Enterprise secondary industry classification method and system based on gating mechanism

Abstract

The present invention relates to the field of data processing technology, and discloses a method and system for enterprise secondary industry classification based on a gating mechanism, including: obtaining enterprise information to be classified; inputting the enterprise information to be classified into a trained enterprise secondary industry classification model, and outputting the enterprise secondary industry classification result; wherein the trained enterprise secondary industry classification model has the following working principle: obtaining the word feature vector of the enterprise information to be classified, and then extracting the context vector of the enterprise information from the word feature vector; then, splicing the word feature vector with the context vector; respectively extracting context features and local significant features from the spliced ​​vector; weighted fusion of the two extracted features, classifying the fused features, and obtaining the final classification result. The present invention reduces the manpower expenditure in the traditional secondary industry classification, shortens the system execution time of the enterprise industry classification, and the classification result is accurate and the system is safe.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for classifying secondary industries of enterprises based on a gating mechanism. Background Art

[0002] The statements in this section merely mention background art related to the present invention and do not necessarily constitute prior art.

[0003] With the development of the information age, more and more enterprises have emerged, and enterprises are also changing dynamically with the development of the economy. With the continuous development of enterprises, their industry categories often involve multiple fields. It is a difficult problem to classify the fields of these enterprises into the national economic industry classification categories. In particular, some policy enterprises require more detailed classification of enterprises, but the existing enterprise classification methods generally focus on understanding the surface model of words, while ignoring the deep mining of the potential information of the text, which leads to a series of classification errors. Secondly, there is a lot of redundant information in the enterprise data, which may involve information from other categories of industries, making the secondary industry classification of enterprises inaccurate.

[0004] In particular, few existing technologies target secondary enterprise classification, and a few target secondary enterprise categories, but the methods used are not accurate and run too slowly. Summary of the invention

[0005] In order to address the deficiencies of the prior art, the present invention provides a method and system for enterprise secondary industry classification based on a gating mechanism; to address the problem of models that focus on understanding the surface of words, while ignoring the deep mining of potential information in enterprise texts, which leads to a series of classification errors. Secondly, it addresses the problem of inaccurate enterprise classification caused by a large amount of redundant information in enterprise data, which may involve information from other categories of industries. In particular, few existing technologies target enterprise secondary classification, and a few target enterprise secondary categories, but the methods used have a low accuracy rate and run too slowly.

[0006] In a first aspect, the present invention provides a method for classifying secondary industries of enterprises based on a gating mechanism;

[0007] The enterprise secondary industry classification method based on the gating mechanism includes:

[0008] Obtain enterprise information to be classified;

[0009] Input the enterprise information to be classified into the trained enterprise secondary industry classification model, and output the enterprise secondary industry classification results;

[0010] Among them, the working principle of the trained enterprise secondary industry classification model is: obtain the word feature vector of the enterprise information to be classified, and then extract the context vector of the enterprise information from the word feature vector; then, splice the word feature vector with the context vector; extract the context features and local significant features of the spliced ​​vector respectively; perform weighted fusion on the two extracted features, classify the fused features, and obtain the final classification result.

[0011] In a second aspect, the present invention provides a secondary industry classification system for enterprises based on a gating mechanism;

[0012] The enterprise secondary industry classification system based on the gating mechanism includes:

[0013] An acquisition module is configured to: acquire enterprise information to be classified;

[0014] A classification module is configured to: input the enterprise information to be classified into the trained enterprise secondary industry classification model, and output the enterprise secondary industry classification result;

[0015] Among them, the working principle of the trained enterprise secondary industry classification model is: obtain the word feature vector of the enterprise information to be classified, and then extract the context vector of the enterprise information from the word feature vector; then, splice the word feature vector with the context vector; extract the context features and local significant features of the spliced ​​vector respectively; perform weighted fusion on the two extracted features, classify the fused features, and obtain the final classification result.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] 1. The enterprise secondary industry classification method based on the gating mechanism proposed in the present invention reduces the manpower expenditure in the traditional secondary industry classification, shortens the system execution time of enterprise industry classification, and the classification results are accurate and the system is safe.

[0018] 2. The parallel computing method of Bilstm, CNN and gating mechanism designed in the classification method proposed in the present invention can not only improve the classification accuracy of the secondary industry, but also effectively reduce the running time of the word context attention mechanism, thereby improving the classification efficiency of the system.

[0019] 3. The present invention provides a method and system for enterprise classification based on the attention mechanism fused with the gating mechanism to solve the problem that the model focuses on understanding the surface of words, but ignores the deep mining of the potential information of enterprise text, which leads to a series of classification errors. Secondly, it solves the problem that there is a large amount of redundant information in enterprise data, which may involve information of other categories of industries, leading to inaccurate enterprise classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0021] Figure 1 It is a structural composition diagram of the present invention. DETAILED DESCRIPTION

[0022] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0023] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0025] In this embodiment, all data is obtained in compliance with laws and regulations and based on the user's consent, and is used legally.

[0026] Embodiment 1

[0027] This embodiment provides a method for classifying secondary industries of enterprises based on a gating mechanism;

[0028] The enterprise secondary industry classification method based on the gating mechanism includes:

[0029] S101: Obtain enterprise information to be classified;

[0030] S102: Input the enterprise information to be classified into the trained enterprise secondary industry classification model, and output the enterprise secondary industry classification result;

[0031] Among them, the working principle of the trained enterprise secondary industry classification model is:

[0032] Obtain a word feature vector of the enterprise information to be classified, and then extract a context vector of the enterprise information from the word feature vector; then, concatenate the word feature vector and the context vector;

[0033] Contextual features and local salient features are extracted from the concatenated vectors respectively; the two extracted features are weighted fused, and the fused features are classified to obtain the final classification result.

[0034] Furthermore, if Figure 1 As shown, the network structure of the enterprise secondary industry classification model includes: word embedding layer;

[0035] The output of the word embedding layer is connected to the input of the attention mechanism layer;

[0036] The output end of the word embedding layer and the output end of the attention mechanism layer are both connected to the input end of the splicing unit;

[0037] The output end of the splicing unit is connected to the input end of the first neural network and the input end of the second neural network respectively;

[0038] The output end of the first neural network is connected to the input end of the first gating unit;

[0039] The output end of the second neural network is connected to the input end of the second gating unit;

[0040] The output end of the first gating unit and the output end of the second gating unit are both connected to the classifier through a fusion unit.

[0041] Furthermore, the word embedding layer is implemented using word2vec, which is used to input the enterprise information to be classified, and perform feature vector extraction on the enterprise information to be classified to extract word feature vectors.

[0042] Exemplarily, the word embedding layer: removes punctuation marks and stop words from the enterprise information, then performs word segmentation on the enterprise information, trains a word embedding model, and Word2vec is used as the word embedding model.

[0043] Using the skip-gram model training mode, the word vector of the context word is predicted by the word vector of the current word, and the word vector of the context word is solved using the conditional probability value of the current word vector.

[0044] In order to calculate the word context vector, the enterprise information is divided into enterprise sentence sets according to punctuation marks.

[0045] The input enterprise information sequence is represented as D = {w 1，1 , w 1，2 , ..., w k，m}, where w k，m Represented as the mth word in the kth sentence.

[0046] Since the sentences in the enterprise sentence set may have different sequence lengths, the sentence length of enterprise information is defined as m, w is the word vector matrix trained by the word2vec model, w∈R d×V , d represents the word embedding dimension, and V represents the number of words in the word list.

[0047] Each word w in the enterprise information sequence k，m There is a corresponding word vector x k，m .

[0048] The enterprise information embedding layer is finally expressed as D = {x 1，1 , x 1，2 , ..., x k，m}, as the input of the attention mechanism layer.

[0049] Furthermore, the attention mechanism layer is used to extract a context vector of enterprise information from a word feature vector.

[0050] For example, in the process of secondary classification of enterprises, each word in the enterprise data contributes differently to the secondary industry classification of the enterprise, and keywords play an important role in the secondary industry classification of the enterprise, more important than other words.

[0051] Therefore, a word vector attention mechanism is introduced to calculate the semantic correlation coefficient between each word and other words in the enterprise information data, and assign word vector weights according to the correlation coefficient.

[0052] The weighted linear combination of word vectors constitutes the word context vector, which is concatenated with the original word vector as a new word vector feature representation.

[0053] In this way, the word context vector can make words with larger weights get more attention, considering that there are less semantic connections between words in sequence.

[0054] In this way, the word context vector can make words with larger weights get more attention.

[0055] Therefore, the enterprise information is divided into sentence sets, and the context vector of the sentence words is calculated to highlight the information of the sentence keywords. Suppose the enterprise information sequence information consists of K sentences, and the length of each sentence is m. i,j is the word vector representation of the jth word in the i-th sentence, and the word context variable g i,j :

[0056]

[0057] α ij,k is the attention weight, α ij,krepresents the correlation coefficient between the jth word and the kth word in the ith sentence.

[0058]

[0059] The score function is used to calculate the correlation coefficient between the current word and other words to measure the words that are closely related to the current word.

[0060] A high score means that the word has a greater weight in the context vector of the current word. The score value is calculated as follows:

[0061]

[0062] v a , w a , w u is the training parameter, and T is the transposed matrix. i，j and x i，k Use w respectively a and w u Perform matrix mapping and transformation. Input the transformation result into the tanh activation function to obtain the connection feature.

[0063] Finally, through v a Get x i，j and x i，k score.

[0064] Furthermore, the concatenation unit is used to concatenate the word feature vector and the context vector of the enterprise information.

[0065] Concatenate the word context vector with the original word vector to get a new word vector representation x i，j =[x i，j , g i，j ]. The sentence sets are merged into a new enterprise information sequence, the length of the enterprise information sequence is n, and the enterprise information sequence is represented by D = {x′1, ..., x′ n}, as the input of the hybrid neural network layer (the first neural network and the second neural network).

[0066] Furthermore, the first neural network is implemented using a bidirectional long short-term memory network BiLSTM, which is used to extract context features from the concatenated vector.

[0067] For example, BiLSTM is used to capture text context features. The BiLSTM unit consists of a forward LSTM unit and a backward LSTM unit, which obtain information from two opposite directions. The LSTM unit consists of a forget gate, an input gate, and an output gate. Given an input x′ at time t t The hidden state h t The calculation is as follows:

[0068] f t =σ(W f [x′ t ,h t-1 ]+b f )

[0069] i t =σ(W i [x′ t ,h t-1 ]+b i )

[0070]

[0071]

[0072] O t =σ(W o [x′ t ,h t-1 ]+b o )

[0073] h t =O t *tanh(C t )

[0074] Among them, f t ,i t , O t They are the states of the forget gate, input gate, and output gate at time t respectively.

[0075] W f , W i , W o , W c is the weight matrix corresponding to each component, b f , b i , b o , b c The deviation vectors corresponding to each component respectively. is the temporary cell state, calculated by the tanh function. C t is the memory cell state at time t, and σ is the sigmoid function. For BiLSTM, the hidden layer outputs of the forward LSTM unit and the backward LSTM unit are and Calculate by the above formulas respectively. Then and Spliced ​​together, we get the hidden layer output h of BiLSTM at time t t , the splicing formula is Finally, the output of BiLSTM is h = {h1, ..., h n}.

[0076] Furthermore, the second neural network is implemented using a convolutional neural network (CNN) to extract significant features of the industry category from the concatenated vector.

[0077] Exemplarily, a convolutional neural network (CNN) is used to extract significant features of industry categories from enterprise information context features. The convolutional neural network (CNN) consists of a convolution layer and a pooling layer, which are used to capture key category label features of industry classification. The convolution kernel and the local window of the input data are used for convolution calculation.

[0078] C i =f(F·h i：i+l-1 +b).

[0079] Among them, F represents the convolution kernel, F∈R l×d′ , where l is the height of the convolution kernel, d′ is the width of the convolution kernel, and its size is the output feature dimension of the BiLSTM unit. b represents the bias parameter, f represents the RELU nonlinear function, and h i：i+l-1 represents the output vector of the BiLSTM hidden layer from i to i+l-1, represents the convolution operation, and C i Represents the result of the convolution calculation.

[0080] The convolution window slides down with a step size of 1 to obtain the local feature vector C = (C1, ..., C i-l+1 ).

[0081] Multiple convolution kernels are used for convolution operation, the number of which is N, thus obtaining N local feature vectors. The maximum pooling layer can retain the key information of the text and extract the maximum eigenvalue, M i =max(C i ), all eigenvalues ​​are combined into M = (M1, ..., M N ) as the output of CNN.

[0082] Furthermore, the first gating unit is implemented using a gating mechanism layer Gate1, which is used to assign weights to context features.

[0083] Furthermore, the second gating unit is implemented using a gating mechanism layer Gate2, which is used to assign weights to significant features of industry categories.

[0084] The gating mechanism layer: Some sentences in the enterprise information may involve other industry categories. CNN may obtain significant features of other industry categories, which will lead to classification errors. Therefore, it is necessary to retain contextual features. According to the different degrees of dependence of each enterprise information sequence on contextual features, a gating device is introduced here to assign weights to the text significant features captured by CNN and the text context features captured by BiLSTM. The combination of these two features further improves the accuracy of industry classification.

[0085] The calculation process of the gate control device is as follows:

[0086] α1=sigmoid(W·H+b)

[0087] α2=sigmoid(W·M+b)

[0088] z=α1·H+α2·M

[0089] The gating mechanism layer consists of a sigmoid function, where W and b are the weight matrix and bias term. α1 and α2 are the weight vectors of BiLSTM output and CNN output, respectively, and the relationship between the two weight vectors is set to α1+α2=1. H is the output of BiLSTM at the last moment, M is the output of CNN, and z is the text feature vector that combines the contextual features of enterprise information and the salient features of text.

[0090] BiLSTM is used to capture the contextual features of enterprise information, and CNN is used to capture the local salient features of enterprise information. Finally, the industry category information in the question text that may be related to other sentences is used to capture the salient features of other industry categories, retaining the contextual feature classification of enterprise information, and using a gating device to assign weights to BiLSTM and CNN output features based on the different dependencies of enterprise information on contextual features, so as to select information features that are more useful for the secondary classification of enterprise industries.

[0091] Furthermore, the fusion unit is used to perform weighted summation of the context features and the significant features of the industry category.

[0092] Furthermore, the classifier is implemented using a Softmax classifier.

[0093] The classification layer: inputs the feature vector z into the fully connected layer, and maps the feature vector to the industry category label space through feature space transformation.

[0094] Finally, the softmax function is used to determine the industry category, where w and b are weights and biases, respectively, and k is the industry category. The model parameters are updated according to the classification results. The formula is as follows:

[0095]

[0096] Furthermore, the training process of the trained enterprise secondary industry classification model includes:

[0097] Constructing a training set; the training set includes: enterprise information with known secondary industry classification labels of the enterprise; wherein the enterprise information refers to the business scope and enterprise information of the enterprise in its production and operation activities;

[0098] The training set is input into the enterprise secondary industry classification model to train the model. When the loss function of the model no longer decreases, or when the number of iterations reaches the set number, the training is stopped to obtain the trained enterprise secondary industry classification model.

[0099] Furthermore, the training set is constructed, and data is classified and marked according to the business scope and enterprise information of each industry in production and operation activities collected according to the set classification standard. The business scope of each enterprise in production and operation activities is obtained through the Internet, and the major categories in the "National Economic Industry Classification" GB / T4754-2017 are marked as secondary industry categories according to similarity. According to this standard, the 97 major categories under the enterprise national economic industry classification are divided into categories, and the obtained enterprise information is marked with the enterprise secondary industry classification according to the newly divided categories.

[0100] Since the categories in the national economic industry classification standard are too broad and the scope of the major categories under the categories is too similar, the formulated standard is to classify the major categories of the national economic industry classification according to similarity; for example:

[0101] 13-Agricultural and sideline food processing industry, 14-Food manufacturing industry, 15-Wine, beverage and refined tea manufacturing industry, 16-Tobacco products industry, 17-Textile industry, 18-Textile clothing and apparel industry, 19-Leather, fur, feather and its products and shoemaking industry, 20-Wood processing and wood, bamboo, rattan, palm and straw products industry, 21-Furniture manufacturing industry, 22-Papermaking and paper products industry, 23-Printing and recording media reproduction industry, 24-Cultural, educational, arts and crafts, sports and entertainment products manufacturing industry are all included in the light textile industry;

[0102] 25-petroleum, coal and other fuel processing industry, 26-chemical raw materials and chemical products manufacturing industry, 27-pharmaceutical manufacturing industry, 28-chemical fiber manufacturing industry, 29-rubber and plastic products industry, 30-non-metallic mineral products industry, 31-ferrous metal smelting and rolling processing industry, 32-non-ferrous metal smelting and rolling processing industry, 33-metal products industry are classified as resource processing industry;

[0103] 34-General equipment manufacturing, 35-Special equipment manufacturing, 36-Automobile manufacturing, 37-Railway, shipbuilding, aerospace and other transportation equipment manufacturing, 38-Electrical machinery and equipment manufacturing, 39-Computer, communication and other electronic equipment manufacturing, 40-Instrumentation manufacturing, 41-Other manufacturing, 42-Waste resources comprehensive utilization, 43-Metal products, machinery and equipment repair industry are classified as machinery and electronics manufacturing;

[0104] The remaining major categories will keep their original category names unchanged, and the remaining major categories refer to 01-agriculture, 02-forestry, 03-animal husbandry, 04-fishery...12-other mining industries; and 44-electricity, heat production and supply, 45-gas production and supply, 46-water production and supply...94-social security, 95-mass organizations, social groups and other member organizations, 96-grassroots mass autonomous organizations and other organizations, 97 international organizations.

[0105] According to this standard, 97 major categories under 20 categories are divided into categories, and the obtained enterprise information is labeled with secondary industry classification according to the newly divided categories.

[0106] Obtain the business scope of each enterprise in production and operation activities through the Internet, and mark the major categories in the national economic industry classification as secondary industry categories based on similarity. According to this standard, 97 major categories under 20 categories are divided into categories, and the obtained enterprise information is labeled for the secondary industry classification of enterprises according to the newly divided categories. Prepare training data sets and verification sets, input the training data sets into the industry classification model for training, and obtain the final industry label classification model based on the Attention BiLSTM and CNN Gate model, referred to as the ABCG model. Obtain the enterprise information to be classified, input the enterprise information to be classified into the final enterprise secondary classification ABCG model for industry classification, and obtain the secondary industry category classification results corresponding to the enterprise information to be classified.

[0107] Embodiment 2

[0108] This embodiment provides a secondary industry classification system for enterprises based on a gating mechanism;

[0109] The enterprise secondary industry classification system based on the gating mechanism includes:

[0110] An acquisition module is configured to: acquire enterprise information to be classified;

[0111] A classification module is configured to: input the enterprise information to be classified into the trained enterprise secondary industry classification model, and output the enterprise secondary industry classification result;

[0112] Among them, the working principle of the trained enterprise secondary industry classification model is: obtain the word feature vector of the enterprise information to be classified, and then extract the context vector of the enterprise information from the word feature vector; then, splice the word feature vector with the context vector; extract the context features and local significant features of the spliced ​​vector respectively; perform weighted fusion on the two extracted features, classify the fused features, and obtain the final classification result.

[0113] It should be noted that the acquisition module and the classification module correspond to steps S101 to S102 in Embodiment 1, and the examples and application scenarios implemented by the modules and the corresponding steps are the same, but are not limited to the contents disclosed in Embodiment 1. It should be noted that the modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0114] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. The enterprise secondary industry classification method based on the gating mechanism is characterized by: include: Obtain enterprise information to be classified; Input the enterprise information to be classified into the trained enterprise secondary industry classification model, and output the enterprise secondary industry classification results; The working principle of the trained enterprise secondary industry classification model is as follows: obtain the word feature vector of the enterprise information to be classified, and then extract the context vector of the enterprise information from the word feature vector; then, concatenate the word feature vector with the context vector; extract the context feature and the local significant feature from the concatenated vector respectively; perform weighted fusion on the two extracted features, classify the fused features, and obtain the final classification result; The network structure of the enterprise secondary industry classification model includes: a word embedding layer; The output of the word embedding layer is connected to the input of the attention mechanism layer; The output end of the word embedding layer and the output end of the attention mechanism layer are both connected to the input end of the splicing unit; The output end of the splicing unit is connected to the input end of the first neural network and the input end of the second neural network respectively; The output end of the first neural network is connected to the input end of the first gating unit; The output end of the second neural network is connected to the input end of the second gating unit; The output end of the first gating unit and the output end of the second gating unit are both connected to the classifier through a fusion unit; The first gating unit is implemented by the gating mechanism layer Gate1, and is used to assign weights to context features; the second gating unit is implemented by the gating mechanism layer Gate2, and is used to assign weights to significant features of industry categories; the fusion unit is used to perform weighted summation of context features and significant features of industry categories; the classifier is implemented by the Softmax classifier.

2. The enterprise secondary industry classification method based on the gating mechanism as claimed in claim 1 is characterized in that: The word embedding layer is implemented by word2vec, which is used to input the enterprise information to be classified, and extract feature vectors from the enterprise information to be classified to extract word feature vectors.

3. The enterprise secondary industry classification method based on the gating mechanism as claimed in claim 1 is characterized in that: The attention mechanism layer is used to extract the context vector of enterprise information from the word feature vector.

4. The enterprise secondary industry classification method based on the gating mechanism as claimed in claim 1 is characterized in that: The concatenation unit is used to concatenate the word feature vector and the context vector of the enterprise information.

5. The enterprise secondary industry classification method based on the gating mechanism as claimed in claim 1 is characterized in that: The first neural network is implemented using a bidirectional long short-term memory network BiLSTM, which is used to extract context features from the concatenated vector.

6. The enterprise secondary industry classification method based on the gating mechanism as claimed in claim 1 is characterized in that: The second neural network is implemented using a convolutional neural network (CNN) and is used to extract significant features of industry categories from the concatenated vectors.

7. The enterprise secondary industry classification method based on the gating mechanism as claimed in claim 1 is characterized in that: The trained enterprise secondary industry classification model; The training process includes: Constructing a training set; the training set includes: enterprise information with known secondary industry classification labels of the enterprise; wherein the enterprise information refers to the business scope and enterprise information of the enterprise in its production and operation activities; The training set is input into the enterprise secondary industry classification model to train the model. When the loss function of the model no longer decreases, or when the number of iterations reaches the set number, the training is stopped to obtain the trained enterprise secondary industry classification model.

8. The enterprise secondary industry classification system based on the gating mechanism includes: An acquisition module is configured to: acquire enterprise information to be classified; A classification module is configured to: input the enterprise information to be classified into the trained enterprise secondary industry classification model, and output the enterprise secondary industry classification result; The working principle of the trained enterprise secondary industry classification model is as follows: obtain the word feature vector of the enterprise information to be classified, and then extract the context vector of the enterprise information from the word feature vector; then, concatenate the word feature vector with the context vector; extract the context feature and the local significant feature from the concatenated vector respectively; perform weighted fusion on the two extracted features, classify the fused features, and obtain the final classification result; The network structure of the enterprise secondary industry classification model includes: a word embedding layer; The output of the word embedding layer is connected to the input of the attention mechanism layer; The output end of the word embedding layer and the output end of the attention mechanism layer are both connected to the input end of the splicing unit; The output end of the splicing unit is connected to the input end of the first neural network and the input end of the second neural network respectively; The output end of the first neural network is connected to the input end of the first gating unit; The output end of the second neural network is connected to the input end of the second gating unit; The output end of the first gating unit and the output end of the second gating unit are both connected to the classifier through a fusion unit; The first gating unit is implemented by the gating mechanism layer Gate1, and is used to assign weights to context features; the second gating unit is implemented by the gating mechanism layer Gate2, and is used to assign weights to significant features of industry categories; the fusion unit is used to perform weighted summation of context features and significant features of industry categories; the classifier is implemented by the Softmax classifier.

Images