiment, by analyzing the problem patterns in historical translations, improvement tips targeted at the translation model are generated to customize personalized optimization guidance for different translation models. In an application scenario, personalized optimization is implemented according to the characteristics of different translation models. Among them, for general large models: strengthen term accuracy and domain knowledge guidance; for domain expert models: improve language fluency and sentence pattern diversity; for language pair optimization models: strengthen cultural adaptation and habitual expressions.

[0063] In a feasible implementation manner, referring to Figure 7 , after step T30, steps E11~E12 can be included: Step E11: Determine the translation performance analysis results based on the comprehensive evaluation results of the special evaluation model committee; Step E12: Based on the translation performance analysis results, dynamically adjust the routing strategy for selecting different translation models in the heterogeneous model pool.

[0064] In this embodiment, based on the comprehensive evaluation results of the special evaluation model committee, the translation performance analysis results, such as the dimensional component scores in the comprehensive evaluation results, are determined, and the routing strategy for selecting different translation models in the heterogeneous model pool is dynamically adjusted. For example, performance monitoring is performed to track the performance of each translation model in different scenarios in real time, identify and discover the optimal model selection mode, and dynamically adjust the model selection weights used in the routing strategy based on the performance.

[0065] In one application scenario, based on the comprehensive evaluation results of the special evaluation model committee, a fine-grained quality evaluation system is constructed. That is, a multi-dimensional quality vector Q = [terminology score, grammar score, logic score, style score, culture score, semantic completeness, domain specialization, ...] is established. Based on the quality vector Q, specific weak areas of each translation model are identified, providing direction for targeted optimization. The changing trends of each dimension of the quality vector Q are tracked to evaluate the effectiveness of optimization measures.

[0066] In one embodiment, when performing multi-dimensional optimization based on the quality vector Q, a contrastive learning-driven improvement mechanism can be adopted. Through systematic comparative analysis, specific improvement guidance is generated, including: high-quality translation analysis: selecting the translation with the best performance in each dimension of the quality vector as a reference standard; gap identification: comparing and analyzing the differences between high-quality translations and translations to be improved in specific dimensions; improvement scheme generation: transforming the results of the comparative analysis into specific optimization prompts and modification suggestions; and pattern extraction: summarizing the common characteristics of high-quality translations to form an optimization knowledge base.

[0067] Reinforcement learning can also be used for route optimization, which involves continuously optimizing route decisions based on quality feedback. The state space is designed as text features + user needs + comprehensive evaluation results + historical quality performance. The action space is designed as: model selection combination + model parameter configuration of translation model + fusion strategy selection. The reward mechanism is designed as: R = quality arbitration weight × quality arbitration score (overall quality score included in the structured detailed report generated by quality assessment) + quality vector consistency weight × quality vector consistency – historical penalty weight × historical translation failure penalty. Thus, the routing strategy is continuously optimized based on cumulative rewards, realizing the iterative update of the routing strategy.

[0068] It should be noted that, among these, quality vector consistency refers to the similarity between the expected quality vector initially assessed when the routing strategy selects the translation model and the actual quality vector obtained by the special evaluation model committee after evaluating the translation results.

[0069] Furthermore, clustering algorithms can be used to identify systematic failure patterns, such as domain-specific terminology errors and grammatical structure problems. This allows for the recording of translation failure cases and the analysis of their characteristics and root causes. Simultaneously, these failure cases can be transformed into training data to optimize routing strategy models.

[0070] Based on the comprehensive evaluation results of the special evaluation model committee, a complete feedback optimization loop is constructed to achieve continuous improvement of the translation model and routing strategy.

[0071] In summary, this application provides an intelligent translation method based on multi-model competitive fusion and feedback-driven approaches. It establishes a complete feedback loop from quality assessment to model optimization, and can be widely applied to professional translation scenarios such as international business cooperation, cross-border technical document management, multilingual academic exchange, and legal document localization. By constructing an intelligent translation architecture based on multi-model competitive fusion and feedback-driven approaches, it achieves end-to-end optimization from terminology database enhancement, dynamic routing selection, semantic-level fusion to quality arbitration. Its innovative use of a specialized evaluation model committee mechanism overcomes the limitations of single-judge models, combining multi-granularity quality assessment and reinforcement learning optimization to significantly improve translation efficiency while ensuring the accuracy of professional terminology and stylistic suitability. It possesses continuous self-evolution capabilities, constantly optimizing model performance through iterative feedback learning, providing various industries with intelligent translation solutions that balance accuracy, professionalism, and efficiency, effectively reducing manual review costs and achieving intelligent control over the entire translation quality process.

[0072] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the translation method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0073] This application also provides a translation device, please refer to... Figure 8 The translation device includes: Module 10 is used to obtain the text to be translated. The second module 20 is used to translate the text to be translated using different translation models, obtain candidate translations from each translation model, and construct semantic graphs of each candidate translation. The third module 30 is used to determine the set of nodes that are semantically matched among the semantic graphs, and to determine the target node in the set of nodes; The fourth module 40 is used to determine the target edges between the target nodes in the semantic relationship edges of each of the semantic graphs; The fifth module 50 is used to construct a fusion graph based on the target node and the target edge, and to generate the target translation of the text to be translated based on the fusion graph.

[0074] In one embodiment, the translation device further includes a sixth module for: Obtain capability description data for each translation model in the heterogeneous model pool, and determine the richness of the capability description data; Determine the routing strategy corresponding to the richness level, and select different translation models from the heterogeneous model pool to translate the text to be translated based on the routing strategy.

[0075] In one embodiment, the third module 30 is further configured to: Determine the node information of each node in the node set, wherein the node information includes the authority of the node source translation model, the local consistency of the semantic graph in which the node is located, and the matching degree between the node content and the terminology database; The confidence level of each node is determined based on the node information, and the target node is determined from each node in the node set based on the confidence level.

[0076] In one embodiment, the fourth module 40 is further configured to: In the semantic relation edges of each of the semantic graphs, determine the candidate edges that the target node has been connected to; Based on the number and / or confidence level of the candidate edges, the target edges between the target nodes are determined from the candidate edges.

[0077] In one embodiment, the translation device further includes a seventh module for: Establish a special evaluation model committee, wherein the special evaluation model committee includes multiple special evaluators; The target translation is evaluated by each of the specialized evaluators in the specialized evaluation model committee, and the preliminary evaluation results of each specialized evaluator are obtained. By integrating the preliminary evaluation results of each of the aforementioned specialized evaluators, the comprehensive evaluation result of the specialized evaluation model committee on the target translation is obtained.

[0078] In one embodiment, the seventh module is further configured to: After the step of obtaining the comprehensive evaluation result of the target translation by the specialized evaluation model committee by integrating the preliminary evaluation results of each of the specialized evaluators: Based on the comprehensive evaluation results of the special evaluation model committee, translation suggestions are generated for each translation model to translate the text to be translated. Referring to the translation prompts, the text to be translated is retranslated using each translation model, and a new fusion graph is constructed based on the retranslated text to generate a new target translation.

[0079] In one embodiment, the seventh module is further configured to: After the step of obtaining the comprehensive evaluation result of the target translation by the specialized evaluation model committee by integrating the preliminary evaluation results of each of the specialized evaluators: Based on the comprehensive evaluation results of the special evaluation model committee, the translation performance analysis results are determined; Based on the translation performance analysis results, the routing strategy for selecting different translation models in the heterogeneous model pool is dynamically adjusted.

[0080] The translation apparatus provided in this application, employing the translation method described in the above embodiments, can solve the technical problem of low translation accuracy when relying on a single translation engine. Compared with the prior art, the beneficial effects of the translation apparatus provided in this application are the same as those of the translation method provided in the above embodiments, and other technical features in the translation apparatus are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0081] This application provides a translation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the translation method in Embodiment 1 above.

[0082] The following is for reference. Figure 9 The diagram illustrates a structural schematic of a translation device suitable for implementing embodiments of this application. The translation device in these embodiments may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 9 The translation device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of this application.

[0083] like Figure 9As shown, the translation device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the translation device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. The communication device 1009 allows the translation device to communicate wirelessly or wiredly with other devices to exchange data. Although the diagram shows translation devices with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented alternatively.

[0084] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0085] The translation device provided in this application, employing the translation method described in the above embodiments, can solve the technical problem of low translation accuracy when relying on a single translation engine. Compared with the prior art, the beneficial effects of the translation device provided in this application are the same as those of the translation method provided in the above embodiments, and other technical features of the translation device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0086] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0087] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0088] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the translation method in the above embodiments.

[0089] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0090] The aforementioned computer-readable storage medium may be included in the translation device; or it may exist independently and not be assembled into the translation device.

[0091] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by the translation device, the translation device causes the translation device to: acquire the text to be translated; translate the text to be translated using different translation models to obtain candidate translations obtained by each translation model, and construct a semantic graph of each candidate translation; determine the set of nodes that semantically match between the semantic graphs, and determine the target node in the node set; determine the target edge between the target nodes in the semantic relationship edges of each semantic graph; construct a fusion graph based on the target node and the target edge, and generate the target translation of the text to be translated based on the fusion graph.

[0092] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0093] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0094] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0095] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for performing the above-described translation method, thereby solving the technical problem of low translation accuracy when relying on a single translation engine. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the translation method provided in the above embodiments, and will not be repeated here.

[0096] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the translation method described above.

[0097] The computer program product provided in this application can solve the technical problem of low translation accuracy when relying on a single translation engine. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the translation method provided in the above embodiments, and will not be repeated here.

[0098] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A translation method, characterized in that, The translation method includes: Get the text to be translated; The text to be translated is translated using different translation models to obtain candidate translations from each model, and a semantic graph of each candidate translation is constructed. Determine the set of nodes that are semantically matched among the semantic graphs, and identify the target node in the set of nodes; In the semantic relation edges of each of the semantic graphs, determine the target edges between the target nodes; A fusion graph is constructed based on the target node and the target edge, and the target translation of the text to be translated is generated based on the fusion graph.

2. The translation method as described in claim 1, characterized in that, The method further includes: Obtain capability description data for each translation model in the heterogeneous model pool, and determine the richness of the capability description data; Determine the routing strategy corresponding to the richness level, and select different translation models from the heterogeneous model pool to translate the text to be translated based on the routing strategy.

3. The translation method as described in claim 1, characterized in that, The step of determining the target node in the node set includes: Determine the node information of each node in the node set, wherein the node information includes the authority of the node source translation model, the local consistency of the semantic graph in which the node is located, and the matching degree between the node content and the terminology database; The confidence level of each node is determined based on the node information, and the target node is determined from each node in the node set based on the confidence level.

4. The translation method as described in claim 1, characterized in that, The step of determining the target edges between the target nodes in the semantic relation edges of each of the semantic graphs includes: In the semantic relation edges of each of the semantic graphs, determine the candidate edges that the target node has been connected to; Based on the number and / or confidence level of the candidate edges, the target edges between the target nodes are determined from the candidate edges.

5. The translation method as described in claim 1, characterized in that, The method further includes: Establish a special evaluation model committee, wherein the special evaluation model committee includes multiple special evaluators; The target translation is evaluated by each of the specialized evaluators in the specialized evaluation model committee, and the preliminary evaluation results of each specialized evaluator are obtained. By integrating the preliminary evaluation results of each of the aforementioned specialized evaluators, the comprehensive evaluation result of the specialized evaluation model committee on the target translation is obtained.

6. The translation method as described in claim 5, characterized in that, The step of obtaining the comprehensive evaluation result of the target translation by the special evaluation model committee by integrating the preliminary evaluation results of each of the special evaluation devices further includes: Based on the comprehensive evaluation results of the special evaluation model committee, translation suggestions are generated for each translation model to translate the text to be translated. Referring to the translation prompts, the text to be translated is retranslated using each translation model, and a new fusion graph is constructed based on the retranslated text to generate a new target translation.

7. The translation method as described in claim 5, characterized in that, The step of obtaining the comprehensive evaluation result of the target translation by the special evaluation model committee by integrating the preliminary evaluation results of each of the special evaluation devices further includes: Based on the comprehensive evaluation results of the special evaluation model committee, the translation performance analysis results are determined; Based on the translation performance analysis results, the routing strategy for selecting different translation models in the heterogeneous model pool is dynamically adjusted.

8. A translation device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the translation method as described in any one of claims 1 to 7.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the translation method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the translation method as described in any one of claims 1 to 7.

Images