CN110287320A - A deep learning multi-category sentiment analysis model combined with attention mechanism - Google Patents

Abstract

The invention relates to a deep learning multi-category sentiment analysis model combined with an attention mechanism, which belongs to the technical field of natural language processing. The invention analyzes the weaknesses of the existing CNN network and LSTM network in text sentiment analysis, and proposes a combination of attention A deep learning multi-category sentiment analysis model based on force mechanism. The model uses the attention mechanism to integrate the local features extracted by the CNN network and the word order features extracted by the LSTM model, and adopts the idea of an integrated model at the classification layer to splicing the emotional features extracted by the CNN network and the LSTM network respectively as the final extraction of the model. emotional characteristics. Through comparative experiments, it is found that the accuracy of the model has been significantly improved.

Description

A deep learning multi-category sentiment analysis model combined with attention mechanism

technical field

The invention belongs to the field of text information processing, and relates to a deep learning multi-category sentiment analysis model combined with an attention mechanism.

Background technique

With the continuous rise of social networks such as Weibo and Twitter, the Internet is not only a source for people to obtain daily information, but also an indispensable platform for people to express their opinions. People comment on hot events in online communities, express opinions on film reviews, and describe product experience, etc., which will generate a large amount of text information with emotional color (such as: emotions, etc.), and effective sentiment analysis of these text information can be more effective. Get a better understanding of the user's interest tendencies and attention levels. However, with the increase of people's attention to network information, a large number of texts with emotional color are produced in the network community every day. If only relying on manual marking, it is far from being able to complete this task, which makes text sentiment analysis a natural A research hotspot in the field of language processing.

With the successful application of deep learning methods in the direction of computer vision, more and more deep learning techniques have also been applied in the direction of natural language processing. The advantage of deep learning is that it can not only automatically extract the features of text, but also has a strong ability to express big data. The current mainstream text sentiment analysis methods based on deep learning mainly include Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN). The accuracy of the sentiment analysis model based on these two methods is very low. lower, mainly due to the following reasons:

First of all, in the process of text sentiment analysis, the convolutional neural network can effectively capture the emotional information of different positions by expanding the size of the convolution kernel, and then obtain the local emotional features of the text. However, in the process of convolution, the contextual relationship between the word order in the text is often ignored. However, in the process of text sentiment analysis, the sequence of word order is very important, and the lack of characteristic information of word order will lead to certain deviations in the results.

Secondly, the cyclic neural network network uses the front and back dependencies to effectively simulate the sequence of text data, and can extract the word order relationship and semantic information of the text, so it can achieve good results in text sentiment analysis. However, when the sample data is long or the language scene is complex, the intervals of useful emotional information vary in size and length, and the performance of Long Short-Term Memory (LSTM) is therefore limited.

The present invention makes full use of attention mechanism, CNN network and LSTM network, proposes and implements a deep learning multi-category sentiment analysis model combined with attention mechanism. This model can effectively improve the accuracy of text sentiment analysis.

Contents of the invention

The present invention proposes a deep learning multi-category sentiment analysis model based on an attention mechanism. This model combines CNN network and LSTM network for emotional feature fusion. First, the multi-scale convolution kernel of the CNN network is used to extract the local features of the text to be analyzed, and then the attention mechanism is used to integrate the local features extracted by the CNN network into the LSTM network. Finally, using the idea of the integrated model, the pooling layer results of the CNN network and the feature extraction results of the LSTM network are spliced as the final model output. Experiments show that the accuracy of the model has been significantly improved in text sentiment analysis.

To achieve the above object, the present invention adopts the following technical solutions:

1. a deep learning multi-category sentiment analysis method combined with attention mechanism, is characterized in that comprising the following steps:

Step (1) data preprocessing

Let the emotional data set be expressed as: G=[(segtxt ₁ ,y ₁ ),(segtxt ₂ ,y ₂ ),...,(segtxt _N ,y _N )], where segtxt _i represents the i-th sample, y _i is the corresponding emotional category label, N represents the number of samples in the data set G, and data preprocessing is performed on the samples in G,

After the data set G is preprocessed, it is expressed as G′=[(seg ₁ ,y ₁ ),(seg ₂ ,y ₂ ),...,(seg _M ,y _M )], where: seg _i is expressed as data The i-th sample in the set G′, y _i is the corresponding emotional category label, and M represents the number of samples in the data set G′;

Step (2) Build the input of the model

For any sample data (seg, y) to be analyzed in the data set G′, it is further refined as:

seg＝[w ₁ ,w ₂ ,w ₃ ,...,w _d ] ^T (1)

y=[0,0,1,...,0] (2)

Among them: w _i ∈ R ^ε refers to the one-hot encoding of the i-th word in the text to be analyzed according to the vocabulary wordList, ε is the size of the vocabulary wordList, and d represents the sentence length of the text. y∈R ^p is the one-hot encoding based on the emotional category, and p represents the number of categories to be classified by the model. Then the word vector embedding matrix of this sample can be expressed as:

X=seg*E ^T (3)

Among them: X∈R ^d×m , X=[x ₁ ,x ₂ ,...,x _d ] ^T is the word vector matrix representation of the text to be analyzed, m is the dimension of the word vector, x _i ∈ R ^m is the The word vector representation of the i-th vocabulary in the text, E is the word vector embedding layer representation;

Step (3) Build a deep learning multi-category sentiment analysis model

The deep learning multi-category sentiment analysis model includes the stage of local feature extraction based on CNN network and the stage of word order relationship feature extraction based on _LSTM network. The results of the feature extraction stage C' _Rnn splicing, that is, the vector [C _Cnn ; C' _Rnn ] is used as the feature vector finally extracted by the model. Then pass the feature vector [C _Cnn ; C' _Rnn ] through the fully connected layer to get the final model output vector where p represents the number of categories to be classified by the model.

The described local feature extraction stage based on the CNN network includes the following:

In the local feature extraction stage, the input is the word vector matrix representation X of the text to be analyzed in formula 3;

The local feature extraction stage is based on the CNN network, which includes two layers, namely a convolutional layer and a pooling layer, in which:

The convolutional layer uses n types of convolution kernels of different scales to convolve the text to be analyzed, and the filters of the same scale convolution kernels are k neurons;

The pooling layer uses the method of the maximum pooling layer to down-sample the vector obtained by convolution and select the local optimal features, so each filter becomes a scalar through the maximum pooling layer, which represents the Optimal emotional characteristics;

The output of the local feature extraction module is C _Cnn =[c ₁ ,c ₂ ,...,c _nk ], that is, the optimal features selected by multiple filters of different sizes in the pooling layer are spliced together C _Cnn =[c ₁ ,c ₂ ,...,c _nk ] as the output of this module, where C _Cnn ∈ R ^nk , nk is the number of all filters in the convolutional layer;

Described word order relation feature extraction stage based on LSTM network, comprises the following contents:

Multi-scale CNN network local feature extraction: Concatenate the convolution results of k filters of the same convolution scale in the convolutional layer in the local feature extraction stage based on the CNN network to obtain a set Z _Cnn , and then combine each of the set Z _Cnn The vector Z _i is input into the GLU mechanism, that is, the gated convolutional network, and the obtained results are recorded as {π ₁ ,π ₂ ,...,π _n } to complete the extraction of local features of the multi-scale CNN network.

Among them, Z _Cnn = {Z ₁ , Z ₂ ,..., Z _n }, Z _i is the concatenation of multiple filter convolution results with a scale of i;

in, Z _i represents the concatenation of k filter convolution results of a certain scale, W ₁ , W ₂ ∈ R ^λ×q is the weight matrix, λ represents the dimension of the corresponding weight matrix, and b ₁ , b ₂ ∈ R ^q is the bias Set amount, σ represents the sigmoid function, π _i ∈ R ^q , q is the output dimension of the LSTM network;

Then, using the attention mechanism, the multi-scale CNN network local feature extraction results {π ₁ ,π ₂ ,...,π _n } are integrated into the LSTM network, and the output result C of the word order relationship feature extraction stage based on the LSTM network is obtained ' _Rnn , ie

in, Indicates the output of the LSTM module corresponding to the last word in the text to be analyzed, Indicates the output of the LSTM module corresponding to the first word in the text to be analyzed. The present invention adopts a bidirectional LSTM model, namely the BiLSTM model,

Using forward propagation, the specific calculation process is as follows:

d is the length of the text to be analyzed, each word sequence in the text corresponds to an LSTM module,,

During forward propagation, the output of the t-1th LSTM module is Then the output of the tth LSTM module Calculated as follows:

in: Is the dot product of two vectors, also known as the scoring function, which is used to calculate the output of the LSTM of the previous word The similarity with the current local feature vector,

Among them: α _t,i ∈ R represents the weight of feature π _i ,

Among them: s _t-1 ∈ R ^q is the weighted result of multiple convolutional features, using s _t-1 instead Combine the word vector x _t of the current word to obtain the output of the current LSTM module The formula is as follows:

Using backpropagation, the specific calculation process is the same as that of forward propagation, and will not be repeated here;

Step (4) Model training: Input the training data into the multi-category sentiment analysis model, use the cross-entropy loss function, adjust the parameters in combination with the backpropagation BP algorithm, and use softmax regression as the classification algorithm to complete the training;

Step (5) Model analysis: Input the text to be analyzed into the trained model, and finally output the sentiment classification result after analyzing the text.

Described pretreatment process comprises the following steps:

1) Word segmentation, deactivation, uppercase to lowercase, and traditional to simplified.

2) Select words whose frequency is greater than or equal to σ in the data set G, and construct a vocabulary wordList={word ₁ ,word ₂ ,...word _ε }, where word _i represents the i-th word in the vocabulary wordlist, and ε represents the data The total number of words in the set G whose word frequency exceeds σ.

3) For each sample in the data set G, if the length is greater than d, delete the sample, and if the length is less than d, fill it with the symbol </>.

The convolutional layer calculation formula of the local feature extraction module based on the CNN network is as follows:

z＝f(∑W ^T *x _i:i+s-1 +b) (8)

Among them: z represents the feature vector obtained by the convolution of a neuron to analyze the text, f(·) represents the activation function, W∈R ^s×m represents the weight matrix of the neuron, the same neuron parameters are shared, s×m represents The size of the convolution kernel, b represents the threshold, x _i:i+s-1 represents the word vector from the i-th word to the i+s-1 word in the text sentence.

The training data is preprocessed data.

The convolution layer of the local feature extraction stage based on the CNN network adopts four convolution kernels of different scales. The training end condition is that the accuracy rate does not change or reaches the set number of iterations.

Description of drawings

Fig. 1 method flowchart of the present invention;

Figure 2 Schematic diagram of the structure of the deep learning multi-category sentiment analysis model combined with the attention mechanism.

Detailed ways

The specific implementation of the present invention will be further described in detail below in conjunction with the diagrams and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

The method that the present invention proposes is to realize by following steps successively:

Step (1) data preprocessing

The emotional language data set is expressed as: G=[(segtxt ₁ ,y ₁ ),(segtxt ₂ ,y ₂ ),...,(segtxt _N ,y _N )], where segtxt _i represents the i-th samples, and y _i is the corresponding emotion category label. N represents the number of samples in the data set G, and the emotion labels are divided into four categories: "joy", "anger", "disgust", and "depression". N is 80,000, and each of the four types of emotion samples is 20,000. Data preprocessing for samples in G includes the following steps:

1) Word segmentation, deactivation, uppercase to lowercase, and traditional to simplified.

2) Select the words whose frequency is greater than or equal to σ in the data set G, and construct a vocabulary wordList={word ₁ , word ₂ ,...word _ε }, where word _i represents the i-th word in the data set G, and ε represents the data The total number of words in the set G whose word frequency exceeds σ. If σ is set to 2, in the final data set G, there are 41763 words with a word frequency greater than or equal to 2, that is, ε is 41763.

3) After the above processing, for each sample in the data set G, if the length is greater than d, delete the sample, and if the length is less than d, fill it with the symbol </>. d takes 64.

After the data set G is preprocessed, it is expressed as G′=[(seg ₁ ,y ₁ ),(seg ₂ ,y ₂ ),...,(seg _M ,y _M )]. Among them: seg _i represents the i-th sample in the data set G′, y _{i represents} the corresponding emotion category label, and M represents the number of samples in the data set G′. The number of samples in the final data set G′ is 73150, and the number of samples of each emotion category is shown in Table 1:

Table 1 Number of samples in each category after preprocessing

Step (2) The input of the model

For any sample data (seg, y) to be analyzed in the data set G′, it is further refined as:

seg＝[w ₁ ,w ₂ ,w ₃ ,...,w _d ] ^T (1)

y=[0,0,1,...,0] (2)

Among them: w _i ∈ R ^ε refers to the one-hot encoding of the i-th word in the analyzed text according to the vocabulary wordList, ε is the size of the vocabulary wordList, and the sentence length d of the text is 64. y∈R ^p is the one-hot encoding based on the emotional category, p represents the number of categories to be classified by the model, and p is 4. Then the word vector embedding matrix of this sample can be expressed as:

X=seg*E ^T (3)

Where: X∈R ^d×m , X=[x ₁ ,x ₂ ,...,x _d ] ^T is the word vector matrix representation of the text to be analyzed, and the word vector dimension m is 256. x _i ∈ R ^m is the word vector representation of the i-th vocabulary in the text, the word vector embedding layer represents E, using Wikipedia's open source word2vec word vector, and then X is used as the input of the network model.

Step (3) Build a deep learning multi-category sentiment analysis model

The deep learning multi-category sentiment analysis model includes the stage of local feature extraction based on CNN network and the stage of word order relationship feature extraction based on _LSTM network. The results of the feature extraction stage C' _Rnn splicing, that is, the vector [C _Cnn ; C' _Rnn ] is used as the feature vector finally extracted by the model. Then pass the feature vector [C _Cnn ; C' _Rnn ] through the fully connected layer to get the final model output vector where p represents the number of categories to be classified by the model.

The described local feature extraction stage based on the CNN network includes the following:

In the local feature extraction stage, the input is the word vector matrix representation X of the text to be analyzed in formula 3;

The local feature extraction stage is based on the CNN network, which includes two layers, namely a convolutional layer and a pooling layer, in which:

The convolution layer uses n types of convolution kernels of different scales to convolve the text to be analyzed, and there are k filters of the same scale convolution kernels, that is, neurons. In the present invention, n is 4, and k is 128.

The pooling layer uses the method of the maximum pooling layer to down-sample the vector obtained by convolution and select the local optimal features, so each filter becomes a scalar through the maximum pooling layer, which represents the Optimal emotional characteristics;

The output of the local feature extraction module is C _Cnn =[c ₁ ,c ₂ ,...,c _nk ], that is, the optimal features selected by multiple filters of different sizes in the pooling layer are spliced together C _Cnn =[c ₁ ,c ₂ ,...,c _nk ] as the output of this module, where C _Cnn ∈ R ^nk , nk is the number of all filters in the convolutional layer, a total of 512;

Described word order relation feature extraction stage based on LSTM network, comprises the following contents:

Multi-scale CNN network local feature extraction: Concatenate the convolution results of k filters of the same convolution scale in the convolutional layer in the local feature extraction stage based on the CNN network to obtain a set Z _Cnn , and then combine each of the set Z _Cnn The vector Z _i is input into the GLU mechanism, that is, the gated convolutional network, and the obtained results are recorded as {π ₁ ,π ₂ ,...,π _n } to complete the extraction of local features of the multi-scale CNN network.

Among them, Z _Cnn = {Z ₁ , Z ₂ ,..., Z _n }, Z _i is the concatenation of multiple filter convolution results with a scale of i;

in, Z _i represents the concatenation of k filter convolution results of a certain scale, W ₁ , W ₂ ∈ R ^λ×q is the weight matrix, λ represents the dimension of the corresponding weight matrix, b ₁ , b ₂ ∈ R ^q is the bias Quantity, σ represents the sigmoid function, π _i ∈ R ^q , q is the output dimension of the LSTM network, and q is 256;

Then, using the attention mechanism, the multi-scale CNN network local feature extraction results {π ₁ ,π ₂ ,...,π _n } are integrated into the LSTM network, and the output result C of the word order relationship feature extraction stage based on the LSTM network is obtained ' _Rnn , ie

in, Indicates the output of the LSTM module corresponding to the last word in the text to be analyzed, Indicates the output of the LSTM module corresponding to the first word in the text to be analyzed. The present invention adopts a bidirectional LSTM model, namely the BiLSTM model,

Using forward propagation, the specific calculation process is as follows:

d is the length of the text to be analyzed, each word sequence in the text corresponds to an LSTM module,,

During forward propagation, the output of the t-1th LSTM module is Then the output of the tth LSTM module Calculated as follows:

in: Is the dot product of two vectors, also known as the scoring function, which is used to calculate the output of the LSTM of the previous word The similarity with the current local feature vector,

Among them: α _t,i ∈ R represents the weight of feature π _i ,

Among them: s _t-1 ∈ R ^q is the weighted result of multiple convolutional features, using s _t-1 instead Combine the word vector x _t of the current word to obtain the output of the current LSTM module The formula is as follows:

Using backpropagation, the specific calculation process is the same as that of forward propagation, and will not be repeated here;

Step (4) Model training: Input the training data into the multi-category sentiment analysis model, use the cross-entropy loss function, combine the backpropagation BP algorithm to adjust the parameters, and use softmax regression as the classification algorithm to complete the training.

Step (5) Model analysis: Input the text to be analyzed into the trained model, and finally output the sentiment classification result after analyzing the text.

The convolutional layer calculation formula of the local feature extraction module based on the CNN network is as follows:

z＝f(∑W ^T *x _i:i+s-1 +b) (8)

Among them: z represents the feature vector obtained by the convolution of a neuron to analyze the text, f(·) represents the activation function, W∈R ^s×m represents the weight matrix of the neuron, the same neuron parameters are shared, s×m represents The size of the convolution kernel, b represents the threshold, x _i:i+s-1 represents the word vector from the i-th word in the text sentence to the i+s-1 word, and s takes [2,3,4, 5] Four different convolution sizes, f(·) using the RELU activation function.

The training data is preprocessed data.

The convolution layer of the local feature extraction stage based on the CNN network adopts four convolution kernels of different scales. The training end condition is that the accuracy rate does not change or reaches the set number of iterations.

1. Experimental analysis

In the test phase, 2,000 emotion corpora were selected for each category of joy, anger, disgust, and depression. Using the accuracy rate Acc (Accuracy) as the evaluation index, the parameters of the model in the test phase remain unchanged, and the results of the test set are shown in Table 2:

Table 2 Comparison of sentiment analysis results

Table 2 shows the comparison of the test results of several models. Among them, Experiment 1 is a general-purpose single-scale CNN network model with a convolution kernel size of 3, Experiment 2 is a general-purpose LSTM network, and Experiment 3 is based on Text Sentiment Analysis Model with Attention Mechanism.

Through the comparative analysis of the experiment, it can be seen that compared with the usual CNN network and LSTM network, the accuracy of the emotional analysis model based on the attention mechanism proposed in this paper is significantly improved, which shows that the method proposed by the present invention can effectively extract the content of the CNN network. The local feature information and the word order feature information of the LSTM network illustrate the effectiveness of the method.

Claims (6)

1. A deep learning multi-classification emotion analysis method combined with an attention mechanism is characterized by comprising the following steps:

step (1) data preprocessing

Let the emotion data set be expressed as: g ═ segtxt₁,y₁),(segtxt₂,y₂),...,(segtxt_N,y_N)]Wherein segtxt_iDenotes the ith sample, y_iThen the corresponding emotion type label is obtained, N represents the number of samples in the data set G, and the samples in G are processedThe pre-processing of the data is carried out,

the dataset G was preprocessed and denoted as G' ═ seg [ ("seg₁,y₁),(seg₂,y₂),...,(seg_M,y_M)]Wherein: seg_iExpressed as the ith sample, y, in the data set G_iThen the label is a corresponding emotion category label, and M represents the number of samples in the data set G';

step (2) input of the constructed model

For any sample data to be analyzed (seg, y) in the data set G', it is further detailed as:

seg＝[w₁,w₂,...,w_i,...,w_d]^T (1)

y＝[0,0,1,...,0] (2)

wherein: w is a_i∈R^εOne-hot coding of the ith word in a text to be analyzed is carried out according to a word list wordList, epsilon is the size of the word list wordList, d represents the sentence length of the text, and y belongs to R^pAccording to one-hot encoding of emotion categories, p represents the number of categories to be classified by the model, and the word vector embedding matrix of the sample can be represented as follows:

X＝seg*E^T (3)

wherein: x is formed by R^d×m，X＝[x₁,x₂,...,x_d]^TFor a word vector matrix representation of the text to be analyzed, m is the dimension of the word vector, x_i∈R^mRepresenting the word vector of the ith vocabulary in the text, and E representing the word vector embedded layer;

step (3) constructing a deep learning multi-classification emotion analysis model

The deep learning multi-classification emotion analysis model comprises a CNN network-based local feature extraction stage, an LSTM network-based word order relation feature extraction stage and a pooling layer result C of the CNN network-based local feature extraction stage_CnnAnd result C 'of language sequence relation feature extraction stage based on LSTM network'_RnnSplicing, i.e. vector [ C_Cnn；C'_Rnn]As the feature vector of the final extraction of the model, and then the feature vector [ C ]_Cnn；C'_Rnn]Obtaining the final model through the full connection layerOutput vectorWhere p represents the number of classes to be assigned to the model,

the local feature extraction stage based on the CNN network comprises the following contents:

inputting a word vector matrix representation X of the text to be analyzed of a formula 3 in a local feature extraction stage;

the local feature extraction stage is based on a CNN network and comprises two layers in total, namely a convolutional layer and a pooling layer, wherein:

the convolution layer adopts n convolution kernels with different scales to convolute the text to be analyzed, and the number of filters of the convolution kernel with the same scale is k, namely, k of each neuron;

in the pooling layer, the vector obtained by convolution is down-sampled by adopting a maximum pooling layer method, and local optimal features are selected, so that each filter becomes a scalar through the maximum pooling layer, and the scalar represents the optimal emotional features in the filter;

the output of the local feature extraction module is C_Cnn＝[c₁,c₂,...,c_nk]Splicing the optimal features selected by a plurality of filters with different sizes in the pooling layer together C_Cnn＝[c₁,c₂,...,c_nk]As output of this module, where C_Cnn∈R^nkAnd nk is the number of all filters in the convolutional layer;

the language order relation feature extraction stage based on the LSTM network comprises the following contents:

multi-scale CNN network local feature extraction: splicing convolution results of k filters with the same convolution scale of convolution layers in the local feature extraction stage based on the CNN network to obtain a set Z_CnnThen set Z_CnnEach vector Z in (a)_iInputting the result into a GLU mechanism, namely, gating a convolution network, and marking the obtained result as { pi₁,π₂,...,π_nFinishing the extraction of the local features of the multi-scale CNN network,

wherein Z is_Cnn＝{Z₁,Z₂,...,Z_n}，Z_iSplicing convolution results of a plurality of filters with the scale i;

wherein,Z_istitching of convolution results of k filters representing a certain scale, W₁，W₂∈R^λ×qFor the weight matrix, λ represents the dimension of the corresponding weight matrix, b₁，b₂∈R^qFor the offset, σ denotes the sigmoid function, π_i∈R^qQ is the output dimension of the LSTM network;

then, extracting the local feature extraction result { pi ] of the multi-scale CNN network by using an attention mechanism₁,π₂,...,π_nIntegrating the language sequence relationship feature extraction stage into an LSTM network to obtain an output result C 'of the language sequence relationship feature extraction stage based on the LSTM network'_RnnI.e. by

Wherein,represents the output of the LSTM module corresponding to the last word in the text to be analyzed,representing the output of the LSTM module corresponding to the first word in the text to be analyzed, the invention uses a bi-directional LSTM model, i.e. a BiLSTM model,

the forward propagation is adopted, and the specific calculation process is as follows:

d is the length of the text to be analyzed, each word in the text sequentially corresponds to an LSTM module,

in the forward propagation process, the output of the t-1 th LSTM module isThen the output of the tth LSTM moduleThe calculation formula is as follows:

wherein:is a dot product of two vectors, also known as a scoring function, and is the output of the LSTM used to calculate the previous wordAnd the similarity of the current local feature vector,

wherein α_t,iEpsilon R represents characteristic pi_iThe weight of (a) is determined,

wherein: s_t-1∈R^qIs a weighted result of a plurality of convolution features, using s_t-1Instead of the formerWord vector x in combination with current word_tEvaluating the output of a current LSTM moduleThe formula is as follows:

the backward propagation is adopted, the specific calculation process is the same as the forward propagation, and the details are not repeated here;

step (4), model training: inputting training data into a multi-classification emotion analysis model, adjusting parameters by adopting a cross entropy loss function and combining a back propagation BP algorithm, and finishing training by using softmax regression as a classification algorithm;

and (5) analyzing a model: and inputting the text to be analyzed into the trained model, and finally outputting the emotion classification result after the text is analyzed.

2. The method for deep learning multi-classification emotion analysis combined with attention mechanism as claimed in claim 1, wherein the preprocessing process comprises the following steps:

1) word segmentation, stop removal, conversion from English capitals to lowercases, conversion from traditional Chinese to simplified Chinese,

2) selecting words with frequency more than or equal to sigma in the data set G, and constructing a vocabulary table word List ═ word₁,word₂,...word_εTherein, word_iRepresenting the ith word in the vocabulary word, epsilon represents the total number of words in the data set G with a word frequency exceeding sigma,

3) for each sample in the data set G, if the length is larger than d, deleting the sample, and if the length is smaller than d, filling the sample with symbols </>.

3. The method for deep learning multi-classification emotion analysis combined with attention mechanism as claimed in claim 1, wherein the convolutional layer calculation formula of the CNN network-based local feature extraction module is as follows:

z＝f(∑W^T*x_i:i+s-1+b) (8)

wherein z represents a feature vector obtained by convolution of a neuron with a text to be analyzed, f (-) represents an activation function, and W is equal to R^s×mWeight matrix representing neurons, shared by the same neuron parameter, sxm represents the size of convolution kernel, b represents threshold, x represents_i:i+s-1A word vector representing the i-th word to the i + s-1 words in the text sentence.

4. The method as claimed in claim 1, wherein the training data is preprocessed data.

5. The method for deep learning multi-classification emotion analysis combined with attention mechanism as claimed in claim 1, wherein the convolutional layer of the local feature extraction stage based on the CNN network employs 4 convolutional kernels with different scales.

6. The method as claimed in claim 1, wherein the end condition of the training is that the accuracy is not changed or the set number of iterations is reached.