Large model generation type multilingual translation teaching system and method, electronic equipment and storage medium

Abstract

The invention discloses a large model generation type multilingual translation teaching system and method, electronic equipment and a storage medium, and relates to the technical field of large model generation type multilingual translation. The system comprises a read-in module and a translation teaching module. The read-in module is used for determining input vocabularies and corresponding languages; the languages comprise Russian, Ukrasan, Thai and the like; the translation teaching module is used for inputting the input vocabularies into a multi-modal non-universal language index model, and performing index reading through the multi-modal non-universal language index model to obtain a corresponding index result in a target translation language; the index result comprises vocabulary translation, sentence translation, paraphrase translation, synonym recommendation and the like; wherein the multi-modal non-universal language index model is constructed based on a Hash graph, an inverted index and a JSP development technology; the inverted index comprises a word pair index, a sentence pair index, a voice index and a synonym index. According to the invention, teaching and communication requirements of users can be met.

Description

Technical Field

[0001] The present invention relates to the technical field of large-model generative multilingual translation, and in particular to a large-model generative multilingual translation teaching system, method, electronic equipment and storage medium. Background Art

[0002] With the development of the times and technology, people's life and social contacts are becoming more and more extensive, and the Internet can connect every corner. On the Internet, relying on English as a common language can only meet the needs of daily communication, which is obviously far from enough. The cultures of all parts of the world are colorful, and if you want to explore and learn in depth, the languages ​​of all places are indispensable media.

[0003] In addition to English and Chinese, various non-common languages ​​are gradually coming into people's sight. For example, the raw materials of factories of multinational companies mostly come from countries with non-common languages. In these countries, common languages ​​such as English and Chinese are not perfectly universal. In the current national trade, there will be obstacles in communication and interaction. At this time, the demand for learning non-common languages ​​is urgently increasing. In today's social life, the need for communication does not only refer to the above situation, but also includes learning needs and social needs, which are becoming more and more prominent. Summary of the invention

[0004] The purpose of the present invention is to provide a large-model generative multilingual translation teaching system, method, electronic device and storage medium that can meet the teaching and communication needs of users.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A large model-generated multilingual translation teaching system, comprising:

[0007] A reading module is used to determine input words and corresponding languages; the languages ​​include Russian, Ukrainian, Thai, etc.;

[0008] The translation teaching module is used to input the input vocabulary into a multimodal non-universal language index model, perform index reading through the multimodal non-universal language index model, and obtain the corresponding index results in the target translation language; the index results include vocabulary translation, sentence translation, interpretation translation and synonym recommendation, etc.; wherein, the multimodal non-universal language index model is constructed based on Hash graph, inverted index and JSP development technology; the inverted index includes word pair index, sentence pair index, phonetic index and synonym index.

[0009] Optionally, the reading module is also connected to the large model generative corpus to provide external corpus resources for reading vocabulary and indexing results.

[0010] Optionally, the translation teaching module specifically includes:

[0011] An indexing submodule, for inputting the input vocabulary into a multimodal non-universal language indexing model, using the multimodal non-universal language indexing model to read the input translation and the corresponding index data in the form of an index file, and directly obtain them from a cache to obtain an index result; the cache stores corpus resources;

[0012] The result display submodule is used to display the vocabulary translation, sentence translation, interpretation translation and synonym recommendation in the display page, and a voice play button is provided for the content with audio under each display content.

[0013] Optionally, the result display submodule is in the form of a Web page.

[0014] The present invention also provides a large model generation multilingual translation teaching method, comprising:

[0015] Determine input vocabulary and corresponding language; the languages ​​include Russian, Ukrainian, Thai, etc.;

[0016] The input vocabulary is input into a multimodal non-universal language index model, and index reading is performed by the multimodal non-universal language index model to obtain a corresponding index result in a target translation language; the index result includes vocabulary translation, sentence translation, interpretation translation and synonym recommendation, etc.; wherein the multimodal non-universal language index model is constructed based on a hash graph, an inverted index and JSP development technology; the inverted index includes a word pair index, a sentence pair index, a phonetic index and a synonym index.

[0017] Optionally, the input vocabulary is read through a large model generative corpus resource; the large model generative corpus resource includes an external corpus resource and corresponding audio.

[0018] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the large model generative multilingual translation teaching method described above.

[0019] The present invention also provides a computer-readable storage medium, which stores a computer program, a generative big model, and language data. When the computer program, the generative big model, and the language data are executed by a processor, the big model generative multilingual translation teaching method as described above is implemented.

[0020] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0021] The present invention discloses a large-model generative multilingual translation teaching system, method, electronic device and storage medium, the system includes a read-in module and a translation teaching module; the read-in module is used to determine input vocabulary and corresponding languages; the languages ​​include Russian, Ukrainian, Thai, etc.; the translation teaching module is used to input the input vocabulary into a multimodal non-universal language index model, and the multimodal non-universal language index model is used to perform index reading to obtain the corresponding index result in the target translation language; the index result includes vocabulary translation, sentence translation, interpretation translation and synonym recommendation; wherein the multimodal non-universal language index model is constructed based on Hash graph, inverted index and JSP development technology; the inverted index includes word pair index, sentence pair index, voice index and synonym index. The present invention can meet the teaching and communication needs of users. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0023] Figure 1 It is a structural schematic diagram of the large model generation multi-language translation teaching system of the present invention;

[0024] Figure 2 This is a schematic diagram of the settings of each layer of the system in this embodiment;

[0025] Figure 3 This is a schematic diagram of the steps of the large-scale model reinforcement learning training in this embodiment;

[0026] Figure 4 Schematic diagram of the index structure and process in this embodiment;

[0027] Figure 5 This is a schematic diagram of the main page layout in this embodiment;

[0028] Figure 6 This is a schematic diagram of the result display layout in this embodiment. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] The purpose of the present invention is to provide a large-model generative multilingual translation teaching system, method, electronic device and storage medium that can meet the teaching and communication needs of users.

[0031] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] like Figure 1 As shown, the present invention provides a large model generation multi-language translation teaching system, comprising:

[0033] A reading module is used to determine input words and corresponding languages; the languages ​​include Russian, Ukrainian, Thai, etc.;

[0034] The translation teaching module is used to input the input vocabulary into a multimodal non-universal language index model, perform index reading through the multimodal non-universal language index model, and obtain the corresponding index results in the target translation language; the index results include vocabulary translation, sentence translation, interpretation translation and synonym recommendation, etc.; wherein, the multimodal non-universal language index model is constructed based on Hash graph, inverted index and JSP development technology; the inverted index includes word pair index, sentence pair index, phonetic index and synonym index.

[0035] like Figure 2 As shown in the figure, the system settings at each level include:

[0036] Electronic device layer: including general computing power (CPU), vector computing power (GPU), high-speed inline network, cloud storage media, etc.

[0037] Language data layer: including vocabulary, sentences, audio and other data in multiple languages ​​such as Russian, Ukrainian, and Thai.

[0038] Large model interface layer: compatible with mainstream large model interfaces, including OpenAI GPT, Google Gemini, MetaLLaMa, Baidu Wenxin, etc.

[0039] Generative technology layer: including text-generated text, text-generated images, content summaries, etc.

[0040] Generative application layer: including translation teaching, translation practice, foreign language research, etc.

[0041] Based on the above system composition, the following embodiments and application processes are provided.

[0042] In this embodiment, a large model is used as the base index model for generating responses to questions or prompts in plain text. These systems use specialized multi-level and multi-faceted neural networks trained on very large natural language text data sets, usually collected from the Internet and other suitable sources. Training a large model is very time-consuming and expensive. Currently, the most common commercially available systems train for several weeks simultaneously on thousands of powerful vector computing power (GPUs) at a cost of millions of dollars. These programs are often referred to as "base models" and have wide applicability and long service life. They can serve as the basis for many different types of domain-specific large models, although it is also entirely possible to interact with them directly.

[0043] After the big model has completed the "basic training" on a large text corpus, it will enter the "cultivation" stage. This includes providing it with a series of examples of how it should respond to "prompts" politely and cooperatively to answer questions, and most importantly, what it is not allowed to say. Figure 3 As shown, in contrast to the initial training step, which is a mostly automated process, this socialization step is done through what is known as reinforcement learning with human feedback. A human reviews the big model's responses to a series of prompts that might provoke inappropriate behavior, and then a person explains to it the problems with the responses or the prohibited content, helping the big model improve.

[0044] After reinforcement learning is complete, the big model takes as input a prompt or question from the user, transforms it, and generates a response. This process is fast and simple compared to the learning step. The big model extends this "guess the next word" technique to longer sequences. The analysis and guessing are not actually done on the words themselves, but on so-called tokens, and these tokens are further expressed in the form of "embeddings" to capture their meaning. The simplified word-level explanation ignores how the big model represents these huge collections of words in a computer. It is impractical to store all possible sequences of thousands of words in any existing computer system. Therefore, the researchers repurposed the tried and true methods of neural networks to reduce these huge collections to a more manageable form. Neural networks were originally applied to solve classification problems. Neural networks must compress data in a way that related inputs produce similar results.

[0045] In this embodiment, the model also uses a Hash graph, which is a specific implementation of the Map interface in the Hash table, and provides all mapping operations. The mapping order of this implementation class is unstable, and it cannot be guaranteed that its order will never change. At the same time, the Hash graph is non-thread-safe, that is, in a multi-threaded environment, operations on the Hash graph will affect the results in other threads. Therefore, the Hash of this embodiment will only run in a single-threaded environment, and locks will be used in special cases to ensure thread safety. In addition, in the application of this system, Null values ​​may appear, and the Hash table does not allow Null values ​​to appear, so this is also an important factor in selecting the Hash graph.

[0046] Hash graphs divide the capacity into many parts, each small capacity is called a bucket, and this implementation assumes that the hash function distributes the elements correctly among the buckets. There are two crucial parameters that affect the performance of hash graph instances - initial capacity and load factor. Capacity is the number of buckets in the hash table. The load factor determines the initially defined capacity of the hash table. When the number of entries is greater than the product of the load factor and the current capacity, the hash graph instance calls the ReHash method to double the capacity. Although a high load factor reduces space overhead, it also increases query costs.

[0047] This embodiment mainly uses the mapping relationship within the HashMap to form an inverted index, which is also the main method used at the current technical level. The capacity of the HashMap can perfectly accommodate corpora of various types of corpora.

[0048] The present embodiment also adopts an index relationship. An index is a storage structure for sorting the values ​​of one or more columns in a relational database table. It is a mapping relationship between the set of values ​​of one or more columns in a table in the database and the corresponding physical address pointing to the value. An index is actually equivalent to the directory of a book. The title in the directory is the attribute of the index (Key in Hash), and the page number corresponding to the title is the value of the index (Value in Hash). People can find the page number of the corresponding content according to the title, that is, find the corresponding value according to the corresponding attribute.

[0049] Inverted index, also known as reverse index, is often used to provide a mapping relationship (Key, Value) between a word or phrase in one or more articles. It is the most commonly used data structure in document retrieval systems. Inverted index mainly consists of two parts: "word dictionary" and "inverted file".

[0050] There are two different inverted index forms:

[0051] A horizontal inverted index of one record, containing a list of documents for each reference to the word.

[0052] A horizontal inverted index of words contains the position of each word in a document. This usage is often used for phrase search. This is also the main innovation of this method using inverted index.

[0053] The first index form is the index form that this system is mainly used to generate, and the second index form is mainly used in the process of generating the corresponding file relationship mapping of the first form.

[0054] This embodiment mainly uses the second index form to generate a mapping of various files related to the words and the files required for the content for translation, forming an index file to facilitate the reading and use of the system.

[0055] In addition, this embodiment also uses JSP, the full name of which is Java ServerPages, which is a dynamic web page development technology that can use tags to insert various codes for use. Unlike JavaScript and PHP, there are obvious boundaries, and the convenience of inserting other languages ​​as internal logic is not high. JSP is an indispensable part of Java EE and is a complete enterprise-level application platform. This means that JSP can implement the most complex applications in the simplest way. In terms of file reading and extraction, this system mainly uses the Java language as the internal logic, and uses the httpservlet package in the javax package to obtain the content of the page. Then, nested JavaScript is used to implement the page layout, so JSP is the best choice.

[0056] Based on the above technologies, the system modules are designed.

[0057] The main service objects of the large model generative multilingual translation teaching system are groups with foreign trade learning, communication learning, foreign language learning, and text translation needs. The main goal of the system platform is to build an online teaching platform with three languages ​​- Russian, Ukrainian, and Thai. Considering the needs of social life and the simplicity of practical use, the system needs to have the functions of vocabulary translation - for daily learning, phrase and sentence translation - for daily teaching and text translation practice needs, mainly suitable for use with relevant business needs, and synonymous word recommendation - mainly for rigorous translation and context use, for users of specific language learning. At the same time, different from the common front-end and back-end interactions and data acquisition methods in the existing market, the innovation of this system lies in reducing the response of the back-end, but reading the specific data in the form of index files and obtaining it directly from the cache, reducing the interaction time between the front-end and back-end in the middle, greatly improving the response time of the system, and increasing the comfort of use.

[0058] In summary, the requirements of the large-scale model-based multilingual translation teaching system are mainly the acquisition, processing and import of corpus data, the construction of the Web platform and the response of the Web to operations. Users select the language on the platform, enter the translation vocabulary, and return the translation results, phrase and sentence translation results, and synonym word recommendation results. The system indexing process is as follows: Figure 4 As shown in the figure. The most critical innovation is to call the large model (LLM) interface before reading the results to obtain the background knowledge content summary and data expansion. In the case of improving the response speed and the small corpus size, using HashMap to build an inverted index can meet the needs. According to system requirements, four index modules need to be built:

[0059] Among them, regarding phrase indexing, by reading all the words in the corpus, a HashMap of the phrase is generated and written into a txt text in a fixed format. The text content can index the corresponding index file for each word in the index file. When needed, the result can be directly read to obtain the corresponding file.

[0060] Regarding the sentence pair index, the words are matched with the obtained sentence corpus file. It is worth mentioning that the sentence needs to be split by word, and then the words are matched. Only when the words exist in the sentence in the form of words can it be a correct result. Otherwise, if the words exist in the sentence in the form of strings, it is an incorrect result. The id corresponding to the content of the file corresponding to the correct result and the word match are written in a fixed format to generate an index file for sentence pair matching, which can be directly read to obtain the corresponding file and the corresponding line of the id in the file.

[0061] Regarding voice indexing, the vocabulary is matched with the file name in the voice library (mp3 file), the corresponding mp3 file path is obtained and written into the txt text in a fixed format, the corresponding voice file is generated, the voice file is directly obtained and played, if there is no corresponding voice file, Null is written, and when the Web platform reads it, it will determine whether there is a voice file, and then confirm whether the play voice button is displayed.

[0062] Regarding synonym index, in this embodiment, the main purpose is to obtain words similar to the word, form a synonym mapping between words, and correspond the synonyms to generate a synonym file. To obtain synonyms, the Levenshtein distance is mainly used for judgment. The Levenshtein distance refers to the number of actions required to convert a string into another string. The actions include deleting a character, moving a character, and inserting a character. Each deletion, insertion, or move is counted as an action. My judgment is that the number of actions is 1, which is a synonym (the same applies to words with similar shapes).

[0063] This embodiment mainly uses HashMap and inverted index, and uses JSP as the main language for building the Web platform, wherein Java is used as the internal operating logic, and JavaScript is nested to write the platform page. The number of processed corpus entries collected is within one million, and HashMap can fully meet this demand. The characteristics of the inverted index can generate a mapping of index attributes and attribute-related files, and save them in the form of plain text files. The corpus is split into a single original text and a Chinese corresponding format by line (i.e., by article), saved in a fixed interval manner, and its corresponding index relationship is generated. The index file is read into the web page cache, and the corresponding translation, sentence, audio file save path and other data can be read and displayed without the participation of the database, maintaining the fastest response rate.

[0064] In the actual system construction, the language data relied on include relatively complete and comprehensive sources such as public corpora of various universities, TED, Wikipedia, and data from OpenSubtitles and other web pages.

[0065] For Russian data acquisition, we selected the existing public corpus resources of Wikipedia to obtain phrase corpus and obtain sql files. The content of the sql file is saved in the form of sql statement input of the database. It is necessary to perform simple processing and extraction through code to extract the Russian words and obtain their corresponding Chinese translation through the existing translation platform. Generate the corresponding corpus xlsx file.

[0066] Use Python code to obtain the corresponding audio of the phrase through the Google translation platform and name it after the original phrase. Convert the initial corpus file into a txt plain text file for easy text processing. Delete repeated words, pure numbers, words with illegal symbols, etc. in the initial text to obtain a cleaner corpus file. Generate a ru_word.txt file with a fixed format (id / tab / vocabulary / tab / audio path / tab / Chinese translation) corresponding to the audio file.

[0067] Wikipedia mainly consists of article corpus. For people who are not proficient in the language, splitting the sentences in the article and matching the translated sentences can easily cause huge errors and is too difficult to meet the needs of this software. Therefore, in order to continue to develop this system, movie lines and translations are obtained through movie lines websites such as OpenSubtitles, which are highly applicable. Because these two files are line-to-line correspondence, they are connected line by line to form an initial sentence pair corpus wrong.txt file.

[0068] After deleting duplicate sentences, illegal sentences, pure digital sentences, and semantically repeated sentences, the correct.txt file, a cleaner sentence pair corpus, is obtained. The collected audio files and mp3 folders are all named after the vocabulary. At this point, the files required for Russian are basically completed.

[0069] For Ukrainian data acquisition, the Ukrainian corpus mainly comes from the OpenSubtitles website. When the Ukrainian corpus is acquired, due to its non-universal language characteristics, some words cannot be found in the reading and writing audio files on many platforms. Therefore, there are many words that do not have corresponding mp3 format files, resulting in the number of lines in the corpus file being more than the mp3 file.

[0070] For Thai data acquisition, the Thai language and corpus acquisition method are basically similar to Russian acquisition, but when processing the corpus, the corresponding files are obtained, but Thai is mixed with a lot of English during use. Therefore, there is a large part of English words or phrases in the initial corpus file, and there is less pure Thai text data. After the initial word file is processed multiple times for repeated word entries, repeated translation entries, pure digital entries, pure English entries, year, month, day, etc., it is mapped with the audio file to generate a relatively quick and useful initial corpus th_word.txt file. The initial files required for the Thai system are basically completed.

[0071] After a simple cleaning, a relatively clean and simple initial prediction file for Ukrainian was obtained. At this point, the files required for Ukrainian are basically completed. The three language files are basically completed, and the index files are generated. In the end, the three languages ​​will obtain three files each, mp3 folders, word pair corpus files and sentence pair corpus files, including mp3, xlsx and txt formats. Among them, Russian has 295,727 mp3 files, 324,003 word pair corpus, and 1,046,256 sentence pair corpus. Ukrainian has 200,845 mp3 files, 215,872 word pair corpus, and 442,079 sentence pair corpus. Thai has 15,855 mp3 files, 2,354 word pair corpus, and 1,149,066 sentence pair corpus.

[0072] Then, the index is generated. The three languages ​​run in one system, using the same business logic and file content format. Therefore, the generated content format and files are also the same. The following examples all take Russian as an example. Because the audio tag of JavaScript is embedded in JSP to play audio files, and the src attribute of audio only recognizes digitally named mp3 files, it is necessary to generate an additional index file corresponding to words and numbers, and modify the file path in ru_word_txt to a digitally named path.

[0073] Levenshtein distance is mainly used to determine the number of operations required for conversion between two strings. Use the Levenshtein distance algorithm to obtain similar words and synonyms with a distance less than or equal to 1. When the operand is 1, the original string is deleted, inserted, or moved once, and then it can be converted. If the distance is 0 when judging synonyms, it is a perfect synonym mapping. Generate the corresponding synonym index ru_semantic.txt file and the similar word index ru_levenshtein.txt file.

[0074] Word and sentence pair files, if you use HashMap nested Value as HashSet, write the mapping relationship between Key (word) and Value (all sentences containing the word) into an index file, the content of this file will be too large, which is not conducive to reading and using. I use the progressive relationship to read the sentence. Inject the id number into each line of the sentence pair corpus file, generate a new sentence pair ru_sen.txt file, and generate another index ru_wordsen.txt file that maps the word and the corresponding sentence id.

[0075] Finally, the development and operation environment of this system is as follows: IDEA and Pycharm are mainly used for development, as well as data acquisition and processing, corpus segmentation, index file generation and other operations. IDEA: runs on multiple platforms, mainly for Java language development, and also includes JSP writing environment. Pycharm: runs on multiple platforms, mainly for Python language development. Tomcat: plug-ins can be installed in IDEA for easy use, as a running server agent. The operating system and running machine are Window10 Professional Edition, i7 processor, 16GB running memory.

[0076] Use JSP to build web pages, because JSP can easily obtain the internal logic written in Java files, and can nest languages ​​such as java and js, so JSP is used as the main language for page construction, and Tomcat is used as the server agent. The main page for language selection and the start page defaults to the Russian word translation teaching system. The first time you enter the web page, there will be an identification to determine whether it is the first time you enter the system web page. If you jump from other web pages in the system, it does not fall into the category of first entry. The first time you enter the server, files such as ru_word, ru_levenshtein, ru_semantic, ru_sen, and ru_wordsen will be pre-loaded. The initial page contains a text input box. Enter the corresponding word text. The server will read the files in the word, levenshtein, semantic, sen, wordsen, and mp3 folders to obtain the corresponding content. And correspondingly, the jump interface of other language translation teaching pages, such as Figure 5 shown.

[0077] The returned results are mainly divided into five parts: the translation part of the corresponding word, the voice playback button part if there is a corresponding voice file, the sentence components, the sentence translation part (by clicking the Russian button or the Chinese button, the corresponding content can be displayed or hidden), the synonyms and similar words recommendation part, such as Figure 6 shown.

[0078] The internal logic mainly includes file IO operations, the use of RandomAccessFile, BufferReader, File, FileInputStream, InputStreamReader and Iterator packages under the java.io package to process and read files. The information acquisition on the server side mainly uses the http package under the javax.servlet package, and the two packages HttpServletRequest and HttpServletResponse to obtain and return data from the Web page. The audio file is saved in the mp3 file under the Web directory and read.

[0079] In summary, after investigating the market demand and analyzing the application scenarios of existing non-universal language translation teaching, we believe that the development prospects of non-universal languages ​​are very broad, so we designed and developed this system. The online translation teaching platform is mainly built for three languages: Russian, Ukrainian and Thai. The corpus is obtained and processed through web pages such as TED and OpenSubtitles. The multimodal data includes voice, Chinese translation, sentence composition, synonyms and similar words. For some incomplete data, the existing mature translation platforms such as Google are combined to supplement the data, such as translation, audio files and sentences. In view of the data size of the corpus, a hash graph is used to build an inverted index, generate an index file and store it in the cache of the Web platform, omit the intervention of the database, improve the response rate and reduce the demand of the online platform on the network. The functions and displays of the recommendation algorithms for words, word translation, corresponding sentences, sentence translation, synonyms and similar words are realized.

[0080] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0081] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A large model generative multilingual translation teaching system, characterized in that: include: The reading module is used to determine the input vocabulary and the corresponding language; The languages ​​mentioned include Russian, Ukrainian and Thai; The translation teaching module is used to input the input vocabulary into a multimodal non-universal language index model, perform index reading through the multimodal non-universal language index model, and obtain the corresponding index results in the target translation language; the index results include vocabulary translation, sentence translation, interpretation translation and synonym recommendation; wherein the multimodal non-universal language index model is constructed based on Hash graph, inverted index and JSP development technology; the inverted index includes word pair index, sentence pair index, phonetic index and synonym index.

2. The large model-based multilingual translation teaching system according to claim 1, characterized in that: The reading module is also connected to the corpus generated by the large model to provide expanded corpus resources for reading vocabulary and indexing results.

3. The large model-based multilingual translation teaching system according to claim 1 is characterized in that: The translation teaching module specifically includes: An indexing submodule, for inputting the input vocabulary into a multimodal non-universal language indexing model, using the multimodal non-universal language indexing model to read the input translation and the corresponding index data in the form of an index file, and directly obtain them from a cache to obtain an index result; the cache stores corpus resources; The result display submodule is used to display the vocabulary translation, sentence translation, interpretation translation and synonym recommendation in the display page, and a voice play button is provided for the content with audio under each display content.

4. The large model-generated multilingual translation teaching system according to claim 3 is characterized in that: The result display submodule is in the form of a web page.

5. A large-model generative multilingual translation teaching method, characterized in that: include: Determine the input vocabulary and the corresponding language; The languages ​​mentioned include Russian, Ukrainian and Thai; The input vocabulary is input into a multimodal non-universal language index model, and index reading is performed by the multimodal non-universal language index model to obtain a corresponding index result in a target translation language; the index result includes vocabulary translation, sentence translation, interpretation translation and synonym recommendation; wherein the multimodal non-universal language index model is constructed based on a hash graph, an inverted index and JSP development technology; the inverted index includes a word pair index, a sentence pair index, a phonetic index and a synonym index.

6. The large model-based generative multilingual translation teaching method according to claim 5, characterized in that: The input vocabulary is read through the corpus resources generated by the large model; the corpus resources generated by the large model include expanded external knowledge corpus resources and corresponding audio.

7. An electronic device, characterized in that: The electronic device comprises a memory, a vector processor, a general-purpose processor and a vector processor, wherein the memory is used to store a computer program, the vector processor is specifically used to process a generative large model, and the general-purpose processor runs the computer program to enable the electronic device to execute the large-model generative multilingual translation teaching method according to any one of claims 5-6.

8. A computer-readable storage medium, characterized in that: It stores computer programs, generative big models, and language data. When the computer programs, generative big models, and language data are executed by a processor, the big model generative multilingual translation teaching method as described in any one of claims 5-6 is implemented.

Images