A multilingual machine translation method based on deep learning

Abstract

The present invention relates to the field of deep learning, and in particular to a multilingual machine translation method based on deep learning. It includes: obtaining and preprocessing input sentences, obtaining a word embedding vector sequence through a pre-training method, and processing the word embedding vector using a semantic tensor generation algorithm to generate a third-order semantic tensor element; performing a nonlinear transformation on the third-order semantic tensor element to obtain a transformed tensor element, and using a dynamic weight-controlled multidimensional semantic feature extraction algorithm to extract features from the transformed tensor element to obtain a core semantic vector; processing the core semantic vector through a target language dynamic mapping mechanism to dynamically generate a target language semantic vector; based on the target language semantic vector, predicting the generation method of the next word through a sentence translation generation algorithm to complete the sentence. It solves the technical problem that in the process of recursively generating semantic vectors, the numerical value is prone to divergence or instability, resulting in semantic representation deviation and reduced generation quality.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and in particular to a multilingual machine translation method based on deep learning. Background Art

[0002] Machine translation is an important branch of natural language processing, dedicated to the automatic conversion of texts between different languages, accurately and fluently translating a source language into another target language. The research on machine translation technology can be traced back to the 1950s, and has undergone multiple evolutions from rule-based methods to statistical methods and, in recent years, deep learning-based methods.

[0003] Rule-based machine translation relies on rules manually formulated by linguistic experts, including lexical, syntactic and semantic rules. It requires a lot of manual intervention and is difficult to adapt to languages ​​with complex grammars and has poor scalability. Statistical machine translation trains translation models through bilingual corpora and uses probability models to derive the most likely translation results. However, it makes simple assumptions about the correspondence between language pairs and is difficult to capture global information at the sentence level. It also requires a large amount of bilingual corpora as training data, which is difficult to work on language pairs with scarce resources. A multilingual machine translation method based on deep learning has brought revolutionary changes to machine translation. Through deep learning models, it can automatically learn the mapping relationship between languages, reduce dependence on manual feature engineering, and better handle long-distance dependencies and contextual relationships. It effectively improves the breadth and depth of application of multilingual machine translation technology, provides efficient solutions for multilingual communication and information dissemination, and will further promote information exchange and cultural integration on a global scale.

[0004] However, existing multilingual machine translation methods have the following technical problems: word embedding models cannot effectively capture the high-order grammatical structure, syntactic dependency and semantic features of input sentences, resulting in a lack of depth and completeness of semantic information in complex tasks; high-dimensional semantic representation is accompanied by the problems of sparsity and high computational complexity, which will increase the storage and computational burden, and sparsity may lead to the loss of detailed semantic features; the existing technology uses a static weight scheme to allocate weights to semantic features, which lacks dynamic adjustment capabilities, making it difficult to cope with diverse input scenarios and context changes, and the feature representation is not flexible enough; semantic alignment cannot accurately capture the global and local semantic features between sentences, resulting in low quality of generated results; in the process of recursively generating semantic vectors, the values ​​are prone to divergence or instability, resulting in semantic representation deviation and reduced generation quality. Summary of the invention

[0005] The present invention provides a multilingual machine translation method based on deep learning to solve the problems that a word embedding model is difficult to effectively capture high-order grammatical structures, syntactic dependencies and semantic features of input sentences, resulting in a lack of depth and completeness of semantic information in complex tasks; high-dimensional semantic representation is accompanied by the problems of sparsity and high computational complexity, which will increase the storage and computational burden, and sparsity may lead to the loss of detailed semantic features; the prior art uses a static weight scheme for weight allocation of semantic features, lacks dynamic adjustment capabilities, and is difficult to cope with diverse input scenarios and context changes, and feature representation is not flexible enough; semantic alignment is difficult to accurately capture global and local semantic features between sentences, resulting in low quality of generated results; in the process of recursively generating semantic vectors, values ​​are prone to divergence or instability, resulting in technical problems such as semantic representation deviation and reduced generation quality.

[0006] A multilingual machine translation method based on deep learning of the present invention specifically includes the following technical solutions:

[0007] A multilingual machine translation method based on deep learning, comprising the following steps:

[0008] S1: Obtain and preprocess the input sentence, obtain the word embedding vector sequence through the pre-training method, and use the semantic tensor generation algorithm to process the word embedding vector to generate the third-order semantic tensor elements;

[0009] S2: Perform nonlinear transformation on the third-order semantic tensor elements to obtain transformed tensor elements, and use the dynamic weight control multi-dimensional semantic feature extraction algorithm to extract features from the transformed tensor elements to obtain the core semantic vector;

[0010] S3: Process the core semantic vector through the target language dynamic mapping mechanism to dynamically generate the target language semantic vector;

[0011] S4: Based on the target language semantic vector, the sentence translation generation algorithm is used to predict the generation method of the next word to complete the sentence.

[0012] Preferably, the S1 specifically includes:

[0013] In the implementation of the semantic tensor generation algorithm, each word embedding vector in the input word embedding vector sequence is processed to form a dynamically adjusted representation; during the adjustment process, the basic semantic features of the word are captured through the basic word vector matrix, which is used as the base layer, and the context information is superimposed on the base layer by weighted summation of the context adjustment matrix; after the context adjustment is completed, the word embedding vector is expanded to a high-dimensional tensor representation to generate a third-order semantic tensor element:

[0014]

[0015] Among them, T ijk Represents the third-order semantic tensor element, reflecting the grammatical structure position information i, syntactic dependency feature j, and semantic features under the semantic feature index k; represents the sum of the semantic contributions of all word embedding vectors in the input sentence; represents a dynamic embedding function; W f represents the basic word vector matrix; x p represents the p-th word embedding vector; α represents the context weight parameter; ∑ q≠p W represents the sum of the semantic contributions of the context words to the p-th word embedding vector in the input sentence; c represents the context adjustment matrix; x q represents the qth word embedding vector; g(i, j, k) represents the weight function.

[0016] Preferably, the S2 specifically includes:

[0017] The nonlinear transformation uses logarithmic function compression and sign function retention technology to dynamically compress the absolute value of the third-order semantic tensor elements and retain the positive and negative sign information of the third-order semantic tensor elements to obtain the transformed tensor elements. The calculation formula is:

[0018] T′ ijk =log2(1+|T ijk |)·sgn(T ijk )

[0019] Among them, T′ ijk Represents the transformed tensor element; log2(1+|T ijk |) represents the logarithmic function; sgn(T ijk ) represents a sign function.

[0020] Preferably, the S2 specifically includes:

[0021] In the implementation process of the dynamic weight-controlled multi-dimensional semantic feature extraction algorithm, the transformed tensor elements are extracted through tensor decomposition, dynamic weight control, feature importance evaluation, and semantic vector combination to obtain semantic features.

[0022] Preferably, the S2 specifically includes:

[0023] In the process of implementing the dynamic weight-controlled multi-dimensional semantic feature extraction algorithm, dynamic weights are assigned to each feature in the tensor decomposition process, and feature importance evaluation is introduced. By calculating the change amount of the tensor in three dimensions, the change rate in the three directions of word grammatical structure position information, syntactic dependency, and semantic feature index is calculated. The tensor decomposition and weight allocation formulas are as follows:

[0024]

[0025] Among them, T′ ijk Represents the transformed tensor element; represents the summation operation of tensor decomposition; r represents the rank of tensor decomposition, that is, the number of features after tensor decomposition; σ l A represents the eigenvalue of the lth decomposition feature; l , B l , C l represents the feature matrix after the lth decomposition, that is, the projection of the tensor in the three dimensions of word position information, syntactic dependency features, and semantic feature index; Represents the outer product operation of a tensor; λ l (i, j, k) represents the dynamic weight control function of the lth decomposition feature.

[0026] Preferably, the S2 specifically includes:

[0027] After completing tensor decomposition and weight assignment, the low-rank feature matrix obtained by tensor decomposition is combined according to the weights to generate the core semantic vector. The formula is as follows:

[0028] u=[λ1·σ1·A1, λ1·σ1·B1, λ1·σ1·C1,...,λ l ·σ l ·A l ,λ l ·σ l ·B l ,λ l ·σ l ·C l , ..., λ r ·σ r ·A r ,λ r ·σ r ·B r ,λ r ·σ r ·C r ]

[0029] Among them, u represents the core semantic vector, that is, the combination of semantic features after decomposition.

[0030] Preferably, the S3 specifically includes:

[0031] The core semantic vector is globally mapped, locally modified, and recursively enhanced through the target language dynamic mapping mechanism to dynamically generate the target language semantic vector. The generation formula is:

[0032]

[0033] Among them, v trepresents the target language semantic vector generated at time t; Represents the modulus of the global semantic mapping item, as a normalization factor; represents the logarithmic function; u represents the core semantic vector, that is, the combination of semantic features after decomposition; represents the global mapping matrix; S represents the local correction matrix; γ represents the recursive coefficient; R(v t-1 ) represents the recursive feature enhancement function.

[0034] Preferably, the S3 specifically includes:

[0035] After each round of recursive generation of the target language semantic vector, the difference between the currently generated target language semantic vector and the target language semantic vector generated in the previous round is evaluated. When the difference is lower than the preset convergence threshold, it is considered that the recursive process has reached convergence. When the difference is higher than and equal to the preset convergence threshold, the next round of mapping will continue until the convergence condition is met. The recursive termination condition formula is:

[0036] ||v t -v t-1 ||<∈

[0037] Among them, ||v t -v t-1 || represents the Euclidean distance between the target language semantic vector generated at time t and the target language semantic vector generated at time t-1; ∈ represents the convergence threshold.

[0038] Preferably, the S4 specifically includes:

[0039] In the implementation process of the sentence translation generation algorithm, the input target language semantic vector and the previous generated word are used to predict the current word; in the process of generating words at each step, the input target language semantic vector is updated according to the current context state; the input vector combination is projected into the target language vocabulary space through the mapping matrix, and the result of the linear mapping is normalized into a probability distribution through the normalization function, and the word with the largest probability is selected as the generated word of the current step.

[0040] Preferably, the S4 specifically includes:

[0041] The calculation formula for the conditional probability of generating a word is:

[0042] P(y m |y m-1 , v t )=softmax(W m ·[y m-1 , v t ])

[0043] Among them, P(y m|y m-1 , v t ) represents the conditional probability of generating the mth word; y m represents the embedding vector of the mth word; softmax represents the normalization function; W m Represents the mapping matrix; [y m-1 , v t ] represents the input vector combination;

[0044] The formula for selecting generated words is:

[0045]

[0046] Among them, y m Represents the embedding vector of the mth word, that is, the generated mth word; Indicates taking the word with the highest generation probability; P(y m =w|y m-1 , v t ) represents the conditional probability of generating the mth word as w; w represents the candidate word.

[0047] The beneficial effects of the technical solution of the present invention are:

[0048] 1. By combining the semantic representation and contextual information of the words in the input sentence, a third-order semantic tensor element is generated, and the representation of each word embedding vector is dynamically adjusted. The grammatical structure position, syntactic dependency and semantic feature index of the word in the sentence are integrated. The weight function is designed using sine function, cosine function and exponential function, so that the high-dimensional tensor can accurately express the grammatical structure and semantic information of the input sentence, enhance the ability to understand complex grammatical structures and semantic relationships, and improve the adaptability to changes in sentence context.

[0049] 2. Perform nonlinear transformation on the elements of the third-order semantic tensor and reduce sparsity through logarithmic function compression, making the representation of the tensor in high-dimensional space more compact, thereby improving the storage and processing efficiency of semantic information, avoiding the dominant influence of extreme values ​​on subsequent processing, and enhancing the stability and reliability of semantic tensor processing. Combined with the sign function retention technology, the positive and negative directionality of the tensor structure is maintained.

[0050] 3. Through tensor decomposition and weight allocation, it can accurately capture the changes in semantic features in three dimensions: word grammatical structure position, syntactic dependency, and semantic feature index, to avoid feature loss. The dynamic weight mechanism adjusts the feature distribution in real time according to the specific context, improves the generalization ability when processing diverse sentences, and low-rank decomposition reduces the computational complexity, ensuring that the feature extraction process strikes a balance between efficiency and precision.

[0051] 4. The global mapping matrix effectively captures the global semantic relationship between the input sentence and the target language, ensuring the accuracy of semantic transfer. The local correction matrix supplements the deficiencies of the global mapping, and refines the expression of the specific grammatical and syntactic features of the target language. By using the recursive enhancement mechanism and introducing the previously generated semantic vectors, the mapping process has higher continuity and stability, avoiding the loss or deviation of semantic information.

[0052] 5. In the word-by-word generation, the context state and semantic information work together to ensure the grammatical logic and semantic coherence of the generated sentences. The normalization function and mapping matrix ensure the rationality of the distribution of vocabulary generation and optimize the efficiency of the generation process. The generated words are selected by maximizing the probability, avoiding the addition of unreasonable words, and improving the accuracy of the generated results. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of a multilingual machine translation method based on deep learning described in the present invention. DETAILED DESCRIPTION

[0054] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all 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.

[0055] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0056] The following is a detailed description of a multi-language machine translation method based on deep learning provided by the present invention in conjunction with the accompanying drawings.

[0057] See attached Figure 1 , which shows a flow chart of a multilingual machine translation method based on deep learning provided by an embodiment of the present invention, the method comprising the following steps:

[0058] S1. Obtain and preprocess the input sentence, obtain the word embedding vector sequence through the pre-training method, and use the semantic tensor generation algorithm to process the word embedding vector to generate the third-order semantic tensor elements;

[0059] The input sentence is obtained and preprocessed, such as word segmentation, standardization, and dictionary mapping. The preprocessing method is a technical means well known to those skilled in the art and will not be described in detail here. The word embedding vector sequence x = [x1, x2, ..., x p , ..., x n ], n represents the number of word embedding vectors in the input sentence, x p represents the p-th word embedding vector, p∈[1,n]; the pre-training method is a technical means well known to those skilled in the art and will not be described in detail here;

[0060] Use the semantic tensor generation algorithm to process the word embedding vector to generate the third-order semantic tensor elements;

[0061] The semantic tensor generation algorithm constructs a high-dimensional tensor that can reflect grammatical structure, syntactic dependency and semantic features by combining the semantic representation of words in the input sentence with context information;

[0062] In order to enhance the contextual adaptability of word embedding vectors, the semantic tensor generation algorithm processes each word embedding vector in the input word embedding vector sequence one by one. Each word embedding vector is combined with the word embedding vector of the context in the same sentence to form a dynamically adjusted representation. During the adjustment process, the basic semantic features of the word are captured by the basic word vector matrix as the initial semantic representation, as the base layer, and the context information is superimposed on the base layer by weighted summation of the context adjustment matrix. The weight of the context information is controlled by the context weight parameter, which is used to adjust the semantic influence strength of the context.

[0063] After the context adjustment is completed, the word embedding vector is further expanded into a high-dimensional tensor representation, which not only contains the semantic information of the word, but also integrates the grammatical structure position, syntactic dependency and semantic features of the word in the sentence. Specifically, the weight adjustment mechanism uses three functions: sine function, cosine function and exponential function; the sine function is used to capture the periodic change characteristics of the word in the sentence and to normalize the length of the input sentence so that the generated features can reflect the relative importance of the position regardless of the length of the input sentence; the cosine function complements the effect of the sine function by introducing an alternating change feature in the representation of grammatical relations, reflecting the complex dependency characteristics of the words in the sentence; the exponential function is used to attenuate the influence of the high-dimensional feature index of the word to ensure that the high index value will not introduce too large errors in the high-dimensional tensor;

[0064] The generation formula of the third-order semantic tensor element is:

[0065]

[0066] Among them, T ijkRepresents the third-order semantic tensor element, reflecting the grammatical structure position information i, syntactic dependency feature j, and semantic features under the semantic feature index k; represents the sum of the semantic contributions of all word embedding vectors (from p=1 to p=n) in the input sentence; W represents a dynamic embedding function, which is used to dynamically adjust the word embedding vector; f represents the basic word vector matrix, which is used to capture the basic semantic features of words as the initial semantic representation and is obtained through word embedding model training; p represents the p-th word embedding vector; α represents the context weight parameter, which is used to adjust the semantic influence of the context. It is adjusted through experiments and its value range is [0.1, 1.0]; ∑ q≠p W represents the sum of the semantic contributions of the context words to the p-th word embedding vector in the input sentence; c represents the context adjustment matrix, which is used to capture the influence of the context words on the p-th word embedding vector; x q represents the qth word embedding vector; g(i, j, k) represents the weight function, which weights the tensor components of different dimensions and generates the third-order tensor weight by combining the position information, syntactic dependency and semantic features. The calculation formula is:

[0067]

[0068] in, Represents the relative position i of the combined word in the sentence, and uses a sine function to periodically reflect the differences in word positions; Represents the syntactic dependency feature j of the combined word, and uses the cosine function to provide the weight of the grammatical relationship; The representation exponential function weights the semantic feature index k, which gradually decreases as the semantic feature index increases, to prevent excessive influence of high-dimensional information on tensor representation;

[0069] The third-order semantic tensor elements not only contain the semantic relations of sentences, but also capture the syntactic dependencies and structural features, providing a comprehensive high-dimensional representation of the input sentences, which is a complete concentration of semantic information.

[0070] S2, by performing nonlinear transformation on the third-order semantic tensor elements, obtaining transformed tensor elements, using a dynamic weight-controlled multi-dimensional semantic feature extraction algorithm to extract features from the transformed tensor elements, and obtaining a core semantic vector;

[0071] In order to solve the problem of excessive sparsity in high-dimensional tensors, enhance stability and retain semantic information, a nonlinear transformation is performed on the third-order semantic tensor elements to obtain the transformed tensor elements. The nonlinear transformation adopts logarithmic function compression and sign function retention technology. The logarithmic function is used to dynamically compress the absolute value of the third-order semantic tensor elements, compress the maximum value to prevent the dominant influence on subsequent processing, stretch the minimum value, and reduce the sparsity in the high-dimensional tensor; the sign function is used to retain the positive and negative sign information of the third-order semantic tensor elements, so as to ensure that the structure of the transformed tensor elements does not lose the semantic direction;

[0072] The transformed tensor elements are calculated as:

[0073] T′ ijk =log2(1+|r ijk |)·sgn(T ijk )

[0074] Among them, T′ ijk Represents the transformed tensor element; log2(1+|T ijk |) represents a logarithmic function, which is used to dynamically compress the absolute values ​​of the third-order semantic tensor elements, compress the maximum value to prevent the dominant influence on subsequent processing, stretch the minimum value, and reduce the sparsity in the high-dimensional tensor; sgn(T ijk ) represents a sign function, which is used to preserve the sign information of the third-order semantic tensor elements, thereby ensuring that the structure of the transformed tensor elements does not lose the semantic direction;

[0075] Use the dynamic weight control multi-dimensional semantic feature extraction algorithm to extract features from the transformed tensor elements and obtain the core semantic vector;

[0076] The dynamic weight control multi-dimensional semantic feature extraction algorithm achieves accurate extraction of transformed tensor elements through tensor decomposition, dynamic weight control, feature importance evaluation, and semantic vector combination to obtain semantic features;

[0077] The tensor decomposition is to represent the semantic tensor as a combination of several low-rank feature matrices. The feature matrices correspond to the projections of the semantic features in different dimensions. Through tensor decomposition, the original high-dimensional semantic tensor is simplified into several easy-to-handle low-rank features. The number of feature matrices is determined by a predefined rank. The size of the rank affects the fineness of the feature decomposition and controls the computational complexity. It can be set according to the specific implementation scenario and is not limited here.

[0078] In order to improve the accuracy of feature extraction, the dynamic weight control multi-dimensional semantic feature extraction algorithm assigns a dynamic weight to each feature during the tensor decomposition process, and introduces feature importance evaluation, which is achieved by calculating the change amount of the tensor in three dimensions point by point, calculating the change rate in the three directions of word position information, syntactic dependency, and semantic feature index, and further amplifying the influence of significant features through square operations, while making the partial derivative value always non-negative; using exponential functions to ensure that the dynamic weight control function value is in (0, 1], which is convenient for normalization processing;

[0079] The tensor decomposition and weight allocation formulas are as follows:

[0080]

[0081] Among them, T′ ijk Represents the transformed tensor element; Represents the summation operation of tensor decomposition; r represents the rank of tensor decomposition, that is, the number of features after tensor decomposition, which is used to determine the fineness of feature decomposition, affecting the computational complexity and the richness of semantic representation. It can be set according to the specific implementation scenario and is not limited here; σ l A represents the eigenvalue of the lth decomposition feature, reflecting the importance of the feature; l , B l , C l represents the feature matrix after the lth decomposition, that is, the projection of the tensor in the three dimensions of word position information, syntactic dependency features, and semantic feature index, capturing the feature patterns in different dimensions; Represents the outer product operation of a tensor; λ l (i, j, k) represents the dynamic weight control function of the i-th decomposition feature, which controls the weight distribution of features at different positions and adapts to the dynamically changing feature importance. The calculation formula is:

[0082]

[0083] Wherein, β represents the smoothing coefficient, which controls the sensitivity of weight regulation and can be set according to the specific implementation scenario and is not limited here; exp represents the exponential function, which ensures that the value of the dynamic weight regulation function is in (0, 1], which is convenient for normalization processing; It represents the feature importance evaluation function, which is used to measure the importance of the feature at the (i, j, k) position. The square operation is used to amplify the influence of the significant feature and make the partial derivative always non-negative. It represents the sensitivity of semantic features to changes in word position information, which is calculated by the gradient of the transformed tensor elements; It represents the sensitivity of semantic features to changes in syntactic dependency features, and is calculated by the gradient of the transformed tensor elements; It represents the sensitivity of the semantic feature to the change of the semantic feature index, which is calculated by the gradient of the transformed tensor element;

[0084] After completing tensor decomposition and weight assignment, all low-rank feature matrices are combined according to the weights to generate the core semantic vector. The formula is as follows:

[0085] u=[λ1·σ1·A1, λ1·σ1·B1, λ1·σ1·C1,...,λ l ·σ l ·A l ,λ l ·σ l ·B l ,λ l ·σ l ·C l , ..., λ r ·σ r Ar, λ r ·σ r ·B r ,λ r ·σ r ·C r ]

[0086] Among them, u represents the core semantic vector, that is, the combination of decomposed semantic features, which provides a low-dimensional representation of semantic features for target language mapping;

[0087] S3, processing the core semantic vector through the target language dynamic mapping mechanism to dynamically generate the target language semantic vector;

[0088] The target language dynamic mapping mechanism is used to globally map, locally modify and recursively enhance the core semantic vector, and dynamically generate the target language semantic vector.

[0089] The target language dynamic mapping mechanism linearly transforms the core semantic vector through a global mapping matrix, wherein the global mapping matrix represents the global semantic relationship and is used to capture the semantic alignment pattern between the input sentence and the target language;

[0090] Apply local correction to the result of global mapping. The local correction step is based on a local correction matrix. The local correction matrix makes fine-grained adjustments to global features through element-by-element operations, which enhances the flexibility of mapping and enables the target language semantic vector to reflect the specific grammatical and syntactic characteristics of the target language.

[0091] In order to make the semantic mapping stable and take into account continuity, a recursive enhancement mechanism is introduced to combine the target language semantic vector generated in the previous round with the result of the current step to form a new feature representation;

[0092] By applying the logarithmic transformation method, the value range of the target language semantic vector can be smoothed to avoid the influence of extremely large or small values ​​on the overall quality of the target language semantic vector. The modulus of the global semantic mapping item is further calculated as a normalization factor to normalize the target language semantic vector to avoid numerical divergence in the recursive process and maintain the calculation stability of the target language semantic vector in subsequent steps.

[0093] The formula for generating the target language semantic vector is:

[0094]

[0095] Among them, v t represents the target language semantic vector generated at time t; The modulus length of the global semantic mapping item is used as a normalization factor to normalize the target language semantic vector to avoid numerical divergence in the recursive process; represents a logarithmic function, which smoothes the change of the target language semantic vector to avoid the value being too large or too small; u represents the core semantic vector, that is, the combination of decomposed semantic features, which provides a low-dimensional representation of the semantic features for target language mapping; represents the global mapping matrix, which captures the mapping relationship between the global features in the core semantic vector and the target semantics. It is obtained by training the global attention mechanism of the deep learning model. It is a technical means well known to those skilled in the art and will not be described in detail here. S represents the local correction matrix, which is used to enhance the local characteristics of the input core semantics on the basis of the global mapping to make the mapping accurate. It is obtained by optimizing the loss function during the deep learning model training. It is a technical means well known to those skilled in the art and will not be described in detail here. u 2 represents the square of the core semantic vector, introduces nonlinear features, and improves the complexity of the target language semantic vector; γ represents the recursive coefficient, which reflects the degree of dependence of the current iteration on the result of the previous iteration. It is used to adjust the intensity of the recursive process and ensure the continuity of semantic generation. It can be set according to the specific implementation scenario and is not limited here; R(v t-1 ) represents a recursive feature enhancement function, which uses the target language semantic vector generated in the previous round to enhance the expression ability of the current target language semantic vector. The calculation formula is:

[0096]

[0097] Among them, v t-1 represents the target language semantic vector generated at time t-1;

[0098] After each round of recursive generation of the target language semantic vector, the difference between the currently generated target language semantic vector and the target language semantic vector generated in the previous round will be evaluated. If the difference is lower than the preset convergence threshold, it is considered that the recursive process has reached convergence. Otherwise, the next round of mapping will continue until the convergence condition is met.

[0099] The recursive termination condition formula is as follows:

[0100] ||v t -v t-1 ||<∈

[0101] Among them, ||v t -v t-1 || represents the Euclidean distance between the target language semantic vector generated at time t and the target language semantic vector generated at time t-1, which is used to measure the degree of change of the target language semantic vector and serves as a criterion for whether the recursion ends; ∈ represents the convergence threshold, which is used to control the recursive termination condition to prevent excessive iterations. It can be set according to the specific implementation scenario and is not limited here;

[0102] S4, based on the target language semantic vector, predict the next word generation method complete sentence through sentence translation generation algorithm;

[0103] Based on the target language semantic vector, the sentence translation generation algorithm gradually predicts the generation method of the next word to complete the sentence;

[0104] The sentence translation generation algorithm predicts the current word using the input target language semantic vector and the generated previous word, and gradually expands the generated word sequence until the sentence is completed;

[0105] In the process of generating words at each step, the input target language semantic vector is updated according to the current context state. The state update is achieved by combining the target language semantic vector with the last word generated. The target language semantic vector provides global semantic information about the input sentence, while the last word generated reflects the local grammatical and semantic state. The input vector combination is further projected into the target language vocabulary space through the mapping matrix, and the result of the linear mapping is normalized into a probability distribution through the normalization function, so that the sum of the generation probabilities of all words is 1. The word with the largest probability is selected as the generated word of the current step, which ensures the accuracy of the generated result and reduces the possibility of unreasonable words being selected.

[0106] The calculation formula for the conditional probability of generating a word is:

[0107] P(y m |y m-1 , v t )=softmax(W m ·[ym-1 , v t ])

[0108] Among them, P(y m |y m-1 , v t ) represents the conditional probability of generating the mth word; y m represents the embedding vector of the mth word; softmax represents the normalization function, which normalizes the result of the linear mapping into a probability distribution so that the sum of the generation probabilities of all words is 1; W m Represents the mapping matrix, which is used to project the input vector combination into the target language vocabulary space; [y m-1 , v t ] represents the input vector combination, which consists of two parts: the embedding vector of the previous word and the target language semantic vector;

[0109] The formula for selecting generated words is:

[0110]

[0111] Among them, y m Represents the embedding vector of the mth word, that is, the generated mth word; Indicates taking the word with the highest generation probability; P(y m =w|y m-1 , v i ) represents the conditional probability of generating the mth word as w; w represents a candidate word, which belongs to the target language vocabulary;

[0112] The currently generated word will be used as input to enter the next generation loop, and the generation probability of the next word will continue to be calculated based on the new context state. Each iteration is based on the current semantic information and context state to ensure the consistency of the sentence's grammatical structure and semantic logic.

[0113] In summary, a multilingual machine translation method based on deep learning has been completed.

[0114] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0115] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0116] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A multilingual machine translation method based on deep learning, characterized in that: The following steps are involved: S1: Obtain and preprocess the input sentence, obtain the word embedding vector sequence through the pre-training method, and use the semantic tensor generation algorithm to process the word embedding vector to generate the third-order semantic tensor elements; S2: Perform nonlinear transformation on the third-order semantic tensor elements to obtain transformed tensor elements, and use the dynamic weight control multi-dimensional semantic feature extraction algorithm to extract features from the transformed tensor elements to obtain the core semantic vector; S3: Process the core semantic vector through the target language dynamic mapping mechanism to dynamically generate the target language semantic vector; S4: Based on the target language semantic vector, the sentence translation generation algorithm is used to predict the generation method of the next word to complete the sentence.

2. The multilingual machine translation method based on deep learning according to claim 1, characterized in that: The S1 specifically includes: In the implementation of the semantic tensor generation algorithm, each word embedding vector in the input word embedding vector sequence is processed to form a dynamically adjusted representation; during the adjustment process, the basic semantic features of the word are captured through the basic word vector matrix, which is used as the base layer, and the context information is superimposed on the base layer by weighted summation of the context adjustment matrix; after the context adjustment is completed, the word embedding vector is expanded to a high-dimensional tensor representation to generate a third-order semantic tensor element: Among them, T ijk Represents the third-order semantic tensor element, reflecting the grammatical structure position information i, syntactic dependency feature j, and semantic features under the semantic feature index k; represents the sum of the semantic contributions of all word embedding vectors in the input sentence; represents a dynamic embedding function; W f represents the basic word vector matrix; x p represents the p-th word embedding vector; α represents the context weight parameter; ∑ q≠p W represents the sum of the semantic contributions of the context words to the p-th word embedding vector in the input sentence; c represents the context adjustment matrix; x q represents the qth word embedding vector; g(i, j, k) represents the weight function.

3. The multilingual machine translation method based on deep learning according to claim 1, characterized in that: The S2 specifically includes: The nonlinear transformation uses logarithmic function compression and sign function retention technology to dynamically compress the absolute value of the third-order semantic tensor elements and retain the positive and negative sign information of the third-order semantic tensor elements to obtain the transformed tensor elements. The calculation formula is: T′ ijk =log2(1+|T ijk |)·sgn(T ijk ) Among them, T′ ijk Represents the transformed tensor element; log2(1+|T ijk |) represents the logarithmic function; sgn(I ijk ) represents a sign function.

4. The multilingual machine translation method based on deep learning according to claim 1, characterized in that: The S2 specifically includes: In the implementation process of the dynamic weight-controlled multi-dimensional semantic feature extraction algorithm, the transformed tensor elements are extracted through tensor decomposition, dynamic weight control, feature importance evaluation, and semantic vector combination to obtain semantic features.

5. The multilingual machine translation method based on deep learning according to claim 4, characterized in that: The S2 specifically includes: In the process of implementing the dynamic weight-controlled multi-dimensional semantic feature extraction algorithm, dynamic weights are assigned to each feature in the tensor decomposition process, and feature importance evaluation is introduced. By calculating the change amount of the tensor in three dimensions, the change rate in the three directions of word grammatical structure position information, syntactic dependency, and semantic feature index is calculated. The tensor decomposition and weight allocation formulas are as follows: Among them, T′ ijk Represents the transformed tensor element; represents the summation operation of tensor decomposition; r represents the rank of tensor decomposition, that is, the number of features after tensor decomposition; σ l A represents the eigenvalue of the lth decomposition feature; l , B l , C l represents the feature matrix after the lth decomposition, that is, the projection of the tensor in the three dimensions of word position information, syntactic dependency features, and semantic feature index; Represents the outer product operation of a tensor; λ l (i, j, k) represents the dynamic weight control function of the lth decomposition feature.

6. The multilingual machine translation method based on deep learning according to claim 5, characterized in that: The S2 specifically includes: After completing tensor decomposition and weight assignment, the low-rank feature matrix obtained by tensor decomposition is combined according to the weights to generate the core semantic vector. The formula is as follows: u=[λ1·σ1·A1,λ1·σ1·B1,λ1·σ1·C1,...,λ l ·s l ·A l ,l l ·s l ·B l ,l l ·s l ·C l ,...,l r ·s r ·A r ,l r ·s r ·B r ,l r ·s r ·C r ] Among them, u represents the core semantic vector, that is, the combination of semantic features after decomposition.

7. The multilingual machine translation method based on deep learning according to claim 1, characterized in that: The S3 specifically includes: The core semantic vector is globally mapped, locally modified, and recursively enhanced through the target language dynamic mapping mechanism to dynamically generate the target language semantic vector. The generation formula is: Among them, v t represents the target language semantic vector generated at time t; Represents the modulus of the global semantic mapping item, as a normalization factor; represents the logarithmic function; u represents the core semantic vector, that is, the combination of semantic features after decomposition; represents the global mapping matrix; s represents the local correction matrix; γ represents the recursive coefficient; R(v t-1 ) represents the recursive feature enhancement function.

8. The multilingual machine translation method based on deep learning according to claim 7, characterized in that: The S3 specifically includes: After each round of recursive generation of the target language semantic vector, the difference between the currently generated target language semantic vector and the target language semantic vector generated in the previous round is evaluated. When the difference is lower than the preset convergence threshold, it is considered that the recursive process has reached convergence. When the difference is higher than and equal to the preset convergence threshold, the next round of mapping will continue until the convergence condition is met. The recursive termination condition formula is: ||v t -v t-1 ||<∈ Among them, ||v t -v t-1 || represents the Euclidean distance between the target language semantic vector generated at time t and the target language semantic vector generated at time t-1; ∈ represents the convergence threshold.

9. The multilingual machine translation method based on deep learning according to claim 1, characterized in that: The S4 specifically includes: In the implementation process of the sentence translation generation algorithm, the input target language semantic vector and the previous generated word are used to predict the current word; in the process of generating words at each step, the input target language semantic vector is updated according to the current context state; the input vector combination is projected into the target language vocabulary space through the mapping matrix, and the result of the linear mapping is normalized into a probability distribution through the normalization function, and the word with the largest probability is selected as the generated word of the current step.

10. The multilingual machine translation method based on deep learning according to claim 9, characterized in that: The S4 specifically includes: The calculation formula for the conditional probability of generating a word is: P(and m |and m-1 ,v t )=softmax(W m ·[and m-1 ,v t ]) Among them, P(y m |y m-1 , v t ) represents the conditional probability of generating the mth word; y m represents the embedding vector of the mth word; softmax represents the normalization function; W m Represents the mapping matrix; [y m-1 , v t ] represents the input vector combination; The formula for selecting generated words is: Among them, y m Represents the embedding vector of the mth word, that is, the generated mth word; Indicates taking the word with the highest generation probability; P(y m =w|y m-1 , v t ) represents the conditional probability of generating the mth word as w; w represents the candidate word.

Images