LLM-based multilingual machine translation automatic method, device and system

Abstract

The application relates to the technical field of language translation, in particular to a multilingual machine translation automatic method, device and system based on an LLM, which specifically comprises the following: for the translation between small languages, when selecting a pivot language, the data volume of parallel corpus jointly involved by the small language and each mainstream language is used to evaluate the initial reference of the mainstream language to the small language, and the geographical proximity between the two countries commonly used by the two languages is referred to; the semantic vector similarity of the vocabulary of the mainstream language and the small language in the parallel corpus is analyzed, and the vocabulary quantity and position difference are analyzed; the initial reference is corrected, the final reference of the mainstream language to the small language is calculated, the final pivot language is determined, and the translation between small languages is performed; the reliability of the parallel corpus data reference evaluation of each mainstream language is improved, a more suitable mainstream language can be selected as the pivot language of the two small languages, and the translation accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of language translation technology, specifically to an automatic method, apparatus, and system for multilingual machine translation based on LLM. Background Technology

[0002] With the rapid development of Large Language Modeling (LLM) and the improved processing capabilities of computers for large-scale data, LLM-based machine translation, compared to traditional machine translation that relies on dictionary matching, achieves faster and more accurate multilingual translations due to its massive training data and ability to read and understand source language contexts. For less commonly taught languages ​​(such as Spanish and Serbian), the training data volume is far smaller than that of more widely used mainstream languages ​​(such as Simplified Chinese and English), resulting in less parallel corpora between these languages. However, the parallel corpora between less commonly taught languages ​​and mainstream languages ​​are much richer. Therefore, current automated machine translation between less commonly taught languages ​​often requires selecting a "hub language" as a bridging path between the two languages ​​to obtain smoother and more stable translation results.

[0003] However, in the translation process between different less commonly spoken languages, the choice of different "hub languages" has a significant impact on translation efficiency and accuracy. Current techniques often directly select the language with the largest amount of parallel corpus data between the source and target languages ​​as the hub language for translation. This approach does not take into account the significant differences in word order rules between languages. The hub language selected solely based on the amount of parallel corpus data between languages ​​may have low similarity to the source and target languages, leading to awkward translations and difficulty in accurately expressing the semantics and emotional intensity of the source language. Summary of the Invention

[0004] To address the aforementioned technical problems, the purpose of this application is to provide an automatic method, apparatus, and system for multilingual machine translation based on LLM, and the specific technical solution adopted is as follows: In a first aspect, embodiments of this application provide an automatic method for multilingual machine translation based on LLM, the method comprising the following steps: Obtain the amount of data in the parallel corpora for each language to distinguish between minor languages ​​and major languages; Based on the amount of data in the parallel corpus involving both minor languages ​​and major languages, and the distance between the countries where minor languages ​​and major languages ​​are frequently used, the initial reference value of major languages ​​for the translation of minor languages ​​is assessed. For each parallel corpus segment involving both mainstream and minor languages ​​in the parallel corpus, we analyze the lexical correspondence between the mainstream and minor languages, and calculate the corpus consistency between the mainstream and minor languages ​​in the parallel corpus segment by combining the semantic vector similarity of the corresponding words and the positional differences of the corresponding words in the segment. Based on the initial reference and the consistency, the final reference of each mainstream language to each minor language is obtained; based on the final reference of each mainstream language to any two minor languages, the hub language suitability of each mainstream language is calculated to determine the final hub language; and combined with a pre-trained multilingual machine translation model, translation is performed between the two minor languages.

[0005] In one embodiment, the method for classifying minority languages ​​and mainstream languages ​​is as follows: All languages ​​are arranged in descending order of the amount of parallel corpus data involved. The first preset number of languages ​​are selected as the mainstream languages, and the remaining languages ​​are selected as minor languages.

[0006] In one embodiment, the process of obtaining the initial reference for translation is as follows: Obtain the data volume of parallel corpora involving any minor language and any major language in the parallel corpus, denoted as the first data volume; obtain the distance between the countries commonly used by any minor language and any major language in the standard world map; The initial reference value of any mainstream language to any minor language is determined based on the first data volume and the distance; the initial reference value is directly proportional to the first data volume and inversely proportional to the distance.

[0007] In one embodiment, the process of obtaining the corpus consistency is as follows: The total number of words in each language in the parallel corpus fragments is counted; the ratio of the minimum to the maximum total number of words in the two languages ​​in the parallel corpus fragments is calculated and denoted as the vocabulary size consistency. The system matches words from two languages ​​in parallel corpus fragments to obtain each matched word pair; it obtains the semantic vector of each word through a word vector model; and it calculates the similarity between the semantic vectors of the two words in a word pair. The difference between the relative positions of words in the two languages ​​in the corresponding content is obtained and denoted as the first difference; The consistency between the two languages ​​in the parallel corpus segments is calculated based on the vocabulary size consistency, the first difference, and the similarity. The corpus consistency is positively correlated with the vocabulary size consistency and the similarity, and negatively correlated with the first difference.

[0008] In one embodiment, the process of obtaining the relative position is as follows: In the corresponding content, the ratio of the position order of each word to the total number of words in the corresponding content is obtained as the relative position of each word.

[0009] In one embodiment, the process of obtaining the final reference is as follows: The mean value of the consistency between the mainstream languages ​​and the minor languages ​​of all parallel corpus fragments in the parallel corpus is calculated and normalized. The normalized value is used as the weight of the initial reference value of each mainstream language to each minor language, so as to obtain the final reference value of each mainstream language to each minor language.

[0010] In one embodiment, the hub language suitability is the product of the final reference value of each major language to any two minor languages.

[0011] In one embodiment, the final hub language is the mainstream language corresponding to the maximum hub language suitability among all mainstream languages.

[0012] Secondly, embodiments of this application also provide an automatic multilingual machine translation device based on LLM, wherein the device stores a computer program, and when the computer program is executed by a processor, it implements the steps of the method described in the first aspect above.

[0013] Thirdly, embodiments of this application also provide an automatic multilingual machine translation system based on LLM, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect above.

[0014] The embodiments of this application have at least the following beneficial effects: This application categorizes languages ​​into minor and major languages ​​based on the amount of parallel corpora involved in each language. For translations between minor languages, when selecting a "hub language," it assesses the initial reference value of the major language for minor language translation by considering the amount of parallel corpora shared by the minor language and major languages, as well as the geographical proximity between the countries where these languages ​​are frequently used. It analyzes the lexical correspondences between the two languages ​​in parallel corpus segments, and, combined with the semantic vector similarity of corresponding words and the differences in their positions within the segments, calculates the corpus consistency between the major and minor languages ​​in the parallel corpus segments. This corrects the initial reference value for translation, calculating the final reference value of the major language for minor languages. It considers the language similarity and word order consistency between the major language and the source and target languages, improving the reliability of the reference value assessment of parallel corpus data for each major language. This allows for the selection of a more suitable major language as the hub language between two minor languages, thus improving translation accuracy. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the steps of an LLM-based automatic multilingual machine translation method provided in one embodiment of this application; Figure 2 This is a schematic diagram illustrating the process of obtaining corpus consistency. Detailed Implementation

[0016] The following description, in conjunction with the accompanying drawings, details the specific solutions for the LLM-based automatic multilingual machine translation method, apparatus, and system provided in this application.

[0017] Please see Figure 1 The diagram illustrates a flowchart of an LLM-based automatic multilingual machine translation method according to an embodiment of this application. The method includes the following steps: Step S1: Obtain the amount of data in the parallel corpus for each language in the parallel corpus, so as to distinguish between minor languages ​​and major languages.

[0018] First, a base model is selected and trained using a parallel corpus database to obtain an initial LLM-based automatic multilingual machine translation model. In this embodiment, the Qwen large model is used as the base model, and the United Nations parallel corpus is used for model training. Implementers may also use other models and other parallel corpora for model training; this application does not impose specific restrictions.

[0019] The parallel corpus used contains a total of [number] language types. Each language has a set of parallel corpora, and the data volume of each language is recorded as follows: , … All languages ​​are classified according to the amount of parallel corpus data they involve. Arranged from highest to lowest quality, languages ​​with the latter half of their parallel corpus data are considered low-resource languages, i.e., minor languages; languages ​​with the former half are considered major languages, which can serve as candidate "hub languages" in the translation process between minor and major languages. It should be noted that this example uses an equal division for the threshold between minor and major languages; however, this threshold can be user-defined, and this application does not impose specific restrictions.

[0020] For the source and target languages ​​required for translation, if neither is a minor language, a direct translation is performed based on the trained model to obtain the translation result; otherwise, if both the source and target languages ​​are minor languages, they are denoted as minor languages. With less commonly spoken languages In order to further select mainstream languages ​​that are suitable as the hub languages ​​for translation between the two.

[0021] Step S2, evaluating the initial translation reference of each mainstream language for each minority language based on the data volume of parallel corpora jointly involved by each minority language and each mainstream language in the parallel corpus, and the distance between the commonly used countries of the minority language and the mainstream language.

[0022] During the development of languages, there is also mutual integration and influence between languages, such as the introduction of foreign words, for example, after "coffee" was introduced into China, it was translated as "kāfēi (coffee)". Since languages influence each other in the process of communication and development, there is a certain degree of similarity between different languages, which can be used as an important indicator for selecting pivot languages. The more similar two languages are, the higher the referenceability of parallel corpora between them. In order to determine the pivot language for translation between minority language a and minority language b, any one mainstream language is taken as , taking mainstream language c and minority language a as an example.

[0023] The data volume of parallel corpora between languages directly reflects its translation referenceability. Two languages with a relatively larger data volume of parallel corpora have more referenceable data for translation, making it more likely to obtain highly referential data for translation work. The data volume of parallel corpora jointly involved by minority language a and mainstream language c in the parallel corpus is obtained and recorded as the first data volume , further the maximum value of the parallel corpus data volume between minority language a and all other mainstream languages is recorded as . Wherein, before calculating , mainstream languages that have no parallel corpus intersection with minority language a are excluded; if the number of remaining mainstream languages after exclusion is 0, a preset general language (such as English) is designated as the default pivot language.

[0024] Different languages are used by different countries and regions around the world. In the process of development, regions that are closer to each other may have a relatively higher degree of cultural integration, and the mutual influence between languages in these regions is also likely to be higher. Languages used in geographically closer regions are more likely to have relatively higher similarity, so geographical proximity is a potential prior feature of language contact. In the absence of complete language family evolution tree data, the geographical distance between countries is used as a reference feature. For the distance between countries, the embodiment of the present application uses the distance between capitals on a standard world map, such as the International One-in-a-Million World Map, as the distance between countries, and records the mainstream language and the minority language the normalized value of the distance between the commonly used countries as , which is specifically the ratio of said distance to the length of said world map. If a certain language has multiple commonly used countries, the country with the largest speaking population is taken as the representative.

[0025] evaluating said mainstream language according to the data obtained above For less commonly spoken languages The initial reference value for translation. Preferably, in this embodiment, the expression for the initial reference value for translation is: In the formula, As the mainstream language For less commonly spoken languages Initial reference for translation; Mainstream languages ​​in the parallel corpus database With less commonly spoken languages The amount of parallel corpus data between them; For less commonly spoken languages The maximum amount of parallel corpus data between other mainstream languages; As the mainstream language With less commonly spoken languages Normalized values ​​of distances between commonly used countries; The value is a preset minimum positive number, used to avoid a denominator of 0. In this embodiment, it is... The value is set to 0.01. In other embodiments of this application, the implementer may set the value according to the actual situation. The value of .

[0026] and The selected minor languages With mainstream languages The amount of parallel corpus data between them and the minority languages The ratio of the largest parallel corpus data among all major languages ​​to the total number of parallel corpora represents the number of major languages. With less commonly spoken languages The larger the ratio of the relative amount of comparable parallel corpora, the stronger the proportion of the dominant language. In less commonly spoken languages The more readily available reference data there is during the translation process, the greater its potential for reference. For the selected minor languages With mainstream languages The distance between the countries where the two languages ​​are most commonly used is considered. A smaller value indicates a closer physical distance between the countries where the two languages ​​are most commonly used, suggesting a greater likelihood of mutual influence and interaction. This also suggests higher similarity between the two languages ​​and greater value for their parallel corpus data. (Mainstream languages) For less commonly spoken languages The greater the reference value of the translation, the better. (Language-specific translation reference value) The amount of parallel corpus data available for reference The revisions yielded an initial reference for translation between the two languages.

[0027] In summary, the mainstream languages For less commonly spoken languages is the initial translation reference. The larger this value is, the higher the initial translation reference of the minor language to be translated obtained by this major language pair is, and the more suitable this major language is as a pivot language.

[0028] Step S3: for each parallel corpus segment commonly involved between a major language and a minor language in the parallel corpus, analyze the lexical correspondence between the major language and the minor language, and calculate the corpus consistency between the major language and the minor language in the parallel corpus segment by combining the semantic vector similarity of corresponding words and the position difference of corresponding words in the segment.

[0029] Different languages have different expression methods. As synonymous sentences between different languages, parallel corpora have different word orders in different languages, such as "subject-verb-object", "subject-object-verb", etc. For example, the Chinese translation of the original Japanese sentence "私 (I, subject) は りんご (apple, object) を 食べます (eat, verb)" is "I (subject) eat (verb) apple (object)". Or there are differences in part of speech used. For example, the Chinese translation of the original English sentence "He is a strong advocate for environmental protection." is "He firmly advocates protecting the environment." In this sentence, "protection (noun)" is synonymous with "protect (verb)" but has different parts of speech. Large differences in word order and parts of speech between parallel corpora can easily lead to poor readability of translation results. Therefore, it is necessary to evaluate the consistency of parallel corpus data to select more suitable major languages. Taking the major language in the parallel corpus and the minor language any parallel corpus segment jointly involved as an example for subsequent analysis.

[0030] If the difference in the number of words used by different languages in a parallel corpus segment is larger while their semantics are consistent, the number of relatively inconsistent words may be larger. Take any parallel corpus segment as the current parallel corpus segment, count the number of words used by each language in the parallel corpus segment, wherein the major language the number of words used is recorded as , the minor language the number of words used is recorded as ; obtain and the larger number of words in is recorded as , the smaller vocabulary size is recorded as .

[0031] Parallel corpora are semantically similar and contain a large number of word pairs with the same semantics. However, words with the same semantics do not necessarily have consistent part-of-speech and positions in sentences. Each word is first vectorized according to a word vector model to obtain semantic vectors of each word. In this embodiment, the mBERT model is used to obtain semantic vectors of each word. There are many existing word vector models, and implementers can also use other word vector models to obtain semantic vectors of each word, which is not specifically limited in the present application.

[0032] For majority languages in the current parallel corpus fragment and each word of minority languages , calculate the semantic similarity between each word of the majority language and each word of the minority language , record the cosine similarity between semantic vectors as semantic similarity . Wherein, the calculation of cosine similarity is well-known in the art, and the specific process will not be repeated herein.

[0033] It should be noted that for the calculation of similarity between semantic vectors, the present application only provides one similarity algorithm. There are many existing similarity algorithms, and implementers can also use other similarity algorithms to calculate the similarity between semantic vectors, which is not specifically limited in the present application.

[0034] Matching words of the majority language and words of the minority language in the current parallel corpus fragment according to the semantic similarity, the matching is performed in a manner that word pairs with maximum semantic similarity are matched preferentially and each word can only be matched once. The specific process is as follows: Taking any combination of one word from the majority language in the current parallel corpus fragment and one word from the minority language as a word group, obtaining the word group with the maximum semantic similarity, and taking the two words in this word group as mutually matched words; eliminating word groups containing any matched word, obtaining the word group with the maximum semantic similarity among the remaining word groups, and taking the two words in this word group as mutually matched words; and so on, until all words in one language in the current parallel corpus fragment are completely matched, so as to obtain a total of pairs of matched word pairs.

[0035] Meanwhile, there exists "lexicalization difference" between different languages. A single word in one language may require multiple words in another language to jointly express its semantics. For example, the Chinese word "喜新厌旧" is translated into English as "abandon the old for the new", this type of word has high translation difficulty, which easily causes the phenomenon of large vocabulary difference between parallel corpora of the two languages mentioned above.

[0036] Furthermore, the position and part of speech of synonyms may not be consistent across different corpora. The more inconsistent the relative positions of synonyms, the more likely there are differences in word order or part of speech usage between their respective languages. In the current parallel corpus fragments, the position order of each word in the corpus content corresponding to the mainstream language c is obtained, and the position order of each word is compared with the total number of words in the corpus content corresponding to the mainstream language c. The ratio represents the relative position of each word in the corpus corresponding to the mainstream language c. That is, the position of the t-th word in this content is t, and its relative position is... .

[0037] Similarly, obtain the position order of each word in the corpus content corresponding to language a in the current parallel corpus segment, and compare it with the total number of words in the corpus content corresponding to language a. The ratio represents the relative position of each word in the corresponding corpus of the minor language 'a'.

[0038] Based on the data obtained above, the mainstream languages ​​in the selected parallel corpus fragments were evaluated. With less commonly spoken languages Consistency between the corpora. Preferably, the expression for consistency is: In the formula, Y represents the dominant language in the current parallel corpus fragments. With less commonly spoken languages Consistency of corpora between them; , These represent the minimum and maximum number of words in these two languages, respectively, within the current parallel corpus segment. The semantic similarity is the i-th matching word pair between the two languages ​​in the current parallel corpus segment. The i-th matching word in the current parallel corpus segment belongs to the mainstream language. The relative positions of words; The i-th matching word in the current parallel corpus segment belongs to a less commonly spoken language. The relative positions of words; The value is a preset minimum positive number, used to avoid a denominator of 0. In this embodiment, it is... The value is set to 0.001. In other embodiments of this application, the implementer may set the value according to the actual situation. The value of . Where, For vocabulary size consistency, This is the first difference.

[0039] h and H are the smaller and larger number of words in each language in the selected parallel corpus fragments. The ratio of the two values ​​ranges from (0,1). The larger the ratio, the closer the number of words in the two languages ​​in the parallel corpus fragments are, the higher the consistency of word count, and the more likely there is a large consistency. For the two languages ​​in this parallel corpus segment For semantic similarity of synonyms, the larger the value, the higher the semantic similarity between the two words, and the more accurate the match of the pair of synonyms. and The first The smaller the difference between the relative positions of synonyms in their respective sentences, the closer the relative positions of the two words are, and the more likely their word order and parts of speech are to be consistent.

[0040] The greater the consistency, the higher the consistency between the corresponding content of the two languages ​​in the selected parallel corpus fragments, and the more suitable it is for providing corpus data reference for translation. The corpus consistency of all parallel corpus fragments involving the two languages ​​is calculated using the above method, and the average value is denoted as . .

[0041] Step S4: Based on the initial reference and the consistency, obtain the final reference of each mainstream language to each minor language; based on the final reference of each mainstream language to any two minor languages, calculate the hub language suitability of each mainstream language to determine the final hub language; and combine the pre-trained multilingual machine translation model to perform translation between the two minor languages.

[0042] The mean Normalization is performed to the range [0.1-1.1] to avoid the normalized value potentially being 0 and affecting subsequent calculations. In this embodiment, the normalization expression can be: ,in, For the normalization function, this embodiment uses a maximum-minimum value normalization function. This value is used as the initial reference value for translation. Make corrections to obtain the final reference. The expression is: In the formula, For the mainstream language C, this serves as the ultimate reference for the less commonly spoken language A. This serves as an initial reference for translating the less commonly spoken language A from the mainstream language C. To represent the fluency and readability of the translation result obtained by using parallel corpora from two languages ​​as translation references, this value is used to correct the initial translation reference, resulting in the corrected final reference. .

[0043] Similarly, calculate the mainstream languages With less commonly spoken languages The final reference between (i.e., target languages) is denoted as To obtain the mainstream language As a minor language With less commonly spoken languages The suitability of the hub language between .

[0044] Based on the above method, the hub language suitability of all mainstream languages ​​is calculated, and the mainstream language with the highest hub language suitability is selected as the final hub language.

[0045] Based on the above steps, the automatic multilingual machine translation model selects a pivot language for the translation process according to the user's chosen source and target languages. It first translates the user-input source text into the pivot language text, and then further translates the pivot language text into the target language text before outputting it to the user. This achieves fast and accurate translation between multiple languages.

[0046] A schematic diagram of the process of obtaining corpus consistency is shown below. Figure 2 As shown.

[0047] Based on the same inventive concept as the above methods, embodiments of this application also provide an automatic multilingual machine translation device based on LLM, wherein the device stores a computer program, and when the computer program is executed by a processor, it implements the steps of any one of the above-described automatic multilingual machine translation methods based on LLM.

[0048] Based on the same inventive concept as the above methods, this application also provides an LLM-based automatic multilingual machine translation system, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described LLM-based automatic multilingual machine translation methods.

[0049] In summary, this application provides an automatic multilingual machine translation method based on LLM. It divides the language into minor and major languages ​​based on the amount of parallel corpora involved in each language. For translations between minor languages, when selecting a "hub language," it assesses the initial reference value of the major language for the minor language translation by considering the amount of parallel corpora involved in both the minor and major languages, and taking into account the geographical proximity between the countries where these languages ​​are frequently used. It analyzes the lexical correspondence between the two languages ​​in parallel corpus segments, and calculates the corpus consistency between the major and minor languages ​​in the parallel corpus segments by combining the semantic vector similarity of corresponding words and the positional differences of corresponding words in the segments. This corrects the initial reference value and calculates the final reference value of the major language for the minor language. It considers the language similarity and word order consistency between the major language and the source and target languages, improving the reliability of the reference value assessment of parallel corpus data for each major language. This allows for the selection of a more suitable major language as the hub language between two minor languages, thus improving translation accuracy.

Claims

1. An automatic multilingual machine translation method based on LLM, characterized in that, The method includes the following steps: Obtain the amount of data in the parallel corpora for each language to distinguish between minor languages ​​and major languages; Based on the amount of data in the parallel corpus involving both minor languages ​​and major languages, and the distance between the countries where minor languages ​​and major languages ​​are frequently used, the initial reference value of major languages ​​for the translation of minor languages ​​is assessed. For each parallel corpus segment involving both mainstream and minor languages ​​in the parallel corpus, we analyze the lexical correspondence between the mainstream and minor languages, and calculate the corpus consistency between the mainstream and minor languages ​​in the parallel corpus segment by combining the semantic vector similarity of the corresponding words and the positional differences of the corresponding words in the segment. Based on the initial reference and the consistency, the final reference of each mainstream language to each minor language is obtained; based on the final reference of each mainstream language to any two minor languages, the hub language suitability of each mainstream language is calculated to determine the final hub language; and combined with a pre-trained multilingual machine translation model, translation is performed between the two minor languages. The process for obtaining the initial reference for the translation is as follows: The amount of data in the parallel corpus involving any minor language and any major language is obtained and denoted as the first data volume. The initial reference value of the translation of any major language to any minor language is determined based on the first data volume and the distance; the initial reference value of the translation is directly proportional to the first data volume and inversely proportional to the distance. The process of obtaining the consistency of the corpus is as follows: The total number of words in each language in the parallel corpus fragments is counted; the ratio of the minimum to the maximum total number of words in the two languages ​​in the parallel corpus fragments is calculated and denoted as the vocabulary size consistency. The words of the two languages ​​in the parallel corpus fragments are matched to obtain each pair of matched words; the similarity between the semantic vectors of the two words in the word pair is calculated. The difference between the relative positions of words in the two languages ​​in the corresponding content is obtained and denoted as the first difference; The consistency between the two languages ​​in the parallel corpus segments is calculated based on the vocabulary size consistency, the first difference, and the similarity. The corpus consistency is positively correlated with the vocabulary size consistency and the similarity, and negatively correlated with the first difference. The process for obtaining the final reference is as follows: The mean value of the consistency between the mainstream languages ​​and the minor languages ​​of all parallel corpus fragments in the parallel corpus is calculated and normalized. The normalized value is used as the weight of the initial reference value of each mainstream language to each minor language, and the final reference value of each mainstream language to each minor language is obtained. The hub language suitability is the product of the final reference value of each mainstream language to any two minor languages.

2. The automatic multilingual machine translation method based on LLM as described in claim 1, characterized in that, The method for classifying minority languages ​​and mainstream languages ​​is as follows: All languages ​​are arranged in descending order of the amount of parallel corpus data involved. The first preset number of languages ​​are selected as the mainstream languages, and the remaining languages ​​are selected as minor languages.

3. The automatic multilingual machine translation method based on LLM as described in claim 1, characterized in that, The process of obtaining the relative position is as follows: In the corresponding content, the ratio of the position order of each word to the total number of words in the corresponding content is obtained as the relative position of each word.

4. The automatic multilingual machine translation method based on LLM as described in claim 1, characterized in that, The final hub language is the mainstream language corresponding to the maximum hub language suitability among all mainstream languages.

5. An automatic multilingual machine translation device based on LLM, wherein the device stores a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the LLM-based automatic multilingual machine translation method as described in any one of claims 1-4.

6. An LLM-based multilingual machine translation automatic system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the LLM-based automatic multilingual machine translation method as described in any one of claims 1-4.

Images