Multi-language real-time semantic alignment translation system based on cross-language pre-training model

Abstract

The invention discloses a multilingual real-time semantic alignment translation system based on a cross-language pre-training model, and relates to the technical field of language translations, and the multilingual real-time semantic alignment translation system comprises the following steps: constructing an original translation statement and a filling analysis statement of the same semantic phrase based on the cross-language pre-training model; using an alignment analysis method to obtain vocabulary structure features; translating the to-be-translated statement by using a cross-language pre-training model based on the vocabulary structure features of the same semantic phrases corresponding to the to-be-translated statement; the method is used for solving the problems that in an existing multilingual real-time semantic alignment translation method, accurate and contextual semantic alignment translation cannot be carried out on specific vocabularies based on the context relation in sentences before and after translation, so that partial vocabularies in the translated sentences do not accord with the meaning of the original text, and the semantics of the original text cannot be accurately expressed.

Description

Technical Field

[0001] This invention relates to the field of language translation technology, specifically to a multilingual real-time semantic alignment translation system based on a cross-language pre-trained model. Background Technology

[0002] Multilingual real-time semantic alignment translation is a technology that completes multilingual translation and semantic matching and alignment in real time. Its core is not word-for-word translation, but ensuring complete semantic equivalence between different languages. At the same time, it can meet the requirements of real-time response and clearly mark the correspondence between the source text and the translation, solving the problems of word-for-word correspondence but semantic deviation in traditional translation, as well as the contradiction between real-time performance and accuracy.

[0003] Existing methods for real-time semantic alignment translation in multilingual languages ​​typically involve collecting multilingual text data from the internet, performing language identification and structured storage, and then using cross-language pre-trained models to achieve semantic alignment matching between sentences in different languages ​​to generate high-quality corpora. While this approach provides a high-quality corpus for real-time semantic alignment translation in multilingual languages, it fails to accurately and contextually align specific words within certain languages ​​based on the contextual relationships between the translated and untranslated sentences. This results in some words in the translated sentences not matching the original meaning and failing to accurately express the original meaning. For example, a multilingual corpus is disclosed in patent application CN120671690A. The automatic construction and translation optimization system automatically collects multilingual text data from the Internet, performs language identification and structured storage, and uses a cross-language pre-trained model to perform semantic vector encoding and alignment matching on sentences in different languages ​​to generate high-quality parallel corpora. However, other improvements to multilingual real-time semantic alignment translation methods usually focus on improving translation performance and data reliability, but they still lack the ability to translate specific words in some languages. They cannot accurately and contextually align specific words based on the contextual relationship between the sentences before and after translation, resulting in some words in the translated sentences that do not match the meaning of the original text and fail to accurately express the meaning of the original text. Therefore, it is necessary to improve the existing multilingual real-time semantic alignment translation methods. Summary of the Invention

[0004] This invention aims to at least partially solve one of the technical problems in the prior art by proposing a multilingual real-time semantic alignment translation system based on a cross-language pre-trained model. This system addresses the problem that existing multilingual real-time semantic alignment translation methods cannot accurately and contextually align specific words based on the contextual relationship between the translated and untranslated sentences, resulting in some words in the translated sentences not matching the meaning of the original text and failing to accurately express the original meaning.

[0005] To achieve the above objectives, this application provides a multilingual real-time semantic alignment translation system based on a cross-language pre-trained model, including a semantic phrase acquisition module, a semantic language analysis module, and a multilingual semantic alignment translation module; The semantic phrase acquisition module is used to acquire multiple translated semantic phrases and construct the original translation sentence and the filling analysis sentence for each semantic phrase based on the cross-language pre-trained model. Here, a semantic phrase is a group of words with the same meaning but different languages. The semantic language analysis module is used to analyze the original translated sentences and the fill-in analysis sentences of each semantic phrase using the alignment analysis method, and to obtain the lexical structure features of each semantic phrase. The multilingual semantic alignment translation module is used to translate sentences by using a cross-language pre-trained model based on the lexical structure features of the semantically related phrases corresponding to the sentences to be translated.

[0006] Furthermore, the semantic phrase acquisition module includes a semantic phrase acquisition unit, which is configured with a semantic phrase acquisition strategy, including: Obtain multiple translated semantic phrases; for any semantic phrase α: denote the language corresponding to the translated words in semantic phrase α as the original language, denote the language obtained by translating the original language in semantic phrase α as the translation language; denote the words corresponding to the original language in semantic phrase α as the original words, and denote the words corresponding to the translation language in semantic phrase α as the translation words; A sentence containing the original vocabulary is obtained based on a cross-language pre-trained model and denoted as the original translation sentence γ. For any original translation sentence γ and any translation vocabulary β, the part of the original translation sentence γ excluding the original vocabulary is translated into the translation language based on the cross-language pre-trained model, and the translated sentence is denoted as the translation analysis sentence.

[0007] Furthermore, the semantic phrase acquisition strategy also includes: Replace the original words in the translation analysis statement with the translated words β, and record the translation analysis statement at this time as the fill analysis statement; Obtain the original translated sentence and the corresponding fill analysis sentences for all translated words, and record them as the aligned translation sentences of the same semantic phrase.

[0008] Furthermore, the semantic language analysis module includes a semantic language analysis unit, which is configured with an alignment analysis method and a semantic language analysis strategy. The alignment analysis method includes: For any semantically similar phrase α: use the sentence analysis sub-method to analyze the original vocabulary of the original translated sentence γ; the sentence analysis sub-method includes: marking the nouns in the original vocabulary as original nouns, and marking the part of the original translated sentence γ other than the original vocabulary as excluded sentences; For any original noun: when the original noun exists in the exclusion statement, the original noun is recorded as a directly translatable noun; when the original noun does not exist in the exclusion statement, the original noun is recorded as a semantically analyzed noun.

[0009] Furthermore, the statement analysis sub-methods also include: Based on a cross-linguistic pre-trained model, the numerals and adjectives that modify semantically analyzed nouns in the original vocabulary are obtained and recorded as modifying numerals and modifying adjectives; For any modifier numeral, obtain the numerals in the exclusion statement that modify the semantically analyzed noun and record them as exclusion modifier numerals; when the sum of the quantities corresponding to all exclusion modifier numerals is not equal to the quantity corresponding to the modifier numeral, the modifier numeral is recorded as a dummy numeral; when the sum of the quantities corresponding to all exclusion modifier numerals is equal to the quantity corresponding to the modifier numeral, the modifier numeral is recorded as an exact numeral. Based on the grammar of the original language, adjectives that are related to dummy index words in the grammar of the original language are recorded as translated adjectives of dummy index words; For any modifying adjective, words with similar meanings to the modifying adjective in the self-exclusion statement are obtained based on a cross-linguistic pre-trained model and recorded as figurative adjectives; words modified by figurative adjectives in the self-exclusion statement are recorded as figurative nouns, and the noun recorded as the most times in the self-exclusion statement is recorded as the pronoun of the semantic analysis noun.

[0010] Furthermore, alignment analysis also includes: For any fill analysis statement corresponding to the original translation statement γ: use the statement analysis sub-method to analyze the translation vocabulary in the fill analysis statement, and mark the nouns in the translation vocabulary as translation nouns during the analysis process; When all translated nouns in the analysis results are directly translatable nouns, the language of the translated nouns is recorded as the perfect language; when the translated nouns in the analysis results are recorded as semantic analysis nouns, the noun translation features of the translated nouns are recorded as the pronouns of the semantic analysis nouns corresponding to the original translated sentence γ after the translated nouns are translated into the original language; the noun modification features of the translated nouns are recorded as the translated adjectives of the pluralistic index words of the semantic analysis nouns corresponding to the original translated sentence γ after the translated nouns are translated into the original language.

[0011] Furthermore, alignment analysis also includes: Within all the fill analysis statements corresponding to the original translated statement γ, when the translated noun is recorded as a semantic analysis noun, the noun translation features and noun modification features of the translated noun are obtained; Obtain the original and translated vocabulary within all semantically related phrases, and use alignment analysis to obtain the noun translation features and noun modification features of the translated nouns within all semantically related phrases.

[0012] Furthermore, semantic language analysis strategies include: The structural feature acquisition method is used to obtain the lexical structural features of the original vocabulary of the same semantic phrase α. The structural feature acquisition method includes: segmenting the original vocabulary of the same semantic phrase α, and placing the part of speech of all words corresponding to the segmented original vocabulary into a set based on the reading order from first to last, and recording the set as the lexical structural features of the same semantic phrase α.

[0013] Furthermore, the multilingual semantic alignment translation module includes a multilingual semantic alignment translation unit, and the multilingual semantic alignment translation configuration includes a multilingual semantic alignment translation strategy, which includes: When translating a statement, the statement to be translated is denoted as statement A, and the language of statement A is denoted as the language to be translated. Using the structural feature acquisition method, obtain the lexical structural features corresponding to all words in statement A. When the lexical structural feature of any phrase B in statement A is the same as the lexical structural feature of any semantic phrase α in the language to be translated, the semantic phrase α is recorded as a reference word. The language obtained after translating statement A is recorded as the target language; the part of statement A except for phrase B is translated into the target language using a cross-language pre-trained model, and the resulting statement is recorded as the statement to be filled.

[0014] Furthermore, multilingual semantic alignment translation strategies also include: For any noun C in phrase B, obtain the fill analysis statement in the aligned translation statement of the reference vocabulary that is in the target language, and denote it as fill statement D; When the alignment analysis method analyzes the fill statement D of the reference vocabulary, and the language of noun C is recorded as the perfect language, the cross-language pre-trained model is used to translate noun C and the words modifying noun C into the target language. Based on the position of noun C and the words modifying noun C in statement A, the translated noun C and the words modifying noun C are filled into the statement to be filled. When the alignment analysis method does not record the language of noun C as the perfect language during the analysis of the fill-in sentence D of the reference vocabulary, after translating noun C into the target language using a cross-language pre-trained model, the noun translation features and noun modification features of noun C in sentence A are obtained based on the method of obtaining the noun translation features and noun modification features of the translated noun in the original translated sentence γ. The translation results corresponding to noun C and words modifying noun C in phrase B are set as follows: the noun translation features and noun modification features of noun C are translated into the target language; based on the position of noun C and words modifying noun C in sentence A, the translated noun C and words modifying noun C are filled into the sentence to be filled.

[0015] The beneficial effects of this invention are as follows: This application first obtains multiple translated synonymous phrases, and then constructs the original translation sentences and fill-in analysis sentences for each synonymous phrase based on a cross-language pre-trained model. The advantage of this is that by constructing the original translation sentences and fill-in analysis sentences for each synonymous phrase based on a cross-language pre-trained model, the linguistic environment of the synonymous phrase can be constructed using the cross-language pre-trained model. This ensures that words of different languages ​​and the same meaning are filled into their appropriate linguistic context, that is, the original translation sentences and fill-in analysis sentences are obtained. This facilitates the analysis of synonymous phrases with the same meaning in subsequent analysis, ensuring that the characteristics of specific words when they are translated are obtained, thereby achieving accurate translation of specific words in synonymous phrases that conforms to the semantics of the original text. This application also uses alignment analysis to analyze the original translated sentences and the fill-in analysis sentences of each semantic phrase, and obtains the lexical structure features of each semantic phrase. Finally, when translating the sentence to be translated, a cross-language pre-trained model is used to translate the sentence based on the lexical structure features of the semantic phrases corresponding to the sentence to be translated. The advantage of this is that by using alignment analysis to analyze the original translated sentences and the fill-in analysis sentences of each semantic phrase, the noun translation features and noun modification features of the nouns within the semantic phrases obtained from the analysis results can be used to translate specific words during the translation process. This achieves accurate and context-appropriate semantic alignment translation of specific words based on the contextual relationship between the sentences before and after translation. Furthermore, by obtaining the lexical structure features of the semantic phrases, specific words can be located within the sentence to be translated, thereby achieving accurate translation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system of the present invention; Figure 2It is a schematic flowchart of the alignment analysis method of the present invention; Figure 3 It is a schematic flowchart of the sentence analysis sub-method of the present invention. Specific implementation manners

[0017] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0018] Please refer to Figure 1 As shown, the present application provides a multilingual real-time semantic alignment translation system based on a cross-lingual pre-trained model, including a same-semantic phrase acquisition module, a same-semantic language analysis module, and a multilingual semantic alignment translation module; The same-semantic phrase acquisition module is used to acquire multiple translated same-semantic phrases, and construct the original translation sentences and padding analysis sentences of each same-semantic phrase based on the cross-lingual pre-trained model. Here, the same-semantic phrases are multiple words with the same semantics and different languages; The same-semantic phrase acquisition module includes a same-semantic phrase acquisition unit, and the same-semantic phrase acquisition unit is configured with a same-semantic phrase acquisition strategy, and the same-semantic phrase acquisition strategy includes: Acquire multiple translated same-semantic phrases; for any same-semantic phrase α: record the language corresponding to the translated word in the same-semantic phrase α as the original language, and record the language obtained by translating the original language in the same-semantic phrase α as the translated language; record the word corresponding to the original language in the same-semantic phrase α as the original word, and record the word corresponding to the translated language in the same-semantic phrase α as the translated word; In the analysis of this embodiment, for example, a translated same-semantic phrase obtained is "He's all talk and no action" after translating the phrase "He talks big but does little". Through analysis, it can be obtained that in the above same-semantic phrase, the language corresponding to the translated word is Chinese, that is, the original language is Chinese; the language obtained by translating the original language is English, that is, the translated language is English; therefore, the original word consists of "He talks big but does little", and the translated word consists of "He's all talk and no action"; A sentence containing the original vocabulary is obtained based on a cross-language pre-trained model and denoted as the original translation sentence γ. For any original translation sentence γ and any translation vocabulary β, the part of the original translation sentence γ excluding the original vocabulary is translated into the translation language based on the cross-language pre-trained model, and the translated sentence is denoted as the translation analysis sentence.

[0019] The semantic phrase acquisition strategy also includes: replacing the position of the original word in the translation analysis statement with the translated word β, and recording the translation analysis statement at this time as the fill analysis statement; Obtain the original translated sentence and the corresponding fill analysis sentences for all translated words, and record them as the aligned translation sentences of the same semantic phrase; In the analysis of this embodiment, for example, by analyzing the original vocabulary and the translated vocabulary, the original translated sentence γ is "He is all talk and no action; although he always makes promises verbally, he hardly ever follows through with deeds." After translating the part of the original translated sentence γ other than the original vocabulary into the target language, the resulting translation analysis sentence is "He is all talk and no action. Though he always makes promises verbally, he hardly ever follows through with deeds." The position of the original vocabulary in the translation analysis sentence is replaced with the translated vocabulary β, that is, replaced with "He's all talk and no action." The resulting aligned translation sentence is "He's all talk and no action. Though he always makes promises verbally, he hardly ever follows through with deeds."

[0020] The semantic language analysis module is used to analyze the original translated sentences and the fill-in analysis sentences of each semantic phrase using the alignment analysis method, and to obtain the lexical structure features of each semantic phrase. The semantic language analysis module includes semantic language analysis units, which are configured with alignment analysis methods and semantic language analysis strategies. Please refer to [link / reference]. Figure 2 As shown, alignment analysis methods include: For any semantically similar phrase α: use the sentence analysis sub-method to analyze the original vocabulary of the original translated sentence γ; please refer to Figure 3 As shown, the sentence analysis sub-method includes: marking the nouns in the original vocabulary as original nouns, and recording the part of the original translated sentence γ other than the original vocabulary as excluded sentences; For any original noun: When the original noun exists in the self-excluding statement, the original noun is recorded as a directly translatable noun; when the original noun does not exist in the self-excluding statement, the original noun is recorded as a semantic analysis noun; In the analysis of this embodiment, for example, when analyzing the above-mentioned synonymous phrases, it can be obtained that the noun in the original vocabulary is "thunder", that is, the original noun contains "thunder"; the self-excluding statement obtained from the original translation statement γ is "Although there are always verbal promises, little is actually done", and through analysis, it can be seen that the original noun "thunder" does not exist in the self-excluding statement, indicating that "thunder" may be a noun that is specially translated during translation. Therefore, "thunder" can be recorded as a semantic analysis noun.

[0021] The statement analysis sub-method further includes: obtaining the numerals and adjectives that modify the semantic analysis nouns in the original vocabulary based on the cross-language pre-trained model, and recording them as modified numerals and modified adjectives; In the analysis of this embodiment, by analyzing the above-mentioned "He talks big but does little", it can be obtained that the modified adjective of the semantic analysis noun "thunder" is "big", and there is no modified numeral; For any modified numeral, obtain the numerals that modify the semantic analysis noun in the self-excluding statement, and record them as self-excluding modified numerals; when the sum of the quantities corresponding to all self-excluding modified numerals is not equal to the quantity corresponding to the modified numeral, the modified numeral is recorded as a virtual numeral; when the sum of the quantities corresponding to all self-excluding modified numerals is equal to the quantity corresponding to the modified numeral, the modified numeral is recorded as an accurate numeral; In the specific implementation process, for example, the original vocabulary for analysis is "all sorts of people", and the original noun in it is "teaching", and through analysis, the modified numeral of "teaching" is "three"; through Chinese semantic analysis, it can be obtained that "three" in "all sorts of people" refers to "all kinds, all colors", so in the self-excluding modified numerals obtained in the self-excluding statement, the sum of the quantities corresponding to all self-excluding modified numerals will not be equal to the quantity "three" corresponding to the modified numeral. Therefore, "three" can be recorded as a virtual numeral, and in subsequent analysis, based on the grammar of Chinese, the adjectives "all kinds, all colors" associated with "three" are recorded as the translation adjectives of the virtual numeral; similarly, "nine" in "all sorts of people" is also a virtual numeral, and the corresponding translation adjective is also "all kinds, all colors". Therefore, when translating "all sorts of people" later, after identifying "all sorts of people" based on the lexical structure characteristics, during the translation process, "three" and "nine" can be uniformly translated as "all kinds, all colors", and after combining the semantics of "teaching" and "class", "all sorts of people" is translated as "all walks of life"; Based on the grammar of the original language, the adjectives in the grammar of the original language that are associated with the virtual numeral are recorded as the translation adjectives of the virtual numeral; For any modifying adjective, obtain words similar in meaning to the modifying adjective in the self-excluding sentence based on a cross-lingual pre-trained model, and record them as virtual adjectives; record the word modified by the virtual adjective in the self-excluding sentence as a virtual noun, and record the noun with the most occurrences of the virtual noun in the self-excluding sentence as the proxy noun for the semantic analysis noun; In the analysis of this embodiment, through the above analysis, the modifying adjective of the semantic analysis noun "thunder" is "big". Therefore, by obtaining the adjective "always" similar in meaning to "big" in the self-excluding sentence "Although always making promises verbally, actually doing very little", that is, the virtual adjective is "always"; the only word modified by "always" in the self-excluding sentence is "making promises verbally", so the virtual noun is "making promises verbally", and "making promises verbally" can be recorded as the proxy noun of the semantic analysis noun "thunder". When translating "He talks big but does little", the semantics of "thunder" can be translated as "making promises verbally", and through semantic transformation, it can be briefly translated as "talk", so "talking big" is translated as "all talk".

[0022] The alignment analysis method further includes: for any filling analysis sentence corresponding to the original translation sentence γ: use the sentence analysis sub-method to analyze the translation vocabulary in the filling analysis sentence, and mark the nouns in the translation vocabulary as translation nouns during the analysis process; When all the translation nouns in the analysis result are directly translatable nouns, record the language of the translation nouns as the perfect language; when the translation nouns in the analysis result are recorded as semantic analysis nouns, record the noun translation feature of the translation nouns as the proxy noun of the corresponding semantic analysis noun in the original translation sentence γ after translating the translation nouns into the original language; record the noun modification feature of the translation nouns as the translation adjective of the virtual numeral of the corresponding semantic analysis noun in the original translation sentence γ after translating the translation nouns into the original language; In the analysis of this embodiment, for example, a filling analysis statement of the original translation statement "He's all thunder but little rain. No matter how confidently he swears he'll get it done, it will most likely fizzle out in the end" obtained from the original translation statement "他这个人雷声大,雨点小,虽然总是口头上保证,实际做的事却很少" is as follows. When analyzing, "thunder" is recorded as a semantic analysis noun. Therefore, the noun translation feature of "thunder" should be that after translating "thunder" into the original language "雷", the referring noun of "雷" is "口头上保证". Also, because "口头上保证" can be briefly translated as "talk" through semantic transformation, during the translation process, "thunder" can be transformed into "talk".

[0023] The alignment analysis method further includes: obtaining, within all the filling analysis statements corresponding to the original translation statement γ, the noun translation feature and the noun modification feature of the translation noun when the translation noun is recorded as a semantic analysis noun. Obtain the original words and translation words within all the same-semantic phrases, and based on the alignment analysis method, obtain the noun translation feature and the noun modification feature of the translation noun within all the same-semantic phrases.

[0024] The same-semantic language analysis strategy includes: using the structural feature acquisition method to obtain the lexical structural feature of the original words of the same-semantic phrase α. The structural feature acquisition method includes: performing word segmentation on the original words of the same-semantic phrase α, and placing the词性 of all the words corresponding to the segmented original words in a set in the order from first to last according to the reading order, and recording the set as the lexical structural feature of the same-semantic phrase α. In the analysis of this embodiment, for example, when analyzing the above original words "他这个人雷声大,雨点小", according to the reading order from first to last, the obtained words are依次 "他", "这", "个人", "雷声", "大", "雨点", "小". The lexical structural feature obtained by obtaining the词性 is [pronoun, demonstrative pronoun, noun, noun, adjective, noun, adjective].

[0025] The multilingual semantic alignment translation module is used to translate the statement to be translated. Based on the lexical structural feature of the same-semantic phrase corresponding to the statement to be translated, use the cross-language pre-trained model to translate the statement to be translated. The multilingual semantic alignment translation module includes a multilingual semantic alignment translation unit. The multilingual semantic alignment translation is configured with a multilingual semantic alignment translation strategy. The multilingual semantic alignment translation strategy includes: When translating the sentence to be translated, denote the sentence to be translated as sentence A, and denote the language of sentence A as the language to be translated; Use the structural feature acquisition method to obtain the lexical structure features corresponding to all the words in sentence A. When the lexical structure features of any phrase B in sentence A are the same as those of any same-semantic phrase α in the language to be translated, denote the same-semantic phrase α as the reference word; In the specific implementation process, for example, in one analysis, the language to be translated is Chinese, and the lexical structure features of phrase B and those of a Chinese same-semantic phrase "He talks big but does little" are both [pronoun, demonstrative pronoun, noun, noun, adjective, noun, adjective]. Then, "He talks big but does little" can be denoted as the reference word; Denote the language obtained after translating sentence A as the target language; use the cross-lingual pre-trained model to translate the part of sentence A except phrase B into the target language, and denote the obtained sentence as the sentence to be filled;

[0026] The multi-lingual semantic alignment translation strategy also includes: for any noun C in phrase B, obtain the filling analysis sentence in the target language in the alignment translation sentence of the reference word, and denote it as filling sentence D; When, in the analysis process of the alignment analysis method for the filling sentence D of the reference word, denote the language of noun C as the perfecting language, use the cross-lingual pre-trained model to translate noun C and the words modifying noun C into the target language, and based on the positions of noun C and the words modifying noun C in sentence A, fill the translated noun C and the words modifying noun C into the sentence to be filled; When, in the analysis process of the alignment analysis method for the filling sentence D of the reference word, the language of noun C is not denoted as the perfecting language, after using the cross-lingual pre-trained model to translate noun C into the target language, obtain the noun translation feature and noun modification feature of noun C in sentence A based on the method of obtaining the noun translation feature and noun modification feature of the translated noun in the original translation sentence γ; In the specific implementation process, for example, if statement A is "He's all talk and no action, he only reads books and never does anything practical," and phrase B is "He's all talk and no action," and the noun C being analyzed is "thunder," then through the analysis in this embodiment, the noun translation feature obtained in the original translation statement γ for the noun "thunder" is "verbal assurance," which is then translated into "talk." Therefore, through the same analysis, the noun translation feature for "thunder" in statement A is "reading books," which is then translated into "talk." Therefore, the translation result for noun C in phrase B can be set as "talk." Since no numerals modifying "thunder" appear in phrase B, "all talk" can be directly translated using a cross-language pre-trained model, that is, "all talk" in statement A can be translated into "all talk." In summary, the translation of statement A can be obtained as "He's all talk and no action. He only reads books and never gets anything practical done." The translation results corresponding to noun C and words modifying noun C in phrase B are set as follows: the noun translation features and noun modification features of noun C are translated into the target language; based on the position of noun C and words modifying noun C in sentence A, the translated noun C and words modifying noun C are filled into the sentence to be filled.

[0027] Working principle: First, multiple translated synonymous phrases are acquired, and the original translation sentence and the fill-in analysis sentence for each synonymous phrase are constructed based on a cross-language pre-trained model. Here, a synonymous phrase is a group of words with the same meaning but different languages. Then, alignment analysis is used to analyze the original translation sentence and the fill-in analysis sentence for each synonymous phrase, and the lexical structure features of each synonymous phrase are obtained. Finally, when translating the sentence to be translated, the cross-language pre-trained model is used to translate the sentence to be translated based on the lexical structure features of the synonymous phrases corresponding to the sentence to be translated.

[0028] Based on the above description of the embodiments, the embodiments of the present invention can be provided as methods, systems, or computer program products. Based on this understanding, the above technical solutions, in essence or in terms of their contribution to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or certain parts of the embodiments.

[0029] In the embodiments provided in this application, it should be understood that the disclosed system or method can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of modules or units is only a logical functional division, and there may be other division methods in actual implementation. Furthermore, multiple modules or units may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interfaces. The indirect coupling or communication connection between systems, modules, and units may be electrical, mechanical, or other forms.

[0030] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A multilingual real-time semantic alignment translation system based on a cross-language pre-trained model, characterized in that, This includes a module for obtaining semantically related phrases, a module for analyzing semantically related languages, and a module for translating semantically related phrases in multiple languages. The semantic phrase acquisition module is used to acquire multiple translated semantic phrases and construct the original translation sentence and the filling analysis sentence for each semantic phrase based on the cross-language pre-trained model. Here, a semantic phrase is a group of words with the same meaning but different languages. The semantic language analysis module is used to analyze the original translated sentences and the fill-in analysis sentences of each semantic phrase using the alignment analysis method, and to obtain the lexical structure features of each semantic phrase. The multilingual semantic alignment translation module is used to translate sentences by using a cross-language pre-trained model based on the lexical structure features of the semantically related phrases corresponding to the sentences to be translated.

2. The multilingual real-time semantic alignment translation system based on a cross-language pre-trained model according to claim 1, characterized in that, The semantic phrase acquisition module includes a semantic phrase acquisition unit, which is configured with a semantic phrase acquisition strategy. The semantic phrase acquisition strategy includes: Obtain multiple translated semantic phrases; for any semantic phrase α: denote the language corresponding to the translated words in semantic phrase α as the original language, denote the language obtained by translating the original language in semantic phrase α as the translation language; denote the words corresponding to the original language in semantic phrase α as the original words, and denote the words corresponding to the translation language in semantic phrase α as the translation words; A sentence containing the original vocabulary is obtained based on a cross-language pre-trained model and denoted as the original translation sentence γ. For any original translation sentence γ and any translation vocabulary β, the part of the original translation sentence γ excluding the original vocabulary is translated into the translation language based on the cross-language pre-trained model, and the translated sentence is denoted as the translation analysis sentence.

3. The multilingual real-time semantic alignment translation system based on a cross-language pre-trained model according to claim 2, characterized in that, Semantic phrase acquisition strategies also include: Replace the original words in the translation analysis statement with the translated words β, and record the translation analysis statement at this time as the fill analysis statement; Obtain the original translated sentence and the corresponding fill analysis sentences for all translated words, and record them as the aligned translation sentences of the same semantic phrase.

4. The multilingual real-time semantic alignment translation system based on a cross-language pre-trained model according to claim 3, characterized in that, The semantic language analysis module includes semantic language analysis units, each configured with alignment analysis methods and semantic language analysis strategies. The alignment analysis methods include: For any semantically similar phrase α: use the sentence analysis sub-method to analyze the original vocabulary of the original translated sentence γ; the sentence analysis sub-method includes: marking the nouns in the original vocabulary as original nouns, and marking the part of the original translated sentence γ other than the original vocabulary as excluded sentences; For any given original noun: if the original noun exists in the exclusion statement, the original noun is recorded as a directly translatable noun; if the original noun does not exist in the exclusion statement, the original noun is recorded as a semantically analyzed noun.

5. The multilingual real-time semantic alignment translation system based on a cross-language pre-trained model according to claim 4, characterized in that, The statement analysis sub-methods also include: Based on a cross-linguistic pre-trained model, the numerals and adjectives that modify semantically analyzed nouns in the original vocabulary are obtained and recorded as modifying numerals and modifying adjectives; For any modifier numeral, obtain the numerals in the exclusion statement that modify the semantically analyzed noun and record them as exclusion modifier numerals; when the sum of the quantities corresponding to all exclusion modifier numerals is not equal to the quantity corresponding to the modifier numeral, the modifier numeral is recorded as a dummy numeral; when the sum of the quantities corresponding to all exclusion modifier numerals is equal to the quantity corresponding to the modifier numeral, the modifier numeral is recorded as an exact numeral. Based on the grammar of the original language, adjectives that are related to dummy index words in the grammar of the original language are recorded as translated adjectives of dummy index words; For any modifying adjective, words with similar meanings to the modifying adjective in the self-exclusion statement are obtained based on a cross-linguistic pre-trained model and recorded as figurative adjectives; words modified by figurative adjectives in the self-exclusion statement are recorded as figurative nouns, and the noun recorded as the most times in the self-exclusion statement is recorded as the pronoun of the semantic analysis noun.

6. The multilingual real-time semantic alignment translation system based on a cross-language pre-trained model according to claim 5, characterized in that, Alignment analysis also includes: For any fill analysis statement corresponding to the original translation statement γ: use the statement analysis sub-method to analyze the translation vocabulary in the fill analysis statement, and mark the nouns in the translation vocabulary as translation nouns during the analysis process; When all translated nouns in the analysis results are directly translatable nouns, the language of the translated nouns is recorded as the perfect language; when the translated nouns in the analysis results are recorded as semantic analysis nouns, the noun translation features of the translated nouns are recorded as the pronouns of the semantic analysis nouns corresponding to the original translated sentence γ after the translated nouns are translated into the original language; the noun modification features of the translated nouns are recorded as the translated adjectives of the pluralistic index words of the semantic analysis nouns corresponding to the original translated sentence γ after the translated nouns are translated into the original language.

7. The multilingual real-time semantic alignment translation system based on a cross-language pre-trained model according to claim 6, characterized in that, Alignment analysis also includes: Within all the fill analysis statements corresponding to the original translated statement γ, when the translated noun is recorded as a semantic analysis noun, the noun translation features and noun modification features of the translated noun are obtained; Obtain the original and translated vocabulary within all semantically related phrases, and use alignment analysis to obtain the noun translation features and noun modification features of the translated nouns within all semantically related phrases.

8. The multilingual real-time semantic alignment translation system based on a cross-language pre-trained model according to claim 7, characterized in that, Synonymous language analysis strategies include: The structural feature acquisition method is used to obtain the lexical structural features of the original vocabulary of the same semantic phrase α. The structural feature acquisition method includes: segmenting the original vocabulary of the same semantic phrase α, and placing the part of speech of all words corresponding to the segmented original vocabulary into a set based on the reading order from first to last, and recording the set as the lexical structural features of the same semantic phrase α.

9. The multilingual real-time semantic alignment translation system based on a cross-language pre-trained model according to claim 8, characterized in that, The multilingual semantic alignment translation module includes multilingual semantic alignment translation units. The multilingual semantic alignment translation configuration includes multilingual semantic alignment translation strategies, which include: When translating a statement, the statement to be translated is denoted as statement A, and the language of statement A is denoted as the language to be translated. Using the structural feature acquisition method, obtain the lexical structural features corresponding to all words in statement A. When the lexical structural feature of any phrase B in statement A is the same as the lexical structural feature of any semantic phrase α in the language to be translated, the semantic phrase α is recorded as a reference word. The language obtained after translating statement A is recorded as the target language; the part of statement A except for phrase B is translated into the target language using a cross-language pre-trained model, and the resulting statement is recorded as the statement to be filled.

10. The multilingual real-time semantic alignment translation system based on a cross-language pre-trained model according to claim 9, characterized in that, Multilingual semantic alignment translation strategies also include: For any noun C in phrase B, obtain the fill analysis statement in the aligned translation statement of the reference vocabulary that is in the target language, and denote it as fill statement D; When the alignment analysis method analyzes the fill statement D of the reference vocabulary, and the language of noun C is recorded as the perfect language, the cross-language pre-trained model is used to translate noun C and the words modifying noun C into the target language. Based on the position of noun C and the words modifying noun C in statement A, the translated noun C and the words modifying noun C are filled into the statement to be filled. When the alignment analysis method does not record the language of noun C as the perfect language during the analysis of the fill-in sentence D of the reference vocabulary, after translating noun C into the target language using a cross-language pre-trained model, the noun translation features and noun modification features of noun C in sentence A are obtained based on the method of obtaining the noun translation features and noun modification features of the translated noun in the original translated sentence γ. The translation results corresponding to noun C and words modifying noun C in phrase B are set as follows: the noun translation features and noun modification features of noun C are translated into the target language; based on the position of noun C and words modifying noun C in sentence A, the translated noun C and words modifying noun C are filled into the sentence to be filled.

Images