Semantically optimized overseas multilingual translation method

Abstract

The invention relates to the field of semantic translation methods, in particular to an overseas multilingual translation method capable of achieving semantic optimization, which comprises the following steps: step 1, a text preprocessing unit performs standardization processing and meta-information extraction on an input text to obtain a source text; 2, performing semantic analysis on the source text by a multi-level semantic analysis unit, and generating a comprehensive multi-dimensional semantic representation through a self-adaptive weighted fusion mechanism; 3, a culture conversion unit recognizes culture sensitive vocabularies and expressions in the source text, culture adaptability conversion is conducted through a cross-culture knowledge base, and it is ensured that a translation result has a proper expression mode in target culture; 4, the consistency constraint unit establishes a global consistency constraint mechanism in terms, pronouns and semantics, and ensures the complete consistency of long document translation through term weight scoring, a global term comparison table, pronouns-antecedents mapping and document-level semantic constraint; and 5, performing multi-dimensional quality evaluation on a translation result.

Description

Technical Field

[0001] This invention relates to the field of semantic translation methods, and more specifically, to a semantically optimized method for multilingual translation overseas. Background Technology

[0002] Existing machine translation technologies are mainly based on neural network models to achieve automatic translation between multiple languages, and mainly include the following two core technical directions: 1. Deep learning-based multilingual neural machine translation: This type of technology uses deep neural network architecture to perform sequence-to-sequence translation tasks. Typical examples include Google's multi-head self-attention mechanism and Facebook's multilingual pre-trained models. These technologies use shared encoders and language-specific decoders to build a unified multilingual representation space, enabling parallel processing of translation tasks for multiple language pairs. Significant progress has been made in language conversion and sequence modeling.

[0003] 2. Context-aware and domain-adaptive translation technology: This technology improves translation quality by introducing document-level contextual information and domain-specific knowledge. It employs a hierarchical encoder to process text information of different granularities, addressing issues such as pronoun resolution and terminology consistency. At the same time, it utilizes transfer learning and domain fine-tuning methods to enable the system to adapt to the translation needs of specific domains and improve the translation accuracy of professional texts.

[0004] However, with deeper applications, it has been found that current machine translation technologies suffer from insufficient deep semantic and cultural adaptability capabilities, as well as deficiencies in context modeling and real-time optimization. Regarding the lack of deep semantic and cultural adaptability, existing multilingual neural machine translation technologies lack depth in semantic understanding. The systems primarily rely on lexical and syntactic level conversions, lacking a complete understanding of deep semantic structures and failing to accurately capture hierarchical relationships and implicit meanings between concepts. In particular, the understanding of the cultural background of both the source and target languages ​​is insufficient, making it difficult to effectively handle culture-specific expressions, idioms, metaphors, and other content. This results in significant deficiencies in semantic coherence and cultural adaptability in the translation results, severely impacting user satisfaction.

[0005] However, existing context-aware technologies have limitations in handling long document translations, and the models struggle to maintain global semantic consistency. They exhibit significant problems in areas such as pronoun reference, terminology consistency, and paragraph-level semantic coherence. Furthermore, the systems lack real-time, multi-dimensional quality assessment mechanisms, making it impossible to dynamically adjust strategies during the translation process. In addition, their computational efficiency is low, resulting in a large average translation latency, which makes it difficult to meet the practical needs of real-time translation, personalized processing, and domain adaptation.

[0006] Therefore, it is necessary to propose a semantically optimized method for multilingual translation overseas to solve the above problems. Summary of the Invention

[0007] To overcome at least one of the defects (deficiencies) of the prior art described above, the present invention provides a semantically optimized method for multilingual translation overseas.

[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A semantically optimized method for multilingual translation overseas, comprising the following steps: Step 1: The text preprocessing unit performs normalization and metadata extraction on the input text to obtain the source text; Step 2: The multi-level semantic analysis unit performs semantic analysis on the source text and generates a comprehensive multi-dimensional semantic representation through an adaptive weighted fusion mechanism; Step 3: The culture conversion unit identifies culturally sensitive words and expressions in the source text and uses a cross-cultural knowledge base to perform culture-adaptive conversion, ensuring that the translation results have appropriate expressions in the target culture. Step 4: The consistency constraint unit establishes a global consistency constraint mechanism for terms, pronouns, and semantics. Through term weight scoring, global term lookup table, pronoun-antecedent mapping, and document-level semantic constraints, it ensures the complete consistency of long document translation. Step 5: Conduct a multi-dimensional quality assessment of the translation results. When the overall score of the multi-dimensional quality assessment is lower than the predetermined threshold, initiate the dynamic optimization process and re-evaluate the optimized translation results to form a closed loop of continuous improvement until the score reaches the predetermined threshold. Furthermore, step 1 includes the following steps: Step 11: The multilingual language recognition module performs language recognition on the content of the input text and determines the corresponding target language; Step 12: The word segmentation unit performs multi-granular word segmentation on the input text, and the sentence segmentation unit performs sentence-level segmentation on the input text. The sentence boundary recognizer identifies the punctuation marks at the end of the sentence. For the boundary cases of ellipses and abbreviations, a hybrid method based on rules and statistics is used for processing, and a unique ID number and location information are generated for each sentence. Step 13: The format processing unit performs format standardization processing on the input text; Step 14: The metadata extraction unit establishes a global metadata record and automatically configures the parameters for subsequent processing based on the metadata content to obtain the source text.

[0009] Furthermore, step 2 includes the following steps: Step 21: The lexical layer analysis unit performs deep semantic analysis on each word, identifies the specific meaning of polysemous words through polysemous word resolution technology, and after resolution using contextual information, identifies the semantic relationships between words and constructs a lexical semantic relationship network. Step 22: The syntactic layer analysis unit performs in-depth analysis of the sentence's grammatical structure and performs part-of-speech tagging to identify the part of speech and grammatical function of each word; Step 23: The semantic layer analysis unit performs deep semantic analysis and uses an external knowledge base to expand semantics and supplement the semantic information of words and concepts; Step 24: The pragmatic layer analysis unit performs semantic analysis of context and cultural background, uses contextual information for disambiguation, and selects the most appropriate expression meaning; Step 25: The fusion unit uses an improved multi-head attention mechanism to perform weighted fusion of semantic features from the four levels. The fusion formula is as follows: F_fused = Σ(i=1 to 4) α_i × V_i + ResNet(V_i); Wherein: V_1 is the lexical layer feature, V_2 is the syntactic layer feature, V_3 is the semantic layer feature, V_4 is the pragmatic layer feature, α_i is the attention weight of each layer, and ResNet represents the residual connection, which preserves the original feature information.

[0010] Furthermore, the cross-cultural knowledge base includes a cultural concept knowledge layer, an idiom and phrase knowledge layer, a cultural symbol knowledge layer, an expression mode knowledge layer, and a domain adaptation knowledge layer. The cultural concept base knowledge layer includes a mapping between culture-specific concepts in the source language culture and their corresponding concepts in the target language culture; The idiom and phrase knowledge layer contains culturally specific phrases of idioms, proverbs, and slang in the source language, which are mapped to corresponding idioms in the target language. The knowledge layer of the cultural symbol database contains symbols, emblems, and metaphors from a specific culture, which are mapped to corresponding symbols in the target culture. The knowledge layer of the expression library contains specific expressions, rhetorical devices, and expression habits in different cultures, which are mapped to the corresponding expressions of the target culture. The domain adaptation knowledge layer addresses cultural differences across different industries and fields in a specialized way.

[0011] Furthermore, step 3 includes the following steps: Step 31: Perform precise matching at the word, phrase, and sentence levels on the source text, query the cross-cultural knowledge base, and directly retrieve the conversion rules for expressions that match perfectly. Step 32: Use rule-based pattern recognition technology to identify cultural expressions with specific patterns; Step 33: Calculate the semantic similarity between the expressions in the source text and the cultural expressions in the knowledge base; Step 34: Use contextual information to identify possible cultural meanings. For expressions that may be misjudged in isolation, make a second judgment using contextual information. Step 35: The conversion unit performs conversion rule queries on the identified cultural expressions, and after confidence assessment and processing, ensures that the translation results have an appropriate expression in the target culture.

[0012] Furthermore, step 4 includes the following steps: Step 41: The terminology recognition unit identifies all key terms from the source text and uses a multi-strategy terminology recognition method to identify words and positions that appear frequently in the source text to form candidate words. The identified candidate words are then matched with the domain terminology database, and the terms are identified based on statistical indicators such as TF-IDF. Step 42: The terminology scoring unit assigns a weight score to the identified terms. The formula for calculating the term weight score is as follows: Term score = 0.25 × frequency of occurrence score + 0.35 × positional importance score + 0.2 × domain universality score + 0.2 × knowledge base confidence score; Step 43: The terminology classification unit classifies the terms according to the scoring results and divides them into high-weight terms, medium-weight terms, and low-weight terms. Step 44: The pronoun recognition unit identifies and marks all pronouns in the source text, the coreference resolution unit resolves coreference for each pronoun and establishes the correspondence between the pronoun and its antecedent, the pronoun translation unit ensures that the referential relationship of pronouns in the target language is consistent with that in the source language when translating pronouns, and the ambiguity handling unit handles possible pronoun referential ambiguities. Step 45: The semantic constraint unit maintains multiple consistency constraints at the document level. When constraints at different levels conflict, priority rules are used for coordination: semantic coherence > terminology consistency > expression fluency, to ensure the complete consistency of long document translation.

[0013] Furthermore, step 5 includes the following steps: Step 51: The multi-dimensional evaluation system scores the translation results from the perspectives of semantic accuracy, fluency of expression, cultural appropriateness, and grammatical standardization. Step 52: Weight the scores for the four dimensions—semantic accuracy, fluency, cultural appropriateness, and grammatical correctness—to obtain the overall quality score. Overall score = w1 × semantic accuracy score + w2 × fluency score + w3 × cultural appropriateness score + w4 × grammatical correctness score Each dimension has a rating range of 0-1.0, and the weights are set as follows: w1=0.35 (semantic accuracy weight), w2=0.25 (expressive fluency weight), w3=0.25 (cultural appropriateness weight), w4=0.15 (grammatical correctness weight); Step 53: When the overall score is lower than the predetermined threshold, the dynamic optimization process is initiated, and the optimized translation results are re-evaluated to form a closed loop of continuous improvement until the score reaches the predetermined threshold. Specifically, if the semantic accuracy score is the lowest, the terminology table and semantic understanding parameters are readjusted; if the fluency score is the lowest, the search scope of translation solutions is expanded to find more natural expressions; if the cultural appropriateness score is the lowest, the application of the cultural knowledge base is enhanced and the cultural conversion rules are adjusted; and if the grammatical accuracy score is the lowest, grammar checks and format standardization are performed.

[0014] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention discloses a semantically optimized multilingual translation method for overseas languages. Through terminology weight scoring and a global terminology lookup table mechanism, it ensures the consistency of high-frequency terminology translation. A cross-cultural knowledge base and confidence assessment mechanism make the conversion of culture-specific expressions more accurate. Furthermore, a multi-dimensional quality assessment and dynamic optimization mechanism enables continuous improvement in translation quality, significantly enhancing the system's real-time processing capabilities. In summary, this technical solution significantly improves the system's real-time performance by generating events and event features in real time, ensuring that hot topics are captured and analyzed instantly. In addition, combined with an intelligent dynamic adjustment mechanism and personalized recommendations based on the RAG model, this solution also has significant advantages in responding to emergencies and providing accurate recommendations. These improvements enable the system to respond to user needs more promptly and accurately, and provide effective public opinion monitoring and hot topic recommendations in the early stages of information dissemination. Attached Figure Description

[0015] Figure 1 This is a flowchart of the semantically optimized overseas multilingual translation method in this invention. Detailed Implementation

[0016] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0017] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation" and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The technical solution of this invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] like Figure 1 As shown, a semantically optimized multilingual translation method for overseas markets includes the following steps: Step 1: The text preprocessing unit performs normalization and metadata extraction on the input text to obtain the source text; Step 2: The multi-level semantic analysis unit performs semantic analysis on the source text and generates a comprehensive multi-dimensional semantic representation through an adaptive weighted fusion mechanism; Step 3: The culture conversion unit identifies culturally sensitive words and expressions in the source text and uses a cross-cultural knowledge base to perform culture-adaptive conversion, ensuring that the translation results have appropriate expressions in the target culture. Step 4: The consistency constraint unit establishes a global consistency constraint mechanism for terms, pronouns, and semantics. Through term weight scoring, global term lookup table, pronoun-antecedent mapping, and document-level semantic constraints, it ensures the complete consistency of long document translation. Step 5: Conduct a multi-dimensional quality assessment of the translation results. When the overall score of the multi-dimensional quality assessment is lower than the predetermined threshold, initiate the dynamic optimization process and re-evaluate the optimized translation results to form a closed loop of continuous improvement until the score reaches the predetermined threshold. In this invention, step 1 includes the following steps: Step 11: The multilingual language recognition module performs language recognition on the content of the input text and determines the corresponding target language; Step 12: The word segmentation unit performs multi-granular word segmentation on the input text, and the sentence segmentation unit performs sentence-level segmentation on the input text. The sentence boundary recognizer identifies the punctuation marks at the end of the sentence. For the boundary cases of ellipses and abbreviations, a hybrid method based on rules and statistics is used for processing, and a unique ID number and location information are generated for each sentence. Step 13: The format processing unit performs format standardization processing on the input text; Step 14: The metadata extraction unit establishes a global metadata record and automatically configures the parameters for subsequent processing based on the metadata content to obtain the source text.

[0019] In this invention, step 2 includes the following steps: Step 21: The lexical layer analysis unit performs deep semantic analysis on each word, identifies the specific meaning of polysemous words through polysemous word resolution technology, and after resolution using contextual information, identifies the semantic relationships between words and constructs a lexical semantic relationship network. Step 22: The syntactic layer analysis unit performs in-depth analysis of the sentence's grammatical structure and performs part-of-speech tagging to identify the part of speech and grammatical function of each word; Step 23: The semantic layer analysis unit performs deep semantic analysis and uses an external knowledge base to expand semantics and supplement the semantic information of words and concepts; Step 24: The pragmatic layer analysis unit performs semantic analysis of context and cultural background, uses contextual information for disambiguation, and selects the most appropriate expression meaning; Step 25: The fusion unit uses an improved multi-head attention mechanism to perform weighted fusion of semantic features from the four levels. The fusion formula is as follows: F_fused = Σ(i=1 to 4) α_i × V_i + ResNet(V_i); Wherein: V_1 is the lexical layer feature, V_2 is the syntactic layer feature, V_3 is the semantic layer feature, V_4 is the pragmatic layer feature, α_i is the attention weight of each layer, and ResNet represents the residual connection, which preserves the original feature information.

[0020] In this invention, the cross-cultural knowledge base includes a cultural concept knowledge layer, an idiom and phrase knowledge layer, a cultural symbol knowledge layer, an expression mode knowledge layer, and a domain adaptation knowledge layer. The cultural concept base knowledge layer includes a mapping between culture-specific concepts in the source language culture and their corresponding concepts in the target language culture; The idiom and phrase knowledge layer contains culturally specific phrases of idioms, proverbs, and slang in the source language, which are mapped to corresponding idioms in the target language. The knowledge layer of the cultural symbol database contains symbols, emblems, and metaphors from a specific culture, which are mapped to corresponding symbols in the target culture. The knowledge layer of the expression library contains specific expressions, rhetorical devices, and expression habits in different cultures, which are mapped to the corresponding expressions of the target culture. The domain adaptation knowledge layer addresses cultural differences across different industries and fields in a specialized way.

[0021] In this invention, step 3 includes the following steps: Step 31: Perform precise matching at the word, phrase, and sentence levels on the source text, query the cross-cultural knowledge base, and directly retrieve the conversion rules for expressions that match perfectly. Step 32: Use rule-based pattern recognition technology to identify cultural expressions with specific patterns; Step 33: Calculate the semantic similarity between the expressions in the source text and the cultural expressions in the knowledge base; Step 34: Use contextual information to identify possible cultural meanings. For expressions that may be misjudged in isolation, make a second judgment using contextual information. Step 35: The conversion unit performs conversion rule queries on the identified cultural expressions, and after confidence assessment and processing, ensures that the translation results have an appropriate expression in the target culture.

[0022] In this invention, step 4 includes the following steps: Step 41: The terminology recognition unit identifies all key terms from the source text and uses a multi-strategy terminology recognition method to identify words and positions that appear frequently in the source text to form candidate words. The identified candidate words are then matched with the domain terminology database, and the terms are identified based on statistical indicators such as TF-IDF. Step 42: The terminology scoring unit assigns a weight score to the identified terms. The formula for calculating the term weight score is as follows: Term score = 0.25 × frequency of occurrence score + 0.35 × positional importance score + 0.2 × domain universality score + 0.2 × knowledge base confidence score; Step 43: The terminology classification unit classifies the terms according to the scoring results and divides them into high-weight terms, medium-weight terms, and low-weight terms. Step 44: The pronoun recognition unit identifies and marks all pronouns in the source text, the coreference resolution unit resolves coreference for each pronoun and establishes the correspondence between the pronoun and its antecedent, the pronoun translation unit ensures that the referential relationship of pronouns in the target language is consistent with that in the source language when translating pronouns, and the ambiguity handling unit handles possible pronoun referential ambiguities. Step 45: The semantic constraint unit maintains multiple consistency constraints at the document level. When constraints at different levels conflict, priority rules are used for coordination: semantic coherence > terminology consistency > expression fluency, to ensure the complete consistency of long document translation.

[0023] In this invention, step 5 includes the following steps: Step 51: The multi-dimensional evaluation system scores the translation results from the perspectives of semantic accuracy, fluency of expression, cultural appropriateness, and grammatical standardization. Step 52: Weight the scores for the four dimensions—semantic accuracy, fluency, cultural appropriateness, and grammatical correctness—to obtain the overall quality score. Overall score = w1 × semantic accuracy score + w2 × fluency score + w3 × cultural appropriateness score + w4 × grammatical correctness score Each dimension has a rating range of 0-1.0, and the weights are set as follows: w1=0.35 (semantic accuracy weight), w2=0.25 (expressive fluency weight), w3=0.25 (cultural appropriateness weight), w4=0.15 (grammatical correctness weight); Step 53: When the overall score is lower than the predetermined threshold, the dynamic optimization process is initiated, and the optimized translation results are re-evaluated to form a closed loop of continuous improvement until the score reaches the predetermined threshold. Specifically, if the semantic accuracy score is the lowest, the terminology table and semantic understanding parameters are readjusted; if the fluency score is the lowest, the search scope of translation solutions is expanded to find more natural expressions; if the cultural appropriateness score is the lowest, the application of the cultural knowledge base is enhanced and the cultural conversion rules are adjusted; and if the grammatical accuracy score is the lowest, grammar checks and format standardization are performed.

[0024] Example In this embodiment, when translating multiple languages ​​from overseas, the first step is to preprocess the input text.

[0025] When a user uses the system, they will import or receive an English technical document (for example, the English technical document involves cloud computing architecture and is about 5,000 words). The system identifies the source language as English and the target language as Chinese, performs word segmentation and sentence segmentation on the text, identifies 245 sentence units, and establishes meta-information records: document type = technical manual, domain = cloud computing, user level = professional translator. Based on this information, the system configures parameters for subsequent processing: the terminology consistency weight is increased to the highest priority, and the quality assessment threshold is set to 0.85.

[0026] The second step involves multi-level semantic understanding and fusion. The system performs multi-level semantic analysis on the entire text. At the lexical level, 78 key technical terms (such as cloud infrastructure, distributed system, etc.) and their semantic relationships are identified. At the syntactic level, the grammatical structure of 245 sentences is analyzed, and 33 complex clauses and 48 parallel structures are identified. At the semantic level, 42 core concepts and their inclusion relationships and 56 event structures are extracted. At the pragmatic level, the actual meaning of these contents in the context of cloud computing and business is understood. The system integrates the analysis results of each level through adaptive weighted fusion.

[0027] In the third step, identify and convert culturally specific expressions. At this time, the system identifies 12 culturally sensitive expressions in the document, including "infrastructure as a service", "cloud-native architecture", etc. The system queries the cross-cultural knowledge base to obtain the corresponding conversion rules. For example, the conversion rule for "infrastructure as a service" is "Infrastructure as a Service (IaaS)", and the confidence level of the knowledge base is 0.95. The conversion rule for "cloud-native" is "cloud-native", and the confidence level is 0.92. For expressions with a confidence level higher than 0.90, directly apply the conversion rule. For expressions with a confidence level between 0.70 and 0.90, adopt a hybrid strategy, that is, provide both the transliteration and free translation of the original expression for selection in the translation.

[0028] In the fourth step, perform multi-level consistency constraint and integration. The specific implementation of the term consistency constraint: The system performs weighted scoring on the 78 identified terms according to the term scoring formula: Term score = 0.25×Occurrence frequency + 0.35×Position importance + 0.2×Domain generality + 0.2×Knowledge base confidence.

[0029] Specific calculation example - "cloud" term: - Occurrence frequency: Appears 47 times in the document, score = 47×0.1 = 4.7 points. - Position importance: Appears in the title 2 times (2 points), and in the main text 45 times (22.5 points), total 24.5 points (exceeding the upper limit of 10 points, counted as 10 points). - Domain generality: Standard term, score 10 points. - Knowledge base confidence: The confidence level for converting to "cloud" is 0.98, score 9.8 points.

[0030] Term score = (0.25×4.7) + (0.35×10) + (0.2×10) + (0.2×9.8) = 1.175 + 3.5 + 2 + 1.96 = 8.635 points; The system incorporates the terms with a score higher than 8.0 (a total of 34 terms) into the global term comparison table and unifies their translations. For example: "cloud" → "云", "service" → "服务", "distributed" → "分布式", etc. The system ensures the translation consistency of these 34 terms in the 245 sentences of the entire document. Terms with a score lower than 8.0 are allowed to be translated flexibly according to the context.

[0031] Then, the system analyzes the pronouns in the document, identifying 67 pronouns (mainly it, they, which, etc.) and establishing a pronoun-antecedent correspondence: for example, "it" in sentence 5 refers to "the cloud platform" in sentence 3, and the system ensures that the corresponding pronoun in the target language correctly refers to "cloud platform" during translation. For possible referential ambiguities (such as a pronoun referring to multiple antecedents), the system marks them, which are then manually reviewed and confirmed. Afterward, the system maintains multiple consistency constraints at the document level: Information integrity: All key information (28 key parameters and technical indicators) in the source document is fully presented in the translation without omission; Logical relationships: The 10 causal relationships and 8 adversative relationships in the source document are kept accurate in the translation; Sentimental tone: The 9 positive expressions (emphasizing the system's advantages) in the source document maintain a consistent tone in the translation; Parameter accuracy: The 15 numerical indicators and technical parameters in the source document are accurate in the translation. The information at these three levels is integrated in stages to ensure complete consistency of the entire document in terms of terminology, pronouns, logical relationships, and parameters.

[0032] The fifth step is multi-dimensional quality assessment and dynamic optimization. After generating the initial translation, the system conducts multi-dimensional quality assessment: Semantic accuracy assessment: The system compares the semantics of the source text with that of the translation for each of the 245 sentences, with a score of 0.94. Fluency assessment: The system evaluates the naturalness of the translation in Chinese technical literature, finding that 12 sentences have low fluency (score <0.85), with an overall score of 0.89. Cultural appropriateness assessment: The system checks the conversion of 12 cultural expressions, finding that 11 conform to the target cultural norms (confidence >0.90), and 1 needs adjustment, with an overall score of 0.92. Grammatical accuracy assessment: The system checks grammar, punctuation, and formatting, finding 3 punctuation errors and 2 formatting inconsistencies, with an overall score of 0.97.

[0033] Calculate the overall score: Overall score = 0.35×0.94 + 0.25×0.89 + 0.25×0.92 + 0.15×0.97; = 0.329 + 0.2225 + 0.23 + 0.1455; = 0.927 points.

[0034] The score was 0.927, which is higher than the predetermined threshold of 0.85.

[0035] Although the comprehensive score has reached the standard, the system found that the expression fluency is the lowest (0.89) according to the scores of each dimension, and initiated local optimization. The system searched for extended translation solutions for these 12 sentences with lower fluency and tried 3-5 different expressions. For example, the preliminary translation of the original sentence "The system provides the capability to process massive datastreams in real time" was optimized to "The system is capable of processing massive data streams in real time" (more concise and in line with the expression habits of Chinese technical documents). After optimization, these 12 sentences were re-evaluated, and the expression fluency score increased to 0.94. The new comprehensive score is: Comprehensive score = 0.35×0.94 + 0.25×0.94 + 0.25×0.92 + 0.15×0.97 = 0.941 points.

[0036] At the same time, the system records the cultural knowledge base rules and term weight parameters adjusted during the optimization process. User feedback results: the satisfaction score is 4.9 points (out of 5), and no further modification is required. The translation processing time for the entire document is 3.8 seconds, with an average of 0.76 seconds per thousand words. The system saves the optimization strategy and new rules of this translation to the knowledge base for subsequent translation of similar technical documents to continuously improve the system performance.

[0037] In the figure, the description of the positional relationship is only for illustrative purposes and should not be construed as a limitation of this patent; obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention and are not intended to limit the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the claims of the present invention.

Claims

1. A method for semantic optimization of overseas multilingual translation, characterized in that: The method comprises the following steps: Step 1, the text preprocessing unit normalizes the input text and extracts meta information to obtain a source text; Step 2, the multi-level semantic analysis unit performs semantic analysis on the source text, and generates a comprehensive multi-dimensional semantic representation through an adaptive weighted fusion mechanism; Step 3, the cultural conversion unit identifies culturally sensitive words and expressions in the source text, and performs cultural adaptability conversion using a cross-cultural knowledge base to ensure that the translation result has a proper expression in the target culture; Step 4, the consistency constraint unit establishes a global consistency constraint mechanism for terms, pronouns and semantics, and ensures the integrity and consistency of long document translation through term weight scoring, global term comparison table, pronoun-antecedent mapping and document-level semantic constraints; Step 5, the translation result is subjected to multi-dimensional quality evaluation, and when the comprehensive score of multi-dimensional quality evaluation is lower than the predetermined threshold, the dynamic optimization process is started, and the optimized translation result is re-evaluated to form a closed loop of continuous improvement until the score reaches the predetermined threshold.

2. The semantically optimizable overseas multilingual translation method as claimed in claim 1, wherein: The step 1 comprises the following steps: Step 11, the multi-language language recognition module recognizes the language of the input text content and determines the corresponding target language; Step 12, the word segmentation unit performs multi-granularity word segmentation processing on the input text, the sentence segmentation unit performs sentence-level segmentation on the input text, the sentence boundary recognizer is used to recognize the punctuation symbols at the end of the sentence, and the hybrid method based on rules and statistics is used to process the boundary conditions of ellipsis and abbreviations, and a unique ID number and position information are generated for each sentence; Step 13, the format processing unit normalizes the format of the input text; Step 14, the meta information extraction unit establishes a global meta information record, and automatically configures the parameters of the subsequent processing according to the meta information content to obtain a source text.

3. The semantically optimizable overseas multilingual translation method as claimed in claim 1, wherein: The step 2 comprises the following steps: Step 21, the lexical layer analysis unit performs deep semantic analysis on each word, identifies the specific meaning of polysemous words through polysemous word resolution technology, and then identifies the semantic relationship between words after resolution using context information, and constructs a lexical semantic relationship network; Step 22, the syntactic layer analysis unit performs deep analysis on the syntactic structure of the sentence and performs part-of-speech tagging to identify the part-of-speech and grammatical function of each word; Step 23, the semantic layer analysis unit performs deep semantic analysis, and uses an external knowledge base to expand semantics to supplement semantic information of words and concepts; Step 24, the pragmatic layer analysis unit performs semantic analysis of context and cultural background, and uses context information for disambiguation to select the most appropriate expression meaning; Step 25, the fusion unit uses an improved multi-head attention mechanism to weight and fuse the semantic features of the four levels, and the fusion formula is: F_fused = Σ(i=1 to 4) α_i × V_i + ResNet(V_i); Where: V_1 is the lexical layer feature, V_2 is the syntactic layer feature, V_3 is the semantic layer feature, V_4 is the pragmatic layer feature, α_i is the attention weight of each level, and ResNet represents the residual connection to retain the original feature information.

4. The semantically optimizable overseas multilingual translation method as claimed in claim 1, wherein: The cross-cultural knowledge base comprises a cultural concept base knowledge layer, an idiom and phrase base knowledge layer, a cultural symbol base knowledge layer, an expression mode base knowledge layer and a field adaptation base knowledge layer; The cultural concept base knowledge layer comprises mapping of cultural specific concepts in a source language culture and corresponding concepts in a target language culture; The idiom and phrase base knowledge layer comprises mapping of cultural specific phrases of idioms, idiomatic expressions and slang in a source language and corresponding idioms in a target language; The cultural symbol base knowledge layer comprises mapping of symbols, symbols and metaphors in a specific culture and corresponding symbols in a target culture; The expression mode base knowledge layer comprises mapping of specific expression modes, rhetorical devices and expression habits in different cultures and corresponding expression modes in a target culture; The field adaptation base knowledge layer is specially processed for cultural differences in different industries and fields.

5. The semantically optimizable overseas multilingual translation method as claimed in claim 1, wherein: In step 3, the following steps are included: Step 31, accurate matching of words, phrases and sentences in the source text is performed, the cross-cultural knowledge base is queried, and for completely matched expressions, conversion rules are directly retrieved; Step 32, a rule-based pattern recognition technique is used to identify cultural expressions with specific patterns; Step 33, the semantic similarity of expressions in the source text and the cultural expressions in the knowledge base is calculated; Step 34, possible cultural meanings are identified using context information, and for expressions that may be misjudged in isolation, secondary judgment is made through context information; Step 35, the conversion unit queries the conversion rules for the identified cultural expressions, and after confidence evaluation and processing, it ensures that the translation result has appropriate expression modes in the target culture.

6. The semantically optimizable overseas multilingual translation method as claimed in claim 1, wherein: In step 4, the following steps are included: Step 41, a term recognition unit identifies all key terms from the source text, and a multi-strategy term recognition method is used to identify candidate words with high frequency and position in the source text, and the identified candidate words are matched with the field term base, and terms are identified based on statistical indicators such as TF-IDF; Step 42, a term scoring unit scores the identified terms, wherein the term weight scoring formula is: Term score = 0.25 × frequency score + 0.35 × position importance score + 0.2 × field generality score + 0.2 × knowledge base confidence score; Step 43, a term classification unit classifies terms according to the scoring results, and divides them into high weight terms, medium weight terms and low weight terms; Step 44, a pronoun recognition unit identifies and labels all pronouns in the source text, a co-reference resolution unit resolves co-reference for each pronoun, and a pronoun translation unit ensures that the reference relationship of pronouns in the target language is consistent with that in the source language, and an ambiguity processing unit processes possible pronoun reference ambiguity; Step 45, a semantic constraint unit maintains multiple consistency constraints at the document level, and when different levels of constraints conflict, priority rules are used for coordination: semantic coherence > term consistency > expression fluency, to ensure the integrity and consistency of long document translation.

7. The semantically optimizable overseas multilingual translation method as claimed in claim 1, wherein: Step 5 includes the following steps: Step 51, the multi-dimensional evaluation system scores the translation results in terms of semantic accuracy, expression fluency, cultural appropriateness, and grammatical correctness respectively; Step 52, the scores of semantic accuracy, expression fluency, cultural appropriateness, and grammatical correctness are weighted and calculated to obtain the comprehensive quality score: Comprehensive score = w1 x semantic accuracy score + w2 x expression fluency score + w3 x cultural appropriateness score + w4 x grammatical correctness score Wherein the score of each dimension ranges from 0 to 1.0, and the weights are set as follows: w1=0.35 (semantic accuracy weight), w2=0.25 (expression fluency weight), w3=0.25 (cultural appropriateness weight), w4=0.15 (grammatical correctness weight); Step 53, when the comprehensive score is lower than the predetermined threshold, the dynamic optimization process is started, and the optimized translation results are re-evaluated to form a continuous improvement closed loop until the score reaches the predetermined threshold, wherein if the semantic accuracy score is the lowest, the term dictionary and semantic understanding parameters are adjusted, if the expression fluency score is the lowest, the translation scheme search range is expanded to find a more natural expression, if the cultural appropriateness score is the lowest, the cultural knowledge base application is enhanced, the cultural conversion rules are adjusted, and if the grammatical correctness score is the lowest, the grammar check and format standardization processing are performed.

Images