Multi-language intelligent translation system based on artificial intelligence

Abstract

The invention relates to the technical field of language translation, and discloses a multi-language intelligent translation system based on artificial intelligence, and the system comprises an acquisition module which is used for receiving externally inputted voice information or text information; the corpus management module is used for receiving the text information transmitted by the acquisition module; the translation quality improving module is used for providing semantic mapping data based on the corpus management module; the privacy and security module is used for receiving the translation result generated by the translation quality improvement module; and the cloud-end cooperation module is used for coordinating task scheduling among the acquisition module, the corpus management module, the translation quality improvement module and the privacy and security module. Semantic mapping data of high resources and scarce languages is generated through a corpus management module and is called by a translation quality improvement module; the translation quality improving module generates target language output through context modeling and multi-stage translation, efficient cooperation of multi-language and scarce language translation can be achieved, and translation efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of language translation, and in particular to a multilingual intelligent translation system based on artificial intelligence. Background Art

[0002] Multilingual intelligent translation system is a solution based on computer technology, especially artificial intelligence technology, which aims to realize automatic translation between different languages. Such systems can accurately convert speech or text content in one language into the target language through technologies such as speech recognition, natural language processing (NLP), and machine translation (MT). Its main uses include cross-language communication, international business, language learning, cultural communication, and multilingual support for technical documents. These systems are widely used in instant translation tools, language translation equipment, cross-border e-commerce platforms, and international intelligent customer service systems, solving the problem of poor information exchange caused by language barriers and helping to promote global development.

[0003] With the rapid development of artificial intelligence technology, multilingual intelligent translation systems have gradually evolved from traditional rule-based translation to statistical-based machine translation, and then to today's neural machine translation (NMT) technology. In particular, the application of neural networks in the field of multilingual translation has significantly improved the quality of translation in terms of semantic understanding and fluency. However, in practical applications, multilingual intelligent translation systems still face many challenges. The scope of application for scarce languages ​​is limited, so there is insufficient support for scarce languages, resulting in limited translation capabilities for scarce languages ​​and limited accuracy of contextual translation. Summary of the invention

[0004] In view of the deficiencies of the prior art, the present invention provides a multilingual intelligent translation system based on artificial intelligence, which solves the problems in the prior art of limited translation capabilities of scarce languages ​​and limited accuracy of contextual translation.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multilingual intelligent translation system based on artificial intelligence, comprising: Acquisition module: used to receive external input voice information or text information, convert voice signals into text input through speech recognition technology, and pass the processed text information to the corpus management module; Corpus management module: receives text information transmitted by the acquisition module, combines the high-resource language corpus and scarce language corpus stored in the system, generates a unified multi-language shared semantic representation, expands the scarce language corpus, and generates semantic mapping data for use by the translation quality improvement module; Translation quality improvement module: Based on the semantic mapping data provided by the corpus management module, it performs context modeling on the input text and generates natural language output in the target language through a multi-stage translation strategy; Privacy and security module: receives the translation results generated by the translation quality improvement module and protects data privacy during the translation process, including localized translation and end-to-end encryption processing; Cloud-end collaboration module: used to coordinate task scheduling among the acquisition module, corpus management module, translation quality improvement module, and privacy and security module, dynamically allocate translation tasks to local or cloud, and ensure translation efficiency and resource optimization.

[0006] Preferably, the acquisition module includes: A sound signal receiving unit, used for receiving external input voice information; A speech recognition module, used to convert the sound information into text input, wherein the speech recognition module is trained using a deep neural network; The text preprocessing module is used to perform word segmentation, denoising and formatting on the collected text information, and output standardized text data to the corpus management module.

[0007] Preferably, the corpus management module includes: Coding units, which are used to generate unified multilingual semantic representations by semantically encoding high-resource language texts and scarce language texts; The generation unit generates scarce language pseudo corpus through the generator of the generative adversarial network to expand the scarce corpus; The discriminant unit uses the discriminator of the generative adversarial network to discriminate the authenticity of the pseudo-corpus and the real corpus, which is used to optimize the quality of the generated corpus; The screening unit calculates the score of the pseudo-corpus through grammatical and semantic rules, and is used to screen out high-quality corpus and store it in the shared semantic space.

[0008] Preferably, the screening unit calculates the grammatical and semantic scores of the pseudo corpus by the following formula:

[0009] Where: represents the grammatical correctness score of the pseudo-corpus, calculated by the pre-trained grammar model, Represents the semantic similarity between the pseudo-corpus and the high-resource language corpus, calculated from the semantic representation generated by the encoding unit, and is the weight coefficient of grammatical and semantic scores, satisfying + =1.

[0010] Preferably, the translation quality improvement module includes: A context modeling unit and a multi-stage translation unit, wherein the context modeling unit is used to calculate the context relationship of the input text through a multi-head self-attention mechanism; The multi-stage translation unit comprises: Semantic extraction module, used to extract the semantic representation of input text; A syntax generation module, used to generate a syntax tree according to the syntax rules of the target language; The natural language generation module is used to convert the syntax tree into target language text.

[0011] Preferably, the multi-head self-attention mechanism calculates the contextual relationship by the following formula:

[0012] Where: is the query matrix, is the key matrix, is the value matrix, The dimension of the key.

[0013] Preferably, the corpus management module uses contrastive learning to optimize the semantic alignment of high-resource languages ​​and scarce languages, and the optimization goal is:

[0014] Where: and are the semantic representations of high-resource languages ​​and scarce languages, respectively. is the cosine similarity, is the temperature parameter.

[0015] Preferably, the corpus management module expands scarce corpus by generating adversarial networks, wherein the generator and the discriminator are optimized by adversarial objectives through the following formula:

[0016] Where: represents a generator, represents the discriminator, is the generated pseudo scarce corpus.

[0017] Preferably, the privacy and security module includes: The localized translation module is used to run the compressed translation model on the user's device to achieve real-time translation of high-frequency short texts; The end-to-end encryption module includes: a data encryption transmission unit, which is used to encrypt user data through TLS; and a homomorphic encryption processing unit, which is used to translate the data in an encrypted state.

[0018] Preferably, the cloud collaboration module dynamically allocates tasks through the following steps: High-frequency short texts are processed preferentially by the local model; Low-frequency long texts or complex translation tasks are handled by cloud servers; Regularly synchronize the translation model weights of the cloud server to the local model to achieve iterative model updates.

[0019] The present invention provides a multi-language intelligent translation system based on artificial intelligence. It has the following beneficial effects: 1. The present invention generates semantic mapping data of high-resource and scarce languages ​​by using unified semantic representation technology through the corpus management module, and provides a shared semantic space for the translation quality improvement module to call; on this basis, the translation quality improvement module generates accurate target language output through context modeling and multi-stage translation strategy, which can effectively realize efficient collaboration of multilingual translation tasks from corpus preparation to translation generation, thereby improving the translation efficiency and accuracy of multiple languages ​​and scarce languages.

[0020] 2. The present invention can achieve efficient expansion and optimization of scarce language corpora through the organic collaboration of the encoding unit, generation unit, discrimination unit and screening unit in the corpus management module: the encoding unit generates a unified semantic representation of high-resource languages ​​and scarce languages ​​through semantic alignment technology; the generation unit generates pseudo-scarce corpora using generative adversarial networks; the discrimination unit strictly evaluates the authenticity of pseudo-corpora; the screening unit further screens high-quality corpora through grammatical and semantic rules and stores them in a shared semantic space for translation models to call. The above process can greatly reduce the limitation of insufficient bilingual corpora of scarce languages, thereby achieving the effect of effectively supporting scarce language translation tasks.

[0021] 3. The present invention uses the localized translation of the privacy and security module and the intelligent task scheduling of the cloud-end collaboration module. The privacy and security module deploys a lightweight translation model on the user device side to achieve real-time translation of high-frequency short texts. At the same time, TLS data encryption transmission and homomorphic encryption technology are used to ensure the security of user data during the translation process. The cloud-end collaboration module assigns low-frequency long text tasks to the cloud for processing according to task complexity and real-time requirements, and regularly synchronizes the optimized cloud model to the local device to achieve iterative updates of the model. The above mechanism can effectively improve the execution efficiency and security of translation tasks, thereby achieving the effect of protecting user data privacy and security. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a system architecture diagram of the present invention; DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the specification 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.

[0024] In order to better understand the present invention, the above contents are described in detail below in conjunction with specific embodiments. Please see attached Figure 1 The embodiment of the present invention provides a multilingual intelligent translation system based on artificial intelligence, including: a collection module: used to receive voice information or text information input from the outside, convert the voice signal into text input through voice recognition technology, and transmit the processed text information to the corpus management module; the corpus management module: receives the text information transmitted by the collection module, combines the high-resource language corpus and the scarce language corpus stored in the system, generates a unified multilingual shared semantic representation, and expands the corpus of the scarce language to generate semantic mapping data for use by the translation quality improvement module; the translation quality improvement module: based on the semantic mapping data provided by the corpus management module, context modeling is performed on the input text, and a natural language output of the target language is generated through a multi-stage translation strategy; the privacy and security module: receives the translation result generated by the translation quality improvement module, protects the data privacy in the translation process, including localized translation and end-to-end encryption processing; the cloud-end collaboration module: coordinates the task scheduling between the collection module, the corpus management module, the translation quality improvement module and the privacy and security module, and dynamically allocates the translation task to the local or cloud to ensure translation efficiency and resource optimization.

[0025] The acquisition module is used to receive voice information or text information input from the outside, and convert the voice signal into text format through voice recognition technology. At the same time, the collected text is pre-processed by word segmentation, denoising, etc. before being transmitted to the corpus management module; the corpus management module receives the text information transmitted by the acquisition module, combines the stored high-resource language and scarce language corpora, generates multi-language shared semantic representations through unified semantic representation technology, further expands the corpus of scarce languages, forms semantic mapping data and provides it for the translation quality improvement module to call; the translation quality improvement module, based on the semantic mapping data provided by the corpus management module, contextually models the input text, and generates fluent and accurate target language output through a multi-stage translation strategy (including semantic extraction, syntax generation and natural language generation); the privacy and security module receives the translation results generated by the translation quality improvement module, protects user data privacy through localized translation and end-to-end encryption technology, and ensures the security of the translation process; the cloud-end collaboration module dynamically coordinates the task scheduling between modules, and allocates high-frequency short texts to local processing first according to the complexity of the task, while allocating low-frequency long texts or complex translation tasks to cloud servers, so as to maximize translation efficiency and optimize resource allocation and utilization.

[0026] The acquisition module includes: A sound signal receiving unit, used for receiving external input voice information; Speech recognition module, used to convert sound information into text input. The speech recognition module is trained using a deep neural network; The text preprocessing module is used to perform word segmentation, denoising and formatting on the collected text information, and output standardized text data to the corpus management module.

[0027] In the present invention, the acquisition module is composed of a sound signal receiving unit, a speech recognition module and a text preprocessing module, which is used to efficiently complete the acquisition and preliminary processing of voice or text information. The sound signal receiving unit can receive voice information input by the user, and supports multiple voice input methods, including receiving real-time voice through a microphone, or receiving audio files through file upload. The unit can perform preliminary processing on the sound signal, including signal enhancement, noise reduction, sampling rate adjustment, etc., to ensure that the voice information can be accurately recognized and converted into text by the subsequent processing module. The core of the sound signal receiving unit lies in its stability and adaptability, which can effectively process audio signals in various voice input scenarios, including voice collection in noisy environments. At the same time, the unit will pass the optimized voice signal to the speech recognition module; The speech recognition module is responsible for converting sound signals into corresponding text information, and uses deep neural networks for model training and reasoning during the recognition process. This module can handle voice input in multiple languages ​​by introducing a pre-trained speech recognition model and combining it with a large-scale speech corpus for optimization. At the same time, it has high robustness for complex pronunciations, dialects, and speech with fast and slow speech. The speech recognition module adopts an end-to-end training method, integrating speech feature extraction, acoustic modeling, and language modeling into a unified framework to improve recognition accuracy and processing speed. The generated text results will be passed to the text preprocessing module. The text preprocessing module further optimizes the recognized text information, including word segmentation, denoising, and formatting of the text. Word segmentation can split continuous text into independent words or phrases, which is convenient for semantic analysis in subsequent modules; denoising is used to remove possible erroneous symbols, repeated content, or invalid characters; formatting ensures the standardization of text data, including unified character encoding and text structure adjustment. The standardized text data after text preprocessing is output to the corpus management module for use in subsequent translation tasks. Therefore, the acquisition module can efficiently and accurately convert the user's input voice or text information into standardized input data suitable for translation tasks; The deep neural network model training and reasoning are as follows: 1. Data preprocessing Before training a deep neural network, the speech data needs to be fully preprocessed to ensure the quality of the data and the generalization ability of the model: Speech feature extraction: Extract features such as Mel-frequency cepstral coefficients (MFCC) and log-Mel filter bank energy (Log-Mel Spectrogram) from the original speech signal. These features can effectively reflect the spectral information of the speech signal and facilitate deep neural network modeling.

[0028] Before extracting features, speech signals are usually subjected to denoising, normalization, and framing operations to divide continuous audio signals into frames of fixed length, and use windowing techniques (such as Hamming windows) to enhance the fine-grained representation capabilities of spectral features.

[0029] Data Augmentation: Use data augmentation techniques to expand speech training datasets, such as adding background noise, changing speech speed and pitch, adding reverberation effects, etc. These techniques help improve the robustness of the model in noisy environments and multi-scenario conditions.

[0030] After data enhancement, the speech labels are aligned (the correspondence between text and speech frames) to ensure that each speech signal has accurate text annotations.

[0031] Data partitioning: The training data is usually divided into training set, validation set and test set, with the ratio usually being 8:1:1. The training set is used for model learning, the validation set is used for model parameter adjustment, and the test set is used for final performance evaluation.

[0032] 2. Deep Neural Network Model Design The deep neural network model of the speech recognition module adopts an end-to-end structure, integrating feature extraction, acoustic modeling, and language modeling functions. The specific model design is as follows: Acoustic Modeling: The acoustic modeling part usually uses a convolutional neural network (CNN) or a long short-term memory network (LSTM) to process the temporal characteristics of the speech signal.

[0033] CNN can extract local spectral features of speech signals and capture high-level time-frequency features through multi-layer convolution stacking.

[0034] LSTM can model the long-term dependencies of speech signals and is suitable for processing continuous speech streams.

[0035] Encoder-Decoder Architecture: Use an encoder-decoder model based on the attention mechanism (such as Transformer) to encode the speech feature sequence into a semantic vector of fixed dimension, and then generate the target text sequence through the decoder.

[0036] The encoder part captures the context of the speech, and the decoder combines the attention mechanism to generate word-by-word text output.

[0037] Language Modeling: The candidate text sequences generated by the acoustic model are input into a language model (such as RNN or BERT) to improve grammatical accuracy and fluency.

[0038] 3. Model training process The training process of the deep neural network is based on speech data and corresponding text annotations, using supervised learning methods. The specific steps are as follows: Loss function design: Use the connection temporal classification (CTC) loss function to align the speech features with the text annotations: CTC calculates the optimal path between the phoneme sequence and the target text sequence through a dynamic programming algorithm.

[0039] For the end-to-end attention model, the cross entropy loss function is used to calculate the degree of match between the generated text and the real text.

[0040] Gradient calculation and back propagation: Based on the loss function, the backpropagation algorithm is used to calculate the gradient of each layer of the network and adjust the network parameters to minimize the loss.

[0041] Update the network parameters using stochastic gradient descent (SGD) or an adaptive optimization algorithm such as Adam.

[0042] Batch training: The speech data is input into the network in batches, and the size of each batch of data is determined by the GPU memory.

[0043] In each training cycle (epoch), the network is iteratively updated multiple times on all training data.

[0044] 4. Model Optimization Strategy In order to further improve the performance of the speech recognition model, the present invention adopts the following optimization strategies: Regularization: The Dropout technique is used to randomly block some neurons to avoid overfitting of the model.

[0045] Use weight decay (L2 regularization) to control the scale of network parameters.

[0046] Pre-training and transfer learning: Use a pre-trained speech model (such as Wav2Vec) as the base model and fine-tune it with a small amount of scarce language data to improve the recognition ability of low-resource languages.

[0047] Learning rate scheduling: A dynamic learning rate adjustment strategy is adopted, such as a large initial learning rate and a gradual reduction of the learning rate in the later stages of training to accelerate convergence and avoid oscillation.

[0048] Model Ensemble: Combine the prediction results of multiple models (such as the CTC model and the end-to-end model) and take a weighted average or voting approach to improve the final recognition accuracy.

[0049] 5. Training Results and Performance Evaluation Evaluation Metrics: The performance of the speech recognition model is evaluated by the character error rate (CER) and word error rate (WER): CER: Indicates the difference between the generated text and the target text at the character level.

[0050] WER: represents the difference between the generated text and the target text at the word level.

[0051] Model iteration: According to the evaluation results of the validation set, the model structure and hyperparameters (such as network depth, convolution kernel size, etc.) are adjusted, and overfitting is avoided through the early stopping strategy.

[0052] Through the above training process, the speech recognition module can efficiently learn the mapping relationship between speech and text, realize accurate recognition of multilingual speech, and provide high-quality text input for subsequent translation tasks.

[0053] The corpus management module includes: Coding units, which are used to generate unified multilingual semantic representations by semantically encoding high-resource language texts and scarce language texts; The generation unit generates scarce language pseudo corpus through the generator of the generative adversarial network to expand the scarce corpus; The discriminant unit uses the discriminator of the generative adversarial network to discriminate the authenticity of the pseudo-corpus and the real corpus, which is used to optimize the quality of the generated corpus; The screening unit calculates the score of the pseudo-corpus through grammatical and semantic rules, and is used to screen out high-quality corpus and store it in the shared semantic space.

[0054] The corpus management module in the present invention includes an encoding unit, a generation unit, a discrimination unit and a screening unit. Each unit works together to realize the generation, optimization and management of multilingual corpora, especially the generation and screening of pseudo corpora of scarce languages, and provide high-quality corpus support for subsequent translation tasks, wherein: 1. Coding unit The main function of the encoding unit is to semantically encode high-resource language texts and scarce language texts to generate unified multilingual semantic representations. The encoding unit uses an encoder based on a multilingual pre-trained model to achieve cross-language sharing and mapping of semantic representations; Specific model type used: The encoding unit is based on the XLM-R (Cross-lingual Language Model-RoBERTa) or mBERT (Multilingual BERT) multilingual pre-trained models. These models learn cross-lingual semantic representations through unsupervised training on large-scale corpora in multiple languages, ensuring that high-resource languages ​​and scarce languages ​​can share semantic space; The encoding unit projects the corpus representation of the high-resource language and the corpus representation of the scarce language into the same vector space by training the alignment model, so that the rich information of the high-resource language can be transferred to the scarce language; Capturing the contextual semantic relationship of the text through a self-attention mechanism (such as the Transformer architecture) to generate accurate multilingual shared semantic representations. This representation is not only used for scarce corpus generation, but also for context modeling of the target language; Function: The encoding unit ensures that the scarce language can use the corpus training results of the high-resource language to improve the accuracy and generalization ability of the corpus representation, and provide high-quality semantic input for the generation unit; 2. Generate Unit The generation unit uses the generator model in the generative adversarial network (GAN) to generate pseudo corpus of scarce languages ​​based on the multilingual semantic representation provided by the encoding unit. The core task of this unit is to generate pseudo corpus that conforms to the grammatical and semantic rules of scarce languages ​​by learning the feature mapping between high-resource languages ​​and scarce languages; Implementation process: The generator receives the multilingual semantic representation generated by the encoding unit and converts it into corpus text of the scarce language; Random noise is added during the generation process to enhance the diversity of the generated corpus and ensure that the pseudo-corpus covers a wider range of language characteristics; 3. Discrimination unit The discriminator in the generative adversarial network (GAN) is used to distinguish the authenticity of the pseudo corpus output by the generation unit and the real scarce corpus, so as to optimize the generation quality of the generation unit. Implementation process: The discriminator receives the generated pseudo-corpus and real corpus as input and determines whether the input corpus conforms to the real characteristics of the scarce language; The discriminant unit improves the quality of generated pseudo corpus by optimizing the adversarial objective function (e.g., minimizing the generation loss of the generator unit and the discriminant loss of the discriminator); Function: The discriminant unit provides feedback information to help the generator unit continuously improve its generation strategy and generate scarce corpus that is closer to the real language characteristics; 4. Screening unit The screening unit evaluates the quality of the generated pseudo-corpus through grammatical and semantic rules, calculates the comprehensive score of the pseudo-corpus, and uses it to screen out high-quality corpora, which are then stored in a shared semantic space for subsequent use. Implementation process: The screening unit uses a pre-trained grammar model to score the grammatical correctness of the pseudo corpus; The semantic similarity between the pseudo-corpus and the high-resource corpus is calculated using the similarity function of the multilingual shared semantic space; Comprehensive grammatical and semantic scoring, and filter out pseudo-corpora that meet quality standards through preset thresholds; Function: The screening unit ensures that the corpus stored in the shared semantic space is of high quality, providing reliable data support for translation model training and translation quality improvement; The corpus management module generates a unified semantic representation through encoding units, generates pseudo-scarce corpus using generation units, and optimizes and controls the quality of the corpus through discrimination units and screening units. In particular, the multilingual pre-trained models (such as XLM-R and mBERT) used by the encoding unit can achieve cross-language semantic sharing, providing key technical support for the expansion of scarce language corpus. Through multi-unit collaboration, this module can effectively solve the problem of insufficient corpus in scarce language translation and provide core corpus support for the entire translation system.

[0055] The screening unit calculates the grammatical and semantic scores of the pseudo-corpus using the following formula:

[0056] Where: represents the grammatical correctness score of the pseudo-corpus, calculated by the pre-trained grammar model, Represents the semantic similarity between the pseudo-corpus and the high-resource language corpus, calculated from the semantic representation generated by the encoding unit, and is the weight coefficient of grammatical and semantic scores, satisfying + =1.

[0057] In the present invention, the screening unit is used to perform quality assessment on the pseudo-corpus output by the generation unit, ensuring that the screened corpus meets high quality standards in terms of grammatical correctness and semantic consistency, and storing the screened corpus in a shared semantic space for subsequent use. First, the screening unit uses a pre-trained grammar model to score the grammatical correctness of the pseudo-corpus. The model is trained with a large amount of multilingual grammatical corpora and can accurately assess whether the pseudo-corpus conforms to the grammatical rules of the target language, including the integrity of grammatical structure, accuracy of word usage, and standardization of sentence patterns. At the same time, the screening unit compares and analyzes the pseudo-corpus with the real corpus to determine the degree of conformity of the pseudo-corpus in complex syntactic structure, phrase connection, and language logic, and generates a grammar score for each pseudo-corpus as one of the important indicators for subsequent screening; In addition, the screening unit also calculates the semantic similarity of the pseudo-corpus in combination with the semantic representation in the multilingual shared semantic space, and evaluates the semantic consistency and expression accuracy of the pseudo-corpus by comparing it with the high-resource language corpus. The semantic representation generated by the encoding unit is a high-dimensional vector that can capture the contextual semantic relationship and language characteristics of the corpus. The screening unit calculates the semantic similarity score between the pseudo-corpus and the high-resource corpus based on these semantic representations. The screening unit combines the grammatical score and the semantic similarity score, and synthesizes the two into a final score according to the weight ratio through a preset weighting coefficient, and sets a threshold to screen the pseudo-corpus. The screening unit can automatically identify and eliminate pseudo-corpora with many grammatical errors or serious semantic deviations, ensuring that the corpus stored in the shared semantic space has high language quality and practical application value.

[0058] The translation quality improvement module includes: Context modeling unit and multi-stage translation unit. The context modeling unit is used to calculate the contextual relationship of the input text through a multi-head self-attention mechanism; A multi-stage translation unit consists of: Semantic extraction module, used to extract the semantic representation of input text; A syntax generation module, used to generate a syntax tree according to the syntax rules of the target language; The natural language generation module is used to convert the syntax tree into target language text.

[0059] In the present invention, the translation quality improvement module includes a context modeling unit and a multi-stage translation unit, and the semantic accuracy and grammatical coherence of the translated text are improved through the collaborative work of the two. The context modeling unit models the contextual relationship of the input text through a multi-head self-attention mechanism. The multi-head self-attention mechanism can capture the interdependencies between words in the text, regardless of whether the words are adjacent or distributed at a distance. Specifically, the unit divides the input text into a series of embedding vectors, and dynamically adjusts the weights between these vectors through a self-attention mechanism to identify the core semantic information and important grammatical associations in the text, thereby generating a context-aware semantic representation. The output result of the context modeling unit is a high-dimensional semantic vector, which can accurately reflect the semantic structure of the input text and provide basic data for the multi-stage translation unit; The multi-stage translation unit includes a semantic extraction module, a syntax generation module and a natural language generation module, and the three cooperate in turn to complete the natural language output of the target language. First, the semantic extraction module extracts key semantic information from the output of the context modeling unit and converts it into the semantic representation of the target language. These semantic representations are the basis of translation generation and can accurately reflect the content and logic of the source text. Then, the syntax generation module generates a corresponding syntax tree according to the grammatical rules and syntactic structure of the target language. This process generates a syntactic structure that conforms to the expression habits of the target language based on the grammatical characteristics of the target language, such as word order, subject-verb-object relationship, modifier position, etc. Finally, the natural language generation module further converts the output of the syntax generation module into a fluent target language text to ensure that the output text is not only grammatically correct, but also has natural language expression and contextual coherence. Through the division of labor and collaboration of the context modeling unit and the multi-stage translation unit, the translation quality improvement module of the present invention can generate high-quality target language text in complex contexts and multilingual translation scenarios.

[0060] The multi-head self-attention mechanism calculates the context relationship through the following formula:

[0061] Where: is the query matrix, is the key matrix, is the value matrix, The dimension of the key.

[0062] In the present invention, a multi-head self-attention mechanism is used to calculate the contextual relationship of the input text, and the dependency and importance between words are identified by dynamically weighting the word embeddings of the input text. Specifically, the input text is first encoded as an embedding vector, and each word is represented as a high-dimensional vector, called a query matrix, a key matrix, and a value matrix. These matrices represent the focus of the word in the sentence (query), the degree of association with other words (key), and the corresponding semantic information (value). The self-attention mechanism calculates the relevance weight of each word in the context by matching the query matrix and the key matrix, and then generates a new context vector representation in combination with the value matrix; From the above, it can be seen that the multi-head self-attention mechanism can capture local and global semantic information in the text. For example, it can identify the associations between long-distance words in a sentence (such as the relationship between subject and predicate) and semantic dependencies within local phrases. The introduction of the multi-head mechanism further improves the expressiveness of the model. Each head can focus on different types of semantic or grammatical relationships, making the semantic modeling of the text more comprehensive and accurate. Ultimately, these context vectors are used as semantic inputs for subsequent translation tasks to ensure that the generated target language text accurately reflects the meaning and structure of the source text. Through the self-attention mechanism, the translation system can significantly improve the translation quality in complex sentences and multilingual contexts.

[0063] The corpus management module uses contrastive learning to optimize the semantic alignment of high-resource languages ​​and scarce languages. The optimization objectives are:

[0064] Where: and are the semantic representations of high-resource languages ​​and scarce languages, respectively. is the cosine similarity, is the temperature parameter.

[0065] In the present invention, the corpus management module realizes semantic alignment between high-resource languages ​​and scarce languages ​​through contrastive learning technology, solving the problem of insufficient corpus of scarce languages. Specifically, the core idea of ​​contrastive learning is to shorten the distance between the representations of high-resource languages ​​and scarce languages ​​under the same semantics, while increasing the distance between the representations of different semantics, so that the rich corpus information of high-resource languages ​​can be effectively transferred to scarce languages. The module first uses encoding units to map the text inputs of high-resource languages ​​and scarce languages ​​into high-dimensional semantic representations, and captures the semantic characteristics of the two languages ​​through a shared semantic space; During the optimization process, the loss function of contrastive learning is defined by calculating the similarity between the semantic representations of high-resource languages ​​and scarce languages, where cosine similarity is used to evaluate the similarity of the two semantic representations. The corpus management module maximizes the similarity between the representations of high-resource languages ​​and scarce languages ​​under the same semantics through alignment training, and adjusts the gradient distribution in contrastive learning by introducing a temperature parameter to improve the semantic alignment effect of the model on scarce languages. This optimization method ensures that the generated scarce language pseudo-corpus can more accurately reflect the target semantics, thereby improving the quality of subsequent translation, while significantly enhancing the system's support for low-resource languages. The corpus management module expands the scarce corpus by generating adversarial networks, where the generator and discriminator are optimized against the objective by the following formula:

[0066] Where: represents a generator, represents the discriminator, is the generated pseudo scarce corpus.

[0067] In the present invention, the corpus management module expands the scarce corpus through the generative adversarial network (GAN), and realizes the generation and optimization of high-quality pseudo-scarce corpus through adversarial training of the generator and the discriminator. Specifically, the task of the generator is to generate pseudo-scarce corpus based on the corpus of the high-resource language. The generated pseudo-corpus needs to be as close to the real scarce language corpus as possible to meet the characteristics of the target language in terms of grammar and semantics. The discriminator is responsible for distinguishing pseudo-corpus from real corpus, and its purpose is to improve the authenticity of the corpus generated by the generator. Through the adversarial training of the two, the generator continuously learns to improve the quality of corpus generation, while the discriminator continuously improves its ability to identify pseudo-corpus through feedback; During the optimization process, the goal of the generator is to minimize the probability of the pseudo-corpus being judged as fake corpus, while the goal of the discriminator is to maximize the probability of correctly distinguishing real corpus from pseudo-corpus. The pseudo-scarce corpus generated by the generator will be input into the discriminator for authenticity evaluation, and the parameters of the generator will be adjusted according to the judgment results to gradually improve the grammatical and semantic consistency of the generated corpus. Through this dynamic adversarial optimization, the generator is able to generate high-quality pseudo-corpus that is closer to the real scarce language, significantly expanding the corpus of scarce languages ​​and providing strong data support for scarce language translation tasks. Ultimately, this method not only improves the coverage of scarce language corpus, but also ensures the language quality and diversity of the generated corpus, effectively enhancing the system's translation capabilities for scarce languages; The corpus management module expands scarce corpus through generative adversarial networks (GANs), and the generator and discriminator are responsible for the generation and quality assessment of pseudo scarce corpus respectively. In order to achieve the generation and optimization of high-quality pseudo corpus, the generator and discriminator adopt a deep neural network architecture, and its specific design is as follows: 1. Generator network structure The generator is responsible for generating pseudo corpus of scarce languages ​​based on corpus of high-resource languages. Its network structure design focuses on the diversity and authenticity of the generated corpus, and is mainly composed of an encoding layer, a decoding layer, and a semantic alignment module.

[0068] Input layer: The generator receives corpus inputs in high-resource languages. These inputs are passed through the encoding unit of the corpus management module to generate multilingual semantic representations. The input dimension is the same as the dimension of the semantic space.

[0069] Encoding layer: The encoding layer consists of multiple layers of stacked convolutional neural networks (CNNs) or self-attention mechanisms (Transformers) to extract high-level semantic features of the input corpus.

[0070] For example, the Transformer encoder can capture the contextual semantic relations of the input corpus and ensure the semantic consistency between the scarce language pseudo-corpus and the high-resource language corpus.

[0071] Decoding layer: The decoding layer uses a recurrent neural network (RNN), a long short-term memory network (LSTM), or a Transformer decoder to decode high-dimensional semantic features into text representations of scarce languages. During the decoding process, the model combines the contextual information of the corpus to generate pseudo corpus that conforms to the characteristics of the scarce language (such as grammatical structure and language rules).

[0072] Semantic alignment module: In the generation process, a semantic alignment module is introduced to optimize the semantic representation of the generated corpus through contrastive learning method to make it consistent with the semantics of the high-resource language.

[0073] This module provides feedback optimization to the generator by calculating the semantic similarity between the generated corpus and the target scarce language.

[0074] Output layer: The output layer generates pseudo scarce language corpus text in the form of a series of words or phrases, ensuring that it is grammatically and semantically consistent with the real corpus of the scarce language.

[0075] 2. Discriminator network structure The task of the discriminator is to distinguish the pseudo corpus generated by the generator from the real scarce corpus. Its design focuses on accurately evaluating the grammatical and semantic consistency of the pseudo corpus. The network structure mainly consists of a feature extraction layer, a classification layer, and a grammatical and semantic fusion module.

[0076] Input layer: The discriminator receives the pseudo corpus generated by the generator and the real scarce corpus as input. The input is in the form of a series of text sequences, which are converted into vector representations through the embedding layer.

[0077] Feature extraction layer: Use multi-layer stacked convolutional neural networks (CNNs) to extract local features of the corpus, such as phrase structure, grammatical patterns, etc. At the same time, combine the bidirectional long short-term memory network (Bi-LSTM) to capture the long-distance dependencies and contextual semantic features of the corpus.

[0078] Syntax-semantics fusion module: This module combines the outputs of the grammatical model and the semantic model to conduct a comprehensive analysis of the grammatical correctness and semantic consistency of the input corpus. The grammatical model evaluates whether the input corpus conforms to the grammatical rules of the scarce language (such as word order, sentence structure, etc.). The semantic model evaluates the degree of semantic alignment of the input corpus with the high-resource language.

[0079] Classification layer: The classification layer consists of a fully connected layer and a Softmax layer, which outputs the discrimination probability of pseudo-corpus and real-corpus. The discriminator improves its ability to distinguish pseudo-corpus from real-corpus by continuously optimizing the classification error.

[0080] 3. Adversarial training between generator and discriminator The generator and discriminator are jointly optimized through the adversarial objective function: The goal of the generator is to generate pseudo corpus that is as close to the real scarce corpus as possible, so that the discriminator cannot distinguish its authenticity.

[0081] The goal of the discriminator is to accurately distinguish between pseudo-corpus and real corpus and maximize the discrimination probability of real corpus.

[0082] Through alternating optimization of the generator and the discriminator, the generator gradually learns the ability to generate high-quality pseudo-corpus, while the discriminator continuously improves its ability to recognize pseudo-corpus, ultimately achieving efficient expansion of scarce language corpus.

[0083] Therefore, the generator focuses on generating high-quality pseudo-corpora that conform to the characteristics of scarce languages ​​from high-resource language corpora, using a deep encoding and decoding architecture and a semantic alignment module. The discriminator strictly evaluates the syntax and semantics of the generated corpora through feature extraction and classification to optimize the performance of the generator. This improves the generation quality of scarce corpora and effectively expands the coverage of corpora, providing strong data support for the translation of scarce languages.

[0084] The Privacy and Security module includes: The localized translation module is used to run the compressed translation model on the user's device to achieve real-time translation of high-frequency short texts; The end-to-end encryption module includes: a data encryption transmission unit, which is used to encrypt user data through TLS; and a homomorphic encryption processing unit, which is used to translate the data in an encrypted state.

[0085] In the present invention, the privacy and security module is used to protect the privacy and security of user data during the translation process. The module design includes a localization translation module and an end-to-end encryption module, which work together to effectively achieve the combination of real-time translation and data protection, including: 1. Localization translation module The localized translation module provides real-time translation of high-frequency short texts by running a compressed translation model on the user's device. In order to reduce the demand for device computing resources, the module uses knowledge distillation technology to compress large cloud-based translation models into lightweight models while keeping translation performance close to the original model. Specifically: Knowledge distillation technology allows the local model to mimic the output behavior of the cloud model (including prediction distribution and translation results) to train a small model with similar capabilities. The localized translation model enables fast translation of short texts on the user's device, effectively reducing the number and duration of data transmission and reducing the risk of user privacy exposure.

[0086] 2. End-to-end encryption module The end-to-end encryption module is responsible for encrypting and protecting user data during the transmission and processing of translation tasks. The design includes a data encryption transmission unit and a homomorphic encryption processing unit.

[0087] Data encryption transmission unit: The data encryption transmission unit adopts the Transport Layer Security Protocol (TLS) to protect the integrity and security of user data during network transmission.

[0088] The TLS protocol prevents man-in-the-middle attacks, data tampering and theft by encrypting communication channels, ensuring that user data will not be illegally obtained when it is transmitted to the cloud for processing.

[0089] Homomorphic encryption processing unit: The homomorphic encryption processing unit uses homomorphic encryption technology to enable user data to be translated while being encrypted. The specific implementation is as follows: Data encryption: Before sending a translation task, the user device performs homomorphic encryption on the original text data to generate encrypted ciphertext.

[0090] Encrypted translation calculation: The cloud-based translation model directly calculates on the encrypted data without decrypting the ciphertext, and performs operations on the ciphertext (such as encoding, context modeling, and translation generation) through homomorphic encryption technology. Since homomorphic encryption allows specific operations (such as addition or multiplication) to be performed on the ciphertext, the cloud-based model can complete the translation task normally and generate encrypted translation results.

[0091] Decrypted output: The translation result is returned to the user's device in encrypted form. The user decrypts it with the private key to obtain the translated text in the target language.

[0092] Core technology implementation: Homomorphic encryption uses partially homomorphic encryption (PHE) or fully homomorphic encryption (FHE) algorithms, such as Paillier encryption, Batched RSA, or Gentry fully homomorphic encryption. In order to balance computational complexity and encryption strength, the module prefers partially homomorphic encryption schemes (such as Paillier encryption), which can support addition operations on encrypted data and are suitable for common tasks such as weighted calculation of model parameters.

[0093] Fully homomorphic encryption is suitable for scenarios that require higher security and supports addition and multiplication operations, but the computational complexity is relatively high. In order to improve efficiency, the module adopts a hierarchical optimization method to split different encryption operations into suitable homomorphic encryption algorithms.

[0094] Through the collaboration of the localized translation module and the end-to-end encryption module, the present invention can achieve full protection of user data from transmission to translation processing. Among them, the specific implementation of homomorphic encryption technology allows user data to always be encrypted when performing translation calculations in the cloud, effectively preventing the leakage of sensitive information, and taking into account both translation efficiency and security.

[0095] The cloud collaboration module dynamically allocates tasks through the following steps: High-frequency short texts are processed preferentially by the local model; Low-frequency long texts or complex translation tasks are handled by cloud servers; Regularly synchronize the translation model weights of the cloud server to the local model to achieve iterative model updates.

[0096] The cloud collaboration module in the present invention optimizes translation efficiency and resource utilization by dynamically allocating translation tasks, and ensures consistency between the local model and the cloud model through a model weight synchronization mechanism. The specific task allocation and weight synchronization methods are as follows: 1. Dynamic Task Allocation The cloud collaboration module assigns translation tasks to local models or cloud servers based on the length, complexity, and real-time requirements of the translation tasks: 1. High-frequency short texts are processed preferentially by the local model: The local model is specifically designed to handle short text tasks that are frequently input by users, reducing network transmission delays and providing a real-time translation experience.

[0097] The local model runs on the user's device and is lightweight and optimized to ensure that it can quickly process short texts and maintain high translation quality.

[0098] 2. Low-frequency long texts or complex translation tasks are processed by cloud servers: For complex sentences, multi-layer nested grammar or tasks with strong domain expertise, the module will be assigned to cloud servers for processing first. The large translation model on the cloud server has stronger semantic parsing capabilities and higher computing power support, ensuring the accuracy and fluency of complex tasks.

[0099] 2. Model Weight Synchronization In order to ensure the consistency of translation capabilities between the local model and the cloud model, the cloud collaboration module implements regular synchronization of translation model weights. The specific algorithm is as follows: 1. Data preparation: The cloud model continuously optimizes weight parameters in translation tasks. The training data may include user feedback annotation data (such as user evaluation of translation results) and new domain corpus. The weight parameters of the cloud model are stored on the server side and decomposed into subsets suitable for the local model.

[0100] 2. Weight screening and lightweighting: The weights of cloud models usually contain a large number of parameters. Direct synchronization to the local machine may lead to excessive storage and computing burdens. The module uses the following strategies to screen and compress weights: Parameter pruning: Delete the parameters in the weight matrix that have little impact on the model and only keep the important weight items.

[0101] Quantization: Quantize the floating-point weight values ​​into low-order integers (such as 8 bits), significantly reducing storage space and computational complexity.

[0102] Distillation: Use the translation results output by the cloud model to train the local model and simplify the weight representation.

[0103] 3. Weight distribution: The cloud collaboration module sends the filtered and lightweight model weights to the user device through a secure transmission protocol (such as TLS).

[0104] The weight update frequency is determined by the computing power of the user device and the network conditions, and is usually updated when the device is idle.

[0105] 4. Local model updates include: Weight replacement: Directly replace the weight values ​​of the local corresponding layer with the pruned and quantized weights.

[0106] Progressive merging: Use weight interpolation to smoothly update local model parameters, assuming that the local weight is , the cloud weight is , then the combined formula is:

[0107] Where: is the merging coefficient, which is usually adjusted dynamically according to the model performance; 5. Consistency Verification After the local model completes the weight update, the module verifies the performance of the new model through a small-scale translation task to ensure the accuracy and robustness of the updated model; Through the weight synchronization algorithm, the present invention can ensure that the local model is lightweight, so that its translation ability is consistent with the cloud model. The weight synchronization mechanism of the cloud collaboration module not only improves the performance of the local model, but also avoids the need to frequently rely on cloud resources, providing users with an efficient and consistent translation experience, while ensuring the optimal use of device resources.

[0108] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multilingual intelligent translation system based on artificial intelligence, characterized in that: include: Acquisition module: used to receive external input voice information or text information, convert voice signals into text input through speech recognition technology, and pass the processed text information to the corpus management module; Corpus management module: receives text information transmitted by the acquisition module, combines the high-resource language corpus and scarce language corpus stored in the system, generates a unified multi-language shared semantic representation, expands the scarce language corpus, and generates semantic mapping data for use by the translation quality improvement module; Translation quality improvement module: Based on the semantic mapping data provided by the corpus management module, it performs context modeling on the input text and generates natural language output in the target language through a multi-stage translation strategy; Privacy and security module: receives the translation results generated by the translation quality improvement module and protects data privacy during the translation process, including localized translation and end-to-end encryption processing; Cloud-end collaboration module: coordinates the task scheduling between the collection module, corpus management module, translation quality improvement module and privacy and security module, and dynamically allocates translation tasks to the local or cloud to ensure translation efficiency and resource optimization.

2. The multilingual intelligent translation system based on artificial intelligence according to claim 1, characterized in that: The acquisition module comprises: A sound signal receiving unit, used for receiving external input voice information; A speech recognition module, used to convert the sound information into text input, wherein the speech recognition module is trained using a deep neural network; The text preprocessing module is used to perform word segmentation, denoising and formatting on the collected text information, and output standardized text data to the corpus management module.

3. The multilingual intelligent translation system based on artificial intelligence according to claim 1, characterized in that: The corpus management module includes: Coding units, which are used to generate unified multilingual semantic representations by semantically encoding high-resource language texts and scarce language texts; The generation unit generates scarce language pseudo corpus through the generator of the generative adversarial network to expand the scarce corpus; The discriminant unit uses the discriminator of the generative adversarial network to discriminate the authenticity of the pseudo-corpus and the real corpus, which is used to optimize the quality of the generated corpus; The screening unit calculates the score of the pseudo-corpus through grammatical and semantic rules, and is used to screen out high-quality corpus and store it in the shared semantic space.

4. The multilingual intelligent translation system based on artificial intelligence according to claim 3, characterized in that: The screening unit calculates the grammatical and semantic scores of the pseudo corpus by the following formula: Where: represents the grammatical correctness score of the pseudo-corpus, calculated by the pre-trained grammar model, Represents the semantic similarity between the pseudo-corpus and the high-resource language corpus, calculated from the semantic representation generated by the encoding unit, and is the weight coefficient of grammatical and semantic scores, satisfying + =1.

5. The multilingual intelligent translation system based on artificial intelligence according to claim 1, characterized in that: The translation quality improvement module includes: A context modeling unit and a multi-stage translation unit, wherein the context modeling unit is used to calculate the context relationship of the input text through a multi-head self-attention mechanism; The multi-stage translation unit comprises: Semantic extraction module, used to extract the semantic representation of input text; A syntax generation module, used to generate a syntax tree according to the syntax rules of the target language; The natural language generation module is used to convert the syntax tree into target language text.

6. The multi-language intelligent translation system based on artificial intelligence according to claim 5, characterized in that: The multi-head self-attention mechanism calculates the context relationship through the following formula: Where: is the query matrix, is the key matrix, is the value matrix, The dimension of the key.

7. The multilingual intelligent translation system based on artificial intelligence according to claim 1, characterized in that: The corpus management module uses contrastive learning to optimize the semantic alignment of high-resource languages ​​and scarce languages. The optimization goal is: Where: and are the semantic representations of high-resource languages ​​and scarce languages, respectively. is the cosine similarity, is the temperature parameter.

8. The multi-language intelligent translation system based on artificial intelligence according to claim 7, characterized in that: The corpus management module expands scarce corpus by generating adversarial networks, where the generator and the discriminator are optimized against the target by the following formula: Where: represents a generator, represents the discriminator, is the generated pseudo scarce corpus.

9. The multi-language intelligent translation system based on artificial intelligence according to claim 1, characterized in that: The privacy and security module includes: The localized translation module is used to run the compressed translation model on the user's device to achieve real-time translation of high-frequency short texts; The end-to-end encryption module includes: a data encryption transmission unit, which is used to encrypt user data through TLS; and a homomorphic encryption processing unit, which is used to translate the data in an encrypted state.

10. The multi-language intelligent translation system based on artificial intelligence according to claim 1, characterized in that: The cloud collaboration module dynamically allocates tasks through the following steps: High-frequency short texts are processed preferentially by the local model; Low-frequency long texts or complex translation tasks are handled by cloud servers; Regularly synchronize the translation model weights of the cloud server to the local model to achieve iterative model updates.

Images