A regional industry chain generation method and device and a storage medium

Abstract

The application discloses a regional industry chain generation method and device and a storage medium. According to a subdivided field, an existing classification system and related literature, a self-defined standardized industry classification system is constructed through different information processing models, and then corresponding regional industry chains are generated according to the system and regional product information, so that the generation of the regional industry chains is more intelligent and automatic, the data is more reliable, the content is more comprehensive and conforms to regional characteristics. Therefore, the regional industry chains generated through the method have higher reference value for governments and related enterprises, and are beneficial to the economic development and the improvement of the collaboration ability of the region.

Description

TECHNICAL FIELD

[0001] The technology belongs to the field of artificial intelligence, and the present application relates to the field of natural language processing, in particular to a regional industry chain generation method and device and storage medium. BACKGROUND

[0002] An industry chain refers to a chain-like associated relationship formed objectively between various industrial departments based on certain technical and economic correlations, according to specific logical relationships and spatial and temporal layout relationships, and is a concept including four dimensions of value chain, enterprise chain, supply and demand chain and spatial chain. A regional industry chain refers to obtaining an industry chain that meets the characteristics of a specific region, analyzing the chain-like correlation between various industrial departments in the internal system of the region, and analyzing the chain-like correlation between various industrial departments. These industry chains can facilitate the subsequent location advantage, industrial synergy and cluster benefit maximization of the region, and are conducive to creating a "location brand" and improving the development of regional economy, so the construction of a regional industry chain is also a research point of particular concern to the government and enterprises.

[0003] However, the current method for identifying and generating an industry chain is relatively traditional, and is often designed only from the perspective of industry classification and manually written by industry researchers of securities companies or research institutes. A related industry chain research report is generally written by staff of a consulting agency based on local experience, corresponding patents and statistical data, and the collected data is scattered and lacks fairness, making it difficult to directly interface with the market, and the generated industry chain has differences in different degrees and directions, which is not conducive to the market value mining of the industry chain and cannot improve the actual resource efficiency of the region.

[0004] With the development of information technology, intelligent systems based on natural language have emerged. There is a need for a construction method to meet the regional characteristics of the industry chain generation scheme, further mine product data and regional potential, improve the efficiency and speed of industrial resource capture, and provide valuable industrial information for the government and enterprises. SUMMARY

[0005] The present application provides a regional industry chain generation method, device and storage medium, which automatically generates a regional industry chain by constructing an industry classification system based on field information, regional information and related literature, and taking regional products as terminal nodes of the industry chain. Further traversal and merging can efficiently provide high-quality and valuable regional industry chains for enterprises and governments.

[0006] A regional industry chain generation method includes the following steps:

[0007] Obtain regional products and standardize the labels as terminal nodes of the regional industry chain;

[0008] Construct a self-defined standardized industry classification system based on the subdivision field, existing classification system and related literature;

[0009] generate a corresponding industry chain based on the self-defined standardized industry classification system and the regional industry chain terminal nodes, and obtain a regional industry chain set;

[0010] perform node traversal and merging on the regional industry chain set to generate a regional industry chain.

[0011] Further, the regional products include:

[0012] a tangible product, which is a product circulating in the regional market and is provided with a level attribute;

[0013] an intangible product, which is a factor and technology specific to the region;

[0014] a potential product, which is a product that can be developed in the future according to regional factors;

[0015] wherein the level attribute of the tangible product is divided by a market position, and the market position includes market share, product standardization degree, and market recognition.

[0016] Further, the self-defined standardized industry classification system includes:

[0017] obtaining the subfield, the existing classification system, and the related literature to construct an ontology layer;

[0018] processing and recognizing the related literature to extract industry chain entity information;

[0019] fusing the industry chain entity information to the ontology layer to generate an industry chain relationship graph;

[0020] auditing the fused industry chain relationship graph to generate a self-defined standardized industry classification system;

[0021] wherein the subfield is a field classification further refined based on the global industry classification standard and combined with regional factors; the existing classification system includes a national economic industry classification, a strategic emerging industry classification, a securities industry classification, a commodity classification, and a self-defined classification of other related industries; and the related literature includes industry research reports, industry patents and papers, listed company reports, and regional industry research reports.

[0022] Further, the processing and recognizing the related literature to extract industry chain entity information includes:

[0023] Customize a label type based on a BIO tagging method and an industry chain feature; the label type includes a regional element Ele, a sub-field Fie, industry information Pro, and industry attribute Attr; a format of the label is: when the label type is the regional element type, a start tag is "B-Ele" and an intermediate tag is "I-Ele"; when the label type is the sub-field type, a start tag is "B-Fie" and an intermediate tag is "I-Fie"; when the label type is the industry information type, a start tag is "B-Pro" and an intermediate tag is "I-Pro"; when the label type is the industry attribute type, a start tag is "B-Attr" and an intermediate tag is "I-Attr";

[0024] Obtain relevant literature of annotated industry chain entity information as training corpus for training an industry chain entity information extraction model;

[0025] The training industry chain entity information extraction model includes:

[0026] Output a word vector of a text through a pre-trained BERT module;

[0027] Calculate a word score through a BiLSTM module;

[0028] Calculate an attention score of each word through an attention mechanism module;

[0029] Perform sequence labeling through a CRF module, calculate a probability of a label sequence, and output a label sequence with the highest probability as a prediction result.

[0030] Further, the outputting of the word vector of the text through the pre-trained BERT module includes:

[0031] Take the training corpus text as an input unit;

[0032] Perform a mask prediction task to output a word-level feature;

[0033] Perform a sentence prediction task to output a sentence-level feature;

[0034] Generate three kinds of representation vectors including a word vector, a sentence vector, and a position vector after task training;

[0035] Sum the three kinds of representation vectors to generate the word vector of the text.

[0036] Further, the BiLSTM module includes:

[0037] Take the word vector as an input unit of the BiLSTM module; set a forward LSTM layer to calculate a feature of context information of the text; and set a reverse LSTM layer to calculate a feature of context information of the text;

[0038] The outputs of the forward LSTM layer and the reverse LSTM layer are concatenated to output a character sequence probability feature vector with contextual feature information;

[0039] The LSTM layer model is as follows: Let vector X t ={X1,X2,……,X n} is the input vector h of the LSTM layer. t It is the output of the LSTM layer, obtained by analyzing the output h of the hidden layer of the LSTM network at time t-1. t-1 and the word vector x at time t t Then calculate the forget gate f t :

[0040] f t =σ(W f *[h t-1 ,x t ]+b f )

[0041] Calculate the input gate i t :

[0042] i t =σ(W i *[h t-1 ,x t ]+b i )

[0043] Calculate the output gate o t :

[0044] o t =σ(W o *[h t-1 ,x t ]+b o )

[0045] In the formula, W f W i and W o These represent the recursive connection weights for the corresponding thresholds, and σ is the sigmoid function. Where tanh is the hyperbolic tangent function, b i b f and b o This is the corresponding offset. M can be calculated using the values ​​of the forget gate, the input gate, time t, and time t-1. t value, M was calculated t =f t *M t-1 +i t *tanh(W m*[h t-1 ,x t ]+b m ), thus the hidden layer state is h. t =o t *tanh(C t The output of the LSTM layer is h = {h1, h2, ..., h}. n}

[0046] Furthermore, the calculation of attention scores for each word via the attention mechanism module includes:

[0047] Input the feature vector output by the BiLSTM module; calculate the score of the labeled sequence based on the feature weight values, and the attention weight vector is... In the formula y i b represents the feature sequence output by the BiLSTM module. ij The relevance probabilities between words in a text are respectively defined as follows:

[0048]

[0049] In the formula p i and p j Let f(p) represent the i-th and j-th words in the document, where N is the number of words in the document. i ,p j ) indicates word p i With word p j The similarity score, f(p) i ,p k ) indicates word p i With any word p in the document k The similarity score.

[0050] Furthermore, the sequence labeling using the CRF module, and the calculation of the probability of the labeled sequence, includes:

[0051] The vector processed by the attention mechanism module is input into the CRF module, and the feature of the i-th vector is labeled as Z. i Z i The weight of the feature is denoted as λ. i Each sentence is denoted as S, j represents the position of a word in the sentence, the labeled sequence is denoted as L, and the labeled sequence is scored in the sentence.

[0052]

[0053] In the formula, scroe(L|S) represents the score of each token sequence in each sentence, l j It is the marker for the current word, l j-1 It is a tag for the previous word, the sentence length is m, and the number of features corresponding to each tag sequence is n;

[0054] Normalization and conversion to probability

[0055]

[0056] Where, ∑ j exp[score[L|S]] represents the sum of all label sequences, P(L|S) is the normalized probability, and L with the largest output probability value is the final label sequence.

[0057] Further, the industry chain entity information is fused into the ontology layer, including:

[0058] Based on the similarity calculation, an information fusion model is constructed to fuse the extracted industry chain entity information with the ontology layer;

[0059] Based on the industry chain entity information extraction model and the information fusion model, an industry chain relationship graph is generated;

[0060] The construction of the information fusion model includes:

[0061] The ontology layer entity information vector is obtained through the BERT model;

[0062] The cosine similarity of the ontology layer entity information vector and the industry chain entity information vector is calculated, and a threshold is set;

[0063] Based on the corpus, the linking rate, linking frequency and topic consistency features of the industry chain entity information and the ontology layer entity information are calculated;

[0064] According to the calculation results and the threshold, the fusion of the industry chain entity information and the ontology layer is completed;

[0065] Wherein, the linking rate is

[0066]

[0067] In the formula, m is the industry chain entity information, e i is the ontology layer entity information, Γ(m,e j ) is the number of times m is linked to entity e i in the extended document set, E c =D[m] is the candidate entity set of m;

[0068] The linking frequency is

[0069]

[0070] In the formula, M D is the set of industry chain entity information identified in the extended document set;

[0071] The topic consistency feature is

[0072]

[0073] In the formula, rel1 is a topic consistency measure based on the ontology layer,

[0074]

[0075] Further, the generating a corresponding industry chain, obtaining a regional industry chain set comprises:

[0076] Training a regional industry chain generation model based on an evaluation function:

[0077] The labeled regional product and the constructed custom standardized industry classification system are used as the input layer of the model;

[0078] The weights of the evaluation function are set, including upstream node weight, midstream node weight and downstream node weight;

[0079] The output layer is a regional industry chain with the terminal node of the regional product industry chain;

[0080] Through the trained regional industry chain generation model, the labeled regional product information is input to generate a corresponding regional industry chain set.

[0081] Further, the node traversal and merging of the regional industry chain set to generate a regional industry chain comprises:

[0082] Selecting a target industry chain according to the regional industry chain set;

[0083] Based on the target industry chain, the nodes of other industry chains are traversed and compared, and the nodes of other industry chains are merged into the target industry chain;

[0084] The merged target industry chain is proofread to generate a regional industry chain.

[0085] A regional industry chain generation device comprises the regional industry chain generation method, comprising:

[0086] An information labeling module is configured to label regional product information;

[0087] A system generation module is configured to establish an industry chain entity information extraction model and an information fusion model, and generate a custom standardized industry classification system;

[0088] A regional industry chain generation module is configured to generate a regional industry chain set;

[0089] A merging module is configured to generate a regional industry chain by traversing and merging the regional industry chain set.

[0090] An electronic device comprising a memory and one or more processors;

[0091] The memory is in communication connection with the one or more processors, and the memory stores instructions executable by the one or more processors, and the instructions are executed by the one or more processors to implement the method for generating a regional industrial chain.

[0092] A computer-readable storage medium having stored thereon computer-executable instructions that, when executed by a computing device, implement the method for generating a regional industrial chain.

[0093] A computer program product comprising a computer program stored on a computer-readable storage medium, the computer program comprising program instructions that, when executed by a computer, implement the method for generating a regional industrial chain.

[0094] The above technical solution has the following advantages or beneficial effects:

[0095] The method for generating a regional industrial chain has the advantages that: the ontology layer is constructed by using the existing classification system, the subfield and the regional information; the industrial chain entity information extraction model is established to extract information from a large number of related literatures and fuse the ontology layer, thereby improving the authority and credibility of the industrial chain information; the deep learning-based method is used to process the text information, thereby improving the work efficiency and the automation level of generating the industrial chain; the self-defined standardized industrial classification system generated after fusion has comprehensive industrial information and regional information; the regional industrial chain generated according to the classification system and the regional product is more perfect and has higher quality; the method provides valuable information for the government and enterprises, and is more conducive to the production and development of regional economy, the creation of regional brands and the improvement of the division of labor and cooperation among industries. BRIEF DESCRIPTION OF DRAWINGS

[0096] Figure 1 is a flowchart of the method for generating a regional industrial chain of the present application;

[0097] Figure 2 is a flowchart of the method for constructing a self-defined standardized industrial classification system of the present application;

[0098] Figure 3 is a flowchart of the method for training an industrial chain entity information extraction model of the present application;

[0099] Figure 4 is a flowchart of the method for constructing an information fusion model of the present application;

[0100] Figure 5 is a structural diagram of a device for generating a regional industrial chain of the present application. DETAILED DESCRIPTION

[0101] In order to make the above objectives, features and advantages of the present application more apparent, specific embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0102] In combination Figure 1 Figure 5 As shown in the drawings, a regional industry chain generation method comprises the following steps:

[0103] S1, obtaining regional products and performing standardized marking as regional industry chain terminal nodes;

[0104] S2, constructing a self-defined standardized industry classification system based on a subdivision field, an existing classification system and related literature;

[0105] S3, generating a corresponding industry chain based on the self-defined standardized industry classification system and the regional industry chain terminal nodes, and obtaining a regional industry chain set;

[0106] S4, performing node traversal and merging on the regional industry chain set to generate a regional industry chain.

[0107] Preferably, the regional products comprise:

[0108] a tangible product, which is a product circulating in a regional market and is provided with a level attribute;

[0109] an intangible product, which is a factor and a technology specific to a region;

[0110] a potential product, which is a product that can be developed in the future according to a regional factor;

[0111] wherein the level attribute of the tangible product is divided by a market position, and the market position comprises a market share, a product standardization degree and a market recognition degree.

[0112] Illustratively, the region can be divided according to provinces and cities, or according to geographical features and climate features. Different regional products exhibit the value of the corresponding region, and are generally products provided by the region and meeting market demand. According to the form, the regional products can be divided into tangible products, intangible products and potential products. The tangible products include generated products such as local famous product enterprises, well-known products with local characteristics, etc.; regional characteristic buildings such as streets and artificial landscapes; and regional natural resources mainly referring to regional natural endowments and other natural products such as resources and minerals. The intangible products include cultural products and social activities such as humanistic atmosphere, trade activities and sports activities. The potential products embody the possible evolution trend and prospect of current products. The regional products can be obtained from official reports, data and relevant documents of local leading enterprises of the region.

[0113] ​Optionally, the product grade classification mainly includes the need set, product set, product category, product line, product type, brand and product item; the product grade differentiation mode includes the superior product, the first-class product and the qualified product; the superior product has the international advanced level, the first-class product has the national general level, and the qualified product is produced according to the current national and industry general standard. Preferably, the product grade is set to at least the first-class product in the industry chain.

[0114] Further, the construction of the self-defined standardized industrial classification system includes:

[0115] S21, acquiring the subdivision field, the existing classification system and the related literature to construct the ontology layer;

[0116] S22, processing and identifying the related literature to extract the industrial chain entity information;

[0117] S23, fusing the industrial chain entity information to the ontology layer to generate the industrial chain relationship graph;

[0118] S24, auditing the fused industrial chain relationship graph to generate the self-defined standardized industrial classification system;

[0119] The subdivision field is based on the global industry classification standard and further refined in combination with regional elements; the existing classification system includes the national economic industry classification, the strategic emerging industry classification, the securities industry classification, the commodity classification and other self-defined classifications of related industries; the related literature includes industry research reports, industry patents and papers, listed company reports and regional industry research reports.

[0120] Illustratively, the global industry classification standard covers developed markets and emerging markets, and can be divided into different granularities according to social development and regional elements, such as freshwater fishing and marine fishing, and marine fishing includes South China Sea fishing ground and Yellow-Bohai Sea fishing ground according to the region; the regional elements include natural elements such as geology, topography, climate, biology and soil, and human elements such as population, culture and economy. The existing classification system information can be obtained through official website published files, authoritative platform analysis reports, research papers, technical information published by enterprises or individuals in the patent system, etc. The relevant files are extracted by searching and screening, and the ontology layer is arranged. Optionally, the subdivision field can be a manually edited technical field, an automatically recommended technical field or a common classification name extracted from scientific and technological literature, and a search formula can be constructed through the subdivision field to screen patents and papers in different fields.

[0121] Preferably, the processing and identifying the related literature to extract the industrial chain entity information includes:

[0122] Defining a label type based on a BIO tagging method and an industry chain feature; the label type includes a regional element Ele, a subdivision field Fie, industry information Pro, and industry attribute Attr; the format of the label is: when the label type is the regional element type, the corresponding start tag is "B-Ele", and the middle tag is "I-Ele"; when the label type is the subdivision field type, the corresponding start tag is "B-Fie", and the middle tag is "I-Fie"; when the label type is the industry information type, the corresponding start tag is "B-Pro", and the middle tag is "I-Pro"; when the label type is the industry attribute type, the corresponding start tag is "B-Attr", and the middle tag is "I-Attr";

[0123] Obtaining relevant literature of annotated industry chain entity information as training corpus for training an industry chain entity information extraction model;

[0124] The training industry chain entity information extraction model includes:

[0125] S221, outputting a word vector of a text through a pre-trained BERT module;

[0126] S222, calculating a word score through a BiLSTM module on the word vector;

[0127] S223, calculating an attention score of each word through an attention mechanism module;

[0128] S224, performing sequence labeling through a CRF module, calculating a probability of a label sequence, and outputting a label sequence with the highest probability as a prediction result.

[0129] Exemplary, the BIO tagging method is to use B to represent the start of an entity word in the text, I to represent the word immediately following the start of the entity word, and O to represent the word that is not an entity word. Based on the industrial chain information, new label categories are added, wherein the industrial information includes industry information and enterprise information, and the industrial attribute includes at least one of the position of the industry in the industrial chain, the upstream, the middle stream and the downstream. For example, "the brand construction of agricultural products in the region has AA brand of XX county", "agricultural products" is marked as "B-Fie, I-Fie, I-Fie", "XX county" is marked as "B-Ele, I-Ele, I-Ele", "AA brand" is marked as "B-Pro, I-Pro, I-Pro, I-Pro", and other words are marked as "O". The above processing of related documents is to extract entity words and the relationship between entity words corresponding to the industrial chain information. At present, different knowledge extraction technologies have been researched. Entity extraction can be roughly divided into three types: rule-based method, traditional machine learning-based method and deep learning-based method; relationship extraction has template-based method, deep learning-based method, supervised learning-based method and weakly supervised learning-based method. This paper establishes a model based on deep learning to extract industrial chain entities and entity relationships.

[0130] Preferably, the outputting word vector of the text by the pre-training BERT module comprises:

[0131] Taking the training corpus text as an input unit;

[0132] Performing a mask prediction task to train and output a word-level feature;

[0133] Performing a sentence prediction task to train and output a sentence-level feature;

[0134] Generating three kinds of representation vectors after task training, including word vector, sentence vector and position vector;

[0135] Summing the three kinds of representation vectors to generate the word vector of the text.

[0136] Exemplarily, the pre-trained BERT module can generate a combination of three feature vectors of the current word, the sentence where the current word is located, and the position according to the input corpus. The word vector represents the feature vector of the current word, the sentence vector represents the feature of the position of the current word in the sentence, and the position vector represents the position feature of the sentence in the text to which the current word belongs. The masking prediction task is to predict the masked words after randomly masking 15% of the words in the input word sequence, and then train the bidirectional language representation vector. The sentence prediction task, also known as the next sentence prediction task, selects sentence A and sentence B as training samples, where B has a 50% chance of being the next sentence of A and a 50% chance of coming from a random sentence in the corpus. The first word of the sentence is marked with [CLS], and [SEP] is used to mark the end of the sentence and the space between the two sentences. The input text is converted into a vector form through the training of the above two prediction tasks. The training corpus is a report, a scientific paper, a patent, etc. with marked industrial chain entity information.

[0137] Preferably, the BiLSTM module is constructed by:

[0138] The word vector is used as the input unit of the BiLSTM module; a forward LSTM layer is set to calculate the features of the context information of the text; and a reverse LSTM layer is set to calculate the features of the context information of the text.

[0139] The output of the forward LSTM layer and the output of the reverse LSTM layer are spliced to output a character sequence probability feature vector with context feature information.

[0140] The LSTM layer model is that a vector X t ={X1,X2,……,X n} is the input of the LSTM layer, and a vector h t is the output of the LSTM layer. The output h t-1 of the LSTM network hidden layer at t-1 and the word vector x t at t are used to calculate a forgetting gate f t .

[0141] f t =σ(W f *[h t-1 ,x t ]+b f )

[0142] An input gate i t is calculated.

[0143] i t =σ(W i *[h t-1 ,x t ]+b i )

[0144] Computing output gate o t :

[0145] o t = σ(W o *[h t-1 ,x t ]+b o )

[0146] wherein W f , W i and W o represent the recursive connection weights of the corresponding threshold, σ is a sigmoid function, wherein tanh is a hyperbolic tangent function, b i , b f and b o are corresponding offsets, the value of the forgetting gate, the value of the input gate, the value at time t and the value at time t-1 can be used to calculate M t value, M t = f t *M t-1 +i t *tanh(W m *[h t-1 ,x t ]+b m ), the state of the hidden layer is h t = o t *tanh(C t ), and the output result of the LSTM layer is h = {h1, h2, …, h n}.

[0147] For example, the BiLSTM module is used to calculate the word vector output by the pre-trained BERT module by using a bidirectional long short-term memory network to obtain the score of each word in the sentence, wherein the sigmoid function and the tanh hyperbolic tangent function are two activation functions, and the output result h includes LSTM forward and backward represent the score of each word in the sentence. Optionally, the LSTM is a unidirectional recurrent neural network, which can only obtain the feature information of the context of the current word, and affects the accuracy of information recognition. The BiLSTM module used in the present application is a bidirectional LSTM network model, which performs text representation in two different directions, obtains the context feature information of the current word, and improves the accuracy of the extraction of industrial chain entity information.

[0148] Preferably, the attention score of each word calculated by the attention mechanism module comprises:

[0149] The feature vector output by the input BiLSTM module; the score of the label sequence is calculated based on the feature weight value, and the attention weight vector is In the formula, y i The feature sequence output by the BiLSTM module is represented by b ij The probability of the correlation between the words in the text is represented by b

[0150]

[0151] In the formula, p i And p j Indicate the i-th word and the j-th word in the document, and N is the number of words in the document. f(p i , p j ) represents the similarity score of word p i and word p j , f(p i , p k ) represents the similarity score of word p i and any word p k in the document.

[0152] For example, due to the long sequence front-end semantic dilution problem of the BiLSTM module, the output feature vector is not accurate enough, and the recognition information may be incorrect, so the attention mechanism module is introduced. The attention mechanism is to calculate the correlation between the words to generate a changing semantic vector, so that the model focuses on the key words and ignores the useless words. The attention mechanism layer mainly calculates the attention weight vector. For example, if the industry chain entity information in "the development of the textile industry" is to be judged, the attention weight vector needs to be calculated according to the formula. Assuming that the result is that the weight of "textile industry" is 0.8, the weight of "of" is 0.1, and the weight of "development" is 0.1, the maximum weight entity is selected, so "textile industry" is selected as the industry chain entity information.

[0153] Preferably, the sequence labeling by the CRF module includes:

[0154] The vector processed by the attention mechanism module is input into the CRF module, and the feature label of the i-th vector is Z i , and the weight value of the Z i feature is λ i , each sentence is S, j is the position of the word in the sentence, the label sequence is L, and the score of the label sequence in the sentence is

[0155]

[0156] In the formula, scroe(L|S) is the score of each label sequence in each sentence, l j is the label of the current word, and l j-1is a mark of the last word, the length of the sentence is m, and the number of features corresponding to each mark sequence is n;

[0157] Normalization processing is performed and converted into probability

[0158]

[0159] wherein ∑ j exp[score[L|S]] represents the sum of all mark sequences, P(L|S) is the normalized probability, and L with the largest output probability value is the final mark sequence.

[0160] For example, the CRF module performs sequence labeling on the vectors processed by the BERT module, the BiLSTM module, and the attention mechanism module, adds constraint rules based on the label format, and further improves the accuracy of industry chain extraction. The constraint rules can be learned in the process of training data, such as "B-Fie, I-Ele, I-Pro" is an incorrect label, the label type should be the same in this mode, and if it is different, it is an incorrect label, and for example, "O, I-Pro" is incorrect, the beginning of the entity should be "B".

[0161] Preferably, the integration of the industry chain entity information into the ontology layer comprises:

[0162] Based on the similarity calculation, an information fusion model is constructed to fuse the extracted industry chain entity information with the ontology layer;

[0163] Based on the industry chain entity information extraction model and the information fusion model, an industry chain relationship graph is generated;

[0164] The construction of the information fusion model comprises:

[0165] S231, obtaining ontology layer entity information vectors through a BERT model;

[0166] S232, calculating the cosine similarity of the ontology layer entity information vectors and the industry chain entity information vectors, and setting a threshold value;

[0167] S233, calculating the linking rate, linking frequency, and topic consistency features of the industry chain entity information and the ontology layer entity information based on a corpus;

[0168] S234, fusing the industry chain entity information and the ontology layer according to the calculation results and the threshold value;

[0169] The linking rate is

[0170]

[0171] In the formula, m is the industry chain entity information, ei is the ontology layer entity information, Γ(m, e j ) is the number of times m is linked to entity e i in the extended document set, E c = D[m] is the candidate entity set of m;

[0172] The link frequency is

[0173]

[0174] In the formula, M D is the set of identified industry chain entity information in the extended document set;

[0175] The topic consistency feature is

[0176]

[0177] In the formula, rel1 is the topic consistency measure based on the ontology layer,

[0178]

[0179] Optionally, the information fusion model integrates and disambiguates the industry chain entity information and the nodes of the ontology layer. The cosine similarity is calculated by calculating the cosine value of two vectors to compare the similarity of two things. The calculation formula is

[0180]

[0181] In the formula, A i and B i are the sub-vectors of vectors A and B, respectively. Since the cosine value ranges from -1 to 1, the closer the cosine value is to 1, the more similar the two vectors are. The entity information with high similarity is integrated. The link rate, link frequency, and topic consistency feature are based on the corpus of identified entities, such as the set of relevant literature identifying industry chain information. The most similar document of the input document d in the corpus is retrieved, and the returned document is called the extended document of d. The similarity measure can apply the three features mentioned above.

[0182] Illustratively, the fusion model also needs to normalize the data to improve the accuracy of fusion. Common methods include syntax regularization and data regularization techniques. The extracted industry chain entity information and ontology layer entity are processed in blocks. Among all possible entity pairs, the candidate entities that may have a correlation are selected to reduce the number of candidate entities and improve the accuracy of matching. Common blocking techniques include adjacent classification methods.

[0183] Optionally, the similarity between entities can also be integrated with the processed data through the calculation of concept similarity auxiliary information, including concept similarity based on concept name, concept similarity based on concept structure and concept similarity based on attribute; the similarity between two words can also be calculated through the graph structure.

[0184] Preferably, the generation of the corresponding industrial chain comprises:

[0185] Training the regional industrial chain generation model based on the evaluation function:

[0186] The labeled regional product and the constructed customized standardized industrial classification system are used as the input layer of the model;

[0187] The weights of the evaluation function are set, including the upstream node weight, the midstream node weight and the downstream node weight;

[0188] The output layer is the regional industrial chain of the terminal node of the regional product industrial chain;

[0189] Through the trained regional industrial chain generation model, the labeled regional product information is input to generate the corresponding regional industrial chain set.

[0190] Illustratively, the process of generating a regional industrial chain through a regional industrial chain generation model is to match the labeled regional product with the nodes in the customized standardized industrial classification system, obtain the optimal solution which can be called the original node, and then retrieve the corresponding upstream, midstream and downstream nodes according to the original node, the relationship between the nodes and the set evaluation function. Among them, the node matching process can use the fusion model or other existing entity linking technology; the evaluation function is to set the weight value of the node according to the regional element Q1, the correlation degree Q2 between the nodes, the attribute A of the node and the distance Q3 between the other nodes and the original node The score is calculated, for example, the attribute of the retrieved node is midstream A2, and the midstream evaluation function is In the formula, and are the corresponding weight values, the weight value of the node is set to different values according to different node attributes, Q1, Q2 and Q3 are set to the corresponding score standards according to different industries and regions, which can be adjusted during model training to obtain the evaluation function of the output better regional industrial chain. Illustratively, if the node attribute is upstream and downstream, it can be represented as A1 and A3 respectively. When setting the weight value of the downstream attribute node, in order to obtain the industrial chain information of the corresponding region, the weight of the regional element is relatively large, while for the midstream and upstream, the correlation between the nodes and the distance value from the original node generally show the relevance of the industry where the regional product is located, so the weight in the evaluation function can be set to be larger, and the weight value of the node can be adjusted according to different situations.

[0191] Preferably, the node traversal and merging of the regional industry chain set generates a regional industry chain, including:

[0192] Selecting a target industry chain according to the regional industry chain set;

[0193] Based on the target industry chain, the nodes of other industry chains are traversed and compared, and the nodes of other industry chains are merged into the target industry chain;

[0194] Proofreading the merged target industry chain to generate a regional industry chain.

[0195] Illustratively, the self-defined standardized industry classification system is constructed by a model and includes industry information, industry chain relationship and regional information in the field; with the development of economy and technology, the industry chain information in different fields and regions will change, and new information data can be input into the constructed model to realize timely updating of the self-defined standardized industry classification system, so that the generated regional industry chain is consistent with the current situation and regional characteristics, and has higher reference value.

[0196] Illustratively, the process of traversal and fusion of the industry chain set can first traverse all nodes by algorithm, remove duplicate information, and then the processed results are audited by experts in the corresponding field to form a complete chain. Among them, the target industry chain is a relatively complete, comprehensive and highly relevant industry chain in the industry chain set; BFS algorithm can be used for traversal, the BFS algorithm is breadth-first traversal, which preferentially traverses the nodes of the same layer of the industry chain, each industry chain can be regarded as a tree to be traversed, starting from a vertex of the tree, accessing all adjacent vertices, after all adjacent nodes are accessed, the next layer of nodes is accessed, and the traversal of all nodes is completed; optionally, an array should also be set to store the access of the nodes, the target industry chain is first traversed and stored, and when traversing other industry chains, it is necessary to judge whether the upstream, midstream and downstream nodes of the target industry chain have the same items, if there are the same items, the duplicate points are discarded, if they are different, they are retained, and the traversal of other nodes is continued. The regional industry chain after traversal and merging eliminates duplicate items and enriches the content of the regional industry chain, and through the audit and processing of experts in the corresponding field, the authority and credibility of the regional industry chain are further improved.

[0197] A regional industry chain generation device, comprising the regional industry chain generation method, comprising:

[0198] M1, information marking module: for marking regional product information;

[0199] M2, system generation module: for establishing an industry chain entity information extraction model and an information fusion model, and generating a self-defined standardized industry classification system;

[0200] M3, regional industry chain generation module: for generating a regional industry chain set corresponding to regional product information;

[0201] M4, merging module: for generating a regional industry chain by traversing and merging the regional industry chain set.

[0202] For example, the information marking module can be provided with an information extraction unit and a marking unit, respectively, for extracting regional product information from regional papers or authoritative materials and marking the extracted content according to a preset standard; the system generation module respectively trains an industry chain entity information extraction model and an information fusion model based on a large amount of corpus, and realizes output of a more comprehensive and detailed self-defined standardized industry classification system through this module; the regional product information marked by the information marking module and the self-defined standardized industry classification system output by the system generation module are taken as inputs of the regional industry chain generation module, a regional industry chain generation model based on an evaluation function is trained, and a set of industry chains with regional products as terminal nodes is output, and a regional industry chain is further generated by traversing and merging the merging module.

[0203] Preferably, since the regional industry chain will produce a new pattern with changes in regional factors, economic development of various industries, environmental factors, technological innovation and many other factors, the above device realizes the generation method of the regional industry chain of the present application, and new industry-related data is input into the device, information extraction and model training are performed again, and a new regional industry chain that conforms to the current information can be output. Therefore, the flexibility of this generation method is higher, and it is not limited by industries and fields, and has good applicability.

[0204] An electronic device, comprising: a memory and one or more processors;

[0205] The memory is in communication connection with the one or more processors, and the memory stores instructions executable by the one or more processors, and when the instructions are executed by the one or more processors, the electronic device is configured to implement the generation method of the regional industry chain as described in any one of the above.

[0206] Specifically, the processor and the memory can be connected through a bus or other means, taking the bus connection as an example. The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations thereof.

[0207] A computer readable storage medium having stored thereon computer executable instructions, which when executed by a computing device, can be used to implement the method for generating an industrial chain of a region as described in any one of the above.

[0208] Those skilled in the art can appreciate that the units and method steps of the examples described in conjunction with the embodiments of the present application can be realized by electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0209] When the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts that essentially contribute to the original technology or the parts of the technical solutions can be embodied in the form of software products, which are stored in a storage medium and include a number of instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the embodiments of the present application.

[0210] In summary, the regional industry chain generation method, device and storage medium of the present application take regional products as the terminal nodes of the industry chain, obtain the information nodes of each industry, the relationship between the nodes and the regional elements corresponding to the industry through the self-defined standardized industry classification system, generate the regional industry chain based on the self-defined standardized industry classification system and regional products, the method adopts the model of deep learning, improves the automation level of the generation of the regional industry chain, the model training and system generation are carried out according to the subdivided fields, the existing classification system, various industry analysis reports and enterprise patents and other data, the information system is relatively standard and complete, the regional elements are embedded, the corresponding regional industry chain is convenient to extract, the regional industry chain with high credibility is generated with high efficiency through the method, important data is provided for the regional development, the cooperation between enterprises is conducive to strengthening and the emergence of new enterprises is promoted, and the government and enterprises are provided with favorable references.

[0211] The above is only the preferred embodiment of the present application, and does not limit the implementation and protection scope of the present application. It should be understood by those skilled in the art that any equivalent replacement and obvious change made by referring to the content of the present application should be included in the protection scope of the present application.

[0212] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

Claims

1. A method for generating a regional industry chain, characterized in that, Includes the following steps: Acquire regional products and standardize their labeling to serve as terminal nodes in the regional industrial chain; A custom standardized industry classification system is constructed based on sub-sectors, existing classification systems, and relevant literature, specifically including: The ontology layer is constructed by acquiring the aforementioned sub-fields, the existing classification system, and the relevant literature. The sub-fields are based on global industry classification standards and further refined by regional factors. The existing classification system includes national economic industry classification, strategic emerging industries classification, securities industry classification, commodity classification, and custom classifications for other industries. The relevant literature includes industry research reports, industry patents and papers, listed company reports, and regional industry research reports. Process and identify relevant documents to extract entity information from the industrial chain; Relevant literature with labeled supply chain entity information is used as training corpus to train the supply chain entity information extraction model; the supply chain entity information extraction model includes: The word vectors of the text are output through a pre-trained BERT module; the output of word vectors of the text through the pre-trained BERT module includes: taking the training corpus text as the input unit; performing a masking prediction task to train and output word-level features; performing a sentence prediction task to train and output sentence-level features; generating three types of representation vectors after task training, including word vectors, sentence vectors and position vectors; and summing the three types of representation vectors to generate word vectors of the text. The word scores are calculated using the BiLSTM module. The BiLSTM module includes: taking the word vectors as input units; setting features for calculating the context information of the text using a forward LSTM layer; setting features for calculating the context information of the text using a backward LSTM layer; and concatenating the outputs of the forward LSTM layer and the backward LSTM layer to output a character sequence probability feature vector with contextual feature information. The attention score for each word is calculated using the attention mechanism module; Sequence labeling is performed using the CRF module, the probability of the labeled sequence is calculated, and the labeled sequence with the highest probability is output as the prediction result; The entity information of the industrial chain is integrated into the ontology layer to generate an industrial chain relationship graph. The merged industry chain relationship map is reviewed, and a custom standardized classification system is generated. Based on the custom standardized industry classification system and the regional industrial chain terminal nodes, corresponding industrial chains are generated to obtain a set of regional industrial chains. The nodes of the aforementioned regional industrial chain set are traversed and merged to generate a regional industrial chain.

2. The method of claim 1, wherein, The processing and identification include: defining custom label types based on the BIO labeling method and industry chain characteristics; the custom labels include regional element (Ele), sub-sector (Fie), industry information (Pro), and industry attribute (Attr); the label format is as follows: when the label type is regional element, the corresponding starting label is "B-Ele" and the intermediate label is "I-Ele"; when the label type is sub-sector classification, the corresponding starting label is "B-Fie" and the intermediate label is "I-Fie"; when the label type is industry information, the corresponding starting label is "B-Pro" and the intermediate label is "I-Pro"; when the label type is industry attribute, the corresponding starting label is "B-Attr" and the intermediate label is "I-Attr".

3. The method for generating a regional industrial chain as described in claim 1, characterized in that, The products in the region include: Tangible products, which are products circulating in regional markets, have grade attributes; Intangible products are elements and technologies unique to a region. Potential products are those that can be developed in the future based on regional factors. The grade attributes of the tangible products are determined by market position, which includes market share, product standardization, and market recognition.

4. The method for generating a regional industrial chain as described in claim 1, characterized in that, The word vectors of the text output by the pre-trained BERT module include: Use the training corpus text as the input unit; Train and output word-level features for a masking prediction task; Train and output sentence-level features for sentence prediction tasks; After training, three types of representation vectors are generated, including word vectors, sentence vectors, and position vectors. The three representation vectors are summed to generate word vectors for the text.

5. The method for generating a regional industrial chain as described in claim 1, characterized in that, The calculation of attention scores for each word through the attention mechanism module includes: Input the feature vector output by the BiLSTM module; calculate the score of the labeled sequence based on the feature weight values, and the attention weight vector is... In the formula y i b represents the feature sequence output by the BiLSTM module. ij The relevance probabilities between words in a text are respectively defined as follows: where p i and p j represent the i-th and j-th word in the document, N is the number of words in the document, f(p i , p j ) represents the similarity score of word p i to word p j , and f(p i , p k ) represents the similarity score of word p i to any word p k in the document.

6. The method for generating a regional industrial chain as described in claim 1, characterized in that, The sequence labeling using the CRF module, and the calculation of the probability of labeling the sequence, includes: The vectors processed by the attention mechanism module are input into the CRF module, and the feature of the i-th vector is marked as Z i , Z i The weight value of the feature is marked as λ i Each sentence is marked as S, j is marked as the position of the word in the sentence, the mark sequence is marked as L, and the score of the mark sequence in the sentence is where scroe(L|S) is the score of each token sequence in each sentence, l j is the token of the current word, j-1 is the token of the previous word, the length of the sentence is m, and the number of features corresponding to each token sequence is n. Normalize and convert to probability where ∑ j exp[score[L|S]] represents the sum of all labeled sequences, P(L|S) is the normalized probability, and the L with the largest output probability value is the final label sequence.

7. The method for generating a regional industrial chain as described in claim 1, characterized in that, The step of traversing and merging nodes in a set of regional industrial chains to generate a regional industrial chain includes: Select the target industry chain based on the industry chain set of the region; Based on the target industry chain, nodes of other industry chains are traversed and compared, and nodes of other industry chains are merged into the target industry chain; The target industrial chain after verification and integration is used to generate a regional industrial chain.

8. A device for generating a regional industrial chain, comprising the method for generating a regional industrial chain as described in any one of claims 1 to 7, characterized in that, include: Information tagging module: used to tag regional product information; System generation module: used to establish an information extraction model and information fusion model for the industrial chain entity, and generate a custom standardized industrial classification system; Regional Industrial Chain Generation Module: Used to generate a collection of regional industrial chains; Merging module: Generates a regional industrial chain by traversing and merging the set of regional industrial chains.

9. An electronic device, characterized in that, include: Memory and one or more processors; The memory is communicatively connected to the one or more processors, and the memory stores instructions that can be executed by the one or more processors. When the instructions are executed by the one or more processors, the electronic device is used to implement the method as described in any one of claims 1-7.

Images