A Chinese-Tibetan translation device and Tibetan translation method based on pre-training model

Through the Chinese-Tibetan translation device based on the pre-trained model, using syllable recognition and paragraph division techniques, the problem of intention capture and nuance in Tibetan translation is solved, and more natural and accurate translation is achieved, which improves the ability of Tibetan translation Reliability and quality.

CN118898259BActive Publication Date: 2024-11-26BEIJING WANGZHI TIANYUAN BIG DATA TECH CO LTD +1
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411394196.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-10-08
Publication Date
2024-11-26
Estimated Expiration
2044-10-08

AI Technical Summary

Technical Problem

Existing machine translation systems are difficult to accurately capture the original intention and nuances when processing Tibetan language, resulting in the translation results being not natural and accurate enough, especially when facing complex sentence patterns, polysynonyms and homophones, it is difficult to improve the credibility of Tibetan translation. .

Method used

Using a Chinese-Tibetan translation device based on a pre-trained model, the semantic loss, part-of-speech distinction and semantic particle size of Tibetan syllables is determined through syllable recognition and paragraph division, and combining multi-scale combination and particle size entropy, gradient translation is performed to improve translation quality.

Benefits of technology

Through gradient hierarchical translation technology, the accuracy and fluency of Tibetan translation are improved, mistranslation and semantic ambiguity are reduced, the naturalness and overall quality of translation results are improved, and the processing ability of complex sentence patterns and long-distance dependence is enhanced. Ensure the consistency and credibility of translation results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118898259B_ABST
    Figure CN118898259B_ABST
Patent Text Reader

Abstract

The present application provides a Chinese-Tibetan translation device and a Tibetan translation method based on a pre-training model, which obtains a Tibetan syllable sequence by identifying the Tibetan character structure, and then determines the semantic granularity between Tibetan syllables by the part-of-speech distinction of the Tibetan syllables and the gradient distance between each Tibetan syllable, and then determines the word contribution of each Tibetan syllable in the Tibetan translation process by the semantic granularity and the loss of syllable recognition; multi-scale combination of Tibetan syllables in the Tibetan syllable sequence to obtain syllable paragraphs, and then determines the granularity entropy of Tibetan syllables in each syllable paragraph, and determines the paragraph priority of each syllable paragraph in the Tibetan translation process by the granular entropy and the semantic loss of the syllable paragraph; and then performs gradient translation on the Tibetan text based on all word contributions and all paragraph priorities. The scheme of the present application can realize the gradient hierarchical translation between Tibetan syllables and syllable paragraphs in the Tibetan text, thereby improving the credibility of the Tibetan translation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of machine translation, and more specifically, to a Chinese-Tibetan translation device and a Tibetan translation method based on a pre-trained model. Background Art

[0002] As an important branch of natural language processing, machine translation (MT) has made remarkable progress in recent years. In particular, the introduction of deep learning and neural network technology has made great progress in accuracy and fluency. Early machine translation systems were mainly based on rules and statistical models, relying on a large amount of bilingual data and pre-set language rules for translation. Such methods work well when processing common sentences, but are often unable to cope with complex sentence patterns and language variants. The emergence of neural machine translation (NMT) technology marks a new era for machine translation. NMT systems use neural networks, especially recursive neural networks (RNNs), convolutional neural networks (CNNs), and more advanced transformer models. Through large-scale parallel computing and deep learning algorithms, they can better capture contextual information and long-distance dependencies, thereby generating more natural and fluent translations.

[0003] There are many problems in the Tibetan translation process, mainly including complex grammatical structure, rich polysemy and homonymy, large differences in word order, and different cultural backgrounds and expression habits. These problems make it difficult for existing machine translation systems to accurately capture the intention and nuances of the original text when processing Tibetan, and the translation results are often not natural and accurate enough. Gradient hierarchical translation of Tibetan texts can gradually parse and process syllables and syllable paragraphs in Tibetan texts. This method can capture the language characteristics of Tibetan more carefully, analyze semantics and structure layer by layer, thereby improving the accuracy and fluency of translation. Gradient hierarchical translation of Tibetan texts can not only solve the problems of complex sentence patterns and long-distance dependencies, but also better handle polysemy and homonyms, thereby reducing mistranslation and semantic ambiguity. Therefore, how to achieve gradient hierarchical translation between Tibetan syllables and syllable paragraphs in Tibetan texts, thereby improving the credibility of Tibetan translation, has become a difficult problem faced by the industry. Summary of the invention

[0004] The present application provides a Chinese-Tibetan translation device and a Tibetan translation method based on a pre-trained model, which can realize gradient hierarchical translation between Tibetan syllables and syllable paragraphs in Tibetan text, thereby improving the credibility of Tibetan translation.

[0005] In a first aspect, the present application provides a Tibetan translation method based on a pre-trained model, comprising the following steps:

[0006] Get the Tibetan text to be translated;

[0007] Performing syllable recognition on the Tibetan character structure in the Tibetan text to obtain a Tibetan syllable sequence, and determining the semantic loss of each Tibetan syllable in the Tibetan syllable sequence based on a pre-trained semantic model;

[0008] Determine the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence, determine the semantic granularity between each Tibetan syllable and all other Tibetan syllables according to all the part-of-speech distinctions and the gradient distances between each Tibetan syllable in the Tibetan syllable sequence, and then determine the word contribution of each Tibetan syllable in the Tibetan translation process through the semantic granularity and semantic loss of each Tibetan syllable;

[0009] Based on the pre-trained paragraph model, all Tibetan syllables in the Tibetan syllable sequence are combined at multiple scales to obtain multiple syllable paragraphs, and then the granular entropy of all Tibetan syllables in each syllable paragraph is determined, and the paragraph priority of each syllable paragraph in the Tibetan translation process is determined by all the granular entropies and the semantic loss of each syllable paragraph;

[0010] The Tibetan text is gradient translated based on all word contributions and all paragraph priorities.

[0011] Preferably, performing syllable recognition on the Tibetan character structure in the Tibetan text to obtain a Tibetan syllable sequence specifically includes:

[0012] Recognizing the Tibetan character structure in the Tibetan text based on a pre-trained syllable recognition model to obtain a plurality of Tibetan syllables;

[0013] The sequence composed of all Tibetan syllables is regarded as a Tibetan syllable sequence.

[0014] Preferably, determining the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence specifically includes:

[0015] Marking the part of speech of each Tibetan syllable in the Tibetan syllable sequence to obtain the syllable part of speech of the corresponding Tibetan syllable, and then determining the part of speech frequency of each syllable part of speech;

[0016] Determine the part-of-speech entropy of each Tibetan syllable based on a preset Tibetan corpus;

[0017] The part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence is determined by the part-of-speech entropy of each Tibetan syllable and the part-of-speech frequency of each syllable part-of-speech.

[0018] Preferably, determining the semantic granularity between each Tibetan syllable and all other Tibetan syllables according to all the part-of-speech distinctions and the gradient distances between the Tibetan syllables in the Tibetan syllable sequence specifically includes:

[0019] Selecting a Tibetan syllable from the Tibetan syllable sequence as a selected Tibetan syllable;

[0020] Determine the gradient distance between the selected Tibetan syllable and all other Tibetan syllables according to the Euclidean distance between each Tibetan syllable in the Tibetan syllable sequence;

[0021] Determining the semantic dependency relationship between the selected Tibetan syllable and all other Tibetan syllables;

[0022] Determining the semantic relevance between the selected Tibetan syllable and all other Tibetan syllables through the gradient distance and the semantic dependency;

[0023] Determining the semantic granularity between the selected Tibetan syllable and all other Tibetan syllables according to the part-of-speech distinction of the selected Tibetan syllable and the semantic association;

[0024] Continue to determine the semantic granularity between the remaining Tibetan syllables and all other Tibetan syllables.

[0025] Preferably, determining the word contribution of each Tibetan syllable in the Tibetan translation process by the semantic granularity and semantic loss of each Tibetan syllable specifically includes:

[0026] Select a Tibetan syllable as the selected Tibetan syllable;

[0027] Obtaining the word frequency of the selected Tibetan syllable;

[0028] Determining an attention factor of a corresponding Tibetan syllable according to the semantic loss of the selected Tibetan syllable and the word frequency;

[0029] The attention factor is compensated for by the semantic granularity of the selected Tibetan syllable to obtain the word contribution of the selected Tibetan syllable in the Tibetan translation process;

[0030] Continue to determine the word contribution of the remaining Tibetan syllables in the Tibetan translation process.

[0031] Preferably, performing multi-scale combination on all Tibetan syllables in the Tibetan syllable sequence based on the pre-trained paragraph model to obtain multiple syllable paragraphs specifically includes:

[0032] Performing multi-scale pooling on the Tibetan syllable sequence based on a pre-trained paragraph model to generate a multi-scale pooling window;

[0033] All Tibetan syllables in the Tibetan syllable sequence are combined through the multi-scale pooling window, and each combined paragraph is used as a syllable paragraph.

[0034] Preferably, the Tibetan syllables refer to Tibetan words in a Tibetan text.

[0035] In a second aspect, the present application provides a Chinese-Tibetan translation device based on a pre-trained model, the Chinese-Tibetan translation device based on the pre-trained model includes a Tibetan translation processing unit, and the Tibetan translation processing unit includes:

[0036] An acquisition module, used to acquire Tibetan text to be translated;

[0037] A processing module, configured to perform syllable recognition on the Tibetan character structure in the Tibetan text to obtain a Tibetan syllable sequence, and determine the semantic loss of each Tibetan syllable in the Tibetan syllable sequence based on a pre-trained semantic model;

[0038] The processing module is further used to determine the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence, determine the semantic granularity between each Tibetan syllable and all other Tibetan syllables according to all the part-of-speech distinctions and the gradient distances between each Tibetan syllable in the Tibetan syllable sequence, and then determine the word contribution of each Tibetan syllable in the Tibetan translation process through the semantic granularity and semantic loss of each Tibetan syllable;

[0039] The processing module is further used to perform multi-scale combination of all Tibetan syllables in the Tibetan syllable sequence based on the pre-trained paragraph model to obtain multiple syllable paragraphs, and then determine the granularity entropy of all Tibetan syllables in each syllable paragraph, and determine the paragraph priority of each syllable paragraph in the Tibetan translation process through all the granularity entropies and the semantic loss of each syllable paragraph;

[0040] An execution module is used to perform gradient translation on the Tibetan text based on all word contributions and all paragraph priorities.

[0041] In a third aspect, the present application provides a computer device, comprising a memory and a processor, wherein the memory stores codes, and the processor is configured to obtain the codes and execute the above-mentioned Tibetan translation method based on the pre-trained model.

[0042] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned Tibetan translation method based on the pre-trained model is implemented.

[0043] The technical solution provided by the embodiments disclosed in this application has the following beneficial effects:

[0044] In an embodiment of the present application, a Tibetan text to be translated is obtained; syllable recognition is performed on the Tibetan character structure in the Tibetan text to obtain a Tibetan syllable sequence, and the semantic loss of each Tibetan syllable in the Tibetan syllable sequence is determined based on a pre-trained semantic model; the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence is determined, and the semantic granularity between each Tibetan syllable and all other Tibetan syllables is determined according to all the part-of-speech distinctions and the gradient distance between each Tibetan syllable in the Tibetan syllable sequence, and then the word contribution of each Tibetan syllable in the Tibetan translation process is determined through the semantic granularity and semantic loss of each Tibetan syllable; all Tibetan syllables in the Tibetan syllable sequence are multi-scale combined based on a pre-trained paragraph model to obtain multiple syllable paragraphs, and then the granularity entropy of all Tibetan syllables in each syllable paragraph is determined, and the paragraph priority of each syllable paragraph in the Tibetan translation process is determined through all the granularity entropies and the semantic loss of each syllable paragraph; the Tibetan text is gradient translated based on all word contributions and all paragraph priorities.

[0045] It can be seen that the present application uses gradient hierarchical translation technology to perform syllable recognition and paragraph division on Tibetan text, which can parse semantic information layer by layer, greatly improving the accuracy and fluency of translation; first, by introducing part-of-speech distinction, it can better handle polysemous words and homophones in Tibetan, thereby reducing mistranslation and semantic ambiguity; secondly, the semantic granularity between Tibetan syllables is determined according to the part-of-speech distinction and the gradient distance between Tibetan syllables. The semantic granularity can more accurately identify and process complex semantic structures, thereby improving the accuracy of the translation results; further, the word contribution of Tibetan syllables in the Tibetan translation process is determined by semantic granularity and semantic loss. The word contribution can better capture the semantic relationship within and between sentences, which facilitates the generation of more coherent and Natural translation, thereby improving the overall quality of translation; then, determining the paragraph priority of syllable paragraphs in the Tibetan translation process based on granular entropy and semantic loss of syllable paragraphs can further optimize the syllable paragraph structure and semantic coherence in the translation process, thereby enabling the translation system to more effectively capture the intent and nuances of the original text. This refined processing method can better cope with complex sentence patterns and long-distance dependencies in Tibetan, thereby ensuring the coherence and naturalness of the translation results; finally, gradient translation of Tibetan texts based on determined word contributions and paragraph priorities can comprehensively improve the credibility and quality of Tibetan translations; in summary, the present application scheme can realize gradient hierarchical translation between Tibetan syllables and syllable paragraphs in Tibetan texts, thereby improving the credibility of Tibetan translations. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is an exemplary flow chart of a Tibetan translation method based on a pre-trained model according to some embodiments of the present application;

[0047] Figure 2 is a schematic diagram of a process for determining part-of-speech distinction according to some embodiments of the present application;

[0048] Figure 3 is a schematic diagram of a process for determining paragraph priority according to some embodiments of the present application;

[0049] Figure 4 is a schematic diagram of an exemplary software module of a Tibetan translation processing unit according to some embodiments of the present application;

[0050] Figure 5 It is a structural schematic diagram of a computer device for implementing a Tibetan translation method based on a pre-trained model according to some embodiments of the present application. DETAILED DESCRIPTION

[0051] The core of the present application is that the present application provides a Chinese-Tibetan translation device and a Tibetan translation method based on a pre-training model. The Tibetan syllable sequence is obtained by identifying the Tibetan character structure, and then the semantic granularity between the Tibetan syllables is determined by the part-of-speech distinction of the Tibetan syllables and the gradient distance between each Tibetan syllable, and then the word contribution of the Tibetan syllables in the Tibetan translation process is determined by the semantic granularity and the loss of syllable recognition; the Tibetan syllables in the Tibetan syllable sequence are multi-scale combined to obtain syllable paragraphs, and then the granularity entropy of the Tibetan syllables in each syllable paragraph is determined, and the paragraph priority of each syllable paragraph in the Tibetan translation process is determined by the granular entropy and the semantic loss of the syllable paragraph; and the Tibetan text is gradient translated based on all word contributions and all paragraph priorities; the scheme can realize the gradient hierarchical translation between Tibetan syllables and syllable paragraphs in the Tibetan text, thereby improving the credibility of the Tibetan translation.

[0052] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods. Figure 1 , which is an exemplary flow chart of a Tibetan translation method based on a pre-training model according to some embodiments of the present application. The Tibetan translation method 100 based on a pre-training model mainly includes the following steps:

[0053] In step 101, a Tibetan text to be translated is obtained.

[0054] In a specific implementation, the Tibetan text refers to Tibetan text data to be translated. The Tibetan text data may be, for example, Tibetan text data obtained by recognizing an image containing Tibetan through a Chinese-Tibetan translation device, which is not specifically limited here.

[0055] In step 102, syllable recognition is performed on the Tibetan character structure in the Tibetan text to obtain a Tibetan syllable sequence, and the semantic loss of each Tibetan syllable in the Tibetan syllable sequence is determined based on a pre-trained semantic model.

[0056] In some embodiments, syllable recognition of the Tibetan character structure in the Tibetan text to obtain a Tibetan syllable sequence may be achieved by the following steps:

[0057] Recognizing the Tibetan character structure in the Tibetan text based on a pre-trained syllable recognition model to obtain a plurality of Tibetan syllables;

[0058] The sequence composed of all Tibetan syllables is regarded as a Tibetan syllable sequence.

[0059] It should be noted that the syllable recognition model in the present application refers to a pre-trained model for identifying Tibetan syllable words in Tibetan text. The syllable recognition model can be obtained by performing Tibetan character recognition training on a large amount of Tibetan text data. The algorithm structure of the syllable recognition model described in the present application adopts a convolutional neural network. In other embodiments, other existing algorithms can also be used for implementation, which is not limited here.

[0060] In specific implementation, the Tibetan character structure in the Tibetan text is recognized based on a pre-trained syllable recognition model to obtain a plurality of Tibetan syllables, which can be achieved in the following manner: the Tibetan text is used as the input of the syllable recognition model, and then a plurality of Tibetan syllables are obtained through the syllable recognition model, wherein the Tibetan syllables refer to Tibetan words in the Tibetan text; a sequence composed of all Tibetan syllables is used as a Tibetan syllable sequence, which can be achieved in the following manner: all Tibetan syllables can be arranged in the order of recognition of the syllable recognition model, and the sequence obtained by the arrangement is used as a Tibetan syllable sequence.

[0061] In some embodiments, determining the semantic loss of each Tibetan syllable in the Tibetan syllable sequence based on the pre-trained semantic model can be achieved by using the following steps:

[0062] Initialize the pre-trained semantic model;

[0063] Using the Tibetan syllable sequence as an input parameter of the pre-trained semantic model;

[0064] The semantic loss of each Tibetan syllable in the Tibetan syllable sequence is evaluated by using the initialized semantic model to obtain the semantic loss of each Tibetan syllable in the Tibetan syllable sequence.

[0065] It should be noted that the semantic model in this application refers to a machine learning model used to understand the semantic information of words and sentences in natural language. Its main function is to capture and represent the semantic relationship in the language, help computers understand the meaning and context in human language, and the semantic loss of Tibetan syllables can be evaluated through the semantic relationship. In the evaluation process, the cosine similarity in the prior art can be used to evaluate the semantic loss of Tibetan syllables. In other embodiments, the Euclidean distance in the prior art can also be used to evaluate the semantic loss of Tibetan syllables, which is not limited here.

[0066] In the specific implementation, firstly, a pre-trained semantic model suitable for processing Tibetan and with low computational complexity is selected, such as DistilBERT, TinyBERT, etc. These models have low computational overhead while maintaining high performance. Secondly, the Tibetan syllable sequence is converted into the input form of the pre-trained semantic model (such as a word embedding vector), and then the semantic representation corresponding to each Tibetan syllable is generated by the pre-trained semantic model (such as a hidden layer vector), and then the cosine similarity is used to measure the semantic loss of each Tibetan syllable. As a preferred embodiment, the cosine similarity is used to measure the semantic loss of each Tibetan syllable. It can be implemented in the following way: the semantic representation of each Tibetan syllable is compared with the target semantic representation, and then the cosine similarity between the semantic representation and the target semantic representation is calculated, and the calculated cosine similarity is used as the semantic loss of the corresponding Tibetan syllable in the Tibetan syllable sequence.

[0067] It should be noted that the semantic loss in this application is an indicator to measure the degree of semantic loss when the semantic model generates Tibetan syllables. It aims to quantify the semantic deviation. The semantic loss is usually used to train the natural language processing (NLP) model so that the text it generates or understands is as close to the semantics of the target text as possible.

[0068] In step 103, the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence is determined, and the semantic granularity between each Tibetan syllable and all other Tibetan syllables is determined based on all the part-of-speech distinctions and the gradient distances between each Tibetan syllable in the Tibetan syllable sequence, and then the word contribution of each Tibetan syllable in the Tibetan translation process is determined through the semantic granularity and semantic loss of each Tibetan syllable.

[0069] In some embodiments, reference Figure 2 As shown, this figure is a schematic diagram of the process of determining the part-of-speech distinction in some embodiments of the present application. In this embodiment, the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence can be determined by the following steps:

[0070] In step 1031, the part of speech of each Tibetan syllable in the Tibetan syllable sequence is marked to obtain the syllable part of speech of the corresponding Tibetan syllable, and then the part of speech frequency of each syllable part of speech is determined;

[0071] In step 1032, the part-of-speech entropy of each Tibetan syllable is determined based on a preset Tibetan corpus;

[0072] In step 1033, the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence is determined by the part-of-speech entropy of each Tibetan syllable and the part-of-speech frequency of each syllable part-of-speech.

[0073] In specific implementation, the part of speech of each Tibetan syllable in the Tibetan syllable sequence is marked to obtain the syllable part of speech of the corresponding Tibetan syllable, and then the part of speech frequency of each syllable part of speech is determined. The following method can be used to achieve this, that is, the existing Tibetan part of speech tagging tool can be used to mark the part of speech of each Tibetan syllable in the Tibetan syllable sequence to obtain the syllable part of speech of the corresponding Tibetan syllable, and then the part of speech frequency of each syllable part of speech is counted; the part of speech entropy of each Tibetan syllable based on the preset Tibetan corpus can be determined in the following method, that is, a Tibetan syllable is selected as the selected Tibetan syllable, all syllable parts of speech of the selected Tibetan syllable are obtained from the preset Tibetan corpus, and then all syllable parts of speech are described by entropy features through information entropy in the prior art, and the obtained information entropy is used as the part of speech entropy of the selected Tibetan syllable, and the part of speech entropy of the remaining Tibetan syllables is further determined; it should be noted that the preset Tibetan corpus is a database obtained by preprocessing a large amount of Tibetan text data, which contains a large amount of Tibetan syllable information and syllable part of speech information.

[0074] It should be noted that the part-of-speech entropy in the present application is a measure of the diversity of the parts of speech of Tibetan syllables in different contexts. The part-of-speech entropy can be used to quickly analyze the diversity of the parts of speech of Tibetan syllables, thereby improving the accuracy of the semantic description. The larger the part-of-speech entropy, the greater the diversity of the parts of speech of Tibetan syllables in different contexts, and the smaller the part-of-speech entropy, the smaller the diversity of the parts of speech of Tibetan syllables in different contexts.

[0075] In specific implementation, determining the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence by the part-of-speech entropy of each Tibetan syllable and the part-of-speech frequency of each syllable part-of-speech can be achieved in the following manner, namely: selecting a Tibetan syllable as the selected Tibetan syllable, obtaining the part-of-speech frequency of the selected Tibetan syllable part-of-speech, wherein each Tibetan syllable corresponds to a part-of-speech (when there are multiple parts-of-speech in a Tibetan syllable, the part-of-speech with the largest part-of-speech frequency in the Tibetan text is used as the part-of-speech of the Tibetan syllable), taking the product of the part-of-speech frequency and the part-of-speech entropy of the selected Tibetan syllable as the part-of-speech distinction of the selected Tibetan syllable, and continuing to determine the part-of-speech distinction of the remaining Tibetan syllables. It should be noted that in the present application, the part-of-speech entropy of the Tibetan syllable is adjusted by the part-of-speech frequency, which is conducive to comprehensively considering the distribution of Tibetan syllables in different contexts, thereby calculating the degree of distinction in different contexts, and helping to improve the accuracy of text analysis.

[0076] It should be noted that the part-of-speech discrimination in the present application is an indicator for measuring the ability of different Tibetan syllables in distinguishing parts of speech. The larger the part-of-speech discrimination, the stronger the ability of different Tibetan syllables in distinguishing parts of speech. The smaller the part-of-speech discrimination, the weaker the ability of different Tibetan syllables in distinguishing parts of speech. By analyzing the part-of-speech discrimination, the accuracy of the language model in processing Tibetan text can be improved, thereby improving applications such as speech recognition and machine translation.

[0077] In some embodiments, determining the semantic granularity between each Tibetan syllable and all other Tibetan syllables according to all part-of-speech distinctions and the gradient distances between the Tibetan syllables in the Tibetan syllable sequence can be achieved by the following steps:

[0078] Selecting a Tibetan syllable from the Tibetan syllable sequence as a selected Tibetan syllable;

[0079] Determine the gradient distance between the selected Tibetan syllable and all other Tibetan syllables according to the Euclidean distance between each Tibetan syllable in the Tibetan syllable sequence;

[0080] Determining the semantic dependency relationship between the selected Tibetan syllable and all other Tibetan syllables;

[0081] Determining the semantic relevance between the selected Tibetan syllable and all other Tibetan syllables through the gradient distance and the semantic dependency;

[0082] Determining the semantic granularity between the selected Tibetan syllable and all other Tibetan syllables according to the part-of-speech distinction of the selected Tibetan syllable and the semantic association;

[0083] Continue to determine the semantic granularity between the remaining Tibetan syllables and all other Tibetan syllables.

[0084] It should be noted that the gradient distance in the present application is an indicator for measuring the semantic gradient differences between Tibetan syllables in Tibetan texts. In specific implementation, the gradient distance between the selected Tibetan syllable and all other Tibetan syllables is determined based on the Euclidean distance between each Tibetan syllable in the Tibetan syllable sequence. This can be achieved in the following manner, namely: obtaining the Euclidean distance between the selected Tibetan syllable and each other Tibetan syllable, and then taking the average of all Euclidean distances as the gradient distance between the selected Tibetan syllable and all other Tibetan syllables. The gradient distance can be used to measure the average semantic similarity between Tibetan syllables.

[0085] In specific implementation, determining the semantic dependency relationship between the selected Tibetan syllable and all other Tibetan syllables can be achieved in the following manner, namely: a pre-trained language model (such as BERT, GPT, etc.) can be used to analyze the context dependency relationship between the selected Tibetan syllable and all other Tibetan syllables, and the context dependency relationship is used as the semantic dependency relationship between the selected Tibetan syllable and all other Tibetan syllables; determining the semantic association between the selected Tibetan syllable and all other Tibetan syllables through the gradient distance and the semantic dependency relationship can be achieved in the following manner, namely: the ratio of the semantic dependency relationship and the gradient distance is used as the semantic association between the selected Tibetan syllable and all other Tibetan syllables, wherein the gradient distance is a non-zero value.

[0086] It should be noted that the semantic relevance in this application is an indicator for measuring the semantic relevance between Tibetan syllables. By combining semantic dependency and gradient distance, the semantic relationship between Tibetan syllables can be more comprehensively reflected. In the translation process, semantic relevance can help identify and maintain the interrelated semantic relationships in the source language, thereby improving the accuracy and coherence of the translation.

[0087] In specific implementation, determining the semantic granularity between the selected Tibetan syllable and all other Tibetan syllables based on the part-of-speech distinction and the semantic association of the selected Tibetan syllable can be achieved in the following manner, namely: the natural exponent of the opposite number of the part-of-speech distinction of the selected Tibetan syllable can be used as an adjustment coefficient, and then the product of the adjustment coefficient and the semantic association is used as the semantic granularity between the selected Tibetan syllable and all other Tibetan syllables. It should be noted that in the present application, the semantic association is adjusted by the part-of-speech distinction, which helps to reduce the semantic fine-grained error introduced by the recognition error of the Tibetan syllable, thereby improving the translation accuracy of the Tibetan language.

[0088] It should be noted that the semantic granularity in this application is an indicator of the degree of refinement of Tibetan syllables in distinguishing and understanding the semantics of other Tibetan syllables. The semantic granularity can more accurately analyze and understand the semantic features and relationships of language units (such as words, phrases, syllables, etc.). In natural language processing tasks, semantic granularity can help the model better understand and utilize contextual information, thereby improving the performance of the model.

[0089] In some embodiments, determining the word contribution of each Tibetan syllable in the Tibetan translation process by the semantic granularity and semantic loss of each Tibetan syllable can be achieved by the following steps:

[0090] Select a Tibetan syllable as the selected Tibetan syllable;

[0091] Obtaining the word frequency of the selected Tibetan syllable;

[0092] Determining an attention factor of a corresponding Tibetan syllable according to the semantic loss of the selected Tibetan syllable and the word frequency;

[0093] The attention factor is compensated for by the semantic granularity of the selected Tibetan syllable to obtain the word contribution of the selected Tibetan syllable in the Tibetan translation process;

[0094] Continue to determine the word contribution of the remaining Tibetan syllables in the Tibetan translation process.

[0095] In specific implementation, obtaining the word frequency of the selected Tibetan syllable can be achieved in the following manner, namely: the word frequency statistics tool in the prior art can be used to count the word frequency of each Tibetan syllable in the Tibetan text, and then the word frequency of the selected Tibetan syllable is obtained; determining the attention factor of the corresponding Tibetan syllable according to the semantic loss of the selected Tibetan syllable and the word frequency can be achieved in the following manner, namely: the ratio of the word frequency to the semantic loss can be used as the attention factor of the selected Tibetan syllable, and the attention factor represents a weight parameter for measuring the importance of the Tibetan syllable in the translation process. Using the word frequency as a parameter for determining the attention factor can consider the proportional linear influence of the word frequency on the importance of the Tibetan syllable in the translation process. Using the semantic loss as another parameter for determining the attention factor can simultaneously consider the inverse linear influence of the semantic loss on the importance of the Tibetan syllable in the translation process, thereby improving the availability of the finally determined attention factor.

[0096] In addition, the contribution compensation of the attention factor is performed through the semantic granularity of the selected Tibetan syllable, and the word contribution of the selected Tibetan syllable in the Tibetan translation process can be achieved in the following way, namely: the natural exponent of the opposite number of the semantic granularity can be used as the contribution compensation coefficient of the selected Tibetan syllable in the translation process, and the contribution compensation coefficient is a compensation parameter used to adjust the word attention mechanism, which reflects the contribution degree of the semantic granularity of the Tibetan syllable to the target language word in the translation process, and then the product of the contribution compensation coefficient and the attention factor is used as the word contribution. In this application, the contribution compensation mechanism can adjust the attention factor so that the model can better pay attention to important Tibetan syllables during the translation process. This method makes the translation system more stable and reliable when processing sentences of different complexities, which helps to promote cultural and information exchanges between Chinese and Tibetan.

[0097] It should be noted that the word contribution in this application is an indicator to measure the semantic information contributed by Tibetan syllables in the translation process. The greater the word contribution, the more semantic information the Tibetan syllables contribute in the translation process. The smaller the word contribution, the less semantic information the Tibetan syllables contribute in the translation process. By evaluating the word contribution, the translation system can identify which Tibetan syllables contribute the most to the overall semantics, thereby giving priority to ensuring the correct translation of these Tibetan syllables and improving the overall translation quality. At the same time, targeted training can be carried out according to the word contribution, focusing on optimizing the translation ability of those high-contribution Tibetan syllables, thereby improving the robustness of the model.

[0098] In step 104, all Tibetan syllables in the Tibetan syllable sequence are multi-scaled combined based on the pre-trained paragraph model to obtain multiple syllable paragraphs, and then the granular entropy of all Tibetan syllables in each syllable paragraph is determined, and the paragraph priority of each syllable paragraph in the Tibetan translation process is determined by all the granular entropies and the semantic loss of each syllable paragraph.

[0099] In some embodiments, multi-scale combination of all Tibetan syllables in the Tibetan syllable sequence based on the pre-trained paragraph model to obtain multiple syllable paragraphs can be achieved by the following steps:

[0100] Performing multi-scale pooling on the Tibetan syllable sequence based on a pre-trained paragraph model to generate a multi-scale pooling window;

[0101] All Tibetan syllables in the Tibetan syllable sequence are combined through the multi-scale pooling window, and each combined paragraph is used as a syllable paragraph.

[0102] It should be noted that the paragraph model in the present application is a machine learning model for dividing Tibetan syllable paragraphs. The paragraph model can be pre-trained with a large amount of Tibetan text data. It can capture the structure and semantic information of the text. Its internal architecture can adopt the Transforme architecture in the prior art. In other embodiments, other existing machine learning architectures can also be used for implementation, which is not limited here.

[0103] In specific implementation, multi-scale pooling is performed on the Tibetan syllable sequence based on a pre-trained paragraph model, and then a multi-scale pooling window is generated. The following method can be used to achieve this: the Tibetan syllable sequence is processed using a pre-trained paragraph model, and then multi-scale pooling is performed through the paragraph model, and then the window obtained by pooling is used as a multi-scale pooling window. The multi-scale pooling refers to pooling syllables at different scales (for example, pooling syllables at different time windows or different text lengths). Multi-scale pooling can help the model capture the characteristics of the syllable sequence at different levels of detail. The size and position of the multi-scale pooling window are based on the syllable characteristics of the syllable sequence. Dynamically adjusted; all Tibetan syllables in the Tibetan syllable sequence are combined by the multi-scale pooling window, and each combined paragraph is used as a syllable paragraph. This can be achieved in the following way, namely: the multi-scale pooling window is dynamically generated according to the syllable characteristics of the syllable sequence, which enables the window to better adapt to different syllable combinations. The size and position of these windows are variable and can be automatically adjusted according to the input Tibetan syllable sequence. The multi-scale pooling window is used to perform multiple combinations of Tibetan syllables in the Tibetan syllable sequence (for example: monosyllable paragraphs, disyllabic paragraphs and trisyllable paragraphs), and these combined Tibetan syllable sequences are used as syllable paragraphs for subsequent processing or analysis.

[0104] In some embodiments, determining the granularity entropy of all Tibetan syllables in each syllable segment may be achieved by using the following steps:

[0105] Select a syllable paragraph as the selected syllable paragraph;

[0106] Obtaining the word frequency of each Tibetan syllable in the selected syllable paragraph;

[0107] determining the semantic complexity of the selected syllable segment;

[0108] Determine the granularity entropy of all Tibetan syllables in the selected syllable paragraph by the word frequency of each Tibetan syllable and the semantic complexity;

[0109] Continue to determine the granular entropy of all Tibetan syllables in the remaining syllable paragraphs.

[0110] In specific implementation, the word frequency of each Tibetan syllable in the selected syllable paragraph can be obtained by the following method, that is, the word frequency of each Tibetan syllable in the selected syllable paragraph can be counted by the existing word frequency statistics tool spaCy; the semantic complexity of the selected syllable paragraph can be determined by the following method, that is, initializing a semantic model, taking the selected syllable paragraph as the input parameter of the semantic model, and then outputting the semantic complexity of the selected syllable paragraph through the semantic model. In this application, the semantic model can adopt Latent Dirichlet Allocation (Latent Dirichlet Allocation). Dirichlet Allocation (LDA) model, the semantic complexity refers to the complexity of the selected syllable paragraph text at the semantic level, reflecting the difficulty of understanding, interpreting and analyzing the content of the selected syllable paragraph text; determining the granular entropy of all Tibetan syllables in the selected syllable paragraph by the word frequency of each Tibetan syllable and the semantic complexity can be achieved in the following way, namely: the word frequency of each Tibetan syllable is described by entropy characteristics through the information entropy in the prior art, so as to obtain the word frequency information entropy of the Tibetan syllable, and then the word frequency information entropy is adjusted through the semantic complexity, and the adjusted word frequency information entropy is used as the granular entropy of all Tibetan syllables in the selected syllable paragraph; it should be noted that in the present application, the word frequency information entropy is adjusted through the semantic complexity, so as to better capture the usage characteristics of syllables in different syntactic structures, so that the granular entropy can more accurately reflect the complexity of the language structure, and the granular entropy obtained in this way not only takes into account the distribution of word frequency, but also captures the information at the semantic level, so that the description is more comprehensive.

[0111] It should be noted that the granular entropy in this application is an indicator to measure the uncertainty of the semantic information content of Tibetan syllables in a syllable paragraph. The granular entropy can be used to evaluate the semantic complexity of the syllable paragraph, thereby helping to understand the distribution and detail of the information in the paragraph. The granular entropy can also be used to optimize the parameters of the translation model, so that the model performs better when processing highly complex semantics, thereby improving the accuracy and coherence of the overall translation.

[0112] In some embodiments, reference Figure 3 As shown, the figure is a schematic diagram of the process of determining paragraph priority in some embodiments of the present application. In this embodiment, the paragraph priority of each syllable paragraph in the Tibetan translation process is determined by all granular entropies and semantic losses of each syllable paragraph, which can be achieved by the following steps:

[0113] In step 1041, for each syllable paragraph, the semantic loss of the syllable paragraph and the granularity entropy of all Tibetan syllables in the syllable paragraph are obtained;

[0114] In step 1042, the granularity feature value of the syllable paragraph is determined by the semantic loss of the syllable paragraph and the granularity entropy of all Tibetan syllables in the syllable paragraph, and then the granularity feature values ​​of all syllable paragraphs are obtained;

[0115] In step 1043, the paragraph priority of each syllable paragraph in the Tibetan translation process is determined according to the proportion of each granular feature value.

[0116] In the specific implementation, first, a syllable paragraph is selected from each syllable paragraph as the selected syllable paragraph, and then the semantic loss of the selected syllable paragraph and the granularity entropy of all Tibetan syllables in the selected syllable paragraph are obtained, wherein the semantic loss of the selected syllable paragraph refers to the sum of the semantic losses of the Tibetan syllables in the selected syllable paragraph; secondly, the product of the semantic loss and the granularity entropy can be used as the granularity feature value of the selected syllable paragraph, and the granularity feature value is an indicator for evaluating the degree of refinement of the syllable paragraph in the Tibetan translation process, and the granularity feature values ​​of the remaining syllable paragraphs are further determined; then, the proportion of each granularity feature value is used as the paragraph priority of the corresponding syllable paragraph, wherein the proportion refers to the ratio of the granularity feature value to the sum of all granularity feature values.

[0117] It should be noted that the paragraph priority in this application refers to the priority level of the syllable paragraph in the Tibetan translation process. The larger the paragraph priority, the higher the priority of the syllable paragraph in the Tibetan translation process. In this application, by giving priority to paragraphs with high semantic complexity and large amount of information, it can ensure that the translation of these paragraphs is more accurate, and improve the quality and readability of the overall translation. At the same time, when training the translation model, the paragraph priority can be used to focus on training those high-priority paragraphs, thereby improving the performance of the model when processing complex semantics.

[0118] In step 105, gradient translation is performed on the Tibetan text based on all word contributions and all paragraph priorities.

[0119] In some embodiments, gradient translation of the Tibetan text based on all word contributions and all paragraph priorities may be implemented by the following steps:

[0120] Sorting the translation of each paragraph in the Tibetan text according to the priority of each paragraph to obtain a paragraph translation sequence;

[0121] According to the contribution of each word, the Tibetan text in each paragraph in the paragraph translation sequence is graded and translated in turn, thereby obtaining a translation paragraph of each paragraph in the Tibetan text;

[0122] The translated text of the Tibetan text is determined according to all the translated paragraphs, and the gradient translation of the Tibetan text is completed.

[0123] In the specific implementation, first, the paragraphs in the Tibetan text are sorted in descending order according to the priority of the corresponding paragraphs as a paragraph translation sequence; secondly, for each paragraph in the paragraph translation sequence, a translation model is initialized, and the proportion of the corresponding word contribution of each Tibetan syllable in the paragraph is used as the translation weight of the corresponding Tibetan syllable in the translation model, wherein the proportion of the word contribution refers to the ratio of the word contribution to the sum of all word contributions, and the translation model is used to complete the hierarchical translation of the paragraph, and the translation result of the translation model is used as the translation paragraph of the paragraph. The translation paragraph of each paragraph in the Tibetan text can be obtained in the above manner; finally, all the translated paragraphs can be arranged according to the position of the corresponding paragraph in the Tibetan text as the translation text of the Tibetan text, and the gradient translation of the Tibetan text is completed; it should be noted that the gradient translation in the present application refers to a method of translating the Tibetan text in stages and levels, which is translated in sequence according to the importance and complexity of the words and paragraphs in the text, and the translation quality and efficiency can be gradually optimized by giving priority to translating paragraphs and words with high importance.

[0124] On the other hand, in some embodiments, the present application provides a Chinese-Tibetan translation device based on a pre-trained model. The Chinese-Tibetan translation device based on a pre-trained model in the present application refers to a device for translating between Tibetan and Chinese. A pre-trained translation model is built into the Chinese-Tibetan translation device. In addition, the Chinese-Tibetan translation device based on the pre-trained model in the present application includes a Tibetan translation processing unit. Figure 4 , which is a schematic diagram of exemplary software modules of a Tibetan translation processing unit according to some embodiments of the present application, the Tibetan translation processing unit 400 includes: an acquisition module 401, a processing module 402 and an execution module 403, which are described as follows:

[0125] Acquisition module 401, in this application, acquisition module 401 is mainly used to acquire Tibetan text to be translated;

[0126] Processing module 402, in the present application, processing module 402 is used to perform syllable recognition on the Tibetan character structure in the Tibetan text, thereby obtaining a Tibetan syllable sequence, and determining the semantic loss of each Tibetan syllable in the Tibetan syllable sequence based on a pre-trained semantic model;

[0127] The processing module 402 in the present application is also used to determine the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence, determine the semantic granularity between each Tibetan syllable and all other Tibetan syllables according to all the part-of-speech distinctions and the gradient distances between each Tibetan syllable in the Tibetan syllable sequence, and then determine the word contribution of each Tibetan syllable in the Tibetan translation process through the semantic granularity and semantic loss of each Tibetan syllable;

[0128] The processing module 402 in the present application is also used to perform multi-scale combination of all Tibetan syllables in the Tibetan syllable sequence based on the pre-trained paragraph model to obtain multiple syllable paragraphs, and then determine the granularity entropy of all Tibetan syllables in each syllable paragraph, and determine the paragraph priority of each syllable paragraph in the Tibetan translation process through all the granularity entropies and the semantic loss of each syllable paragraph;

[0129] Execution module 403, in this application, execution module 403 is mainly used to perform gradient translation on the Tibetan text based on all word contributions and all paragraph priorities.

[0130] In addition, the present application also provides a computer device, which includes a memory and a processor, the memory stores code, and the processor is configured to obtain the code and execute the above-mentioned Tibetan translation method based on the pre-trained model.

[0131] In some embodiments, reference Figure 5 , which is a schematic diagram of the structure of a computer device for implementing a Tibetan translation method based on a pre-trained model according to some embodiments of the present application. The Tibetan translation method based on a pre-trained model in the above embodiment can be Figure 5 The computer device 500 shown in the figure is implemented, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503 and at least one communication interface 504.

[0132] The processor 501 may be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0133] The communication bus 502 may be used to transmit information between the above-mentioned components.

[0134] The memory 503 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, an optical disc storage (including a compressed optical disc, a laser disc, an optical disc, a digital versatile disc, a Blu-ray disc, etc.), a magnetic disk or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of an instruction or data structure and can be accessed by a computer, but is not limited thereto. The memory 503 may exist independently and be connected to the processor 501 via the communication bus 502. The memory 503 may also be integrated with the processor 501.

[0135] The memory 503 is used to store the program code for executing the solution of the present application, and the execution is controlled by the processor 501. The processor 501 is used to execute the program code stored in the memory 503. The program code may include one or more software modules. The Tibetan translation method based on the pre-training model in the above embodiment can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0136] The communication interface 504 uses any transceiver or other device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0137] In a specific implementation, as an embodiment, a computer device may include multiple processors, each of which may be a single-CPU processor or a multi-CPU processor. The processor here may refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0138] The above-mentioned computer device may be a general-purpose computer device or a special-purpose computer device. In a specific implementation, the computer device may be a desktop computer, a portable computer, a network server, a PDA (personal digital assistant), a mobile phone, a tablet computer, a wireless terminal device, a communication device or an embedded device. The embodiment of the present application does not limit the type of computer device.

[0139] In addition, the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned Tibetan translation method based on the pre-trained model.

[0140] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0141] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A Tibetan translation method based on a pre-trained model, characterized in that: The steps include: Get the Tibetan text to be translated; Performing syllable recognition on the Tibetan character structure in the Tibetan text to obtain a Tibetan syllable sequence, and determining the semantic loss of each Tibetan syllable in the Tibetan syllable sequence based on a pre-trained semantic model; Determine the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence, determine the semantic granularity between each Tibetan syllable and all other Tibetan syllables according to all the part-of-speech distinctions and the gradient distances between each Tibetan syllable in the Tibetan syllable sequence, and then determine the word contribution of each Tibetan syllable in the Tibetan translation process through the semantic granularity and semantic loss of each Tibetan syllable; Based on the pre-trained paragraph model, all Tibetan syllables in the Tibetan syllable sequence are combined at multiple scales to obtain multiple syllable paragraphs, and then the granular entropy of all Tibetan syllables in each syllable paragraph is determined, and the paragraph priority of each syllable paragraph in the Tibetan translation process is determined by all the granular entropies and the semantic loss of each syllable paragraph; The Tibetan text is gradient translated based on all word contributions and all paragraph priorities.

2. The method according to claim 1, characterized in that Performing syllable recognition on the Tibetan character structure in the Tibetan text to obtain a Tibetan syllable sequence specifically includes: Recognizing the Tibetan character structure in the Tibetan text based on a pre-trained syllable recognition model to obtain a plurality of Tibetan syllables; The sequence composed of all Tibetan syllables is regarded as a Tibetan syllable sequence.

3. The method according to claim 1, characterized in that Determining the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence specifically includes: Marking the part of speech of each Tibetan syllable in the Tibetan syllable sequence to obtain the syllable part of speech of the corresponding Tibetan syllable, and then determining the part of speech frequency of each syllable part of speech; Determine the part-of-speech entropy of each Tibetan syllable based on a preset Tibetan corpus; The part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence is determined by the part-of-speech entropy of each Tibetan syllable and the part-of-speech frequency of each syllable part-of-speech.

4. The method according to claim 1, characterized in that Determining the semantic granularity between each Tibetan syllable and all other Tibetan syllables according to all the part-of-speech distinctions and the gradient distances between the Tibetan syllables in the Tibetan syllable sequence specifically includes: Selecting a Tibetan syllable from the Tibetan syllable sequence as a selected Tibetan syllable; Determine the gradient distance between the selected Tibetan syllable and all other Tibetan syllables according to the Euclidean distance between each Tibetan syllable in the Tibetan syllable sequence; Determining the semantic dependency relationship between the selected Tibetan syllable and all other Tibetan syllables; Determining the semantic relevance between the selected Tibetan syllable and all other Tibetan syllables through the gradient distance and the semantic dependency; Determining the semantic granularity between the selected Tibetan syllable and all other Tibetan syllables according to the part-of-speech distinction of the selected Tibetan syllable and the semantic association; Continue to determine the semantic granularity between the remaining Tibetan syllables and all other Tibetan syllables.

5. The method according to claim 1, characterized in that Determining the word contribution of each Tibetan syllable in the Tibetan translation process by the semantic granularity and semantic loss of each Tibetan syllable specifically includes: Select a Tibetan syllable as the selected Tibetan syllable; Obtaining the word frequency of the selected Tibetan syllable; Determining an attention factor of a corresponding Tibetan syllable according to the semantic loss of the selected Tibetan syllable and the word frequency; The attention factor is compensated for by the semantic granularity of the selected Tibetan syllable to obtain the word contribution of the selected Tibetan syllable in the Tibetan translation process; Continue to determine the word contribution of the remaining Tibetan syllables in the Tibetan translation process.

6. The method according to claim 1, characterized in that Based on the pre-trained paragraph model, all Tibetan syllables in the Tibetan syllable sequence are combined at multiple scales to obtain multiple syllable paragraphs, specifically including: Performing multi-scale pooling on the Tibetan syllable sequence based on a pre-trained paragraph model to generate a multi-scale pooling window; All Tibetan syllables in the Tibetan syllable sequence are combined through the multi-scale pooling window, and each combined paragraph is used as a syllable paragraph.

7. The method according to claim 1, characterized in that The Tibetan syllables refer to Tibetan words in Tibetan texts.

8. A Chinese-Tibetan translation device based on a pre-trained model, the Chinese-Tibetan translation device based on a pre-trained model comprising a Tibetan translation processing unit, characterized in that: The Tibetan translation processing unit comprises: An acquisition module, used to acquire Tibetan text to be translated; A processing module, configured to perform syllable recognition on the Tibetan character structure in the Tibetan text to obtain a Tibetan syllable sequence, and determine the semantic loss of each Tibetan syllable in the Tibetan syllable sequence based on a pre-trained semantic model; The processing module is further used to determine the part-of-speech distinction of each Tibetan syllable in the Tibetan syllable sequence, determine the semantic granularity between each Tibetan syllable and all other Tibetan syllables according to all the part-of-speech distinctions and the gradient distances between each Tibetan syllable in the Tibetan syllable sequence, and then determine the word contribution of each Tibetan syllable in the Tibetan translation process through the semantic granularity and semantic loss of each Tibetan syllable; The processing module is further used to perform multi-scale combination of all Tibetan syllables in the Tibetan syllable sequence based on the pre-trained paragraph model to obtain multiple syllable paragraphs, and then determine the granularity entropy of all Tibetan syllables in each syllable paragraph, and determine the paragraph priority of each syllable paragraph in the Tibetan translation process through all the granularity entropies and the semantic loss of each syllable paragraph; An execution module is used to perform gradient translation on the Tibetan text based on all word contributions and all paragraph priorities.

9. A computer device, comprising a memory and a processor, wherein the memory stores a code, characterized in that: The processor is configured to obtain the code and execute the Tibetan translation method based on the pre-training model as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the Tibetan translation method based on the pre-training model as described in any one of claims 1 to 7 is implemented.

Citation Information

Cited By

  • Tibetan written language character-voice conversion method based on voice structure characteristics

    CN121789641A

  • A method for converting Tibetan written language phonetics based on phonetic structure features

    CN121789641B