CN111651593A - Text emotion analysis method based on word vector and word vector mixed model - Google Patents

Abstract

The invention provides a text emotion analysis method based on a word vector and word vector mixed model, aiming at the problems that emotion information expression is insufficient, words are considered only and other text characteristics are ignored in the current text emotion analysis, and the method comprises the following steps: firstly, preprocessing a Chinese data set, and training a Word vector and a Word vector matrix by using Word2 Vec; then, the word vectors and the word vectors are used as input data and are respectively sent into a Convolutional Neural Network (CNN) and a bidirectional long-short term memory network (BilSTM) for feature extraction; two attention layers are introduced to learn important text features; and finally, combining the text features extracted by the two channels, and classifying the output by using a classification layer. The method provided by the invention has significance and superiority on Chinese data sets.

Description

Text emotion analysis method based on word vector and word vector mixed model

Technical Field

The invention provides a text emotion analysis method based on a word vector and word vector mixed model, and relates to the field of text emotion analysis.

Background

In recent years, with the rapid development of the internet industry, a plurality of new media appear, a large amount of data such as user comments, brands, emotions, politics, viewpoints and the like can be obtained through the internet, emotion analysis is a special text mining work, attitudes or emotions of people are extracted from given texts, at present, text emotion analysis is an important research direction in the field of Natural Language Processing (NLP), aims at unstructured information, mines deep-level emotions or tendencies of the implications of the information, is more and more emphasized by academic circles and industrial circles, is different from images and voices, and has own characteristics in many aspects.

The main task of the text sentiment analysis is to analyze, process, summarize and judge the text with sentiment colors. At present, text emotion analysis methods mainly comprise two methods, namely a method based on an emotion dictionary and a method based on machine learning, wherein the emotion analysis technology based on machine learning is used for text analysis, although good effects are achieved, the methods cannot effectively express complex function calculation, data features need to be selected manually, generalization capability is weak, deep learning can automatically learn important data features from original data, various complex tasks are processed, advantages in modeling, explanation, expression capability, optimization and the like are obvious, the deep learning is applied to the field of text emotion analysis, and research and development of text emotion analysis are greatly promoted.

In order to overcome the defects of traditional machine learning and an emotion dictionary-based method algorithm, a deep learning algorithm is utilized to process an NLP task, a Recurrent Neural Network (RNN) and a Convolutional Neural Network (CNN) are the most widely used Network models in a text emotion analysis task, however, in a text, the influence of a single word vector on the emotion polarity of the whole text is only considered, so that the semantic acquisition of the text is insufficient, the two Neural networks are the overall characteristics of the learning text, and the Neural Network structure with a vertical structure cannot effectively and comprehensively extract the text characteristics with a deeper level.

The invention provides a text emotion analysis method based on a Word vector and Word vector mixed model, which comprises the steps of constructing a Word vector and Word vector based mixed model (BilSTM-CNN-attribute), carrying out Word segmentation, Word filtering, Word stopping and other preprocessing on a Chinese data set, training a Word vector and a Word vector matrix by using Word2Vec, and then respectively sending the Word vector and the Word vector as input data into a convolutional neural network and a bidirectional Long and short term Memory network (Bi-directional Long and short term Memory, BilSTM) for feature extraction; compared with a deep learning network which only considers word vector characteristics generally, the method can fully extract local characteristics and sequence information of the text, solve the problem of semantic multi-level, and can learn the important information characteristics of the text through the attention mechanism, and the accuracy of the method can reach 92.67% on a text data set.

Disclosure of Invention

In view of the above, the main objective of the present invention is to combine the advantages of the convolutional neural network and the two-way long-short term memory network, use the word vector and the word vector as the input of the model at the same time, and add the attention layer after that to extract the important text information features, thereby improving the accuracy of the text emotion analysis.

In order to achieve the purpose, the technical scheme provided by the invention is as follows:

the text emotion analysis method based on the word vector and word vector mixed model comprises the following steps:

step 1, preprocessing a Chinese data set, and simultaneously training a Word vector and a Word vector matrix by using Word2 Vec;

step 2, training the word vector matrix x_1:lAnd word vector matrix w_1:lAs the input characteristics of the bidirectional long and short term memory network, the sequence characteristics of the text are learned, and the attention layer is accessed to optimize the characteristic vector;

step 3, training the word vector matrix x_1:lAnd word vector matrix w_1:lAs the input features of the convolutional neural network, performing convolution and pooling operations to learn local features of the text, and accessing an attention mechanism to obtain deep features of the text;

step 4, extracting the features s from the convolutional neural network layer with the attention mechanism_cAnd feature vectors s extracted by a bidirectional long-term and short-term memory network layer introducing an attention mechanism_lAnd fusing, inputting a softmax classification layer for classification, wherein the positive value is 1, the negative value is 0, and comparing and calculating with the text label to obtain the text classification accuracy.

In summary, the invention integrates the convolutional neural network and the bidirectional long and short term memory network based on the word vector and the word vector, and introduces the attention mechanism to learn important text information, which is essentially to use the word vector and the word vector as the input of the convolutional neural network and the bidirectional long and short term memory network at the same time, use the bidirectional long and short term memory network to learn sequence information, obtain local features of the text by means of the convolutional neural network, and then use two attention layers to identify important features, thereby improving the accuracy of model emotion classification.

Description of the drawings:

FIG. 1 is a schematic general flow chart of a text emotion analysis method based on a word vector and word vector hybrid model according to the present invention;

FIG. 2 is a schematic flow chart of training Word vectors and Word vectors using Word2 Vec;

FIG. 3 is a schematic flow chart of sequence feature extraction using a BilSTM network;

FIG. 4 is a schematic diagram of a process for extracting local features using a CNN network;

FIG. 5 is a schematic diagram of a process of calculating the text classification accuracy by using a Softmax classification layer for classification;

FIG. 6 shows the accuracy results obtained after experiments using the Chinese data set;

fig. 7 is an overall construction diagram of a word vector and word vector hybrid model according to the present invention.

The specific implementation mode is as follows:

the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, it is obvious that the examples are for illustration and not for limiting the embodiments of the present invention, the present invention can also be implemented by other different specific embodiments, and all other embodiments obtained by those skilled in the art without any inventive work are within the scope of the present invention.

Fig. 1 is a general flow diagram of a text emotion analysis method based on a word vector and word vector hybrid model according to the present invention, and as shown in fig. 1, the text emotion analysis method based on a word vector and word vector hybrid model according to the present invention includes the following steps:

step 1, preprocessing a Chinese data set, and simultaneously training a Word vector and a Word vector matrix by using Word2 Vec;

step 2, training the word vector matrix x_1:lAnd word vector matrix w_1:lAs the input characteristics of the bidirectional long and short term memory network, the sequence characteristics of the text are learned, and the attention layer is accessed to optimize the characteristic vector;

step 3, training the word vector matrix x_1:lAnd word vector matrix w_1:lAs the input features of the convolutional neural network, performing convolution and pooling operations to learn local features of the text, and accessing an attention mechanism to obtain deep features of the text;

step 4, extracting the features s from the convolutional neural network layer with the attention mechanism_cAnd feature vectors s extracted by a bidirectional long-term and short-term memory network layer introducing an attention mechanism_lAnd fusing, inputting a softmax classification layer for classification, wherein the positive value is 1, the negative value is 0, and comparing and calculating with the text label to obtain the text classification accuracy.

Fig. 2 is a schematic flow chart of training Word vectors and Word vectors using Word2Vec, as shown in fig. 2, in step 1, a preprocessing operation is performed on a chinese data set, and Word vectors and Word vector matrices are simultaneously trained using Word2Vec, including the following steps:

step 11, preprocessing the text to improve the quality of word vectors, including a Stop word and a word segmentation process, wherein the specific process of the Stop word refers to selecting a work-of-the-job Stop word list, eliminating Stop Words (Stop Words) in a data set, the Stop Words refer to certain Words which have high frequency and are meaningless in the text, such as pronouns, mood-aid Words, prepositions, conjunctions and the like, because the text format is not standardized enough, the Words without special meaning are filtered out in the preprocessing stage, interference of irrelevant information is reduced, the quality of the word vectors is greatly improved, and the specific process of the word segmentation refers to an accurate word segmentation mode of Jieba word segmentation which can accurately separate Chinese texts according to reading habits according to a specific dictionary;

step 12, using a Word2vec tool opened by Google corporation for Word vector training work, training Word vectors and Word vectors by using a Skip-gram model in the Word2vec tool, training preprocessed corpus into 128-dimensional Word vectors and Word vectors, and respectively representing a Word vector matrix and a Word vector matrix as follows:

wherein the content of the first and second substances,

representing the join operator, l representing the length of the sentence, and taking a constant value of 60, x_1:lWord vector matrix, w, representing 60 × 128_1:lRepresenting a 60 × 128 word vector matrix.

FIG. 3 is a flow chart of sequence feature extraction using a two-way long-short term memory networkFIG. 3 shows that in step 2, the trained word vector matrix x_1:lAnd word vector matrix w_1:lAs the input features of the bidirectional long-short term memory network, the sequence features of the text are learned, and the attention layer optimization feature vector is accessed, and the method comprises the following steps:

step 21, converting the word vector matrix x_1:lAnd word vector matrix w_1:lRespectively inputting the results into a BilSTM network for training, extracting features by using the BilSTM network to obtain sequence features of texts, improving the convergence of the network, and obtaining a hidden state h of a character vector of the BilSTM network_wiHidden state h of sum word vector_xiA matrix x of word vectors to be converted_1:lInputting the result into a BilSTM network for training to obtain a hidden state h of a BilSTM network word vector_xiThe formula is expressed as:

f_t＝σ(w^fx_1:l+u^fh_t-1+b^f)

i_t＝σ(wⁱx_1:l+uⁱh_t-1+bⁱ)

o_t＝σ(w^ox_1:l+u^oh_t-1+b^o)

wherein sigma is sigmoid activation function, and the output is [0,1 ]]And determines how much information can pass through, tanh is a hyperbolic tangent function,

operator symbol representing a matrix multiplication operation, forgetting gate f_tDeciding to "forget" information that is not important in the back-propagation, input gate i_tDetermines the information to be updated

Output gate o_tDetermining from the current cell state c_tOutput to hidden layer state h_xtThe content of (1). Current cell state c_tThat is, the past cell state c_t-1Merging with new memory, w⁽ⁱ⁾And u⁽ⁱ⁾Representing weights in the data processing process, a matrix w of word vectors to be converted_1:lInputting the result into a BilSTM network for training to obtain a hidden layer state h of a BilSTM network word vector_wiThe formula is expressed as:

f_t＝σ(w^fw_1:l+u^fh_t-1+b^f)

i_t＝σ(wⁱw_1:l+uⁱh_t-1+bⁱ)

o_t＝σ(w^ow_1:l+u^oh_t-1+b^o)

wherein sigma is sigmoid activation function, and the output is [0,1 ]]And determines how much information can pass through, tanh is a hyperbolic tangent function,

operator symbol representing a matrix multiplication operation, forgetting gate f_tDeciding to "forget" information that is not important in the back-propagation, input gate i_tDetermines the information to be updated

Output gate o_tDetermining from the current cell state c_tOutput to hidden layer state h_wtThe content of (1). Current cell state c_tThat is, the past cell state c_t-1Merging with new memory, w⁽ⁱ⁾And u⁽ⁱ⁾Representing the weight in the data processing process;

step 22, the hidden layer state h of the BiLSTM network word vector obtained in the step 21 is used for_wiHidden state h of sum word vector_xiThe fusion is performed by using a point multiplication mode, and the formula is expressed as follows:

h_i＝[h_wi·h_xi]

step 23, hiding layer state h in the BilSTM network_iNonlinear transformation u_iIs expressed as

u_i＝tanh(w_wh_i+b_w)

Wherein, w_wAnd b_wRepresenting a weight matrix and a weight vector;

step 24, using softmax function to pair u_iNormalization is performed to obtain an attention matrix, i.e., output weighting coefficients α of the BilSTM layer_iThe formula is expressed as:

wherein u is_wThe initialized weight matrix is shown;

step 25, output weight coefficient α of BilSTM layer obtained in step 24_iAnd hidden layer state h obtained in step 22_iMultiplying to obtain a feature vector s extracted from the BilSTM layer and introducing an attention mechanism_lThe formula is expressed as:

wherein, α_iExpressed is the output weight coefficient, h_iThe hidden state obtained in step 22 is shown.

FIG. 4 is a schematic diagram of a process of extracting local features by using a convolutional neural network, as shown in FIG. 4, in step 3, a word vector matrix x to be trained_1:lAnd word vector matrix w_1:lThe method is used as the input feature of the convolutional neural network, the local feature of the text is learned through convolution and pooling operation, and the attention mechanism is accessed to obtain the deep feature of the text, and the method comprises the following steps:

step 31, training the word vector matrix x_1:lAnd word vector matrix w_1:lThe input signals are respectively input into a CNN network for training, and the convolution operation is performed by using a linear filter on an input matrix, which can be expressed as follows:

C_xi＝f(W·x_1:l+b)

C_wi＝f(W·w_1:l+b)

wherein b ∈ R denotes a bias vector, W denotes a convolution kernel with a height h of 3, 4, 5 and a width d of 128, f denotes a nonlinear activation function, and a ReLu function is used as the activation function, which is expressed as:

f(x)＝max(0,x)

step 32, for the trained word vector matrix x_1:lAnd word vector matrix w_1:lAfter convolution operation, a word vector characteristic diagram C is obtained_xSum word vector feature map C_w

C_x＝[C_x1,C_x2,…,C_xl]

C_w＝[C_w1,C_w2,…,C_wl]

Combining the local features extracted by the convolution kernels, and selecting 128 convolution kernels to obtain a plurality of feature maps representing different feature information for extracting more text features;

step 33, after completing the operation of step 32, matching the word vector feature map C_xSum word vector feature map C_wPerforming maximum pooling operation to obtain word vector output characteristics c_xiSum word vector output feature c_wiReducing the characteristic dimension, the formula is expressed as:

c_xi＝max(C_x)

c_wi＝max(C_w)

then outputting the word vector to the feature c at the fusion layer_xiSum word vector output feature c_wiThe point multiplication mode is used for fusion, and the formula is expressed as follows:

c_i＝[c_xi·c_wi]

step 34, using tanh function to compare the characteristic c obtained in step 33_iNonlinear transformation u_ciThe formula is expressed as:

u_ci＝tanh(w_wc_i+b_w)

wherein, w_wAnd b_wRepresenting the weight matrix and weight vector, and then using the Softmax function to pair u_ciNormalization is carried out to obtain an output weight coefficient α of the CNN layer_ciThe formula is expressed as:

wherein u is_cwThe initialized weight matrix is shown;

step 35, outputting the weight coefficient α of the convolution neural layer obtained in step 34_ciWith the feature c obtained in step 33_iMultiplying to obtain text features extracted after the CNN passes through the attention layer, and expressing the text features as a vector s_c：

Wherein, α_ciThe output weight coefficients are indicated.

FIG. 5 is a schematic diagram of a process of classifying by using a Softmax classification layer and calculating the text classification accuracy, and in step 4, features s extracted by a convolutional neural network layer with attention mechanism introduced_cAnd feature vectors s extracted by a bidirectional long-short term memory network layer introducing an attention mechanism_lFusing, inputting a softmax classification layer for classification, positively setting the classification to be 1 and negatively setting the classification to be 0, and comparing and calculating with the text label to obtain the text classification accuracy, wherein the method comprises the following steps:

step 41, extracting features s of the convolutional neural layer with attention mechanism_cAnd feature vectors s extracted by a bidirectional long-short term memory network layer introducing an attention mechanism_lAnd accessing a full connection layer by using a point-product fusion mode to obtain output characteristics:

x＝[s_l·s_c]

step 42, adopting a Dropout strategy, wherein the main idea is that during model training, a part of the method is randomly selected and temporarily discarded from the network, namely, the neural units are temporarily inactivated and do not participate in parameter updating operation any more, and the Dropout rate is set to be 0.5, namely, half of the neurons do not participate in calculation in each iteration;

step 43, taking the feature x obtained in step 41 as an input of the classification layer, and calculating the probability size p of each text belonging to different categories by using a softmax function, which can be described as the following formula:

where the text is divided into 2 categories, w_kAnd b_kIs the weight and bias of the layer;

and 44, judging the text type, judging that the probability value p belongs to a larger type, positively being 1, and negatively being 0, and comparing the probability value p with the text label to calculate to obtain the text accuracy.

Examples

In the embodiment, real Chinese comments collected from the Internet are adopted, and the text emotion is analyzed by using a text emotion analysis method based on a word vector and word vector mixed model, and the specific steps are as follows:

1. preprocessing a text set, using a crust word segmentation, then performing word-off-stop processing, endowing the text with a label, wherein the positive is 1, the negative is 0,

2. vectorizing the text, training Word vectors and Word vectors by using a Word2Vec tool, setting the dimensions of the Word vectors and the Word vectors to be 128, fixing the length of the sentence to be 60, obtaining a Word vector matrix and a Word vector matrix,

3. the word vector matrix and the word vector matrix are respectively used as the input characteristics of the two-way long and short term memory network and the convolutional neural network, wherein the convolutional kernel size of the convolutional neural network is set to be 3, 4 and 5, the number of the convolutional kernels is 128, the number of the hidden layer neurons of the two-way long and short term memory network is set to be 128, in order to prevent overfitting, the Dropout rate is set to be 0.5,

4. after the bidirectional long-short term memory network and the convolutional neural network are respectively accessed to the attention layer to extract important characteristic information, the size of the attention layer is consistent with the size of the extracted characteristic of the channel,

5. the features obtained by the two channels are simultaneously input into the full connection layer to be combined, and finally input into the classification layer to be classified, so that a correct rate result shown in fig. 6 is obtained, wherein the horizontal axis represents the number of experimental iterations, and the vertical axis represents the correct rate, wherein the W-CNN model marked as a circle, the W-BilSTM model marked as an arrow, the W-CNN-BilSTM model marked as a pentagon, the W-ATCNN-ATBilSTM model marked as an inverted triangle, and the model marked as a square are the models constructed by the method of the present invention, and the correct rate of the models on the data set is up to 92.67%.

Claims (5)

1. The text emotion analysis method based on the word vector and word vector mixed model is characterized by comprising the following steps of:

step 1, preprocessing a Chinese data set, and simultaneously training a Word vector and a Word vector matrix by using Word2 Vec;

step 2, training the word vector matrix x_1:lAnd word vector matrix w_1:lAs the input characteristics of the bidirectional long and short term memory network, the sequence characteristics of the text are learned, and the attention layer is accessed to optimize the characteristic vector;

step 3, training the word vector matrix x_1:lAnd word vector matrix w_1:lAs the input features of the convolutional neural network, performing convolution and pooling operations to learn local features of the text, and accessing an attention mechanism to obtain deep features of the text;

step 4, extracting the features s from the convolutional neural network layer with the attention mechanism_cAnd feature vectors s extracted by a bidirectional long-short term memory network layer introducing attention mechanism_lAnd fusing, inputting a softmax classification layer for classification, wherein the positive value is 1, the negative value is 0, and comparing and calculating with the text label to obtain the text classification accuracy.

2. The text emotion analysis method based on the Word vector and Word vector hybrid model as claimed in claim 1, wherein in step 1, preprocessing operation is performed on the chinese data set, and Word vector matrix are trained simultaneously using Word2Vec, comprising the steps of:

step 11, preprocessing the text to improve the quality of word vectors, including the processes of word stop and word segmentation;

step 12, using a Word2vec tool open by Google corporation for Word vector training work, using a Skip-gram model in the Word2vec tool to train Word vectors and Word vectors, and training the preprocessed corpus into 128-dimensional Word vectors and Word vectors, wherein a Word vector matrix and a Word vector matrix are respectively expressed as:

wherein the content of the first and second substances,

representing the join operator, l representing the length of the sentence, and taking a constant value of 60, x_1:lWord vector matrix, w, representing 60 × 128_1:lRepresenting a 60 × 128 word vector matrix.

3. The method for emotion analysis of text based on mixed model of word vector and word vector as claimed in claim 1, wherein in step 2, the trained word vector matrix x is used_1:lAnd word vector matrix w_1:lAs an input feature of BilSTM, learning a sequence feature of a text and accessing an attention layer optimization feature vector, comprising the following steps:

step 21, converting the word vector matrix x_1:lAnd word vector matrix w_1:lRespectively inputting the two-way long and short term memory network for training, extracting the features by using the two-way long and short term memory network to obtain the sequence features of the text, improving the convergence of the network and obtaining the hidden layer state h of the word vector of the two-way long and short term memory network_wiHidden state h of sum word vector_xiA matrix x of word vectors to be converted_1:lInputting the two-way long and short term memory network to train to obtain the hidden layer state h of the word vector of the two-way long and short term memory network_xiThe formula is expressed as:

f_t＝σ(w^fx_1:l+u^fh_t-1+b^f)

i_t＝σ(wⁱx_1:l+uⁱh_t-1+bⁱ)

o_t＝σ(w^ox_1:l+u^oh_t-1+b^o)

h_xt＝o_tΘtanh(c_t)

wherein sigma is sigmoid activation function, and the output is [0,1 ]]And determining how much information can pass, wherein tanh is a hyperbolic tangent function, theta operation symbols represent matrix multiplication operation, and the gate f is forgotten_tDeciding to "forget" information that is not important in the back-propagation, input gate i_tDetermines the information to be updated

Output gate o_tDetermining from the current cell state c_tOutput to hidden layer state h_xtThe content of (1). Current cell state c_tThat is, the past cell state c_t-1Merging with new memory, w⁽ⁱ⁾And u⁽ⁱ⁾Representing weights in the data processing process, a matrix w of word vectors to be converted_1:lInputting the two-way long and short term memory network for training to obtain the hidden layer state h of the word vector of the two-way long and short term memory network_wiThe formula is expressed as:

f_t＝σ(w^fw_1:l+u^fh_t-1+b^f)

i_t＝σ(wⁱw_1:l+uⁱh_t-1+bⁱ)

o_t＝σ(w^ow_1:l+u^oh_t-1+b^o)

h_wt＝o_tΘtanh(c_t)

wherein sigma is sigmoid activation function, and the output is [0,1 ]]And determining how much information can pass, wherein tanh is a hyperbolic tangent function, theta operation symbols represent matrix multiplication operation, and the gate f is forgotten_tDeciding to "forget" information that is not important in the back-propagation, input gate i_tDetermines the information to be updated

Output gate o_tDetermining from the current cell state c_tOutput to hidden layer state h_wtThe content of (1). Current cell state c_tThat is, the past cell state c_t-1Merging with new memory, w⁽ⁱ⁾And u⁽ⁱ⁾Representing weights in the data processing process;

step 22, the hidden layer state h of the word vector of the bidirectional long and short term memory network obtained in step 21 is used_wiHidden state h of sum word vector_xiFusing in a point-by-point manner; :

h_i＝[h_wi·h_xi]

step 23, hiding layer state h in the bidirectional long-short term memory network_iNonlinear transformation u_i，

u_i＝tanh(w_wh_i+b_w)

Wherein, w_wAnd b_wRepresenting a weight matrix and a weight vector;

step 24, using softmax function to pair u_iPerforming normalization operation to obtain attention matrix, i.e. output weight coefficient α of bidirectional long-short term memory network_i；

Wherein u is_wThe initialized weight matrix is shown;

step 25, output weight coefficient α of the bidirectional long and short term memory network layer obtained in step 24_iAnd hidden layer state h obtained in step 22_iMultiplying to obtain feature vector s extracted by bidirectional long-short term memory network layer with attention mechanism_l；

Wherein, α_iThe output weight coefficients are indicated.

4. The method for emotion analysis of text based on mixed model of word vector and word vector as claimed in claim 1, wherein in step 3, the trained word vector matrix x is used_1:lAnd word vector matrix w_1:lThe method is used as the input feature of a convolutional neural network, carries out convolution and pooling operations, learns the local feature of a text, and accesses an attention mechanism to acquire the deep feature, and comprises the following steps:

step 31, training the word vector matrix x_1:lAnd word vector matrix w_1:lRespectively inputting the input signals into a CNN network for training, and performing convolution operation by using a linear filter on an input matrix;

C_xi＝f(W·x_1:l+b)

C_wi＝f(W·w_1:l+b)

wherein, C_xiRepresenting the word vector characteristics, C, of the word vector after the convolution operation_wiThe character of the word vector obtained after the convolution operation is shown, b ∈ R shows an offset vector, W shows a convolution kernel with the height h being 3, 4, 5 and the width d being 128, f shows a nonlinear activation function, and a ReLu function is used as the activation function, which is expressed as:

f(x)＝max(0,x)

step 32, for the trained word vector matrix x_1:lAnd word vector matrix w_1:lAfter convolution operation, a word vector characteristic diagram C is obtained_xSum word vector feature map C_w，

C_x＝[C_x1,C_x2,…,C_xl]

C_w＝[C_w1,C_w2,…,C_wl]

Combining the local features extracted by the convolution kernels, and selecting 128 convolution kernels to obtain a plurality of feature maps representing different feature information for extracting more text features;

step 33, after completing the operation of step 32, matching the word vector feature map C_xSum word vector feature map C_wPerforming maximal pooling operation to obtain word vector inputGo out of characteristic c_xiSum word vector output feature c_wiThe method can reduce the characteristic dimension of the image,

c_xi＝max(C_x)

c_wi＝max(C_w)

then the fusion layer outputs the word vector to the feature c_xiSum word vector output feature c_wiThe point-by-point fusion is used,

c_i＝[c_xi·c_wi]

step 34, using tanh function to compare the characteristic c obtained in step 33_iNonlinear transformation u_ci，

u_ci＝tanh(w_wc_i+b_w)

Wherein, w_wAnd b_wRepresenting the weight matrix and weight vector, and then using the Softmax function to pair u_ciNormalization is carried out to obtain an output weight coefficient α of the CNN layer_ci；

Wherein u is_cwThe initialized weight matrix is shown;

step 35, outputting the weight coefficient α of the convolutional neural network layer obtained in the step 34_ciWith the feature c obtained in step 33_iMultiplying to obtain text features extracted after the convolutional neural network passes through the attention layer, and expressing the text features as a vector s_c，

Wherein, α_ciThe output weight coefficients are indicated.

5. The method for analyzing emotion of text based on mixed model of word vector and word vector as claimed in claim 1, wherein in step 4, the features s extracted from the convolutional neural network layer with attention mechanism introduced_cAnd introduction ofFeature vector s extracted by bidirectional long-short term memory network layer of attention mechanism_lFusing, inputting a softmax classification layer for classification, positively setting the classification to be 1 and negatively setting the classification to be 0, and comparing and calculating the classification with a text label to obtain the text classification accuracy, wherein the method comprises the following steps:

step 41, extracting features s of the convolutional neural network layer with attention mechanism_cAnd feature vectors s extracted by a bidirectional long-short term memory network layer introducing an attention mechanism_lInputting the data into a full connection layer by using a point-by-point fusion mode to obtain output characteristics;

x＝[s_l·s_c]

step 42, using a Dropout strategy, randomly selecting a part of the model to temporarily discard the part of the model from the network during model training, and not participating in parameter updating operation any more, and setting the Dropout rate to be 0.5, namely half of neurons in each iteration do not participate in calculation;

step 43, taking the feature x obtained in step 41 as an input of the classification layer, and calculating the probability size p of each text belonging to different categories by using a softmax function, which can be described as the following formula:

where the text is divided into 2 categories, w_kAnd b_kIs the weight and bias of the layer;

and 44, judging the text type, judging that the probability value p belongs to a larger type, positively being 1, and negatively being 0, and comparing the probability value p with the text label to calculate to obtain the text accuracy.

Images