Multi-factor evolutionary algorithm optimization neural network model language recognition method and application

Abstract

The application discloses a language recognition method for optimizing a neural network model by using a multi-factor evolution algorithm and application, and belongs to the technical field of language recognition, and comprises the following steps: S1, embedding an external knowledge base in a data set to generate mixed data for data enhancement; S2, performing word embedding and regularization processing on original data by using a pre-trained word vector model, and standardizing a processing result to obtain a standard training data set; S3, optimizing hyperparameters of a neural network model by using a multi-factor evolution algorithm; S4, detecting and classifying an intention by using a multi-level classification neural network model; and S5, comparing model classification output results with performance measurement standards, and finally realizing language recognition. Through the above manner, the application is based on an optimization strategy of an evolution algorithm, searches an entire network structure to obtain optimal hyperparameters, and thus the reliability, stability and accuracy of the model are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of language recognition, in particular to a language recognition method based on a multi-factor evolutionary algorithm optimized neural network model and application thereof. BACKGROUND

[0002] With the rapid development of artificial intelligence, language recognition technology has gradually become an important research field. Intelligent language assistants are widely used in language interactive technologies, such as smart homes, intelligent customer service, intelligent language search, etc. Due to the existence of various factors in language information, such as diversification, dialect differences, and complexity of intent, various challenges are faced in the research process of language recognition. Therefore, developing a remote service assistant system for language recognition technology is crucial to improve the accuracy and reliability of language recognition.

[0003] Existing language recognition and intent classification technologies have high misrecognition rates, slow response speeds, and difficulty in covering multiple languages, multiple accents, and multiple dialects when implementing complex scenarios. Even worse, they may not fully understand the user's intent or recognize the wrong intent. For example, the semantic enhancement Gaussian mixture method (SEG) has a Gaussian mixture distribution speech embedding, which can inject dynamic class semantic data into the Gaussian mean value, further promoting downstream outlier recognition. However, this method has high dependency and Gaussian distribution compliance restrictions, making its practical performance poor.

[0004] Therefore, the present application designs a language recognition method based on a multi-factor evolutionary algorithm optimized neural network model and application thereof to solve the above problems. SUMMARY

[0005] In view of the above-mentioned shortcomings of the prior art, the present application provides a language recognition method based on a multi-factor evolutionary algorithm optimized neural network model and application thereof.

[0006] To achieve the above purpose, the present application realizes the following technical solutions:

[0007] The language recognition method based on a multi-factor evolutionary algorithm optimized neural network model comprises the following steps:

[0008] S1, embedding an external knowledge base in the data set to generate mixed data for data augmentation;

[0009] S2, using a pre-trained word vector model to perform word embedding and regularization processing on the original data, and standardizing the processing result to obtain a standard training data set;

[0010] S3, using a multi-factor evolutionary algorithm to optimize the hyperparameters of the neural network model;

[0011] S4, detecting and classifying the intention by using a multi-level classification neural network model;

[0012] S5, comparing the model classification output result with the performance measurement standard, and finally realizing language recognition.

[0013] Further, in step S1, the specific steps of data enhancement include:

[0014] Define batch xi For a specific batch sample for sample xi, batch yi is the label corresponding to the batch sample; λ is the mixing coefficient calculated by the beta distribution of parameters α and β, defined as:

[0015] λ=Beta(α,β) (1)

[0016] mixed_batch x =λ*batch x1 +(1-λ)*batch x2 (2)

[0017] mixed_batch y =λ*batch y1 +(1-λ)*batch y2 (3)

[0018] Where Beta(·) is the beta distribution, mixed_batch x is the mixed batch sample, mixed_batch y is the label corresponding to the mixed sample.

[0019] Further, in step S2, the word embedding step specifically includes:

[0020] 1) Using the global co-occurrence matrix, the GloVe model can generate word vector representation V={v1,v2...v n}, where v i represents the vector of the i-th word in the vocabulary; in addition, A, S, R and H represent the index sets of antonyms, synonyms and related words respectively; the first vector space y is adjusted by the objective function and set constraint to obtain a new vector space V′, where V′={v′1,v′2...v′ n};

[0021] 2) Synonym enrichment, by improving the cosine similarity between vectors, introducing synonyms close to each other in the vector space, defined as:

[0022]

[0023] where SE denotes the enriched set of synonym pairs, δ = 1, and τ(x) = max(0, x);

[0024] 3) Antonym enrichment, which makes antonyms far away from each other by reducing the cosine similarity between vectors to 0, is defined as:

[0025]

[0026] where AE denotes the enriched set of antonym pairs, δ = 1, and τ(x) = max(0, x);

[0027] 4) Related word enrichment, which makes the distance between words with similar meanings in the vector space as small as possible, and the objective function increases the cosine similarity between related words to 0 or greater than 0, ensuring that there is a certain degree of distinction between related words and synonyms, is defined as:

[0028]

[0029] where RE denotes the enriched set of related word pairs, δ is fixed at 0, and τ(x) = max(0, x).

[0030] Further, it also includes adjusting the word vector using constraints, including the following steps:

[0031] The transformed vector space is bent to the original vector space, and the objective function is used to maintain the cosine similarity between adjacent values, and the initial word vector is greater than or equal to the cosine similarity between them, which is defined as:

[0032]

[0033] where M denotes the number of words in the vocabulary, N(i) denotes the i-th adjacent group, N(i) involves words in V, and the cosine similarity between the i-th word is higher than 0.8;

[0034] The objective function is the linear integral of the enrichment of synonyms, antonyms and related words and the cosine similarity, which is defined as:

[0035] E(V, V') = SE(V') + AE(V') + RE(V') + VSP(V, V') (8)

[0036] The objective function is minimized by the stochastic gradient descent method of the transformed vector V' and the initial vector V.

[0037] Further, step S3 adopts a multi-factor evolution algorithm to optimize input entropy value, input weight value, number of hidden layers, number of hidden layer nodes and learning rate of the neural network model; other individuals except the current optimal individual are adjusted according to the position of the current optimal individual (pbest) and the global optimal position (gbest) of the population; the specific steps include:

[0038] 1) generate N initial individuals, each individual is composed of a position vector and a velocity vector; for example, for the i-th individual in the population, its position vector is wherein represents the input entropy value, represents the input weight value, represents the number of hidden layers, represents the number of hidden layer nodes, represents the learning rate; its velocity vector is wherein and are random numbers between [-2, 2];

[0039] 2) the position of the i-th individual in the t-th generation is represented by the position of the individual is adjusted by the velocity , and is defined as:

[0040]

[0041]

[0042] wherein, β represents the visibility coefficient a1 and a2r p and r g represent the Euclidean distance between and pbest and and gbest respectively, g ∈ [0, 1] is a variable coefficient;

[0043] 3) iterative adjustment of g;

[0044]

[0045] wherein, g max = 1, g min = 0, iter is the current iteration number, iter max is the maximum iteration number;

[0046] 4) calculation of the current optimal position pbest;

[0047]

[0048] wherein, fitness is the training error of the multi-stage neural network;

[0049] 5) update of individual speed;

[0050]

[0051] where d represents the attraction coefficient, r e [-1, 1];

[0052] 6) the speed update calculation formula of fitness similar individual is:

[0053]

[0054] where, is the speed of individual i in the tth generation, is the position of individual i in the tth generation, is the position of individual j in the tth generation; a2 and β respectively represent the attraction constant and the visibility coefficient, r mf represents and the Euclidean distance between them, fl represents the random walking coefficient, r e [-1, 1].

[0055] Further, after the speed update of the fitness similar individual, the positions of individual i and individual j and are similar, which leads to falling into local optimal solution. In order to enable the individual to search the entire solution space more widely and have a greater opportunity to find the global optimal solution, the diversity and exploration ability of the individual need to be increased, and the randomness operation of the crossover operator is introduced;

[0056] The crossover operator calculation process includes the following specific steps:

[0057] 1) the crossover operation of individual i and individual j with similar fitness is as follows:

[0058]

[0059] where L e [0, 1];

[0060] 2) add random disturbance to the generated individual i in the t+1th generation to improve the local search ability, that is:

[0061]

[0062] where σ represents the standard deviation, N n represents uniform distribution.

[0063] Further, step S4 specifically includes:

[0064] 1) initialize the score value of the selected feature and the corresponding weight value;

[0065]

[0066] where E i represents the input entropy value, W i represents the corresponding weight value;

[0067] 2) multiply the input with an arbitrarily selected weight vector and sum up completely;

[0068]

[0069] where R represents the weighted sum value;

[0070] 3) estimate the activation function (AF) defined as:

[0071]

[0072] where f represents the sigmoid activation function;

[0073] 4) obtain the hidden layer measurement defined as:

[0074]

[0075] where Y i represents the measurement, A i represents the bias value, W i represents the weight between the input and hidden layer, G i represents the AF applied changed value.

[0076] Further, it also includes error signal normalization, the steps are as follows:

[0077] 1) re-execute the specific steps of step S4 on all layers of the neural network model, estimate the result unit by adding the weight of each input signal to obtain the resulting layer neuron value;

[0078] Ou i = A i +∑P i W i (13)

[0079] where P i represents the value of the layer resulting in an output of 1, W i represents the weight of the hidden layer, Ou i represents the output unit;

[0080] 2) calculate the deviation between the output result and the target value to produce an error signal E r ;

[0081] E r = Ta i -Ou i (14)

[0082] Among them, Ta i Indicates the target output, Ou i Indicates the classification of current output;

[0083] 3) Weighting is applied to the target value in the error signal E. r The relative error δ is calculated based on this. i ;

[0084] δ i =E r [f(Ou i (15)

[0085] 4) Perform further weighted correction, defined as:

[0086] wc i =βδ i (E i (16)

[0087] Among them, wc i For the correction value, β represents the momentum term, E i Denotes the input vector, δ i This represents the error distributed throughout the network.

[0088] Furthermore, step S5 specifically includes:

[0089] This invention also discloses the application of a language recognition method for optimizing neural network models using a multi-factor evolutionary algorithm in the optimization of hyperparameters of neural network models.

[0090] The model's prediction results are quantified to differentiate from the target results, and the matching between the model's classification output and the real labels in the test set is determined. Finally, intent recognition is achieved through the model's performance metrics, precision, and recall.

[0091] Beneficial effects

[0092] This invention is based on the optimization strategy of evolutionary algorithms. It obtains the optimal hyperparameters by searching the entire network structure, thereby improving the reliability, stability and accuracy of the model.

[0093] This invention is based on a highly efficient and accurate deep learning model, which can accurately classify complex and diverse input data sources and has excellent characteristics such as high efficiency, accuracy and stability; it is conducive to improving the language recognition and understanding capabilities of intelligent language assistants and further promoting the development and application of artificial intelligence technology.

[0094] This invention can cover specific problems in fields such as speech recognition and natural language processing, and its application can improve people's life and work efficiency, bring a more intelligent experience, and has better practicality. Attached Figure Description

[0095] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0096] Figure 1 This is a schematic diagram illustrating the principle of the language recognition method for optimizing a neural network model using a multi-factor evolutionary algorithm, as described in this invention. Detailed Implementation

[0097] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0098] The present invention will be further described below with reference to embodiments.

[0099] Example 1

[0100] Please refer to the instruction manual appendix. Figure 1 A language recognition method for optimizing neural network models using a multi-factor evolutionary algorithm includes the following steps:

[0101] S1. Embed external knowledge bases into the dataset to generate hybrid data for data augmentation;

[0102] In step S1, the specific steps of data augmentation include:

[0103] Define batch xi For a specific batch of samples xi, batch yi Here, α represents the label corresponding to this batch of samples; λ is the mixing coefficient calculated from the beta distribution of parameters α and β, defined as:

[0104] λ=Beta(α,β) (1)

[0105] mixed_batch x =λ*batch x1 +(1-λ)*batch x2 (2)

[0106] mixed_batchy = λ * batch y1 + (1 - λ) * batch y2 (3)

[0107] where Beta(·) is the Beta distribution, mixed_batch x is the mixed batch sample, mixed_batch y is the label corresponding to the mixed sample;

[0108] S2, word embedding and regularization processing are performed on the original data by using a pre-trained word vector model, and the processing result is standardized to obtain a standard training data set;

[0109] In step S2, the word embedding step specifically includes:

[0110] 1) By using the global co-occurrence matrix, the GloVe model can generate word vector representation V = {v1, v2...v n}, where v i represents the vector of the i-th word in the vocabulary; in addition, A, S, R and H represent the index sets of antonyms, synonyms and related words respectively; for example: if (i, j) is a component of S, then the i-th word and the j-th word in the vocabulary are synonyms; the first vector space y is adjusted by the objective function and set constraints to obtain a new vector space V', where V' = {v'1, v'2...v' n};

[0111] 2) Synonym enrichment, by improving the cosine similarity between vectors, synonyms close to each other are introduced in the vector space, defined as:

[0112]

[0113] where SE represents the set of enriched synonym pairs, δ = 1, τ(x) = max(0, x);

[0114] 3) Antonym enrichment, by reducing the cosine similarity between vectors to 0 to make antonyms far away from each other, defined as:

[0115]

[0116] where AE represents the set of enriched antonym pairs, δ = 1, τ(x) = max(0, x);

[0117] 4) related word enrichment, the distance between words with similar meanings in the vector space is as small as possible, and the objective function increases the cosine similarity between related words to 0 or greater than 0, ensuring that related words have a certain degree of differentiation with synonyms, which is defined as:

[0118]

[0119] Where RE represents the set of enriched related word pairs, δ is fixed at 0, and τ(x) = max(0, x);

[0120] Preferably, it also includes adjusting the word vector using constraints, including the following steps:

[0121] Bending the transformed vector space to the original vector space, using the objective function to maintain the cosine similarity between adjacent values, the initial word vector is greater than or equal to the cosine similarity between them, which is defined as:

[0122]

[0123] Where M represents the number of words in the vocabulary, N(i) represents the i-th adjacent group, N(i) involves words in V, and the cosine similarity between the i-th word is higher than 0.8;

[0124] The objective function is the linear integral of the enrichment and cosine similarity of synonyms, antonyms and related words, which is defined as:

[0125] E(V, V') = SE(V') + AE(V') + RE(V') + VSP(V, V') (8)

[0126] The objective function is minimized by the stochastic gradient descent method of transformed vector V' and initial vector V;

[0127] S3, using a multi-factor evolutionary algorithm (MFO) to optimize the hyperparameters of a neural network model (DLMNN);

[0128] The multi-factor evolutionary algorithm is used to optimize the input entropy value, input weight value, number of hidden layers, number of hidden layer nodes and learning rate of the neural network model. Except for the current optimal individual, other individuals are adjusted according to the position of the current optimal individual (pbest) and the global optimal position of the population (gbest). The specific steps include:

[0129] 1) Generate N initial individuals, each individual consists of a position vector and a velocity vector. For example, for the i-th individual in the population, the position vector is Where represents the input entropy value, represents the input weight value, represents the number of hidden layers, This represents the number of hidden layer nodes. This represents the learning rate. Its velocity vector is... in and All are random numbers in the range [-2, 2].

[0130] 2) with This represents the position of the i-th individual in generation t, measured by velocity. Adjusting an individual's position is defined as:

[0131]

[0132]

[0133] Where β represents the visibility coefficients a1 and a2r p and r g They represent Between and pbest and The Euclidean distance between gbest and gbest, where g∈[0,1] are variable coefficients;

[0134] 3) Iterative adjustment of g;

[0135]

[0136] Among them, g max =1, g min =0, iter is the current iteration number, iter max This represents the maximum number of iterations.

[0137] 4) Calculation of the current optimal position pbest;

[0138]

[0139] 5) Individual speed updates;

[0140]

[0141] Where d represents the attraction coefficient, and r∈[-1,1];

[0142] 6) The formula for calculating the rate update of individuals with similar fitness is:

[0143]

[0144] in, Let i be the velocity of individual i in generation t. Let i be the position of individual i in generation t. Let a be the position of individual j in generation t; a2 and β represent the attraction constant and visibility coefficient, respectively, and r mf express and Euclidean distance between and, fl represents a random walk coefficient, r∈[-1,1].

[0145] Preferably, after the speed update of the fitness similar individual, the positions of individual i and individual j are and Comparison similar, lead to fall into local optimal solution, in order to individual can more widely search the entire solution space, have greater opportunity to find global optimal solution, need to increase the diversity and exploration ability of individual, introduce the randomness operation of crossover operator;

[0146] The specific steps of the crossover operator calculation process are as follows:

[0147] 1) The crossover operation of individual i and individual j with similar fitness is as follows:

[0148]

[0149] Wherein, L∈[0,1];

[0150] 2) Add random disturbance to the generated individual i of the t+1 generation to improve the local search ability, that is:

[0151] γ′ n =γ n +σN n (0,1) (24)

[0152] Wherein, σ represents standard deviation, N n Indicates uniform distribution;

[0153] S4, using a multi-level classification neural network model to detect and classify the intent; Specifically, the following steps are included:

[0154] 1) Initialize the score value of the selected feature and its corresponding weight value;

[0155]

[0156] Wherein, E i Indicates input entropy value, W i Indicates the corresponding weight value;

[0157] 2) Multiply the input with any selected weighted vector and sum up completely;

[0158]

[0159] Wherein, R represents the weighted sum value;

[0160] 3) Estimate the activation function (AF), defined as:

[0161]

[0162] where f denotes a sigmoid activation function;

[0163] 4) Hidden layer measurement acquisition, defined as:

[0164] Y i = A i +∑G i W i (12)

[0165] where Y i denotes the measurement, A i denotes the bias value, W i denotes the weight between the input and hidden layers, G i denotes the value of the AF application change;

[0166] Preferably, it also comprises error signal normalization, with the following steps:

[0167] 1) Re-executing the specific steps of step S4 on all layers of the neural network model, by adding the weight of each input signal to estimate the resulting layer neuron value;

[0168] Ou i = A i +∑P i W i (13)

[0169] where P i denotes the value of the layer that results in an output of 1, W i denotes the weight of the hidden layer, and Ou i denotes the output unit;

[0170] 2) Calculating the deviation of the output result from the target value to produce an error signal E r ;

[0171] E r = Ta i -Ou i (14)

[0172] where Ta i denotes the target output, and Ou i denotes the classification current output;

[0173] 3) Weighting against the target value, calculating the relative error δ r based on the error signal E i ;

[0174] δ i = E r [f(Ou i )] (15)

[0175] 4) Perform further weight correction, defined as:

[0176] wc i = beta * delta i (E i ) (16)

[0177] where wc i is the correction value, beta represents the momentum term, E i represents the input vector, delta i represents the error distributed in the network.

[0178] S5, compare the model classification output result with the performance measurement standard, and finally realize language recognition; specifically comprising the following steps:

[0179] Quantify the difference between the model prediction result and the target result, and determine the matching condition of the classification output result of the model and the true label in the test set, and finally perform intent recognition through the performance indicators, precision and recall rate of the model.

[0180] The application is based on the optimization strategy of evolutionary algorithm, and the best hyperparameter is obtained by searching the entire network structure, so as to improve the reliability, stability and accuracy of the model.

[0181] The application is based on an efficient and accurate deep learning model, which can accurately classify complex and diversified input data sources, and has excellent characteristics such as high efficiency, accuracy and stability; it is beneficial to improve the language recognition and understanding ability of intelligent language assistants, and further promotes the development and application of artificial intelligence technology.

[0182] The application can cover specific problems in the fields of language recognition and natural language processing, and its application can improve people's work and life efficiency, bring more intelligent experience, and has better practicality.

[0183] The above embodiments are only used to illustrate the technical solutions of the application, but not limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements will not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A language recognition method of optimizing a neural network model by a multi-factor evolutionary algorithm, characterized by, The method comprises the following steps: S1, embedding an external knowledge base in the data set to generate mixed data for data enhancement; S2, using a pre-trained word vector model to perform word embedding and regularization processing on the original data, and standardizing the processing result to obtain a standard training data set; S3, using a multi-factor evolution algorithm to optimize the hyperparameters of the neural network model; The multi-factor evolution algorithm is used to optimize the input entropy value, input weight value, number of hidden layers, number of hidden layer nodes and learning rate of the neural network model; In addition to the current optimal individual, other individuals are adjusted according to the position of the current optimal individual and the global optimal position of the population, and the specific steps include: 1) generate N initial individuals, each individual consists of a position vector and a velocity vector; for the i-th individual in the population, its position vector is where represents the input entropy value, represents the input weight value, represents the number of hidden layers, represents the number of hidden layer nodes, represents the learning rate; its velocity vector is where and are random numbers between [-2, 2]. 2) with denotes the position of the ith individual of the tth generation, which is updated by the velocity adjusting the individual position, defined as: where β represents the visibility coefficient a1and a2r p and r g respectively represent the Euclidean distance between and pbest, and gbest, g ∈ [0, 1] is a variable coefficient; 3) Iterative adjustment of g; where g max = 1, g min = 0, iter is the current iteration number, iter max is the maximum iteration number; 4) Calculation of the current optimal position pbest; Wherein, fitness is the training error of the multi-level neural network; 5) Update of the individual speed; Wherein, d represents an attraction coefficient, and r [belongs to] [-1, 1]; 6) The speed update calculation formula of the fitness similar individual is: wherein, is the velocity of the individual i of the generation t, is the position of the individual i at the generation t, is the position of the individual j at the generation t; a2and β represent the attraction constant and the visibility coefficient, respectively, r mf denotes and the Euclidean distance between and fl denotes the random walk coefficient, r e [-1, 1]; S4, using a multi-level classification neural network model to detect and classify the intention; S5, comparing the model classification output result with the performance measurement standard to finally realize language recognition.

2. The language recognition method of claim 1, wherein the multi-factor evolutionary algorithm optimizes the neural network model. In step S1, the specific steps of data enhancement include: Definition of batch xi For a specific batch sample, xi, batch yi is the label corresponding to this batch sample; λ is the mixing coefficient calculated from the beta distribution of parameters α, β, defined as: Lambda = Beta (alpha, beta) (1) mixed_batch x = lambda * batch x1 + (1 - lambda) * batch x2 (2) mixed batch y = lambda * batch y1 + (1 - lambda) * batch y2 (3) where Beta(·) is the Beta distribution, mixed batch x is the mixed batch sample, mixed batch y is the label corresponding to the mixed sample.

3. The language recognition method of claim 1, wherein the multi-factor evolutionary algorithm optimizes the neural network model. In step S2, the word embedding step specifically includes: 1) Using the global co-occurrence matrix, the GloVe model can generate word vector representations V = {vi, v2... v n} where vi i represents the vector of the i-th word in the vocabulary; in addition, define A, S, R and H to represent the index set of antonyms, synonyms and related words respectively; the first vector space y is adjusted by the objective function and set constraints to obtain a new vector space V', where V' = {v'1, v'2... v' n}. 2) Synonym enrichment, by improving the cosine similarity between vectors, introducing synonyms close to each other in the vector space, defined as: Wherein, SE represents the set of enriched synonym pairs, delta = 1, and tau (x) = max (0, x); 3) Antonym enrichment, by reducing the cosine similarity between vectors to 0 to make antonyms far away from each other, defined as: Wherein, AE represents the set of enriched antonym pairs, delta = 1, and tau (x) = max (0, x); 4) Related word enrichment, the distance between words with similar meanings in the vector space is as small as possible, and the target function increases the cosine similarity between related words to 0 or more than 0, ensuring that related words and synonyms have a certain degree of distinction, which is defined as: Wherein, RE represents the set of enriched related word pairs, delta is fixed as 0, and tau (x) = max (0, x).

4. The language recognition method of claim 3, wherein the multi-factor evolutionary algorithm optimizes the neural network model. It also includes adjusting the word vector using constraints, including the following steps: Bend the transformed vector space to the original vector space, use the target function to keep the cosine similarity between adjacent values, and the initial word vector is greater than or equal to the cosine similarity between them, defined as: Wherein, M represents the number of words in the vocabulary, N (i) represents the i-th adjacent group, N (i) involves words in V, and the cosine similarity between the i-th word is higher than 0.8; The target function is the linear integral of the enrichment and cosine similarity of synonyms, antonyms and related words, defined as: E (V, V') = SE (V') + AE (V') + RE (V') + VSP (V, V') (8) The target function is minimized by the stochastic gradient descent method of the transformed vector V' and the initial vector V.

5. The language recognition method of claim 1, wherein the multi-factor evolutionary algorithm optimizes the neural network model. After updating the speed of the fitness similar individual, a crossover operator calculation process is introduced; The crossover operator calculation process has the following specific steps: 1) The crossover operation of individual i and individual j with similar fitness is shown as follows: wherein L ∈ [0, 1]; 2) Adding random disturbance to the generated individual i of the t+1 generation to improve the local search ability, that is: where σ denotes the standard deviation, N n denotes a uniform distribution.

6. The language recognition method of claim 1, wherein the multi-factor evolutionary algorithm optimizes the neural network model. Step S4 specifically comprises: 1) Initialize the score value of the selected feature and its corresponding weight value; wherein E i represents an input entropy value, W i represents a corresponding weight value; 2) Multiply the input with the arbitrarily selected weighted vector to obtain the complete summation; wherein R represents the weighted summation value; 3) Estimate the activation function (AF), defined as: wherein f represents the sigmoid activation function; 4) Hidden layer measurement acquisition, defined as: Y i = A i +∑G i W i (12) where Y i represents the measurement result, A i represents the deviation value, W i represents the weight between the input and hidden layers, G i represents the value of the AF application change.

7. The language recognition method of claim 6, wherein the multi-factor evolutionary algorithm optimizes the neural network model. Further comprising error signal normalization, the steps are as follows: 1) Re-execute the specific steps of step S4 on all layers of the neural network model, estimate the result unit by adding the weight of each input signal to obtain the layer neuron value; Ou i = A i +∑P i W i (13) where P i represents the value of the layer that results in an output of 1, W i represents the weights of the hidden layer, Ouis i represents an output unit; 2) Calculate the error signal E by biasing the output result with the target value r ; E r = Ta i -Ou i (14) wherein Ta i represents the target output, Ou i represents the classification current output; 3) weighting against the target value, on the basis of the error signal E r ; calculating the relative error δ i ; δ i = E r [f(Ou i )](15) 4) Further weighting correction, defined as: wc i = βδ i (E i ) (16) where wc i is a correction value, β represents a momentum term, E i represents an input vector, δ i represents an error distributed in the network.

8. The language recognition method of claim 1, wherein the multi-factor evolutionary algorithm optimizes the neural network model. Step S5 specifically comprises: Differentially quantize the model prediction result and the target result, determine the matching condition of the classification output result of the model and the real label in the test set, and finally perform intent recognition through the performance indicators, precision and recall rate of the model.

9. Application of the language recognition method of the neural network model optimized by the multi-factor evolutionary algorithm according to any one of claims 1-8 in the hyperparameter optimization of the neural network model.

Images