Target enterprise screening method, terminal and storage medium based on the industry

Abstract

The present invention relates to the field of information classification technology, and discloses a method, terminal and storage medium for screening target enterprises based on the industry to which they belong. The screening method first obtains relevant information of the enterprise to be screened, and the relevant information includes the business scope, the industry to which the enterprise to be screened, and several patent texts; then constructs a text set of the enterprise to be screened based on the relevant information; performs semantic feature extraction on each group of multi-dimensional texts in the text set of the enterprise to be screened; inputs the semantic feature vector extracted from each group of multi-dimensional texts into a trained classification model, and obtains an emerging industry category corresponding to each group of multi-dimensional texts; based on the emerging industry categories corresponding to all multi-dimensional texts, generates a classification result of the emerging industry to which the enterprise to be screened belongs, and judges whether the enterprise to be screened meets the target conditions based on the classification result. The present invention improves the accuracy of enterprise screening.

Description

Technical Field

[0001] The present invention relates to the technical field of information classification, and in particular to a method for screening target enterprises based on the industry to which they belong, as well as a computer terminal and a computer-readable storage medium applying the method. Background Art

[0002] Due to the current situation of "large number" (large proportion of total enterprises) and "lack of credit" of SMEs, financial institutions face great challenges in screening target customers. Financial institutions tend to provide financial support to SMEs that fit the strategic layout and have high development potential.

[0003] Most of the existing methods obtain qualified technology companies from the list of companies with relevant qualifications for strategic emerging industries disclosed by relevant departments. However, many high-quality companies have not yet obtained the title of technology-based enterprises certified by relevant departments due to various reasons such as their short establishment period. Therefore, the number of companies screened out by existing methods is small, and most of them cannot be located in specific emerging industry categories. The more traditional idea is to summarize the corresponding keywords of emerging industries and screen out companies with relevant qualifications through rule matching. However, the keyword method is too rigid and very easy to miss. At the same time, there is a phenomenon of "taking things out of context", which leads to the misclassification of categories, and the actual effect on industrial implementation is not ideal.

[0004] Therefore, how to more quickly and accurately identify and screen high-quality science and technology enterprises that meet the requirements from the massive group of small, medium and micro enterprises has become a problem that needs to be solved urgently. Summary of the invention

[0005] In order to solve the technical problems existing in the prior art, the present invention provides a method, terminal and storage medium for screening target enterprises based on the industry to which they belong. The present invention can comprehensively utilize information such as patents, business scope, and industry to which they belong to process the classification tasks of emerging industries, and accurately identify and screen target enterprises that meet the requirements based on the classification results.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention discloses a method for screening target enterprises based on the industries they belong to, comprising the following steps, namely S1 to S5.

[0008] S1. Obtain relevant information of the enterprise to be screened, including the business scope, industry to which the enterprise belongs, and several patent texts.

[0009] S2. Construct a text collection of the enterprise to be screened based on the relevant information, and the text collection includes several groups of multi-dimensional texts; each group of multi-dimensional texts includes the business scope of the enterprise to be screened, the industry to which it belongs, and the patent product name and abstract corresponding to a patent text.

[0010] S3. Extract semantic features from each group of multi-dimensional texts in the text collection of the screened enterprise.

[0011] S4. Input the semantic feature vector extracted from each group of multi-dimensional texts into the trained classification model to obtain the probability distribution of the group of multi-dimensional texts belonging to each category in a custom emerging industry classification system, thereby obtaining an emerging industry category corresponding to each group of multi-dimensional texts.

[0012] S5. Based on the emerging industry categories corresponding to all multi-dimensional texts, generate the classification results of the emerging industries to which the enterprises to be screened belong, and determine whether the enterprises to be screened meet the target conditions based on the classification results.

[0013] As a further improvement of the above solution, the customized emerging industry classification system includes multiple first-level emerging industry categories; each first-level emerging industry category includes multiple subordinate second-level emerging industry categories.

[0014] In step S4, after calculating the probability distribution of each group of multi-dimensional texts belonging to each category in the self-defined emerging industry classification system, the secondary emerging industry category with the largest probability is selected as the emerging industry category corresponding to the group of multi-dimensional texts.

[0015] As a further improvement of the above scheme, in step S5, if the probability that any group of multi-dimensional texts belongs to each category in the self-defined emerging industry classification system is lower than the preset probability threshold, the group of multi-dimensional texts is marked as non-emerging industry texts; if the proportion of non-emerging industry texts in the text set of the enterprise to be screened is P 1 is greater than a preset upper limit of the ratio c ,Right now P 1> c , then the enterprise to be screened is determined to be a non-target enterprise, that is, it does not meet the target conditions; where, , J1 is the number of non-emerging industry texts in the text set of the enterprise to be screened, and J is the total number of multi-dimensional texts of the enterprise to be screened.

[0016] As a further improvement of the above scheme, the classification results of the emerging industries to which the enterprise belongs include three labels: the first emerging industry label, the second emerging industry label and the third emerging industry label.

[0017] In step S5, if P 1≤ c, then the enterprise to be screened is determined to be a target enterprise, that is, it meets the target conditions, and statistical analysis is performed on the emerging industry categories corresponding to the J2 group multi-dimensional texts of the enterprise to be screened to obtain the proportion of each emerging industry category P 2. Then, the proportion of each emerging industry category P 2 and a preset ratio interval [ a 1, a 2] Make a comparison and make the following decisions:

[0018] Will satisfy P 2> a The emerging industry category of 2 is used as the first emerging industry label for the enterprises to be screened.

[0019] Will satisfy a 1≤ P 2≤ a The emerging industry category of 2 is used as the second emerging industry label for the enterprises to be screened.

[0020] Will satisfy P 2< a 1’s emerging industry category is used as the third emerging industry label for the enterprises to be screened.

[0021] in, , J2 is the number of multi-dimensional texts in the text set of the enterprise to be screened that are not marked as non-emerging industry texts, that is, J2=J-J1, x is the number of times each emerging industry category appears in the multidimensional text of group J2.

[0022] As a further improvement of the above scheme, after the statistical analysis, if the enterprise to be screened has the first emerging industry label and / or the second emerging industry label, the enterprise is determined to be the first target enterprise; if the enterprise to be screened does not have the first emerging industry label and the second emerging industry label, the enterprise is determined to be the second target enterprise.

[0023] As a further improvement of the above solution, in step S2, the text set of the enterprise to be screened is represented as T , , t j Indicates j Group multi-dimensional text, j ∈[1,J]; J is the total number of multidimensional texts, that is, the total number of patent texts of the enterprise to be screened, J>1; multidimensional text , Indicates the business scope of the enterprise to be screened, Indicates the industry to which the enterprise to be screened belongs. Indicates j The name of the patent product corresponding to the patent text, Indicatesj The abstracts of patent texts are obtained by using a generative model to obtain the patent product name corresponding to each patent text.

[0024] As a further improvement of the above solution, in step S4, the training method of the classification model includes the following steps, namely S41~S45.

[0025] S41. Constructing enterprise classification data set; the enterprise classification data set consists of a sample text set of several enterprises T’ , and each T’ It is composed of the label set of the actual industry category corresponding to all multi-dimensional sample texts; among them, T’ Each group of multi-dimensional sample texts in is constructed by referring to steps S1~S2.

[0026] S42. Extract the semantic feature vector of each group of multi-dimensional sample texts; k The group of multi-dimensional sample texts is represented as , k ∈[1,K], K is T’ The total number of multi-dimensional sample texts in is obtained using the vector representation model t k The semantic feature vector of ; , , , Sequentially indicates k The business scope feature vector, industry feature vector, patent product name feature vector, and patent abstract feature vector of the group of multi-dimensional sample texts.

[0027] S43. The factor decomposition bilinear pooling method is used to fuse the business scope feature vector and the industry feature vector of each group of multidimensional sample texts to obtain the enterprise features of each group of multidimensional sample texts, and the patent product name vector and the patent abstract feature vector of each group of multidimensional samples are fused to obtain the patent features of each group of multidimensional sample texts.

[0028] S44. Introduce the attention fusion mechanism to obtain the overall characteristics of adaptive fusion based on the enterprise characteristics and patent characteristics of each group of multi-dimensional sample texts; k The overall feature vector of the group of multi-dimensional sample texts is , the expression is as follows:

[0029] ;

[0030] ;

[0031] ;

[0032] In the formula, and Respectively k Group the enterprise feature vectors and patent feature vectors of multi-dimensional sample texts; for The latent feature representation of for The weight matrix of for The corresponding bias vector; for The latent feature representation of for The weight matrix of for The corresponding bias vector; for The weight matrix of represents the hyperbolic tangent function, Represents an exponential function.

[0033] Softmax is selected as the activation function to calculate the probability distribution of each group of multi-dimensional sample texts belonging to each category in the self-defined emerging industry classification system, and the cross entropy function is selected as the loss function. The training is implemented by back propagation and gradient descent. The optimization direction is to make the value of the cross entropy function smaller. The expression formula is as follows:

[0034] ;

[0035] ;

[0036] In the formula, p For multi-dimensional sample text t k The probability distribution of belonging to each category in the self-defined emerging industry classification system; W p is the probability distribution p The weight matrix of b p is the probability distribution p The bias vector of Loss Indicates loss; represents the total number of multi-dimensional sample texts in the training set divided by the enterprise classification dataset, Indicates the training set The probability value corresponding to the industry category to which the data actually belongs, .

[0037] As a further improvement of the above solution, in step S43, k Enterprise feature vector of group multi-dimensional sample text The expression is:

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] In the formula, Represents intermediate features In Dimension The value of , Intermediate feature Dimensions; Represents vector transpose; represents the learnable weight matrix, , which means that the matrix Factorize into two low-rank matrices and ; is a vector of all 1s, represents the Hadamard product; represents the bias variable; and Respectively indicated by and A third-order tensor composed of

[0043] , ;

[0044] and Respectively indicate and The two-dimensional matrix obtained by tiling the second and third dimensions of Indicates the use of a size of One-dimensional non-overlapping window pairs Perform summing and pooling operations; express After power law normalization, the function To determine The symbol of; function Indicates the Euclidean distance.

[0045] No. k Patent feature vectors of multi-dimensional sample texts The expression is:

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] in, Represents intermediate features In Dimension The value of represents the learnable weight matrix, , which means that the matrix Factorize into two low-rank matrices and ; and Respectively indicated by and A third-order tensor composed of , ; and Respectively indicate that and The two-dimensional matrix obtained by tiling the second three-dimensional; express Results after power law normalization.

[0051] The present invention also discloses a computer terminal, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the target enterprise screening method based on the industry to which it belongs as described above are implemented.

[0052] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the steps of the target enterprise screening method based on the industry to which it belongs as described above are implemented.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. Existing methods generally only use the company name, business scope, and introduction to directly derive the company's strategic emerging label, without considering the important position of scientific and technological achievements in the identification of strategic emerging labels, which is too "arbitrary". The screening method disclosed in the present invention introduces patent information to characterize the company's scientific and technological achievements, and at the same time adopts a smaller granularity. By combining each patent of the company with the business scope and the industry to which it belongs to create a set of corresponding multi-dimensional texts, and combining a custom emerging industry classification system to classify the emerging industries to which the company belongs, it is then comprehensively judged based on all classification results whether the company meets the target conditions, thereby avoiding the problem of a single classification being too "arbitrary" and improving the accuracy of screening.

[0055] 2. The present invention gradually obtains the feature representation of each group of multi-dimensional texts in a hierarchical progressive manner. First, the two groups of features are refined and fused by factor decomposition and bilinear pooling to obtain enterprise features and patent features; then the interactive process of enterprise features and patent features is constructed by adaptive fusion; finally, the overall features of each group of multi-dimensional texts are obtained. In this way, the accuracy of feature representation is effectively improved.

[0056] 3. This screening method can directly screen out some potential non-high-quality science and technology innovation enterprises in the early stage by assessing the proportion of non-emerging industry texts in the enterprises to be screened. On this basis, by statistically analyzing the single-category distribution of emerging industry texts, further detailed screening of target enterprises is carried out. Whether they belong to the first target enterprise or the second target enterprise can be used as an evaluation indicator for users (financial institutions) to provide a reliable basis for their judgment on whether to provide financial services to target enterprises and the type of services.

[0057] 4. This screening method determines the classification system in a user-defined manner, which can make the industry categories in the classification results easy to apply in corporate products, thereby helping financial institutions make reference decisions.

[0058] 5. The screening method can use a generative model to efficiently summarize the patent product name as a classification feature. Whether it is manual review or customer use, the patent product name is a very intuitive classification evidence, which can help users quickly judge the accuracy of the emerging industry classification and has strong interpretability. In addition, as a by-product of the present invention, the patent product name can be subsequently applied to the industrial chain series. For example, financial institutions hope to find companies in the new energy industry chain. Through fuzzy queries on product names, such as "new energy", it is very convenient to retrieve companies in the industry chain, understand whether they are major, related or suspected, and trace the corresponding patents.

[0059] 6. The computer terminal and computer-readable storage medium disclosed in the present invention can apply the above screening method and can produce the same beneficial effects as the above screening method, which will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a flow chart of the method for screening target enterprises based on their industries in Example 1 of the present invention.

[0061] Figure 2 It is a schematic diagram of the structure of a computer terminal in Embodiment 2 of the present invention. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] Example 1

[0064] See also Figure 1 , an embodiment of the present invention provides a method for screening target enterprises based on the industry they belong to, comprising the following steps, namely S1~S5.

[0065] S1. Obtain relevant information of the enterprise to be screened, including the business scope, industry to which the enterprise belongs, and several patent texts.

[0066] In this embodiment, the relevant information of the enterprise to be screened can be captured by means of web crawlers or the like, or it can be obtained through the internal database of the user (such as a financial institution). Among them, the business scope and the industry to which it belongs are mainly used for the initial screening of the enterprise in terms of scale and field; the patent text can be an authorized and currently valid patent disclosure text, which can ensure the reliability of subsequent classification. It should be noted that the business scope and the industry to which it belongs can generally be obtained from some third-party enterprise query websites. If an enterprise does not have any patent texts that meet the above requirements, or the number of patent texts is lower than a certain threshold, the enterprise can be directly judged as a non-target enterprise.

[0067] S2. Construct a text collection of the enterprise to be screened based on the relevant information, and the text collection includes several groups of multi-dimensional texts; each group of multi-dimensional texts includes the business scope of the enterprise to be screened, the industry to which it belongs, and the patent product name and abstract corresponding to a patent text.

[0068] The text collection of the enterprise to be screened can be expressed as T , , t j Indicates j Group multi-dimensional text, j ∈[1,J]; J is the total number of multidimensional texts, that is, the total number of patent texts of the enterprise to be screened, J>1; multidimensional text , Indicates the business scope of the enterprise to be screened, Indicates the industry to which the enterprise to be screened belongs. Indicates j The name of the patent product corresponding to the patent text, Indicates j Abstract of the patent text.

[0069] It should be noted that since the same patent text can include multiple groups of subject names, some of which have lengthy attributives, traditional solutions can only extract some fragmentary information from the text and then simply combine them together. It is very easy for product names to be redundant, messy, incoherent, or even "incomprehensible", which not only affects the effect of subsequent classification models, but also makes it impossible to directly apply them downstream. Therefore, generative models (such as large language models such as ChatGPT) can be used to summarize the patent product names. For example, the patent name "A rainwater pipe desilting robot based on path planning" can be summarized as "industrial robot" or a patent product name with more concise attributives through a generative model. Such a name has a certain degree of generalization and representativeness, which helps financial institutions quickly screen companies in a certain field or companies related to a certain product.

[0070] In this embodiment, in order to ensure the reliability and accuracy of the generated patent product name, a generative model with a UNILM structure can be used for fine-tuning training. Specifically, a large amount of patent sample data is first collected, and the initial version of the patent product name is generated through large language models such as ChatGPT. After manual review, the final corpus in the form of "patent sample-patent product name" is obtained, and then put into the generative model with a UNILM structure for training and fine-tuning, and finally a product name generator dedicated to this embodiment is obtained. For a new patent text, it only needs to be input into the generator to obtain the patent product name and apply it downstream.

[0071] S3. Extract semantic features from each group of multi-dimensional texts in the text collection of the screened enterprise.

[0072] In some embodiments, a pre-trained model can be used to extract semantic features from each group of multi-dimensional texts to obtain a semantic feature vector for each group of multi-dimensional texts. Pre-trained models have been widely used in deep learning, which can help the model converge quickly and improve the final effect of the model. In this embodiment, the Albert-base model is selected as the pre-trained model. The advantage of this model is that it is lightweight enough and the extracted feature vectors are of high quality, which can meet the high quality and high speed requirements of industrial implementation.

[0073] S4. Input the semantic feature vector extracted from each group of multi-dimensional texts into the trained classification model to obtain the probability distribution of the group of multi-dimensional texts belonging to each category in a custom emerging industry classification system, thereby obtaining an emerging industry category corresponding to each group of multi-dimensional texts.

[0074] In this embodiment, the classification model training method includes the following steps, namely S41~S45.

[0075] S41. Constructing enterprise classification data set; the enterprise classification data set consists of a sample text set of several enterprises T’ , and each T’ It is composed of the label set of the actual industry category corresponding to all multi-dimensional sample texts; among them, T’ Each group of multi-dimensional sample texts in is constructed by referring to steps S1~S2.

[0076] S42. Extract the semantic feature vector of each group of multi-dimensional sample texts; k The group of multi-dimensional sample texts is represented as , k ∈[1,K], K is T’ The total number of multi-dimensional sample texts in is obtained using the vector representation model t k The semantic feature vector of ; , , , Sequentially indicates k The business scope feature vector, industry feature vector, patent product name feature vector, and patent abstract feature vector of the group of multi-dimensional sample texts.

[0077] S43. The factor decomposition bilinear pooling method is used to fuse the business scope feature vector and the industry feature vector of each group of multidimensional sample texts to obtain the enterprise features of each group of multidimensional sample texts, and the patent product name vector and the patent abstract feature vector of each group of multidimensional samples are fused to obtain the patent features of each group of multidimensional sample texts.

[0078] In this embodiment, k Enterprise feature vector of group multi-dimensional sample text The expression is:

[0079] ;

[0080] ;

[0081] ;

[0082] ;

[0083] In the formula, Represents intermediate features In Dimension The value of , Intermediate feature Dimensions; Represents vector transpose; represents the learnable weight matrix, , which means that the matrix Factorize into two low-rank matrices and ; is a vector of all 1s, represents the Hadamard product; represents the bias variable; and Respectively indicated by and A third-order tensor composed of

[0084] , ;

[0085] and Respectively indicate that and The two-dimensional matrix obtained by tiling the second and third dimensions of Indicates the use of a size of One-dimensional non-overlapping window pairs Perform summing and pooling operations; express After power law normalization, the function To determine The symbol of; function Indicates taking the Euclidean distance;

[0086] No. k Patent feature vectors of multi-dimensional sample texts The expression is:

[0087] ;

[0088] ;

[0089] ;

[0090] ;

[0091] in, Represents intermediate features In Dimension The value of represents the learnable weight matrix, , which means that the matrix Factorize into two low-rank matrices and ; and Respectively indicated by and A third-order tensor composed of , ; and Respectively indicate that and The two-dimensional matrix obtained by tiling the second three-dimensional; express Results after power law normalization.

[0092] S44. Introduce the attention fusion mechanism to obtain the overall characteristics of adaptive fusion based on the enterprise characteristics and patent characteristics of each group of multi-dimensional sample texts; k The overall feature vector of the group of multi-dimensional sample texts is , the expression is as follows:

[0093] ;

[0094] ;

[0095] ;

[0096] In the formula, for The latent feature representation of for The weight matrix of for The corresponding bias vector; for The latent feature representation of for The weight matrix of for The corresponding bias vector; for The weight matrix of represents the hyperbolic tangent function, Represents an exponential function.

[0097] S45. Softmax is selected as the activation function to calculate the probability distribution of each group of multi-dimensional sample texts belonging to each category in the self-defined emerging industry classification system, and the cross entropy function is selected as the loss function. The training is implemented by back propagation and gradient descent. The optimization direction is to make the value of the cross entropy function smaller. The expression formula is as follows:

[0098] ;

[0099] ;

[0100] In the formula, p For multi-dimensional sample text t kThe probability distribution of belonging to each category in the self-defined emerging industry classification system; W p is the probability distribution p The weight matrix of b p is the probability distribution p The bias vector of Loss Indicates loss; represents the total number of multi-dimensional sample texts in the training set divided by the enterprise classification dataset, Indicates the training set The probability value corresponding to the industry category to which the data actually belongs, .

[0101] A classifier is constructed by combining a nonlinear fully connected layer and a softmax activation function to map the feature vector into the probability value of the corresponding category.

[0102] In this embodiment, the enterprise classification data set can be divided into a training set, a validation set, and a test set according to an 8:1:1 ratio. The maximum number of iterations epoch_number is set to 30, and the loss function Loss is trained using back propagation and gradient descent. In this embodiment, the gradient descent method uses a learning rate , exponential decay rate , The Adam optimization algorithm stops training when the number of iterations reaches epoch_number.

[0103] According to the probability distribution , choose the maximum probability value The corresponding category is used as the final classification label, and the probability value As confidence output.

[0104] In this embodiment, the user-defined emerging industry classification system may include multiple first-level emerging industry categories; each first-level emerging industry category includes multiple subordinate second-level emerging industry categories.

[0105] For example:

[0106] The first-level emerging industry categories are recorded as C={C1, C2, C3, C4, C5, C6, C7, C8, C9}, with a total of nine, namely, new generation information technology industry (C1), high-end equipment manufacturing industry (C2), new materials industry (C3), biological industry (C4), new energy vehicle industry (C5), new energy industry (C6), energy conservation and environmental protection industry (C7), aerospace industry (C8), and marine equipment industry (C9).

[0107] C1={C11, C12, C13, C14}, which are the next generation information network industry (C11), core electronics industry (C12), artificial intelligence (C13), and digital creative technology equipment manufacturing (C14).

[0108] C2={C21, C22}, which are intelligent manufacturing equipment industry (C21) and rail transit equipment industry (C22).

[0109] C3={C31, C32, C33, C34, C35, C36}, which are respectively advanced steel materials (C31), advanced non-ferrous metal materials (C32), advanced petrochemical and chemical new materials (C33), advanced inorganic non-metallic materials (C34), high-performance fibers and products and composite materials (C35), and cutting-edge new materials (C36).

[0110] C4={C41, C42, C43, C44, C45}, which are biopharmaceutical industry (C41), biomedical engineering industry (C42), bio-agriculture and related industries (C43), biomass energy industry (C44), and other biological industries (C45).

[0111] C5={C51, C52, C53}, which respectively refer to the manufacturing of new energy vehicle complete vehicles (C51), the manufacturing of new energy vehicle devices and accessories (C52), and the new energy vehicle related facilities (C53).

[0112] C6={C61, C62, C63, C64, C65, C66, C67}, which are respectively nuclear power industry (C61), wind energy industry (C62), solar energy industry (C63), biomass energy and other new energy industries (C64), smart grid industry (C65), coal-based new fuel manufacturing (C66), and hydrogen energy industry (C67).

[0113] C7={C71, C72, C73}, which are high-efficiency energy-saving industry (C71), advanced environmental protection industry (C72), and resource recycling industry (C73).

[0114] C8={C81, C82}, which are aviation equipment industry (C81) and satellite and application industry (C82).

[0115] C9={C91, C92, C93, C94, C95}, which are respectively marine equipment manufacturing (C91), deep-sea oil drilling equipment manufacturing (C92), other marine-related equipment and product manufacturing (C93), marine environment monitoring and detection equipment manufacturing (C94), and seawater desalination activities (C95).

[0116] Of course, in other embodiments, more levels of subcategories can be customized according to user needs.

[0117] The customized emerging industry classification system of the present invention can be modified and set with reference to documents such as the "Reference Relationship Table of Strategic Emerging Industries Classification and International Patent Classification (2021)", or it can be completely recreated by the user according to actual needs. The reason why the classification system is determined in a user-defined manner is mainly because the classification system in the above-mentioned documents is different from the actual needs of financial institutions, which makes some categories corresponding to the IPC classification number difficult to apply in corporate products.

[0118] After calculating the probability distribution of each group of multi-dimensional texts belonging to each category in the self-defined emerging industry classification system, the secondary emerging industry category with the largest probability is selected as the emerging industry category corresponding to the group of multi-dimensional texts.

[0119] S5. Based on the emerging industry categories corresponding to all multi-dimensional texts, generate the classification results of the emerging industries to which the enterprises to be screened belong, and determine whether the enterprises to be screened meet the target conditions based on the classification results.

[0120] In step S5, if the probability that any group of multi-dimensional texts belongs to each category in the self-defined emerging industry classification system is lower than the preset probability threshold (which can be 0.5), the group of multi-dimensional texts is marked as non-emerging industry texts; if the proportion of non-emerging industry texts in the text set of the enterprise to be screened is P 1 is greater than a preset upper limit of the ratio c ,Right now P 1> c , then the enterprise to be screened is determined to be a non-target enterprise, that is, it does not meet the target conditions; where, , J1 is the number of non-emerging industry texts in the text collection of the enterprise to be screened.

[0121] In this embodiment, the upper limit of the ratio may be 60%. P 1> c When , it indicates that the enterprise to be screened has more non-emerging industry texts, indicating that the probability that the enterprise to be screened is not a high-quality science and technology enterprise is relatively high. The reason may be that it involves some traditional industries (food, clothing, etc.). At this time, the enterprise to be screened is judged as a non-target enterprise.

[0122] The classification results of the emerging industries to which an enterprise belongs include three labels: the first emerging industry label, the second emerging industry label and the third emerging industry label.

[0123] It should be noted that the first emerging industry label can represent the main emerging industries of an enterprise, the second emerging industry label can represent the secondary emerging industries of an enterprise, and the third emerging industry label can represent other emerging industries of an enterprise.

[0124] like P1≤ c , then the enterprise to be screened is determined to be a target enterprise, that is, it meets the target conditions, and statistical analysis is performed on the emerging industry categories corresponding to the J2 group multi-dimensional texts of the enterprise to be screened to obtain the proportion of each emerging industry category P 2. Then, the proportion of each emerging industry category P 2 and a preset ratio interval [ a 1, a 2] Make a comparison and make the following decisions:

[0125] Will satisfy P 2> a The emerging industry category of 2 is used as the first emerging industry label for the enterprises to be screened.

[0126] Will satisfy a 1≤ P 2≤ a The emerging industry category of 2 is used as the second emerging industry label for the enterprises to be screened.

[0127] Will satisfy P 2< a 1’s emerging industry category is used as the third emerging industry label for the enterprises to be screened.

[0128] in, , J2 is the number of multi-dimensional texts in the text set of the enterprise to be screened that are not marked as non-emerging industry texts, that is, J2=J-J1, x is the number of times each emerging industry category appears in the multidimensional text of group J2.

[0129] In this embodiment, a It can be set to 20%, b It can be set to 50%. In some embodiments, it can also be adjusted to other sizes according to actual conditions.

[0130] After the statistical analysis, if the enterprise to be screened has the first emerging industry label and / or the second emerging industry label, the enterprise is determined to be the first target enterprise; if the enterprise to be screened does not have the first emerging industry label and the second emerging industry label, the enterprise is determined to be the second target enterprise.

[0131] It should be noted that when the enterprise to be screened is determined to be the first target enterprise, it can indicate that the investment proportion (layout) of the enterprise to be screened in a new industry is relatively large, and it can also indirectly reflect that it is more professional in a certain type of product; and when the enterprise to be screened is determined to be the second target enterprise, it can indicate that the industrial distribution of the enterprise to be screened is relatively comprehensive and balanced. In some embodiments, whether it belongs to the first target enterprise or the second target enterprise can be used as an evaluation indicator for users (financial institutions) to provide a reliable basis for their judgment on whether to provide financial services to the target enterprise and the type of service.

[0132] Example 2

[0133] like Figure 2 As shown, this embodiment provides a computer terminal, which includes at least one processor 101 and a memory 102 connected to the at least one processor 101. The processor 101 is used to store a computer program. The processor 101 is used to implement the target enterprise screening method based on the industry to which it belongs as described above when executing the computer program.

[0134] In this embodiment, the specific connection medium between the processor 101 and the memory 102 is not limited. Figure 2 In the example, the processor 101 and the memory 102 are connected via the bus 100. The bus 100 is connected to the memory 102 via the bus 100. Figure 2 The connections between other components are shown by thick lines, which are only schematic and not limiting. The bus 100 can be divided into an address bus, a data bus, a control bus, etc. For convenience of representation, Figure 2 In the figure, only one thick line is used, but it does not mean that there is only one bus or one type of bus. Alternatively, the processor 101 can also be called a controller, and there is no limitation on the name.

[0135] In this embodiment, the memory 102 stores instructions that can be executed by at least one processor 101. By executing the instructions stored in the memory 102, the at least one processor 101 can execute the target enterprise screening method based on the industry discussed above.

[0136] Among them, the processor 101 is the control center of the device, and can use various interfaces and lines to connect the various parts of the entire control device. By running or executing instructions stored in the memory 102 and calling data stored in the memory 102, the various functions of the device and process data, the device can be monitored as a whole.

[0137] In one possible design, the processor 101 may include one or more processing units, and the processor 101 may integrate an application processor and a modem processor, wherein the application processor mainly processes an operating system, a user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the modem processor may not be integrated into the processor 101. In some embodiments, the processor 101 and the memory 102 may be implemented on the same chip, and in some embodiments, they may also be implemented separately on separate chips.

[0138] The processor 101 may be a general-purpose processor, such as a central processing unit (CPU), a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component, and may implement or execute the methods, steps, and logic block diagrams disclosed in this embodiment. A general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the target enterprise screening method based on the industry disclosed in this embodiment may be directly embodied as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor 101.

[0139] The memory 102 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The memory 102 may include at least one type of storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory, a random access memory (Random Access Memory, RAM), a static random access memory (Static Random Access Memory, SRAM), a programmable read-only memory (Programmable Read Only Memory, PROM), a read-only memory (Read Only Memory, ROM), an electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, EEPROM), a magnetic memory, a disk, an optical disk, etc. The memory 102 is any other medium that can be used to carry or store a desired program code in the form of an instruction or data structure and can be accessed by a computer, but is not limited thereto. The memory 102 in this embodiment can also be a circuit or any other device that can realize a storage function, for storing program instructions and / or data.

[0140] By programming the processor 101, the code corresponding to the target enterprise screening method based on the industry described in the above embodiment can be fixed into the chip, so that the chip can execute the code when running. Figure 1The steps of the target enterprise screening method based on the industry in the embodiment shown are as follows: How to design and program the processor 101 is a technology well known to those skilled in the art and will not be described in detail here.

[0141] Example 3

[0142] This embodiment provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the program is executed by a processor, the steps of the target enterprise screening method based on the industry to which it belongs as described above are implemented.

[0143] The computer-readable storage medium may include flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the storage medium may be an internal storage unit of a computer device, such as a hard disk or memory of the computer device. In other embodiments, the storage medium may also be an external storage device of a computer device, such as a plug-in hard disk equipped on the computer device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. Of course, the storage medium may also include both an internal storage unit of a computer device and an external storage device thereof. In this embodiment, the memory is generally used to store an operating system and various application software installed on the computer device, etc. In addition, the memory may also be used to temporarily store various types of data that have been output or are to be output.

[0144] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for screening target enterprises based on their industry, characterized in that: The following steps are involved: S1. Obtain relevant information of the enterprise to be screened, including the business scope, industry and several patent texts of the enterprise to be screened; S2. construct a text set of the enterprise to be screened based on the relevant information, the text set including several groups of multi-dimensional texts; each group of multi-dimensional texts includes the business scope and industry of the enterprise to be screened, and the name and abstract of the patent product corresponding to a patent text; In step S2, the text set of the enterprise to be screened is represented as T, where T = {t1, t2, ..., t j ,...,t J }, t j represents the jth group of multidimensional texts, j∈[1,J]; J is the total number of multidimensional texts, that is, the total number of patent texts of the enterprise to be screened, J>1; multidimensional text Indicates the business scope of the enterprise to be screened, Indicates the industry to which the enterprise to be screened belongs. Indicates the name of the patent product corresponding to the jth patent text, represents the abstract of the jth patent text; wherein the generative model is used to obtain the patent product name corresponding to each patent text; S3. Extracting semantic features from each group of multi-dimensional texts in the text collection of the enterprise to be screened; S4. Inputting the semantic feature vector extracted from each group of multi-dimensional texts into the trained classification model to obtain the probability distribution of the group of multi-dimensional texts belonging to each category in a custom emerging industry classification system, thereby obtaining an emerging industry category corresponding to each group of multi-dimensional texts; In step S4, the training method of the classification model includes the following steps: S41. Construct an enterprise classification data set; the enterprise classification data set is composed of a sample text set T' of several enterprises, and a label set of the actual industry category corresponding to all multi-dimensional sample texts in each T'; wherein each group of multi-dimensional sample texts in T' is constructed in accordance with the method of steps S1 to S2; S42. Extract the semantic feature vector of each group of multi-dimensional sample texts; wherein the kth group of multi-dimensional sample texts is represented as K is the total number of multi-dimensional sample texts in T', and the vector representation model is used to obtain t k The semantic feature vector of The business scope feature vector, industry feature vector, patent product name feature vector, and patent abstract feature vector of the k-th group of multi-dimensional sample texts are represented in sequence; S43. Adopt the factor decomposition bilinear pooling method, fuse the business scope feature vector and industry feature vector of each group of multi-dimensional sample texts to obtain the enterprise features of each group of multi-dimensional sample texts, and fuse the patent product name vector and patent abstract feature vector of each group of multi-dimensional samples to obtain the patent features of each group of multi-dimensional sample texts; S44. Introduce the attention fusion mechanism to obtain the overall features of adaptive fusion based on the enterprise features and patent features of each group of multi-dimensional sample texts; where the overall feature vector of the kth group of multi-dimensional sample texts is h k , the expression is as follows: In the formula, u k and k are the enterprise feature vector and patent feature vector of the kth group of multi-dimensional sample texts respectively; for u k The latent feature representation of for The weight matrix of for The corresponding bias vector; for o k The latent feature representation of for The weight matrix of for The corresponding bias vector; h k The weight matrix of ; tanh(·) represents the hyperbolic tangent function, and exp(·) represents the exponential function; S45. Softmax is selected as the activation function to calculate the probability distribution of each group of multi-dimensional sample texts belonging to each category in the self-defined emerging industry classification system, and the cross entropy function is selected as the loss function. The training is implemented by back propagation and gradient descent. The optimization direction is to make the value of the cross entropy function smaller. The expression formula is as follows: p=softmax(W p ·h k +b p ); In the formula, p is the multi-dimensional sample text t k The probability distribution of belonging to each category in the self-defined emerging industry classification system; W p is the weight matrix of probability distribution p; b p is the bias vector of the probability distribution p; Loss represents the loss; Φ represents the total number of multi-dimensional sample texts in the training set divided by the enterprise classification data set, p φ Represents the probability value corresponding to the actual industry category of the φth data in the training set, φ∈[1,Φ]; S5. Based on the emerging industry categories corresponding to all multi-dimensional texts, generate the classification results of the emerging industries to which the enterprises to be screened belong, and determine whether the enterprises to be screened meet the target conditions based on the classification results.

2. The target enterprise screening method based on the industry according to claim 1 is characterized in that: The self-defined emerging industry classification system includes multiple first-level emerging industry categories; each first-level emerging industry category includes multiple subordinate second-level emerging industry categories; In step S4, after calculating the probability distribution of each group of multi-dimensional texts belonging to each category in the self-defined emerging industry classification system, the secondary emerging industry category with the largest probability is selected as the emerging industry category corresponding to the group of multi-dimensional texts.

3. The target enterprise screening method based on the industry according to claim 2 is characterized in that: In step S5, if the probability that any group of multi-dimensional texts belongs to each category in the self-defined emerging industry classification system is lower than a preset probability threshold, the group of multi-dimensional texts is marked as non-emerging industry texts; If the proportion of non-emerging industry texts P1 in the text set of the enterprise to be screened is greater than a preset upper limit c, that is, P1>c, the enterprise to be screened is determined to be a non-target enterprise, that is, it does not meet the target condition; wherein, J1 is the number of non-emerging industry texts in the text set of the enterprise to be screened, and J is the total number of multi-dimensional texts of the enterprise to be screened.

4. The target enterprise screening method based on the industry according to claim 3 is characterized in that: The classification results of the emerging industries to which an enterprise belongs include three labels: the first emerging industry label, the second emerging industry label, and the third emerging industry label; In step S5, if P1≤c, the enterprise to be screened is determined to be a target enterprise, that is, it meets the target condition, and statistical analysis is performed on the emerging industry categories corresponding to the J2 group of multi-dimensional texts of the enterprise to be screened to obtain the proportion P2 of each emerging industry category, and then the proportion P2 of each emerging industry category is compared with a preset ratio interval [a1, a2], and the following decisions are made respectively: The emerging industry category that satisfies P2>a2 is used as the first emerging industry label of the enterprise to be screened; The emerging industry categories that satisfy a1≤P2≤a2 are used as the second emerging industry labels for the enterprises to be screened; The emerging industry categories that satisfy P2<a1 are used as the third emerging industry labels of the enterprises to be screened; in, J2 is the number of multidimensional texts in the text set of the enterprise to be screened that are not marked as non-emerging industry texts, that is, J2=J-J1, and x is the number of times each emerging industry category appears in the multidimensional text group J2.

5. The target enterprise screening method based on the industry according to claim 4 is characterized in that: After the statistical analysis, if the enterprise to be screened has the first emerging industry label and / or the second emerging industry label, the enterprise is determined to be the first target enterprise; If the enterprise to be screened does not have the first emerging industry label and the second emerging industry label, the enterprise will be determined as the second target enterprise.

6. The target enterprise screening method based on the industry according to claim 1 is characterized in that: In step S43, the enterprise feature vector u of the kth group of multi-dimensional sample texts k The expression is: In the formula, Represents intermediate features The value in dimension i, i∈[1,I], I is the intermediate feature Dimensions; Represents vector transpose; represents the learnable weight matrix, Represents the matrix Factorize into two low-rank matrices and is a vector of all 1s, ⊙ represents the Hadamard product; b k,i represents the bias variable; and Respectively indicated by and A third-order tensor composed of and Respectively indicate that and The function SumPooling[V,G] represents the summing and pooling operation of V using a one-dimensional non-overlapping window of size G; express After power law normalization, the function sign(·) is used to determine The symbol of ; the function ||·|| means taking the Euclidean distance; The patent feature vector o of the kth group of multi-dimensional sample texts k The expression is: in, Represents intermediate features The value in dimension i; represents the learnable weight matrix, Represents the matrix Factorize into two low-rank matrices and and Respectively indicated by and A third-order tensor composed of and Respectively indicate that and The two-dimensional matrix obtained by tiling the second three-dimensional; express Results after power law normalization.

7. A computer terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method for screening target enterprises based on their industries as described in any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method for screening target enterprises based on their industries as described in any one of claims 1 to 6 are implemented.

Images