An assisted translation method and system based on aerospace corpus fine-tuning

Abstract

This application relates to an auxiliary translation method and system based on aerospace-related corpus fine-tuning. The method includes: acquiring a bilingual text set in the aerospace field; constructing a bilingual parallel corpus based on the bilingual text set; conducting corpus analysis and extracting text translation features based on the bilingual parallel corpus; inputting the text translation features into a machine translation model for corpus fine-tuning training to generate an auxiliary translation model; acquiring the aerospace-related text to be translated; and using the auxiliary translation model to generate an auxiliary translation text sequence. This method can accurately adapt to aerospace-related text translation scenarios, improve the professionalism, accuracy, and efficiency of aerospace-related text translation, and solve the problem of insufficient adaptability of general translation models in the aerospace field.

Description

Technical Field

[0001] This invention belongs to the field of aerospace text machine translation, and in particular relates to an auxiliary translation method and system based on aerospace corpus fine-tuning. Background Technology

[0002] With the rapid development of natural language processing and machine translation technologies, general-purpose machine translation models trained on large-scale corpora have been widely applied. These technologies can handle translation tasks of everyday language and general texts, featuring fast translation speed and broad language coverage. In technology-intensive professional fields such as aerospace, a large number of technical documents containing specialized terminology, fixed expressions, and specific sentence structures are generated. Currently, translation tasks for such specialized texts mainly rely on general-purpose translation models, or are supplemented by human post-editing.

[0003] In traditional technologies, general translation models typically treat aerospace texts the same as general texts, failing to fully consider the specificities of this field in terms of vocabulary, syntax, and style. Meanwhile, purely human translation or post-editing methods heavily rely on the translator's expertise and experience, resulting in inefficiency, high costs, and difficulty in maintaining consistent quality.

[0004] However, current general-purpose translation models are not well-suited for the aerospace field, struggling to accurately understand and translate complex technical terms, abbreviations, and specific sentence structures. This results in significant deficiencies in the professionalism and accuracy of the translations, and may even lead to ambiguity. Traditional manual methods, on the other hand, cannot meet the demands of large-scale, high-efficiency translation. Therefore, there is an urgent need for a machine translation solution that can deeply integrate aerospace knowledge and accurately adapt to specialized scenarios. Summary of the Invention

[0005] Therefore, it is necessary to provide an auxiliary translation method and system based on aerospace corpus fine-tuning that can solve the above problems.

[0006] Firstly, this application provides an auxiliary translation method based on fine-tuning of aerospace-related corpora, including:

[0007] Acquire bilingual text sets in the aerospace field, and construct a bilingual parallel corpus based on the bilingual text sets;

[0008] Based on a bilingual parallel corpus, corpus analysis was performed to extract text translation features;

[0009] Text translation features are input into the machine translation model, and the machine translation model is fine-tuned and trained with corpus to generate an auxiliary translation model;

[0010] The system acquires aerospace-related text to be translated and generates an auxiliary translation text sequence based on this text using an auxiliary translation model.

[0011] In one embodiment, a bilingual parallel corpus is constructed based on a bilingual text set, including:

[0012] Based on keywords in the aerospace field, the text positions of the keywords are retrieved in a bilingual text set to obtain keyword location data;

[0013] Based on keyword location data, contextual semantic association is determined according to grammatical rules to identify the start and end boundaries of text semantics;

[0014] Based on the semantic start and end boundaries of the text, each bilingual text in the bilingual text set is divided to obtain source language semantic units and target language semantic units;

[0015] Calculate the semantic similarity between semantic units of the source language and semantic units of the target language;

[0016] Based on semantic similarity, source language semantic units are matched with target language semantic units to obtain parallel semantic units;

[0017] By integrating parallel semantic units, a bilingual parallel corpus is obtained.

[0018] In one embodiment, based on a bilingual parallel corpus, corpus analysis is performed to extract text translation features, including:

[0019] Based on a bilingual parallel corpus, corpus feature analysis is performed to obtain corpus analysis data, which includes extracted lexical feature data, syntactic feature data, discourse feature data, and style feature data.

[0020] Based on corpus analysis data, a multi-level feature model is constructed, which includes four levels: the first level is the lexical feature layer, the second level is the syntactic feature layer, the third level is the discourse feature layer, and the fourth level is the style feature layer.

[0021] Based on a multi-level feature model, inter-level feature mapping matching is used to extract inter-level related feature data;

[0022] Using lexical feature data, syntactic feature data, discourse feature data, and style feature data as nodes, and combining the hierarchical structure of a multi-level feature model, the association paths between nodes are constructed through inter-layer associated feature data to generate an initial text feature tree;

[0023] Based on the initial text feature tree, node-level verification and associated path consistency verification are performed to obtain the verified text feature tree, which is then used as the text translation feature.

[0024] In one embodiment, text translation features are input into a machine translation model, and the machine translation model is fine-tuned and trained using corpus data to generate an auxiliary translation model, including:

[0025] The text translation features are input into the machine translation model, and the model translation features corresponding to the text translation features are extracted.

[0026] By comparing the text translation features with the model translation features, the differences in features of nodes, hierarchical structure, and associated paths are obtained;

[0027] Based on the feature differences, the level of feature difference is determined by the difference threshold.

[0028] From the bilingual parallel corpus, based on the number of corpora set according to the feature difference level, the first bilingual parallel corpus set corresponding to the feature differences is extracted, and the machine translation model is specifically adjusted and trained using the first bilingual parallel corpus set to obtain the intermediate translation model.

[0029] Based on a bilingual parallel corpus, a random sampling method was used to extract a second bilingual parallel corpus set according to the number of corpora set based on the feature difference level. Test samples and control samples were then generated based on the second bilingual parallel corpus set.

[0030] An intermediate translation model is used to translate the test samples, generating intermediate translation sequences;

[0031] Calculate the translation performance evaluation index between the intermediate translation sequence and the control sample;

[0032] Based on translation effectiveness evaluation metrics, the parameters of the intermediate translation model are optimized, and the optimized intermediate translation model is used as an auxiliary translation model.

[0033] In one embodiment, source language semantic units are matched with target language semantic units based on semantic similarity to obtain parallel semantic units, including:

[0034] Select semantic unit combinations whose semantic similarity between source language semantic units and target language semantic units is higher than the similarity threshold to obtain candidate semantic matching pairs;

[0035] Based on the candidate semantic matching pairs, extract the writing time data of the corresponding bilingual texts from the bilingual text set to obtain the temporal correlation data of each candidate semantic matching pair;

[0036] Based on temporal correlation data, the order of the writing time of the bilingual texts corresponding to each candidate semantic matching pair is determined, and the translation temporal ranking result is obtained.

[0037] Based on the translation time sequence ranking results, combined with candidate semantic matching pairs, translation association source tracing data is generated;

[0038] Based on translation-related source data, semantic consistency checks are performed on candidate semantic matching pairs to obtain qualified semantic matching pairs, which are then identified as parallel semantic units.

[0039] In one embodiment, the formula for calculating the translation performance evaluation index between the intermediate translation sequence and the control sample is as follows:

[0040]

[0041] MX is the translation effectiveness evaluation index. The semantic similarity between the intermediate translated sequence and the control sample. The normalized edit distance between the dependency syntax trees of the intermediate translated sequence and the control sample is used. This represents the style consistency score of the intermediate translated sequence relative to the control sample at the style feature layer. The fluency score for the intermediate translation sequence. and These are learnable parameters.

[0042] In one embodiment, an auxiliary translation model is used to generate an auxiliary translation text sequence, including:

[0043] The aerospace text to be translated is input into the auxiliary translation model. A multi-level feature model is used to extract the lexical features, syntactic features, discourse features and style features of the text to be translated. Based on the lexical features, syntactic features, discourse features and style features of the text to be translated, a feature tree of the text to be translated is generated.

[0044] The target text feature tree is obtained by performing hierarchical feature similarity matching between the feature tree of the text to be translated and the verified text feature tree of the text translation features.

[0045] Based on the association path of the target text feature tree, the corresponding target language semantic unit is located in the bilingual parallel corpus to obtain a set of candidate target language semantic units;

[0046] Using the feature tree of the text to be translated as the structural constraint and the set of candidate target language semantic units as candidate translation materials, an auxiliary translation text sequence is generated.

[0047] Secondly, this application also provides an auxiliary translation system based on fine-tuning of aerospace corpus, including:

[0048] The bilingual corpus building module is used to acquire bilingual text sets in the aerospace field and build a bilingual parallel corpus based on the bilingual text sets;

[0049] The corpus feature extraction module is used to perform corpus analysis based on a bilingual parallel corpus and extract text translation features;

[0050] The translation model training module is used to input text translation features into the machine translation model, fine-tune the machine translation model with corpus, and generate an auxiliary translation model.

[0051] The translation text generation module is used to acquire aerospace-related text to be translated, and based on the aerospace-related text to be translated, it uses an auxiliary translation model to generate an auxiliary translation text sequence.

[0052] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described auxiliary translation method based on aerospace corpus fine-tuning.

[0053] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described auxiliary translation method based on aerospace corpus fine-tuning.

[0054] The aforementioned assisted translation method and system, based on fine-tuning of aerospace-related corpora, can accurately collect aerospace-specific bilingual corpora by acquiring a bilingual text set in the aerospace field and constructing a bilingual parallel corpus, providing dedicated data support for domain-specific translation. By conducting corpus analysis and extracting text translation features based on this bilingual parallel corpus, it can deeply explore the unique translation characteristics of aerospace texts, compensating for the shortcomings of general models in adapting to domain features. Inputting these text translation features into a machine translation model for corpus fine-tuning training to generate an assisted translation model allows the model to deeply align with the translation needs of the aerospace field, improving the accuracy of translating specialized terminology and specific sentence structures. Using the assisted translation model to generate assisted translation sequences for the aerospace text to be translated can significantly improve the efficiency and professionalism of aerospace-related text translation, effectively solving the problems of insufficient domain adaptability of general translation models, lack of professional translation results, and low efficiency and high cost of manual translation. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of the assisted translation method based on aerospace corpus fine-tuning of the present invention;

[0057] Figure 2 This diagram illustrates the matching steps of parallel semantic units in the aerospace-based corpus-based assisted translation method of the present invention.

[0058] Figure 3 This is a structural diagram of the auxiliary translation system based on aerospace corpus fine-tuning of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0060] In one embodiment, such as Figure 1 As shown, an assisted translation method based on aerospace corpus fine-tuning is provided. This embodiment illustrates the application of this method to an auxiliary terminal. It is understood that this method can also be applied to servers, and to architectures including both terminals and servers, and implemented through interaction between the terminal and server. In application scenarios, when engineers, translators, or researchers in the aerospace field process technical documents, manuals, or reports on auxiliary terminals (such as personal computers or workstations), and there is a need to quickly and accurately translate professional foreign language materials into Chinese, or translate Chinese results into foreign languages, the user submits the aerospace text to be translated through the terminal. The terminal sends the text data to a server that has deployed the assisted translation model. After receiving the request, the server performs feature extraction, corpus matching, and sequence generation to generate a professional assisted translation text sequence, and returns the result to the terminal for presentation to the user. In another implementation, the entire system can also be fully integrated into a high-performance terminal, where the entire process from corpus management and model operation to result generation is completed locally to meet offline or high data security translation needs.

[0061] In this embodiment, the method includes the following steps:

[0062] S01: Obtain a bilingual text set in the aerospace field, and construct a bilingual parallel corpus based on the bilingual text set.

[0063] Optionally, the auxiliary terminal can collect and clean an initial bilingual text set from authoritative sources such as standard literature databases, technical manuals, patent documents, and academic papers in the aerospace field, after authorization. This text set includes source language texts and their corresponding target language translations. The auxiliary terminal can preprocess and segment the collected bilingual texts to form a series of processable source language and target language text segments as preliminary semantic correspondence units. The auxiliary terminal can use rule-based, statistical, or deep learning model-based methods to align and match these text segments to identify and establish accurate translation correspondences between source language units and target language units, generating parallel sentence pairs or chunk pairs. The auxiliary terminal can systematically integrate, deduplicate, and standardize all aligned and verified parallel language pairs, forming a high-quality domain-specific bilingual parallel corpus that can directly serve machine translation model training and evaluation, providing a data foundation for subsequent in-depth domain feature mining and model fine-tuning.

[0064] S02, based on a bilingual parallel corpus, performs corpus analysis and extracts text translation features.

[0065] Optionally, text translation features refer to structured knowledge representations learned from bilingual parallel corpora that can be used to guide and optimize the machine translation process. The auxiliary terminal can employ natural language processing techniques to analyze bilingual text pairs in the corpus. This analysis can cover multiple dimensions from surface to deep levels (e.g., vocabulary, phrases, syntactic structure, semantic units, discourse logic, and text style). The auxiliary terminal can extract discriminative patterns, rules, or statistics from the analysis results. This information collectively constitutes a feature set reflecting the characteristics of domain-specific translation. The auxiliary terminal can then filter, normalize, and structurally integrate these feature sets to generate a machine-readable, computable, and effectively generalizable knowledge representation. Essentially, these text translation features represent the inherent patterns of bilingual conversion in the aerospace field, providing core guiding information for subsequent domain adaptation and optimization of machine translation models.

[0066] S03. Input the text translation features into the machine translation model, fine-tune the machine translation model with corpus, and generate an auxiliary translation model.

[0067] Optionally, the auxiliary terminal can inject text translation features as domain prior knowledge into a pre-trained general machine translation model. Through a fine-tuning training process based on domain corpus, the model can acquire specialized translation capabilities for the aerospace field. Here, the machine translation model refers to a neural network model trained on a large-scale general corpus that possesses basic cross-language conversion capabilities; corpus fine-tuning training refers to the process of iteratively optimizing some or all parameters of the general model using the bilingual parallel corpus and guided by the text translation features. Its general implementation framework is as follows: the auxiliary terminal can input or integrate text translation features into the machine translation model in the form of feature vectors, structured constraints, or loss function components, guiding the model to focus on specific vocabulary, sentence structures, and styles in the aerospace field; the auxiliary terminal can use bilingual parallel corpora as training data to drive the model to perform forward computation and backpropagation, minimizing the model's prediction error on the domain corpus (such as cross-entropy loss) while simultaneously aligning with the domain rules implied by the text translation features, adjusting and updating the model's internal parameters, and transferring and adapting general translation knowledge to the aerospace field; after multiple rounds of iterative optimization, this machine translation model becomes a dedicated model with significantly improved accuracy, professionalism, and consistency in translating aerospace texts, i.e., an auxiliary translation model. This process realizes the transformation from a general model to a domain expert model.

[0068] S04: Obtain the aerospace-related text to be translated, and generate a translated text sequence based on the aerospace-related text using an auxiliary translation model.

[0069] Optionally, after the assisted translation model is trained, the auxiliary terminal can deploy the assisted translation model in a real-world translation application environment to handle specific domain translation tasks. Here, the aerospace-related text to be translated refers to a source language text fragment or document input by the user or received by the system that needs to be translated into the target language. Generating the assisted translation text sequence refers to the process of automatically translating the input text using the finely tuned and optimized assisted translation model and outputting the corresponding target language text. The auxiliary terminal can receive aerospace-related text to be translated through data acquisition methods such as file reading, interface reception, or interface input. The auxiliary terminal can transmit the aerospace-related text to be translated as input data to the auxiliary translation model. Based on the aerospace-related text translation features and parameters internalized during the corpus fine-tuning training phase, the auxiliary translation model automatically analyzes, understands, and converts the input text (such as encoding the input text, extracting attention-based contextual features, and decoding and generating the target language sequence in combination with domain knowledge). The auxiliary translation model can output a target language text sequence that is highly adapted to the aerospace-related context in terms of vocabulary, syntax, style, and professional accuracy, i.e., the auxiliary translation text sequence, completing a domain-specific machine-assisted translation from the source language to the target language.

[0070] The aforementioned assisted translation method, based on fine-tuning of aerospace-related corpora, accurately collects aerospace-specific bilingual corpora by acquiring a bilingual text set and constructing a bilingual parallel corpus, providing dedicated data support for domain-specific translation. By conducting corpus analysis and extracting text translation features based on this bilingual parallel corpus, it can deeply explore the unique translation characteristics of aerospace texts, compensating for the shortcomings of general models in adapting to domain features. Inputting these text translation features into a machine translation model for corpus fine-tuning training generates an assisted translation model, allowing the model to deeply align with the translation needs of the aerospace field and improve the accuracy of translating specialized terminology and specific sentence structures. Using the assisted translation model to generate assisted translation sequences for the aerospace text to be translated can significantly improve the efficiency and professionalism of aerospace-related text translation, effectively solving the problems of insufficient domain adaptability of general translation models, lack of professional translation results, and low efficiency and high cost of manual translation.

[0071] In one embodiment, a bilingual parallel corpus is constructed based on a bilingual text set, including:

[0072] S11. Based on keywords in the aerospace field, retrieve the text location of keywords in the bilingual text set to obtain keyword positioning data;

[0073] S12, based on keyword location data, determine the contextual semantic association according to grammatical rules, and identify the semantic start and end boundaries of the text.

[0074] S13. Based on the semantic start and end boundaries of the text, the bilingual texts in the bilingual text set are divided to obtain source language semantic units and target language semantic units.

[0075] S14, Calculate the semantic similarity between semantic units of the source language and semantic units of the target language;

[0076] S15. Based on semantic similarity, the source language semantic units are matched with the target language semantic units to obtain parallel semantic units.

[0077] S16, integrate parallel semantic units to obtain a bilingual parallel corpus.

[0078] Specifically, in S11, the auxiliary terminal can retrieve the text location of keywords in the bilingual text set based on keywords in the aerospace field, thereby obtaining keyword positioning data. Keywords refer to core terms, technical abbreviations, or fixed expressions with specific professional meanings within the aerospace technology field; keyword positioning data refers to the set of locations (such as line numbers, sentence indices, and character offsets) of these keywords in the source and target language texts of the bilingual text, located and recorded through string matching or regular expression techniques. This step provides initial anchor points for subsequent semantic unit segmentation.

[0079] In S12, the auxiliary terminal can use the keyword location as the starting point for analysis. Based on the respective grammatical rules of the source and target languages ​​(dependency parsing, constituent parsing, etc.), it expands its scan to the context (left and right boundaries) of the keyword until a complete, semantically independent smallest unit is identified. For example, in English, it might analyze down to the boundary of a clause centered on a predicate verb and containing its necessary arguments; in Chinese, it might determine a short sentence or meaning group expressing a complete proposition. The semantic start and end boundaries of the text refer to the indices of the beginning and end positions of a semantically relatively complete text segment determined through this process.

[0080] In section S13, the auxiliary terminal can segment each bilingual text in the bilingual text set based on the identified semantic start and end boundaries, obtaining source language semantic units and target language semantic units. The auxiliary terminal can calculate the semantic boundaries for each pair of source and translated texts, cutting the continuous text stream into a series of discrete segments. These segments, i.e., semantic units, are semantically broader blocks than sentences and words, forming the basic units for subsequent alignment and matching. This segmentation method ensures that subsequent processing is performed at the semantically complete segment level, improving the accuracy of corpus alignment.

[0081] In step S14, the auxiliary terminal can calculate the semantic similarity between the source language semantic units and the target language semantic units. For example, using a pre-trained multilingual sentence embedding model (such as the Sentence-BERT or LaBSE model based on the Transformer architecture), the source language semantic units and the target language semantic units are encoded as vector representations in a high-dimensional space. The auxiliary terminal can quantify the degree of semantic association between these two vector representations by calculating the cosine similarity. Semantic similarity is a numerical value between 0 and 1; a higher value indicates that the two semantic units are more similar in meaning.

[0082] In S15, the auxiliary terminal can match source language semantic units with target language semantic units based on the calculated semantic similarity to obtain parallel semantic units. The auxiliary terminal can set an adjustable similarity threshold. For each source language semantic unit, the auxiliary terminal can search for units with a semantic similarity higher than the threshold among all semantic units in its corresponding target language text, forming candidate matching pairs. To further ensure the accuracy of the matching, the auxiliary terminal can introduce temporal correlation data into the document for verification, such as using the translation direction inferred from the chronological order of the bilingual texts to perform logical consistency filtering on the candidate matching pairs, selecting semantic unit pairs that are highly semantically corresponding and conform to the translation temporal logic. These verified semantic unit pairs that are mutually related in translation are identified as parallel semantic units.

[0083] In S16, the auxiliary terminal can integrate all parallel semantic units obtained through matching and verification to obtain a bilingual parallel corpus. For example, all parallel semantic units are deduplicated, standardized in format (e.g., unified encoding, removal of redundant whitespace characters), have alignment identifiers added, and are persistently stored according to a specific data structure (e.g., stored in TMX translation memory exchange format or a database table containing alignment indexes). This results in a structured bilingual parallel corpus, a dedicated bilingual alignment data resource for the aerospace field, providing a reliable data foundation for subsequent corpus analysis and model fine-tuning.

[0084] In one embodiment, based on a bilingual parallel corpus, corpus analysis is performed to extract text translation features, including:

[0085] S21. Based on a bilingual parallel corpus, corpus feature analysis is performed to obtain corpus analysis data, which includes extracted lexical feature data, syntactic feature data, discourse feature data, and style feature data.

[0086] S22, based on corpus analysis data, constructs a multi-level feature model, which includes four levels: the first level is the lexical feature layer, the second level is the syntactic feature layer, the third level is the discourse feature layer, and the fourth level is the style feature layer.

[0087] S23, based on a multi-level feature model, extracts inter-level related feature data through inter-level feature mapping matching;

[0088] S24 uses lexical feature data, syntactic feature data, discourse feature data, and style feature data as nodes, and combines the hierarchical structure of the multi-level feature model to construct the association path between nodes through inter-layer association feature data, generating an initial text feature tree;

[0089] S25. Based on the initial text feature tree, perform node-level verification and associated path consistency verification to obtain the verified text feature tree, and use the verified text feature tree as the text translation feature.

[0090] For example, in S21, the auxiliary terminal can perform corpus feature analysis based on a bilingual parallel corpus to obtain structured corpus analysis data. This analysis is multi-dimensional and aims to comprehensively characterize the translation characteristics of aerospace texts: Lexical feature data mainly includes professional terms and entity names (such as spacecraft models and component codes) and their cross-linguistic correspondences obtained through terminology identification and statistics; syntactic feature data includes typical sentence structures, phrase collocation patterns, and word order transformation rules extracted after performing dependency parsing or constituent parsing on the bilingual corpus; discourse feature data can be obtained by analyzing inter-sentence connectors, referential relationships, and paragraph thematic coherence, representing the logical progression and information organization of the text; style feature data can be extracted using quantitative linguistic methods, such as sentence length distribution, frequency of passive voice use, and modal verb density, to quantify the formal, objective, and precise stylistic features of technical texts.

[0091] In S22, the auxiliary terminal can construct a hierarchical, multi-level feature model based on corpus analysis data. This model comprises four progressively abstracted and integrated layers: the first layer is the lexical feature layer, directly carrying atomic-level translation unit information such as terms and entities; the second layer is the syntactic feature layer, organizing intra-sentence grammatical relationships such as phrase structures and sentence templates; the third layer is the discourse feature layer, depicting the logical connections and information flow structure across sentences; and the fourth layer is the style feature layer, summarizing the overall language usage preferences and norms of the text. This hierarchical structure embodies a progressive feature abstraction process from local vocabulary to overall style.

[0092] In S23, the auxiliary terminal can extract inter-layer related feature data based on a multi-level feature model through inter-level feature mapping matching. The auxiliary terminal can analyze and establish constraints and co-occurrence relationships between features at different levels. For example, specific core professional vocabulary (lexical layer) often appears in specific syntactic structures (syntactic layer); combinations of certain syntactic structures are closely related to causal, adversative, and other logical relationships in a text (textual layer); and the formality of the overall style (style layer) constrains the selection of features at each level. The auxiliary terminal can extract these key "inter-layer related feature data" by calculating the co-occurrence probability and conditional distribution between cross-level features or by learning their joint representation through neural networks. These data reveal how features at different language levels work together to achieve domain-adaptive translation.

[0093] The auxiliary terminal can use extracted lexical, syntactic, discourse, and style features as basic nodes, with a hierarchical structure of a multi-level feature model as its skeleton. It can then utilize inter-level correlation features as "edges" or "association paths" connecting nodes at different levels to construct a structured initial text feature tree. Within this tree, lower-level nodes (such as a specific term) can point to and constrain upper-level nodes (such as the term's grammatical context and common discourse functions) through correlation paths, integrating discrete features into an organic, structured network that represents domain-specific translation knowledge.

[0094] In S25, the auxiliary terminal can verify and optimize the generated initial text feature tree to obtain a verified text feature tree, which serves as the final text translation feature. Verification can include two aspects: first, node-level verification, ensuring that each feature node is correctly categorized into one of the four levels: lexical, syntactic, discourse, and stylistic. For example, verifying that a node labeled as a syntactic feature truly represents a structural pattern rather than a single word; second, consistency verification of association paths, checking whether the paths connecting nodes at different levels conform to linguistic common sense and strong association patterns statistically derived from parallel corpora, removing or correcting low-probability or contradictory associations. Through these two verification steps, noise and errors in the initial text feature tree are filtered out, enhancing the inherent consistency and reliability of its representation of domain translation knowledge. The resulting verified text feature tree is a machine-understandable and usable text translation feature distilled from the domain corpus. It encapsulates the translation constraints and patterns of aerospace text at multiple levels, including lexical, syntactic, discourse, and stylistic aspects, in a structured form, guiding subsequent model fine-tuning.

[0095] In one embodiment, text translation features are input into a machine translation model, and the machine translation model is fine-tuned and trained using corpus data to generate an auxiliary translation model, including:

[0096] S31, Input the text translation features into the machine translation model and extract the model translation features corresponding to the text translation features;

[0097] S32 compares the text translation features with the model translation features to obtain the feature differences of nodes, hierarchical structure and associated paths;

[0098] S33, Based on the feature differences, determine the feature difference level through the difference threshold;

[0099] S34. From the bilingual parallel corpus, extract the first bilingual parallel corpus set corresponding to the feature differences according to the number of corpora set according to the feature difference level, and use the first bilingual parallel corpus set to specifically adjust and train the machine translation model to obtain the intermediate translation model.

[0100] S35, based on a bilingual parallel corpus, uses a random sampling method to extract a second bilingual parallel corpus set according to the number of corpora set based on the feature difference level, and generates test samples and control samples based on the second bilingual parallel corpus set;

[0101] S36, The test sample is translated using an intermediate translation model to generate an intermediate translation sequence;

[0102] S37, Calculate the translation performance evaluation index between the intermediate translation sequence and the control sample;

[0103] S38. Based on the translation effect evaluation index, the parameters of the intermediate translation model are optimized, and the optimized intermediate translation model is used as an auxiliary translation model.

[0104] Specifically, in S31, the auxiliary terminal can input the text translation features into a pre-trained general machine translation model and extract the model translation features corresponding to the text translation features. Model translation features refer to the multi-level vector representations formed within the machine translation model (usually a Transformer-based neural machine translation model) for the input text, implicit in its network layers (such as encoder output, attention weights, or decoder hidden states). During extraction, the auxiliary terminal can use the structured knowledge carried by the text translation features (i.e., the verified text feature tree) as a query or guiding signal, and through a specific adapter network or feature projection layer, extract vectorized representations corresponding to the vocabulary, syntax, discourse, and style levels from the internal activation of the machine translation model to obtain the model's current internal understanding features of aerospace knowledge.

[0105] In S32, the auxiliary terminal can perform a multi-dimensional comparison between externally injected text translation features and model translation features extracted from the model's internal structure. This comparison is performed simultaneously at three levels: nodes in the feature tree (representing specific feature instances such as terms and sentence structures), hierarchical structure (the organizational relationships of the four layers: vocabulary, syntax, discourse, and style), and association paths (the constraint relationships between features at different levels). The auxiliary terminal can quantify a set of feature difference data by calculating the cosine distance between corresponding node vectors, comparing differences in hierarchical attention distribution, and analyzing differences in the strength of association paths. This data indicates the gap between the internalized knowledge of the general model and the ideal translation features in the aerospace field at various fine-grained levels.

[0106] In S33, the auxiliary terminal can determine the level of feature difference based on the calculated feature differences using a preset difference threshold. The system sets threshold ranges for indicators such as node similarity and structural similarity (e.g., [0, 0.3) for slight difference, [0.3, 0.6) for moderate difference, and [0.6, 1.0] for significant difference), mapping abstract difference values ​​to discrete operational guidance labels for difference levels. This step transforms continuous difference analysis results into classification information that can be used to guide subsequent differential training strategies.

[0107] In S34, the auxiliary terminal can extract a first bilingual parallel corpus set corresponding to the feature differences from the constructed bilingual parallel corpus, based on different corpus quantity strategies set according to the aforementioned feature difference levels, and use it for targeted training. For feature dimensions with a difference level of "significant" (e.g., the translation of certain aerospace proper nouns), the auxiliary terminal can retrieve and extract a large number of relevant sentence pairs containing this feature from the corpus to form a reinforced training set; for dimensions with "moderate" or "slight" differences, a smaller number of corpora are extracted. Using this first bilingual parallel corpus set as training data, the model is trained with a small learning rate or for specific layer parameters through one (or several) rounds of targeted gradient update training, focusing on adjusting the model parameters related to high-difference features, to obtain an intermediate translation model that is initially adapted to domain features.

[0108] In S35, the auxiliary terminal can use a random sampling method based on a bilingual parallel corpus to extract a "second bilingual parallel corpus" with no overlap in content, according to the proportion of another set of corpora set based on the feature difference level. The source language part of this corpus is used as the "test sample", and the corresponding target language part is used as the "control sample" (i.e., the standard answer), forming a test set for the model's performance.

[0109] In S36 and S37, the auxiliary terminal can use an intermediate translation model to translate the test sample, generate an intermediate translation sequence, and calculate the translation effect evaluation index between the intermediate translation sequence and the control sample. This index can comprehensively consider four aspects: semantic similarity, syntactic tree edit distance, style consistency, and fluency. Through learnable parameters and weighted balancing, it can comprehensively and quantitatively evaluate the quality of the translation result in domain adaptability.

[0110] In S38, the auxiliary terminal can optimize the parameters of the intermediate translation model based on the calculated translation performance evaluation index. Specifically, the auxiliary terminal can use the translation performance evaluation index as a training objective (or as part of the loss function) to further iteratively optimize the intermediate translation model on a second bilingual parallel corpus (or a validation set derived from it). The auxiliary terminal can adjust the model parameters through backpropagation to maximize the translation performance evaluation index, that is, to make the translation generated by the model more closely resemble domain standards in terms of semantics, syntax, style, and fluency. When the translation performance evaluation index stabilizes on the validation set or reaches a preset threshold, the optimization stops, and the resulting model is an auxiliary translation model that is deeply adapted to the aerospace field and possesses high professional translation capabilities.

[0111] In one embodiment, such as Figure 2 As shown, based on semantic similarity, source language semantic units are matched with target language semantic units to obtain parallel semantic units, including:

[0112] S41, Select semantic unit combinations whose semantic similarity between source language semantic units and target language semantic units is higher than the similarity threshold to obtain candidate semantic matching pairs;

[0113] S42, Based on the candidate semantic matching pairs, extract the writing time data of the corresponding bilingual texts from the bilingual text set to obtain the temporal correlation data of each candidate semantic matching pair;

[0114] S43. Based on temporal correlation data, determine the order of the writing time of the bilingual texts corresponding to each candidate semantic matching pair, and obtain the translation temporal ranking result.

[0115] S44, based on the translation time sequence ranking results, combined with candidate semantic matching pairs, to generate translation association source tracing data;

[0116] S45. Based on the translation association source data, perform semantic consistency verification on the candidate semantic matching pairs, obtain the qualified semantic matching pairs, and determine the qualified semantic matching pairs as parallel semantic units.

[0117] For example, in S41, the auxiliary terminal can perform preliminary screening of all source language semantic units and target language semantic units based on semantic similarity. The auxiliary terminal can set a configurable similarity threshold (which can be adjusted according to experience and actual application needs), and select combinations of source language units and target language units with semantic similarity higher than this threshold to form candidate semantic matching pairs.

[0118] In S42, the auxiliary terminal can extract the corresponding writing time data from the original documents constituting the bilingual text set based on the candidate semantic matching pairs, thus obtaining the temporal correlation data of each candidate semantic matching pair. The auxiliary terminal can trace back to the original bilingual text file (such as a technical report or patent document) to which the candidate semantic matching pair belongs based on its index, and parse the creation, publication, or last revision time of the text—that is, the writing time data—from the file's metadata, version history, or related database records. The candidate semantic matching pairs associate this time information with the candidate pairs to form temporal correlation data, which records the real-world timestamps of each pair of potentially translatable text fragments.

[0119] In S43, the auxiliary terminal can determine the chronological order of the source and target language texts corresponding to each candidate semantic matching pair based on temporal correlation data. By comparing the writing times of a pair of texts, the auxiliary terminal can infer the possible translation direction (e.g., if the Chinese text was written in January 2023 and its corresponding English text was written in March 2023, logically the Chinese text is more likely to be the source text and the English text the translation). The auxiliary terminal can perform this chronological comparison on all candidate pairs to obtain the translation time sequence ranking result, which assigns a directional constraint to each candidate matching pair from a temporal logical perspective.

[0120] In S44, the auxiliary terminal can generate translation association tracing data based on the translation temporal ranking results and the original candidate semantic matching pairs, thereby integrating temporal logic with semantic association. For example, for a candidate pair, the system records that "the semantic similarity between source language unit A (written in T1) and target language unit B (written in T2) is S, and T1 is earlier than T2," thus generating a comprehensive information record containing semantics, temporal sequence, and inference direction, i.e., translation association tracing data. This data provides multi-dimensional evidence for verifying the rationality and consistency of the matching pairs.

[0121] In S45, based on the generated translation-related source data, the candidate semantic matching pairs are subjected to semantic consistency verification. The core of the verification is to eliminate logical contradictions. For example, if the time sequence shows that text A is earlier than text B, theoretically A is the source text and B is the translation, then the semantic correspondence between A and B in the candidate pair must be highly consistent and the translation direction must be reasonable. The auxiliary terminal can combine more detailed language rules for checking, such as checking whether the translation of professional terms conforms to domain conventions and whether sentence transformation is common under the inferred translation direction. In addition, the auxiliary terminal can also mark or filter candidate pairs with temporal logical conflicts (such as the translation being written earlier than the inferred source text) or semantic relationships that seem far-fetched under a specific translation direction. Only candidate semantic matching pairs that pass the consistency verification based on the source data are judged as "verified semantic matching pairs" and identified as high-quality "parallel semantic units" that can be used to build a domain corpus. This step, by introducing the strong constraint of time sequence, effectively improves the accuracy and reliability of automatically extracting parallel corpora from unaligned bilingual texts and reduces ambiguous matching that may be caused by semantic similarity calculation.

[0122] In one embodiment, S51, the formula for calculating the translation effect evaluation index between the intermediate translation sequence and the control sample is as follows:

[0123]

[0124] MX is the translation effectiveness evaluation index. The semantic similarity between the intermediate translated sequence and the control sample. The normalized edit distance between the dependency syntax trees of the intermediate translated sequence and the control sample is used. This represents the style consistency score of the intermediate translated sequence relative to the control sample at the style feature layer. The fluency score for the intermediate translation sequence. and These are learnable parameters.

[0125] Specifically, the core purpose of this computational formula is to provide a comprehensive, quantitative, and differentiable evaluation criterion to measure the performance of the intermediate translation model on specialized text translation tasks in the aerospace field. This formula can be used to decompose translation quality into four dimensions: semantic accuracy, syntactic fidelity, style fit, and language fluency, and organically integrate them. This provides a clear, calculable learning objective or loss signal consistent with domain translation quality goals for the parameter optimization process during the fine-tuning phase of the machine translation model, thereby guiding the model to improve in a balanced manner across multiple key quality attributes in accordance with the requirements of the specialized domain.

[0126] The overall structure of this calculation formula and the relationship between its variables reflect considerations of multi-objective trade-offs and normalization design. Structurally, it consists of two parts: a numerator and a denominator. The numerator integrates semantic and syntactic dimensions of evaluation. Measuring the closeness of intermediate translated sequences to control samples in terms of deep meaning, including semantic similarity. A higher value indicates more accurate semantic delivery; To assess the similarity between sentences in terms of surface syntactic structure, the normalized dependency tree edit distance is subtracted from 1. To achieve, A smaller value indicates less difference in syntactic structure, resulting in a higher score. The learnable parameter α is used here to dynamically adjust the relative weights of semantic fidelity and syntactic faithfulness in the final evaluation. The denominator focuses on language quality and domain fit. This is used to measure the degree to which translation outputs conform to domain standards (control samples) in terms of stylistic features (such as formality, terminology consistency, and sentence complexity distribution). The higher the value, the more consistent the style. Assess the fluency and grammatical correctness of the translated text itself. A higher value indicates better fluency. The learnable parameter β is used to balance the importance of style consistency and language fluency. The calculation divides the numerator (semantic-syntactic integrated score) by the denominator (style fluency integrated cost). Its design logic is that an excellent domain translation not only needs to be semantically and syntactically correct (numerator as large as possible), but also needs to use fluent language that conforms to domain norms (denominator as small as possible), thereby maximizing the overall evaluation score. This benefit / cost ratio effectively incentivizes the model to improve content accuracy while also considering the professionalism and naturalness of expression.

[0127] The core variables in this calculation formula, such as Its reliable computation directly relies on style feature data extracted from bilingual parallel corpora and constructed multi-level feature models. The style consistency comparison is not based on a general language model, but rather on a quantitative index system representing the stylistic features of aerospace texts, distilled from domain-specific corpora. The intermediate translation sequence evaluated by this formula is generated by an intermediate translation model that has undergone targeted tuning and training. The calculated translation performance evaluation index can be directly used as the target for parameter optimization of the intermediate translation model. Through backpropagation training guided by maximizing the target, the model is further adjusted so that the translation knowledge represented by its internal parameters is more aligned with the domain translation rules (including multi-level constraints such as vocabulary, syntax, discourse, and style) extracted from the corpus and characterized by text translation features and multi-level feature models. Therefore, this evaluation formula is the core bridge connecting domain knowledge analysis (feature extraction) and model parameter optimization. It enables the abstract translation features mined from the corpus to be transformed into computable gradient signals, accurately guiding the transformation and convergence of general models into high-quality domain-specific models (auxiliary translation models).

[0128] In one embodiment, an auxiliary translation model is used to generate a translated text sequence, including:

[0129] S61. Input the aerospace text to be translated into the auxiliary translation model, use a multi-level feature model to extract the lexical features, syntactic features, discourse features and style features of the text to be translated, and generate a feature tree of the text to be translated based on the lexical features, syntactic features, discourse features and style features of the text to be translated.

[0130] S62, perform hierarchical feature similarity matching between the feature tree of the text to be translated and the verified text feature tree of the text translation features to obtain the target text feature tree;

[0131] S63, based on the association path of the target text feature tree, locate the corresponding target language semantic unit in the bilingual parallel corpus to obtain a set of candidate target language semantic units;

[0132] S64 generates a sequence of translated texts using the feature tree of the text to be translated as a structural constraint and the set of candidate target language semantic units as candidate translation materials.

[0133] For example, in S61, the auxiliary terminal can input the aerospace-related text to be translated into a pre-trained auxiliary translation model and call a multi-level feature model to process the input text, extracting its features at four levels: lexical, syntactic, discourse, and stylistic. For instance, at the lexical feature level, domain terminology dictionaries and contextual encodings are used to identify and label specialized terms and entities in the text; at the syntactic feature level, a dependency parser is used to analyze sentence structure and obtain grammatical relationships such as subject-verb-object and phrase collocation patterns; at the discourse feature level, the logical structure of the text is constructed by analyzing inter-sentence connectors, pronoun resolution, and topic coherence; and at the stylistic feature level, its stylistic attributes are quantified by statistical features such as sentence length and voice distribution. The features extracted by the auxiliary terminal are organized according to the hierarchical structure defined by the multi-level feature model and connected by the constraints revealed by the inter-layer correlation feature data to construct a structured feature tree for the text to be translated. This tree is a domain-specific and structured representation of the input text and serves as a blueprint for subsequent translation operations.

[0134] In S62, the auxiliary terminal can perform hierarchical feature similarity matching between the feature tree of the text to be translated and the verified text feature tree, which is learned and solidified from the domain corpus during the training phase of the auxiliary translation model, and serves as the text translation feature. This matching process is performed layer by layer: the auxiliary translation model can calculate the similarity in vector space between the lexical feature set of the text to be translated and the lexical layer nodes of the text translation features, and accordingly find one or more most similar node clusters in the lexical layer of the text translation feature tree; based on these node clusters, the auxiliary translation model can perform similar similarity calculations and node matching upwards at the syntax layer, discourse layer, and style layer. Through this bottom-up hierarchical progressive matching, the auxiliary translation model can locate one or more subtrees or node paths in the text translation feature tree that best fit the text to be translated in terms of overall feature composition. These located parts are collectively referred to as the target text feature tree. The target text feature tree is essentially the most relevant and referential translation template or template retrieved by the model from its learned domain translation knowledge base for the current input text.

[0135] In S63, the auxiliary terminal can locate and retrieve data in a bilingual parallel corpus based on the association path of the target text feature tree. Since the text translation feature tree, its nodes, and paths all have index mapping relationships with parallel semantic units in the bilingual parallel corpus, the auxiliary terminal can quickly retrieve, based on the nodes and paths corresponding to the target text feature tree, all target language semantic units in the corpus that contributed these features during training. These retrieved units constitute a candidate target language semantic unit set, which is a potential translation fragment library highly adapted to the current input text in terms of vocabulary selection, sentence structure, discourse logic, and text style.

[0136] In S64, the auxiliary terminal can generate an assisted translation text sequence using the feature tree of the text to be translated as a structural constraint and the set of candidate target language semantic units as candidate translation materials. This process is conducted through a constraint-based generation method: on the one hand, the auxiliary terminal can use the feature tree of the text to be translated as a structural constraint that must be followed during the generation process, ensuring that the generated translation faithfully adheres to the domain feature structure of the source text in terms of vocabulary, syntax, discourse, and style; on the other hand, the auxiliary terminal can use the set of candidate target language semantic units as the preferred candidate translation materials. The auxiliary terminal can operate under this dual guidance: employing an assisted translation model, referencing the hierarchical structure of the feature tree of the text to be translated to plan the overall framework and organization of the translation, while simultaneously selecting the most matching phrases, sentence structures, and even paragraph blocks from the set of candidate target language semantic units, and smoothly connecting and combining them through a neural network. This method, combining structural constraints and material retrieval, ensures that the generated translation text sequence not only accurately conveys the professional semantics of the source text but also deeply aligns with the professional norms and stylistic features of the aerospace field in terms of language expression, achieving high-quality, domain-adaptive machine-assisted translation.

[0137] The aforementioned assisted translation method based on aerospace corpus fine-tuning involves acquiring a bilingual text set from the aerospace field and constructing a dedicated bilingual parallel corpus for the aerospace field based on keyword localization, semantic boundary recognition, semantic unit segmentation, semantic similarity matching, and temporal association verification. This corpus accurately collects bilingual alignment data specific to the aerospace field, providing dedicated data support for domain-specific translation and addressing the lack of aerospace-specific corpus support in general translation models from the source. Based on this parallel corpus, multi-dimensional feature analysis of vocabulary, syntax, discourse, and style is conducted. A multi-level text feature tree is constructed and verified to extract aerospace-specific translation features, enabling in-depth mining of the unique patterns in aerospace text translation and compensating for the deficiencies of general translation models in adapting to the vocabulary, syntax, discourse, and style features of the aerospace field. The extracted text is then translated... By inputting features into a general machine translation model, comparing feature differences, and retrieving corresponding corpora according to the difference level for targeted fine-tuning training, and combining comprehensive translation effect evaluation indicators to complete model parameter optimization and generate an auxiliary translation model, the model can be deeply adapted to the translation needs of aerospace professionals. This significantly improves the accuracy and consistency of translating professional terms, abbreviations, and specific sentence structures, effectively avoiding translation ambiguity. Using this auxiliary translation model to extract features, perform hierarchical matching, and locate semantic units in the corpus of aerospace texts to be translated, generating professional auxiliary translation text sequences, it can significantly improve the translation efficiency of aerospace professional texts, reduce the cost of manual translation and the dependence on translators' professional knowledge, and solve the technical problems of insufficient adaptability of general translation models in the aerospace field, lack of translation professionalism, and low efficiency and high cost of manual translation.

[0138] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0139] Based on the same inventive concept, this application also provides an auxiliary translation system based on aerospace corpus fine-tuning for implementing the aforementioned auxiliary translation method based on aerospace corpus fine-tuning. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more auxiliary translation system embodiments based on aerospace corpus fine-tuning provided below can be found in the limitations of the auxiliary translation method based on aerospace corpus fine-tuning described above, and will not be repeated here.

[0140] In one exemplary embodiment, such as Figure 3 As shown, an auxiliary translation system based on fine-tuning of aerospace corpus is provided, including:

[0141] The bilingual corpus building module 101 is used to acquire bilingual text sets in the aerospace field and build a bilingual parallel corpus based on the bilingual text sets;

[0142] The corpus feature extraction module 102 is used to perform corpus analysis based on a bilingual parallel corpus and extract text translation features;

[0143] The translation model training module 103 is used to input text translation features into the machine translation model, fine-tune the machine translation model with corpus, and generate an auxiliary translation model.

[0144] The translation text generation module 104 is used to acquire the aerospace field text to be translated, and to generate an auxiliary translation text sequence based on the aerospace field text to be translated using an auxiliary translation model.

[0145] In one embodiment, the bilingual corpus building module 101 can also be used for:

[0146] Based on keywords in the aerospace field, the text positions of the keywords are retrieved in a bilingual text set to obtain keyword location data;

[0147] Based on keyword location data, contextual semantic association is determined according to grammatical rules to identify the start and end boundaries of text semantics;

[0148] Based on the semantic start and end boundaries of the text, each bilingual text in the bilingual text set is divided to obtain source language semantic units and target language semantic units;

[0149] Calculate the semantic similarity between semantic units of the source language and semantic units of the target language;

[0150] Based on semantic similarity, source language semantic units are matched with target language semantic units to obtain parallel semantic units;

[0151] By integrating parallel semantic units, a bilingual parallel corpus is obtained.

[0152] In one embodiment, the corpus feature extraction module 102 can also be used for:

[0153] Based on a bilingual parallel corpus, corpus feature analysis is performed to obtain corpus analysis data, which includes extracted lexical feature data, syntactic feature data, discourse feature data, and style feature data.

[0154] Based on corpus analysis data, a multi-level feature model is constructed, which includes four levels: the first level is the lexical feature layer, the second level is the syntactic feature layer, the third level is the discourse feature layer, and the fourth level is the style feature layer.

[0155] Based on a multi-level feature model, inter-level feature mapping matching is used to extract inter-level related feature data;

[0156] Using lexical feature data, syntactic feature data, discourse feature data, and style feature data as nodes, and combining the hierarchical structure of a multi-level feature model, the association paths between nodes are constructed through inter-layer associated feature data to generate an initial text feature tree;

[0157] Based on the initial text feature tree, node-level verification and associated path consistency verification are performed to obtain the verified text feature tree, which is then used as the text translation feature.

[0158] In one embodiment, the translation model training module 103 can also be used for:

[0159] The text translation features are input into the machine translation model, and the model translation features corresponding to the text translation features are extracted.

[0160] By comparing the text translation features with the model translation features, the differences in features of nodes, hierarchical structure, and associated paths are obtained;

[0161] Based on the feature differences, the level of feature difference is determined by the difference threshold.

[0162] From the bilingual parallel corpus, based on the number of corpora set according to the feature difference level, the first bilingual parallel corpus set corresponding to the feature differences is extracted, and the machine translation model is specifically adjusted and trained using the first bilingual parallel corpus set to obtain the intermediate translation model.

[0163] Based on a bilingual parallel corpus, a random sampling method was used to extract a second bilingual parallel corpus set according to the number of corpora set based on the feature difference level. Test samples and control samples were then generated based on the second bilingual parallel corpus set.

[0164] An intermediate translation model is used to translate the test samples, generating intermediate translation sequences;

[0165] Calculate the translation performance evaluation index between the intermediate translation sequence and the control sample;

[0166] Based on translation effectiveness evaluation metrics, the parameters of the intermediate translation model are optimized, and the optimized intermediate translation model is used as an auxiliary translation model.

[0167] In one embodiment, the bilingual corpus building module 101 can also be used for:

[0168] Select semantic unit combinations whose semantic similarity between source language semantic units and target language semantic units is higher than the similarity threshold to obtain candidate semantic matching pairs;

[0169] Based on the candidate semantic matching pairs, extract the writing time data of the corresponding bilingual texts from the bilingual text set to obtain the temporal correlation data of each candidate semantic matching pair;

[0170] Based on temporal correlation data, the order of the writing time of the bilingual texts corresponding to each candidate semantic matching pair is determined, and the translation temporal ranking result is obtained.

[0171] Based on the translation time sequence ranking results, combined with candidate semantic matching pairs, translation association source tracing data is generated;

[0172] Based on translation-related source data, semantic consistency checks are performed on candidate semantic matching pairs to obtain qualified semantic matching pairs, which are then identified as parallel semantic units.

[0173] In one embodiment, the translation text generation module 104 can also be used for:

[0174] The aerospace text to be translated is input into the auxiliary translation model. A multi-level feature model is used to extract the lexical features, syntactic features, discourse features and style features of the text to be translated. Based on the lexical features, syntactic features, discourse features and style features of the text to be translated, a feature tree of the text to be translated is generated.

[0175] The target text feature tree is obtained by performing hierarchical feature similarity matching between the feature tree of the text to be translated and the verified text feature tree of the text translation features.

[0176] Based on the association path of the target text feature tree, the corresponding target language semantic unit is located in the bilingual parallel corpus to obtain a set of candidate target language semantic units;

[0177] Using the feature tree of the text to be translated as the structural constraint and the set of candidate target language semantic units as candidate translation materials, an auxiliary translation text sequence is generated.

[0178] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the auxiliary translation method based on aerospace corpus fine-tuning as described above.

[0179] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0180] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0181] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. An auxiliary translation method based on fine-tuning of aerospace corpus, characterized in that, The method includes: Acquire a bilingual text set in the aerospace field, and construct a bilingual parallel corpus based on the bilingual text set; Based on the bilingual parallel corpus, corpus analysis is performed to extract text translation features; The text translation features are input into the machine translation model, and the machine translation model is fine-tuned and trained with corpus to generate an auxiliary translation model; Obtain the aerospace field text to be translated, and based on the aerospace field text to be translated, use the auxiliary translation model to generate an auxiliary translation text sequence.

2. The method according to claim 1, characterized in that, The construction of a bilingual parallel corpus based on the bilingual text set includes: Based on keywords in the aerospace field, the text positions of the keywords are retrieved from the bilingual text set to obtain keyword location data; Based on the keyword location data, contextual semantic association is determined according to grammatical rules to identify the start and end boundaries of text semantics. Based on the semantic start and end boundaries of the text, each bilingual text in the bilingual text set is divided to obtain source language semantic units and target language semantic units; Calculate the semantic similarity between the source language semantic unit and the target language semantic unit; Based on the semantic similarity, the source language semantic unit is matched with the target language semantic unit to obtain parallel semantic units; The parallel semantic units are integrated to obtain the bilingual parallel corpus.

3. The method according to claim 2, characterized in that, The process of analyzing the bilingual parallel corpus and extracting text translation features includes: Based on the bilingual parallel corpus, corpus feature analysis is performed to obtain corpus analysis data, which includes extracted lexical feature data, syntactic feature data, discourse feature data, and style feature data. Based on the corpus analysis data, a multi-level feature model is constructed, which includes four levels: the first level is the lexical feature layer, the second level is the syntactic feature layer, the third level is the discourse feature layer, and the fourth level is the style feature layer. Based on the multi-level feature model, inter-level correlation feature data is extracted through inter-level feature mapping matching; Using the lexical feature data, the syntactic feature data, the discourse feature data, and the style feature data as nodes, and combining the hierarchical structure of the multi-level feature model, the association paths between nodes are constructed through the inter-layer association feature data to generate an initial text feature tree; Based on the initial text feature tree, node-level verification and associated path consistency verification are performed to obtain the verified text feature tree, and the verified text feature tree is used as the text translation feature.

4. The method according to claim 3, characterized in that, The step of inputting the text translation features into a machine translation model, fine-tuning the machine translation model with corpus data, and generating an auxiliary translation model includes: The text translation features are input into the machine translation model to extract the model translation features corresponding to the text translation features; The text translation features are compared with the model translation features to obtain the feature differences of the nodes, the hierarchical structure, and the associated paths; Based on the aforementioned feature differences, the level of feature difference is determined using a difference threshold. From the bilingual parallel corpus, according to the number of corpora set according to the feature difference level, a first bilingual parallel corpus set corresponding to the feature difference is extracted, and the machine translation model is specifically adjusted and trained using the first bilingual parallel corpus set to obtain an intermediate translation model. Based on the bilingual parallel corpus, a random sampling method is used to extract a second bilingual parallel corpus set according to the number of corpora set based on the feature difference level, and test samples and control samples are generated based on the second bilingual parallel corpus set. The intermediate translation model is used to translate the test sample to generate an intermediate translation sequence; Calculate the translation performance evaluation index between the intermediate translation sequence and the control sample; Based on the translation performance evaluation index, the parameters of the intermediate translation model are optimized, and the optimized intermediate translation model is used as the auxiliary translation model.

5. The method according to claim 2, characterized in that, The step of matching the source language semantic unit with the target language semantic unit based on the semantic similarity to obtain parallel semantic units includes: Semantic unit combinations with a semantic similarity higher than a similarity threshold between the source language semantic units and the target language semantic units are selected to obtain candidate semantic matching pairs; Based on the candidate semantic matching pairs, the writing time data of the corresponding bilingual texts are extracted from the bilingual text set to obtain the temporal correlation data of each candidate semantic matching pair; Based on the temporal correlation data, the order of the writing time of the bilingual texts corresponding to each candidate semantic matching pair is determined to obtain the translation temporal ranking result; Based on the translation time sequence ranking results, and combined with the candidate semantic matching pairs, translation association source tracing data is generated; Based on the translation association source data, the candidate semantic matching pairs are subjected to semantic consistency verification to obtain qualified semantic matching pairs, and the qualified semantic matching pairs are determined as the parallel semantic units.

6. The method according to claim 4, characterized in that, The formula for calculating the translation effectiveness evaluation index between the intermediate translation sequence and the control sample is as follows: Wherein, MX is the translation performance evaluation index. The semantic similarity between the intermediate translated sequence and the control sample. The normalized edit distance is the dependency parsing distance between the intermediate translated sequence and the control sample. The style consistency score of the intermediate translation sequence relative to the control sample on the style feature layer. The fluency score of the intermediate translated sequence. and These are learnable parameters.

7. The method according to claim 3, characterized in that, The step of generating the auxiliary translation text sequence using the auxiliary translation model includes: The aerospace text to be translated is input into the auxiliary translation model. The multi-level feature model is used to extract the lexical features, syntactic features, discourse features, and style features of the text to be translated. Based on the lexical features, syntactic features, discourse features, and style features of the text to be translated, a feature tree of the text to be translated is generated. The target text feature tree is obtained by performing hierarchical feature similarity matching between the feature tree of the text to be translated and the verified text feature tree of the text translation features. Based on the association path of the target text feature tree, the corresponding target language semantic unit is located in the bilingual parallel corpus to obtain a set of candidate target language semantic units; The auxiliary translation text sequence is generated using the feature tree of the text to be translated as a structural constraint and the set of candidate target language semantic units as candidate translation materials.

8. An auxiliary translation system based on fine-tuning of aerospace corpus, characterized in that, The system includes: The bilingual corpus building module is used to acquire bilingual text sets in the aerospace field and build a bilingual parallel corpus based on the bilingual text sets. The corpus feature extraction module is used to perform corpus analysis and extract text translation features based on the bilingual parallel corpus. The translation model training module is used to input the text translation features into the machine translation model, fine-tune the machine translation model with corpus, and generate an auxiliary translation model. The translation text generation module is used to acquire aerospace-related text to be translated, and based on the aerospace-related text to be translated, to generate an auxiliary translation text sequence using the auxiliary translation model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images