Enterprise intelligence push method and device based on multi-dimensional label vector

Abstract

The present application discloses a method and device for pushing enterprise intelligence based on multi-dimensional tag vectors, which relates to the field of information processing technology. It pushes document intelligence in combination with multi-dimensional tags, making the recommendation more accurate, avoiding the omission of valuable document intelligence, achieving targeted document intelligence recommendation, and improving the recommendation efficiency and accuracy. The method includes: determining the target user to be pushed for enterprise intelligence, constructing the character feature vector of the target user; obtaining a document vector database, which is constructed using multiple document intelligence labeled with explicit tags and implicit tags, each explicit tag labeled with document intelligence is labeled according to the subject word of the document intelligence, and each implicit tag labeled with document intelligence is labeled after the document intelligence is classified according to the enterprise's multi-dimensional classification system; using the character feature vector, performing a search operation in the document vector database, and pushing the retrieved multiple target document intelligence to the target user.

Description

Technical Field

[0001] The present application relates to the field of information processing technology, and in particular to a method and device for pushing enterprise intelligence based on multi-dimensional label vectors. Background Art

[0002] In today's rapidly changing business environment, enterprises have an increasing demand for knowledge intelligence, and with the rapid development of big data and artificial intelligence technology, knowledge management and intelligence services are gradually transforming towards intelligence and personalization. In recent years, personalized recommendation systems have achieved remarkable results in the Internet field. By analyzing user behavior, interest preferences and other data, they provide users with tailored content recommendations, greatly improving user experience and information acquisition efficiency. Therefore, personalized recommendation systems have also been transplanted into internal intelligence services of enterprises, actively recommending comprehensive and accurate intelligence, knowledge, etc. to enterprise employees, so that enterprise employees can obtain accurate enterprise-related intelligence from massive knowledge.

[0003] In related technologies, enterprise intelligence push systems with personalized recommendation functions mostly make recommendations based on user historical behavior or simple demographic characteristics. However, most users in an enterprise have specific needs, and their professional knowledge backgrounds and departments are different. This simple recommendation method will lead to insufficient pertinence and relevance of recommendation results, low recommendation efficiency and accuracy, and easy neglect of valuable content, making it difficult to effectively support the company's innovative decisions and business development. Summary of the invention

[0004] In view of this, the present application provides a method and device for pushing enterprise intelligence based on multi-dimensional label vectors, the main purpose of which is to solve the problems of insufficient pertinence and relevance of current recommendation results, low recommendation efficiency and accuracy, easy neglect of valuable content, and difficulty in effectively supporting the innovative decision-making and business development of enterprises.

[0005] According to the first aspect of the present application, a method for pushing enterprise intelligence based on a multi-dimensional tag vector is provided, the method comprising:

[0006] Determine the target user to whom enterprise intelligence is to be pushed, and construct a character feature vector of the target user;

[0007] Acquire a document vector database, wherein the document vector database is constructed using a plurality of document intelligences annotated with explicit tags and implicit tags, wherein each explicit tag annotated with the document intelligence is annotated according to a subject word of the document intelligence, and each implicit tag annotated with the document intelligence is annotated after the document intelligence is classified according to a multi-dimensional classification system of the enterprise;

[0008] The character feature vector is used to perform a search operation in the document vector database, and the retrieved multiple target document information is pushed to the target user.

[0009] Optionally, before determining the target user to be pushed enterprise intelligence and constructing the character feature vector of the target user, the method further includes:

[0010] Collecting a plurality of corpus data and a plurality of core documents related to the enterprise to which the target user belongs, and creating a subject word library for the enterprise based on the plurality of corpus data and the keywords and subject words corresponding to each of the core documents;

[0011] The subject word library is used as an explicit tag library, the explicit tag library is used to annotate the multiple core documents, and the annotated multiple core documents are used to train an explicit tagging model;

[0012] Creating an implicit label library according to the multi-dimensional classification system associated with the enterprise, using the implicit label library to label the multiple core documents, and using the labeled multiple core documents to train an implicit label labeling model;

[0013] Acquire the plurality of document information under the enterprise, determine the keywords and subject terms corresponding to each of the document information, and input each of the document information and the corresponding keywords and subject terms into the explicit label marking model, so that the explicit label marking model marks an explicit label for each of the document information;

[0014] Determining abnormal explicit tags that are not related to the enterprise among the explicit tags annotated for each of the document information based on the subject word library, and replacing the abnormal explicit tags with default tags;

[0015] Inputting each of the document information into the implicit label marking model, so that the implicit label marking model marks an implicit label for each of the document information;

[0016] The explicit labels annotated with each of the document intelligence are combined to form a label phrase corresponding to each of the document intelligence, and the label phrase corresponding to each of the document intelligence is converted into a label vector. The document vector database is constructed using the converted label vector and the annotated multiple document intelligence.

[0017] Optionally, creating a subject word library for the enterprise based on the plurality of corpus data and the keywords and subject words corresponding to each of the core documents includes:

[0018] Using the plurality of corpus data, construct a seed subject word library for the enterprise;

[0019] The TF-IDF word frequency method is used to analyze the content of each core document and determine the corresponding keywords for each core document;

[0020] Using the pre-trained BERT model, determine the corresponding keyword for each core document;

[0021] Filter the keywords and subject terms corresponding to each of the core documents, delete the keywords or subject terms that are not relevant to the enterprise, and integrate the keywords and subject terms retained in each of the core documents to form an initial word library;

[0022] Comparing the words included in the initial word library with the words included in the seed theme word library semantically, determining the words to be supplemented in the initial word library, the words to be supplemented being words whose semantic similarity between the words in the initial word library and the words in the seed theme word library is greater than a similarity threshold;

[0023] The words to be supplemented are added to the seed subject word library, and the supplemented seed subject word library is used as the subject word library.

[0024] Optionally, the pre-trained BERT model is used to determine corresponding keywords for each core document, including:

[0025] For each of the core documents, encoding the core document using the pre-trained BERT model to extract a semantic vector representation of the core document;

[0026] Performing word segmentation processing on the core document to obtain a plurality of word segmentation phrases, and extracting a word vector corresponding to each of the word segmentation phrases;

[0027] The cosine similarity between the semantic vector representation and the word vector corresponding to each of the segmentation phrases is calculated, and the segmentation phrase with the largest cosine similarity among the multiple segmentation phrases is used as the subject word corresponding to the core document.

[0028] Optionally, the step of using the implicit tag library to tag the plurality of core documents, and using the annotated plurality of core documents to train an implicit tag tagging model includes:

[0029] Using the implicit tag library to annotate the plurality of core documents to obtain the annotated plurality of core documents;

[0030] Preprocessing the original texts of the annotated core documents, wherein the preprocessing includes word segmentation, stop word removal, and text format conversion;

[0031] Calculating the word frequency vectors of the processed original texts of the plurality of core documents, and generating a text feature vector of each of the core documents based on the word frequency vectors;

[0032] A shallow neural network is used to learn the correspondence between the text feature vector of each core document and the annotated implicit label to obtain the implicit label annotation model.

[0033] Optionally, constructing the character feature vector of the target user includes:

[0034] Constructing a character portrait feature model of the target user based on the basic user information and historical interaction information of the target user;

[0035] Referring to the character portrait feature model, determining at least one designated document having an interactive relationship with the target user, obtaining a plurality of explicit tags annotated by the at least one designated document, and using the plurality of explicit tags as a plurality of temporal attribute tags;

[0036] Calculating a time sharp reduction coefficient for each of the time series attribute labels, and calculating a label weight for each of the time series attribute labels using the time sharp reduction coefficient corresponding to each of the time series attribute labels;

[0037] According to the calculated label weights, the plurality of time series attribute labels are divided into a preset number of label groups;

[0038] Using a random function, extracting a plurality of target time series attribute labels whose number meets a quantity threshold corresponding to the label group in each label group;

[0039] Extracting a static attribute label of the target user from the character attribute and professional attribute of the target user indicated by the character portrait feature model;

[0040] The static attribute label and the multiple time series attribute labels are used to construct a character label group of the target user, and the character label group is vector-converted to obtain a character feature vector of the target user.

[0041] Optionally, the using the character feature vector to perform a search operation in the document vector database and push the retrieved multiple target document intelligence to the target user includes:

[0042] Based on the character feature vector, searching for multiple candidate document information in the document vector database;

[0043] According to the historical browsing situation of the target user, a first intermediate value is set for each candidate document information, wherein, for each candidate document information, when the historical browsing situation indicates that the candidate document information has been browsed, the first intermediate value corresponding to the candidate document information is set to 0, and when the historical browsing situation indicates that the candidate document information has not been browsed, the first intermediate value corresponding to the candidate document information is set to 1;

[0044] Obtaining a preset gravity coefficient, and using the page views, publishing time, and the preset gravity coefficient of each candidate document information to calculate a second intermediate value for each candidate document information;

[0045] Calculate a ranking weight for each candidate document information based on the first intermediate value and the second intermediate value corresponding to each candidate document information;

[0046] Sorting the plurality of candidate document information according to the ranking weight corresponding to each candidate document information to obtain a ranking result;

[0047] A preset push quantity is determined, the plurality of target document information is extracted from the sorting result according to the preset push quantity, and the plurality of target document information is pushed to the target user.

[0048] According to the second aspect of the present application, there is provided a device for pushing enterprise intelligence based on a multi-dimensional tag vector, the device comprising:

[0049] A construction module, used to determine the target user to be pushed enterprise intelligence and construct a character feature vector of the target user;

[0050] An acquisition module is used to acquire a document vector database, wherein the document vector database is constructed using a plurality of document intelligences annotated with explicit tags and implicit tags, wherein each explicit tag annotated with the document intelligence is annotated according to a subject word of the document intelligence, and each implicit tag annotated with the document intelligence is annotated after the document intelligence is classified according to a multi-dimensional classification system of an enterprise;

[0051] The push module is used to use the character feature vector to perform a search operation in the document vector database and push the retrieved multiple target document information to the target user.

[0052] Optionally, the construction module is also used to collect multiple corpus data and multiple core documents related to the enterprise to which the target user belongs, and create a subject word library for the enterprise based on the multiple corpus data and the keywords and subject words corresponding to each of the core documents; use the subject word library as an explicit tag library, use the explicit tag library to annotate the multiple core documents, and use the annotated multiple core documents to train an explicit tag annotation model; create an implicit tag library based on the multi-dimensional classification system associated with the enterprise, use the implicit tag library to annotate the multiple core documents, and use the annotated multiple core documents to train an implicit tag annotation model; obtain the multiple document information under the enterprise, determine the keywords and subject words corresponding to each of the document information, and annotate each of the document information Document information and corresponding keywords and subject terms are input into the explicit label annotation model so that the explicit label annotation model annotates explicit labels for each of the document information; based on the subject word library, abnormal explicit labels that are not related to the enterprise are determined in the explicit labels annotated for each of the document information, and the abnormal explicit labels are replaced with default labels; each of the document information is input into the implicit label annotation model so that the implicit label annotation model annotates implicit labels for each of the document information; the explicit labels annotated for each of the document information are combined to form a label phrase corresponding to each of the document information, and the label phrase corresponding to each of the document information is converted into a label vector, and the document vector database is constructed using the converted label vector and the annotated multiple document information.

[0053] Optionally, the construction module is used to use the multiple corpus data to construct a seed subject vocabulary for the enterprise; use the TF-IDF word frequency method to analyze the document content of each of the core documents, and determine corresponding keywords for each of the core documents; use the pre-trained BERT model to determine corresponding subject words for each of the core documents; filter the keywords and subject words corresponding to each of the core documents, delete the keywords or subject words that are not related to the enterprise, and integrate the keywords and subject words retained by each of the core documents to form an initial vocabulary; perform a semantic comparison between the words included in the initial vocabulary and the words included in the seed subject vocabulary, and determine the words to be supplemented in the initial vocabulary, wherein the semantic similarity between the words in the initial vocabulary and the words included in the seed subject vocabulary is greater than a similarity threshold; add the words to be supplemented to the seed subject vocabulary, and use the supplemented seed subject vocabulary as the subject vocabulary.

[0054] Optionally, the construction module is used to encode each of the core documents using the pre-trained BERT model to extract the semantic vector representation of the core document; perform word segmentation on the core document to obtain multiple word segmentation phrases, and extract the word vector corresponding to each of the word segmentation phrases; calculate the cosine similarity between the semantic vector representation and the word vector corresponding to each of the word segmentation phrases, and use the word segmentation phrase with the largest cosine similarity among the multiple word segmentation phrases as the subject word corresponding to the core document.

[0055] Optionally, the construction module is used to use the implicit tag library to annotate the multiple core documents to obtain the annotated multiple core documents; preprocess the original texts of the annotated multiple core documents, and the preprocessing includes word segmentation, stop word removal and text format conversion; calculate the word frequency vectors of the original texts of the multiple core documents after processing, and generate a text feature vector of each of the core documents based on the word frequency vector; use a shallow neural network to learn the correspondence between the text feature vector of each of the core documents and the annotated implicit tags to obtain the implicit tag annotation model.

[0056] Optionally, the construction module is used to construct a character portrait feature model of the target user based on the basic user information and historical interaction information of the target user; determine at least one designated document that has an interactive relationship with the target user with reference to the character portrait feature model, obtain multiple explicit tags annotated by the at least one designated document, and use the multiple explicit tags as multiple time series attribute tags; calculate a time sharp reduction coefficient for each of the time series attribute tags, and calculate a label weight for each of the time series attribute tags using the time sharp reduction coefficient corresponding to each of the time series attribute tags; divide the multiple time series attribute tags into a preset number of label groups according to the calculated label weights; use a random function to extract multiple target time series attribute tags in each of the label groups whose number meets the number threshold corresponding to the label group; extract the static attribute label of the target user from the character feature attributes and professional attributes of the target user indicated by the character portrait feature model; use the static attribute label and the multiple time series attribute labels to construct a character label group of the target user, and perform vector conversion on the character label group to obtain a character feature vector of the target user.

[0057] Optionally, the push module is used to retrieve multiple candidate document intelligence in the document vector database based on the character feature vector; set a first intermediate value for each of the candidate document intelligence according to the historical browsing situation of the target user, wherein, for each of the candidate document intelligence, when the historical browsing situation indicates that the candidate document intelligence has been browsed, the first intermediate value corresponding to the candidate document intelligence is set to 0, and when the historical browsing situation indicates that the candidate document intelligence has not been browsed, the first intermediate value corresponding to the candidate document intelligence is set to 1; obtain a preset gravity coefficient, and use the browsing volume, release time and preset gravity coefficient of each of the candidate document intelligence to calculate a second intermediate value for each of the candidate document intelligence; calculate a ranking weight for each of the candidate document intelligence based on the first intermediate value and the second intermediate value corresponding to each of the candidate document intelligence; sort the multiple candidate document intelligence according to the ranking weight corresponding to each of the candidate document intelligence to obtain a ranking result; determine a preset push quantity, extract the multiple target document intelligence from the ranking result according to the preset push quantity, and push the multiple target document intelligence to the target user.

[0058] According to a third aspect of the present application, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the methods described in the first aspect when executing the computer program.

[0059] According to a fourth aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of any one of the methods described in the first aspect are implemented.

[0060] Through the above technical scheme, the present application provides a method and device for pushing enterprise intelligence based on multi-dimensional label vectors. The present application determines the target users to be pushed enterprise intelligence, constructs the character feature vectors of the target users, and obtains a document vector database. The document vector database is constructed using multiple document intelligence labeled with explicit labels and implicit labels. The explicit label labeled with each document intelligence is labeled according to the subject words of the document intelligence, and the implicit label labeled with each document intelligence is labeled after the document intelligence is classified according to the multi-dimensional classification system of the enterprise. The character feature vector is used to perform a search operation in the document vector database, and the retrieved multiple target document intelligence is pushed to the target users. The document intelligence is pushed in combination with multi-dimensional labels, so that the knowledge recommendation in the business scenario is more accurate. While ensuring that intelligence outside the technical field is not recommended, it can also avoid missing valuable document intelligence, realize targeted document intelligence recommendation, and improve recommendation efficiency and accuracy.

[0061] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0063] Figure 1 A schematic diagram of a method for pushing enterprise intelligence based on a multi-dimensional tag vector provided in an embodiment of the present application is shown;

[0064] Figure 2A A schematic diagram of another method for pushing enterprise intelligence based on a multi-dimensional tag vector provided in an embodiment of the present application is shown;

[0065] Figure 2B A schematic diagram of a process flow of another enterprise intelligence push method based on a multi-dimensional tag vector provided in an embodiment of the present application is shown;

[0066] Figure 3 A schematic diagram of the structure of a device for pushing enterprise intelligence based on a multi-dimensional tag vector provided in an embodiment of the present application is shown;

[0067] Figure 4 A schematic diagram of the device structure of a computer device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0068] The exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0069] The embodiment of the present application provides a method for pushing enterprise intelligence based on a multi-dimensional tag vector, such as Figure 1 As shown, the method includes:

[0070] 101. Determine the target users to whom enterprise intelligence is to be pushed and construct the character feature vectors of the target users.

[0071] The technical solution of this application can be applied to the intelligence push system. In the embodiment of this application, the target users of the enterprise intelligence push need to be clarified first. The target users can be employees in the enterprise. For example, for enterprises in the field of scientific and technological research and development, the target users can be key decision-makers such as research and development managers and technical directors.

[0072] Among them, different target users have different interests and needs. Therefore, in the embodiment of the present application, the multi-dimensional information of the target users, such as the user industry background, job role, interest preferences and information needs, will be comprehensively analyzed, and these user characteristics will be quantified to form a computable vector representation as the character feature vector of the target user, so as to achieve accurate identification and matching of user needs, and provide a basis for the accurate push of subsequent literature intelligence.

[0073] 102. Obtain a document vector database.

[0074] In the embodiment of the present application, the document vector database is constructed by using multiple document intelligences marked with explicit tags and implicit tags, including a large number of document intelligences marked with explicit tags and implicit tags, which are derived from multiple channels such as academic journals, patent databases, and industry reports of enterprises. Among them, the explicit tags marked by each document intelligence are marked according to the subject words of the document intelligence, and are used to quickly locate the content of the document. The implicit tags marked by each document intelligence are marked after the document intelligence is classified according to the multi-dimensional classification system of the enterprise, and are used to quickly determine the technical field to which the document belongs and the department to which the document-related knowledge matches. When constructing the document vector database, it is necessary to use text processing technology to vectorize the document content and document category, ensure that the relevant content of each document intelligence can be stored in the document vector database in the form of a high-dimensional vector, while improving the efficiency of document retrieval, through the dual marking of explicit tags and implicit tags, the retrieval and push of document intelligence are more accurate, effectively improving the pertinence and effectiveness of intelligence services.

[0075] 103. Utilize the character feature vector to perform a search operation in the document vector database, and push the retrieved multiple target document information to the target user.

[0076] After constructing the character feature vector and document vector database of the target user, the search operation can be performed in the database, and the retrieved document information can be pushed to the target user as the target document information, thereby realizing intelligent and personalized document information push.

[0077] The method provided in the embodiment of the present application determines the target user to whom enterprise intelligence is to be pushed, constructs a character feature vector of the target user, and obtains a document vector database. The document vector database is constructed using multiple document intelligence labeled with explicit labels and implicit labels. The explicit label labeled with each document intelligence is labeled according to the subject word of the document intelligence, and the implicit label labeled with each document intelligence is labeled after the document intelligence is classified according to the multi-dimensional classification system of the enterprise. The character feature vector is used to perform a search operation in the document vector database, and multiple retrieved target document intelligence is pushed to the target user. The document intelligence is pushed in combination with multi-dimensional labels, so that knowledge recommendation in business scenarios is more accurate. While ensuring that intelligence outside the technical field is not recommended, it can also avoid missing valuable document intelligence, achieve targeted document intelligence recommendation, and improve recommendation efficiency and accuracy.

[0078] Further, as a refinement and extension of the specific implementation of the above embodiment, in order to fully illustrate the specific implementation process of this embodiment, the embodiment of the present application provides another enterprise intelligence push method based on a multi-dimensional tag vector, such as Figure 2A As shown, the method includes:

[0079] 201. Collect multiple corpus data and multiple core documents related to the enterprise to which the target user belongs.

[0080] The technical solution of the embodiment of the present application can be applied to the intelligence push system. The intelligence push system can be applied to various enterprises and organizations, especially in areas where a large amount of professional knowledge and the latest research results need to be quickly acquired and applied. The intelligence push system can intelligently push intelligence knowledge and documents closely related to its business according to the personalized needs of enterprise employees. Whether it is scientific researchers who need to track the latest scientific research trends and technological advances, market personnel who need to understand industry trends and competitor information, or managers who need to master policies, regulations and market trends, the intelligence push system can provide accurate and efficient information services to ensure that enterprise employees can quickly find what they need in massive information, thereby improving decision-making efficiency.

[0081] In order to push document intelligence, the information push system needs to collect a wide range of information documents and diversified corpus data that are closely related to the target user's company. These data come from a wide range of sources, including but not limited to academic journals, doctoral and master's theses, industry information, policies and regulations, internal research topics and results of enterprises, etc. In addition, in order to improve the depth and breadth of semantic understanding, it is also necessary to collect language resources such as entries, semantic dictionaries, and Chinese-English dictionaries. These resources together constitute the knowledge basis for document intelligence push.

[0082] During the collection process, the intelligence push system will use natural language processing technology and data mining algorithms to pre-process the collected documents and data, including text cleaning, word segmentation, part-of-speech tagging, etc., to facilitate subsequent analysis and modeling. Based on these pre-processed data, the intelligence push system will further screen out core documents so that these core documents can be used for model training in the future, providing a solid foundation for subsequent semantic analysis and personalized recommendations.

[0083] 202. Create a thesaurus for the enterprise based on multiple corpus data and the keywords and subject terms corresponding to each core document.

[0084] In the embodiment of the present application, the intelligence push system will create a subject word library for the enterprise based on multiple corpus data and the keywords and subject words corresponding to each core document. Among them, since there may be some inconsistencies between subject words and keywords, in order to obtain document features more comprehensively, subject words and keywords are extracted according to different algorithms. The following describes the process of creating a subject word library:

[0085] First, use multiple corpus data to build a seed subject vocabulary for the enterprise.

[0086] Subsequently, the TF-IDF (Term Frequency-Inverse Document Frequency) method was used to analyze the content of each core document and determine the corresponding keywords for each core document. TF-IDF is a commonly used text mining and information retrieval technology used to evaluate the importance of a word to a document set or one of the documents in a corpus. Therefore, the TF-IDF method determines the importance of a word by counting the frequency of occurrence of the word in the document and its inverse document frequency, thereby extracting keywords for each core document.

[0087] In order to gain a deeper understanding of the subject of the document, the intelligence push system uses a pre-trained BERT (Bidirectional Encoder Representations from Transformers) model to determine the corresponding subject words for each core document. BERT is a pre-trained language representation model that can capture the contextual information of vocabulary and provide powerful semantic understanding capabilities. Specifically, when determining the subject words, for each core document, the intelligence push system will use the pre-trained BERT model to encode the core document, extract the semantic vector representation of the core document, and perform word segmentation on the core document to obtain multiple word segmentation phrases, and extract the word vector corresponding to each word segmentation phrase, continue to calculate the cosine similarity between the semantic vector representation and the word vector corresponding to each word segmentation phrase, and use the word segmentation phrase with the largest cosine similarity among the multiple word segmentation phrases as the subject word corresponding to the core document. In actual application, since the KeyBert (Keyword and Keyphrase Extraction using BERT) algorithm combines BERT and N-gram word segmentation technology, it can be used to efficiently extract key phrases or sentences from long documents. Therefore, based on the KeyBert algorithm, we can first use Bert to extract the document vector to obtain the semantic vector representation of the document, then use N-gram words / phrases to segment and extract word vectors, and then use vector cosine similarity to find the topic words or topic words and sentences that are most similar to the document.

[0088] Next, the intelligence push system will filter the keywords and subject terms corresponding to each core document, delete keywords or subject terms that are not relevant to the enterprise, remove content that is not related to the enterprise, and integrate the keywords and subject terms retained in each core document to form an initial vocabulary.

[0089] Then, the intelligence push system makes a semantic comparison between the words included in the initial vocabulary and the words included in the seed subject vocabulary, determines the words to be supplemented in the initial vocabulary, adds the words to be supplemented to the seed subject vocabulary, and uses the supplemented seed subject vocabulary as the subject vocabulary. The words to be supplemented are words whose semantic similarity between the words in the initial vocabulary and the words in the seed subject vocabulary is greater than the similarity threshold. That is to say, the intelligence push system adds the words in the initial vocabulary with high semantic similarity to the seed subject vocabulary (i.e., words to be supplemented) to the seed subject vocabulary through semantic comparison, and continuously supplements them, gradually expands and updates the seed subject vocabulary, so as to enrich and improve the subject vocabulary of the enterprise, provide a solid foundation for subsequent vocabulary extraction and subject analysis, and help improve the accuracy of literature analysis and intelligence push.

[0090] 203. Build explicit tag library and implicit tag library.

[0091] In an embodiment of the present application, the intelligence push system will use the subject word library as the explicit tag library. The explicit tags included in the explicit tag library are common keywords and subject words, which can directly reflect the main features and fields of the document or content. Subsequently, the intelligence push system will create an implicit tag library based on the multi-dimensional classification system associated with the enterprise. The implicit tags included in the implicit tag library are knowledge systems associated with the enterprise's business, such as technical field systems, job systems, training systems, document systems, etc. The extraction of implicit tags requires a deep understanding of the business logic and knowledge system of the enterprise. They are often not directly reflected on the content surface, but are obtained through certain analysis, induction and mapping processes. These implicit tags can reveal deeper content associations and user needs.

[0092] By building explicit tag libraries and implicit tag libraries, the intelligence push system can more comprehensively capture and represent document features and character features, thereby enhancing the accuracy and depth of the intelligence push system. Explicit tags ensure the basic relevance of recommended content, while implicit tags further improve the accuracy and personalization of recommendations. In actual applications, the explicit tag library and implicit tag library may include the following content as shown in Table 1:

[0093] Table 1

[0094]

[0095]

[0096] 204. An explicit label library is used to annotate multiple core documents, and the annotated multiple core documents are used to train an explicit label annotation model. An implicit label library is used to annotate multiple core documents, and the annotated multiple core documents are used to train an implicit label annotation model.

[0097] In the embodiment of the present application, the intelligence push system uses the explicit tag library and the implicit tag library to annotate multiple core documents, and trains the explicit tag annotation model and the implicit tag annotation model based on these annotation data.

[0098] When training the explicit label annotation model, the intelligence push system first uses Bert's word segmenter to segment the core documents, then loads the BERT-Base-Uncased (BERT basic case-insensitive version) model for data pre-training, and fine-tunes the model for classification, thereby obtaining an explicit label annotation model that can automatically label text with keyword tags.

[0099] When training the implicit label annotation model, it is necessary to use the implicit label library to annotate multiple core documents to obtain multiple annotated core documents. Then, the original texts of the annotated multiple core documents are preprocessed, including word segmentation, stop word removal, and text format conversion. Next, the word frequency vectors of the original texts of the processed multiple core documents are calculated, and based on the word frequency vectors, the text feature vectors of each core document are generated. Finally, a shallow neural network is used to learn the correspondence between the text feature vectors of each core document and the annotated implicit labels to obtain the implicit label annotation model.

[0100] Specifically, in the process of actual application, an enterprise often covers multiple classification systems. These systems are independent and complementary to each other, and together they build a diversified structure of enterprise knowledge. Therefore, enterprises will establish multiple implicit tag libraries based on these classification systems in order to more comprehensively manage and utilize knowledge resources. In this case, the FastText model can be used to pre-train the implicit tag libraries formed by each internal classification system. During the pre-training process, the FastText model will first segment the original text, remove stop words, and convert uppercase and lowercase letters to ensure the consistency and accuracy of text processing. Then, the word frequency vector is calculated using n-gram technology. n-gram is a language model based on adjacent word combinations. It can capture the contextual relationship between words, thereby more accurately reflecting the content characteristics of the text. By calculating the word frequency vector through n-gram, a comprehensive and detailed description of the text content can be provided. Finally, these word frequency vectors will be further converted into text feature vectors as inputs of shallow neural networks. Shallow neural networks usually use a single hidden layer fully connected network structure, which can efficiently learn and classify feature vectors. Through continuous training and optimization, the model can gradually grasp the correspondence between text and implicit labels, thereby realizing automatic classification and labeling of text.

[0101] It should be noted that if an enterprise has different classification systems in multiple dimensions, the embodiment of the present application needs to construct an implicit label annotation model for each classification system. In this way, when a new text needs to be classified, its corresponding feature vector can be calculated based on the content characteristics of the text, and it can be classified into multiple classification systems of the enterprise, which not only improves the efficiency and accuracy of text processing, but also provides a more convenient and efficient way to manage and utilize enterprise knowledge.

[0102] 205. Based on the explicit labeling model and the implicit labeling model, explicit labeling and implicit labeling are performed on multiple document information of the enterprise.

[0103] In the embodiment of the present application, the information push system will perform explicit labeling and implicit labeling on multiple document intelligence under the enterprise based on the explicit labeling model and the implicit labeling model.

[0104] Specifically, the information push system will obtain multiple document intelligence under the enterprise, determine the keywords and subject words corresponding to each document intelligence, and input each document intelligence and the corresponding keywords and subject words into the explicit label annotation model, so that the explicit label annotation model annotates explicit labels for each document intelligence. Among them, the process of determining the keywords and subject words of each document intelligence is consistent with the process of determining keywords and subject words described in the above step 202, and will not be repeated here; in addition, referring to the content in the above Table 1, the information push system will annotate the document intelligence with explicit labels such as keywords, subject words, and secondary subject words. Since the explicit label annotation model is derived based on a large amount of training data, it may sometimes annotate some abnormal explicit labels that are not related to the enterprise. In order to deal with these abnormal labels, the information push system will screen out labels that are not related to the enterprise from the annotated explicit labels based on the existing subject word library, and use default labels (such as "other subject words") to replace these abnormal labels, thereby ensuring that the annotated explicit labels are related to the actual business of the enterprise and improving the accuracy and practicality of the labels. In addition to the explicit label annotation, the information push system will also input each document intelligence into the implicit label annotation model.

[0105] 206. Using the annotated literature information, a literature vector database is constructed.

[0106] In the embodiment of the present application, the information push system will use the annotated multiple document intelligence to build a document vector database. Among them, the information push system will combine the explicit labels annotated by each document intelligence to form a label phrase corresponding to each document intelligence; specifically, the title, keywords, subject words and other core meanings of the document can be merged to form a more descriptive and distinguishing label phrase, which can more comprehensively reflect the theme and content of the document intelligence and facilitate subsequent processing and analysis.

[0107] Then, using natural language processing technology, such as the Bert model, the label phrase corresponding to each document intelligence is converted into a label vector. By converting the label phrase into a label vector, the intelligence push system can convert the content of the document intelligence into a form that can be understood and processed by computers, providing a basis for subsequent data analysis and mining.

[0108] Finally, the document vector database is constructed using the converted label vectors and the annotated document intelligence. The document vector database can store a large amount of document intelligence and its corresponding label vectors, and supports fast retrieval and matching. By constructing a document vector database, the information push system can achieve efficient management and utilization of document intelligence, perform fast retrieval and matching, and improve the utilization efficiency of document intelligence.

[0109] 207. Determine the target users to whom the enterprise intelligence is to be pushed, and construct the character feature vectors of the target users.

[0110] After the preparation of the literature intelligence data is completed through the above process, the intelligence push system will determine the target users to be pushed for enterprise intelligence, construct the character feature vector of the target users, and analyze the target users so as to push the literature intelligence matching them to the target users. Among them, the intelligence push system will first build a character portrait feature model of the target users based on the basic user information and historical interaction information of the target users. When building the character portrait feature model, the intelligence push system will determine the basic attributes, professional attributes, interest attributes and collaborative generalization attributes of the target users by analyzing the basic user information and historical interaction information of the target users. Among them, the basic attributes cover the basic information of the target users such as educational experience, language ability, hobbies, age, gender, etc.; the professional attributes record the target users' positions, project experience, technical fields, department information, etc. in detail; the interest attributes directly reflect the reading preferences of the target users through their browsing, downloading, collection, sharing and other behaviors of articles; the collaborative generalization attributes further expand the potential interest areas of the target users by mining the articles read by people in the same department, the articles read by people they follow, and what articles people who have read the articles have also read. The intelligence push system will extract the above-mentioned character feature attributes and represent them in a labeled form, and finally integrate them to form a character portrait feature model of the target user. The character portrait feature model not only comprehensively reflects the personal characteristics and interest preferences of the target user, but also provides accurate target positioning and personalized push strategies for subsequent intelligence push, ensuring that the pushed literature intelligence can closely fit the needs and interests of the target user.

[0111] After the document vector database and character portrait model are constructed, in order to further solve the challenges of huge data push volume, long-tail data phenomenon and uneven data distribution, the embodiment of the present application will perform efficient dimensionality reduction processing on the character feature labels, aiming to unify and simplify the complex features of the characters in different dimensions, and only divide them according to their popularity to form a character label group of the target user, and then perform vector conversion on the character label group to obtain the character feature vector of the target user. The process of generating the character feature vector is described below:

[0112] First, the information push system will refer to the character portrait feature model, determine at least one designated document that has an interactive relationship with the target user, obtain multiple explicit tags annotated by at least one designated document, and use the multiple explicit tags as multiple time-series attribute tags. Specifically, the core subject word tags and secondary subject word tags of the document information can be extracted from the document information involved in the interest attribute and the collaborative generalization attribute to form a time-series attribute tag.

[0113] In the embodiment of the present application, a preset time sharp reduction coefficient is set for the time series attribute label, and the time sharp reduction coefficient can be calculated using an exponential decay function, as shown in the following formula 1:

[0114] Formula 1: y=ae^{-bx}

[0115] Among them, y is the time-decreasing coefficient, a is a constant, which is the preset value when t=0; b is also a constant, which is the decay rate, which is a preset value; e is the base of the natural logarithm, which is the mathematical constant e, which is an irrational number, approximately equal to 2.71828.

[0116] After calculating the time sharp reduction coefficient for each of the time series attribute labels by the above formula 1, the time sharp reduction coefficient corresponding to each of the time series attribute labels can be used to calculate the label weight for each of the time series attribute labels. Specifically, since different literature reports may mark the same explicit label, the time series attribute labels may be repeated. Therefore, when calculating the label weight, for each time series attribute label, its label weight is equal to the sum of the weights of the repeated labels. The specific label weight can be calculated by the following formula 2:

[0117] Formula 2:

[0118] In Formula 2, Sum(T) is used to represent the label weight; T k (S k ) is the label score of T label hot and cold score matrix, which is a preset value. The label scores of each attribute label in the character portrait model can be found in the following Table 2; p is the secondary coefficient preset for the explicit label indicating the secondary subject word. When calculating the label weight for the explicit label indicating the core subject word, there is no need to consider the p value; ae^{-bx} refers to the time sharp reduction coefficient calculated by formula 1 above.

[0119] Table 2

[0120]

[0121] In this way, through the above process, the label weight can be calculated for each time series attribute label, and according to the calculated label weight, multiple time series attribute labels can be divided into a preset number of label groups. Specifically, when grouping, it can be divided into 3 label groups, among which, the time series attribute labels with label weight values ​​greater than or equal to 6 are divided into [extremely hot label group]; the time series attribute labels with label weight values ​​less than 6 and greater than or equal to 3 are divided into [hot label group]; the time series attribute labels with label weight values ​​less than 3 are divided into [unpopular label group].

[0122] After completing the grouping of time series attribute labels, you can use the random function to extract multiple target time series attribute labels in each label group whose number meets the corresponding quantity threshold of the label group. In actual application, you can use the random function Rand(n) to extract 4 labels from the [extremely popular label group], 4 labels from the [popular label group], and 2 labels from the [unpopular label group].

[0123] Next, since the character portrait feature model has other attributes, the intelligence push system will continue to extract the static attribute label of the target user from the character feature attributes and professional attributes of the target user indicated by the character portrait feature model, use the static attribute label and multiple time series attribute labels to construct the character label group of the target user, and perform vector conversion on the character label group to obtain the character feature vector of the target user. Specifically, the Bert model can be used to convert the character label group into a character feature vector.

[0124] 208. Utilize the character feature vector to perform a search operation in the document vector database, and push the retrieved multiple target document information to the target user.

[0125] In the embodiment of the present application, the information push system will use the character feature vector to perform a search operation in the document vector database, and push the retrieved multiple target document intelligence to the target user. Among them, since the number of pushed document intelligence is limited, the information push system will further screen the matched document intelligence so that those valuable document intelligence can be pushed to the target user in a limited manner. The specific process of screening multiple target document intelligence is as follows:

[0126] First, the intelligence push system will retrieve multiple candidate document intelligence in the document vector database based on the character feature vector.

[0127] Subsequently, according to the historical browsing situation of the target user, the information push system sets a first intermediate value for each candidate document information. The first intermediate value is used to filter out the candidate document information that the target user has browsed. Therefore, for each candidate document information, when the historical browsing situation indicates that the candidate document information has been browsed, the information push system sets the first intermediate value corresponding to the candidate document information to 0, and when the historical browsing situation indicates that the candidate document information has not been browsed, the information push system sets the first intermediate value corresponding to the candidate document information to 1. In the embodiment of the present application, f 1(g) is used to represent the first intermediate value.

[0128] Next, the information push system obtains the preset gravity coefficient, and uses the page views, release time and preset gravity coefficient of each candidate document information to calculate the second intermediate value for each candidate document information. The second intermediate value can be calculated using the following formula 3:

[0129] Formula 3:

[0130] Among them, g is used to refer to a certain document information, f 2(g) is the second intermediate value calculated for the document information, B is the number of views of the document information, T is the release time of the document information, and G is the preset gravity coefficient, and the value of G can be 1.8.

[0131] After calculating the first intermediate value and the second intermediate value corresponding to each candidate document information, the information push system calculates a ranking weight for each candidate document information based on the first intermediate value and the second intermediate value corresponding to each candidate document information. The ranking weight can be calculated using the following formula 4:

[0132] Formula 4: G = f1(g)*f2(g)

[0133] Among them, G is used to refer to the ranking weight of the document information g, f1(g) is the first intermediate value of the document information, and f2(g) is the second intermediate value calculated for the document information.

[0134] Finally, the information push system will sort the multiple candidate document intelligence according to the sorting weight corresponding to each candidate document intelligence, obtain the sorting result, and determine the preset push quantity, such as 300, and extract multiple target document intelligence from the sorting result according to the preset push quantity, and push the multiple target document intelligence to the target user. Specifically, the candidate document intelligence with the preset push quantity that ranks in front in the sorting result can be extracted as the target document intelligence for push.

[0135] In summary, the logical process of the enterprise intelligence push method based on the multi-dimensional tag vector proposed in the embodiment of the present application is summarized as follows, including steps S1 to S10:

[0136] See also Figure 2B S1 collects documents and corpus related to enterprise intelligence to form an intelligence document library; S2 extracts seed thesaurus; S3 constructs explicit tag library and implicit tag library, and uses keywords and subject words in the explicit tag library to expand the seed thesaurus to form the enterprise's thesaurus; S4 constructs explicit tag annotation model and implicit tag annotation model; S5 automatically annotates all document intelligence in the intelligence document library using explicit tag annotation model and implicit tag annotation model; S6 performs vector conversion on the label annotated for each document intelligence to form a document vector database; S7 constructs a character portrait model for the target user; S8 performs dimensionality reduction processing on the character feature label indicated by the character portrait model to construct the character feature vector of the target user; S9 determines the function used to calculate the sorting weight; S10 uses the character feature vector to perform a search operation in the document vector database, and selects the target document intelligence for push according to the sorting weight function.

[0137] The method provided in the embodiment of the present application determines the target user to whom enterprise intelligence is to be pushed, constructs a character feature vector of the target user, and obtains a document vector database. The document vector database is constructed using multiple document intelligence labeled with explicit labels and implicit labels. The explicit label labeled with each document intelligence is labeled according to the subject word of the document intelligence, and the implicit label labeled with each document intelligence is labeled after the document intelligence is classified according to the multi-dimensional classification system of the enterprise. The character feature vector is used to perform a search operation in the document vector database, and multiple retrieved target document intelligence is pushed to the target user. The document intelligence is pushed in combination with multi-dimensional labels, so that knowledge recommendation in business scenarios is more accurate. While ensuring that intelligence outside the technical field is not recommended, it can also avoid missing valuable document intelligence, achieve targeted document intelligence recommendation, and improve recommendation efficiency and accuracy.

[0138] Further, as Figure 1 In the specific implementation of the method, the embodiment of the present application provides an enterprise intelligence push device based on a multi-dimensional tag vector, such as Figure 3 As shown, the device includes: a construction module 301, an acquisition module 302 and a push module 303.

[0139] The construction module 301 is used to determine the target user to be pushed enterprise intelligence and construct the character feature vector of the target user;

[0140] The acquisition module 302 is used to acquire a document vector database, wherein the document vector database is constructed using a plurality of document intelligences annotated with explicit tags and implicit tags, wherein each explicit tag annotated with the document intelligence is annotated according to a subject word of the document intelligence, and each implicit tag annotated with the document intelligence is annotated after the document intelligence is classified according to a multi-dimensional classification system of an enterprise;

[0141] The push module 303 is used to use the character feature vector to perform a search operation in the document vector database, and push the retrieved multiple target document information to the target user.

[0142] In a specific application scenario, the construction module 301 is also used to collect multiple corpus data and multiple core documents related to the enterprise to which the target user belongs, and to create a subject word library for the enterprise based on the multiple corpus data and the keywords and subject words corresponding to each of the core documents; to use the subject word library as an explicit tag library, to annotate the multiple core documents using the explicit tag library, and to train an explicit tag annotation model using the annotated multiple core documents; to create an implicit tag library based on the multi-dimensional classification system associated with the enterprise, to annotate the multiple core documents using the implicit tag library, and to train an implicit tag annotation model using the annotated multiple core documents; to obtain the multiple document information under the enterprise, to determine the keywords and subject words corresponding to each of the document information, and to Each of the document intelligence and the corresponding keywords and subject terms are input into the explicit label annotation model so that the explicit label annotation model annotates an explicit label for each of the document intelligence; based on the subject word library, abnormal explicit labels that are not related to the enterprise are determined in the explicit labels annotated for each of the document intelligence, and the abnormal explicit labels are replaced with default labels; each of the document intelligence is input into the implicit label annotation model so that the implicit label annotation model annotates an implicit label for each of the document intelligence; the explicit labels annotated for each of the document intelligence are combined to form a label phrase corresponding to each of the document intelligence, and the label phrase corresponding to each of the document intelligence is converted into a label vector, and the document vector database is constructed using the converted label vector and the annotated multiple document intelligence.

[0143] In a specific application scenario, the construction module 301 is used to construct a seed subject vocabulary for the enterprise using the multiple corpus data; use the TF-IDF word frequency method to analyze the document content of each of the core documents and determine corresponding keywords for each of the core documents; use the pre-trained BERT model to determine corresponding subject words for each of the core documents; filter the keywords and subject words corresponding to each of the core documents, delete the keywords or subject words that are not related to the enterprise, and integrate the keywords and subject words retained by each of the core documents to form an initial vocabulary; perform a semantic comparison between the words included in the initial vocabulary and the words included in the seed subject vocabulary, and determine the words to be supplemented in the initial vocabulary, wherein the semantic similarity between the words in the initial vocabulary and the words in the seed subject vocabulary is greater than the similarity threshold; add the words to be supplemented to the seed subject vocabulary, and use the supplemented seed subject vocabulary as the subject vocabulary.

[0144] In a specific application scenario, the construction module 301 is used to encode each of the core documents using the pre-trained BERT model to extract the semantic vector representation of the core document; perform word segmentation on the core document to obtain multiple word segmentation phrases, and extract the word vector corresponding to each of the word segmentation phrases; calculate the cosine similarity between the semantic vector representation and the word vector corresponding to each of the word segmentation phrases, and use the word segmentation phrase with the largest cosine similarity among the multiple word segmentation phrases as the subject word corresponding to the core document.

[0145] In a specific application scenario, the construction module 301 is used to use the implicit tag library to annotate the multiple core documents to obtain the annotated multiple core documents; preprocess the original texts of the annotated multiple core documents, and the preprocessing includes word segmentation, stop word removal and text format conversion; calculate the word frequency vectors of the original texts of the multiple core documents after processing, and generate a text feature vector of each of the core documents based on the word frequency vector; use a shallow neural network to learn the correspondence between the text feature vector of each of the core documents and the annotated implicit tags to obtain the implicit tag annotation model.

[0146] In a specific application scenario, the construction module 301 is used to construct a character portrait feature model of the target user based on the basic user information and historical interaction information of the target user; determine at least one designated document that has an interactive relationship with the target user with reference to the character portrait feature model, obtain multiple explicit tags annotated by the at least one designated document, and use the multiple explicit tags as multiple time series attribute tags; calculate a time sharp reduction coefficient for each of the time series attribute tags, and calculate a label weight for each of the time series attribute tags using the time sharp reduction coefficient corresponding to each of the time series attribute tags; divide the multiple time series attribute tags into a preset number of label groups according to the calculated label weights; use a random function to extract multiple target time series attribute tags in each of the label groups whose number meets the number threshold corresponding to the label group; extract the static attribute tag of the target user from the character feature attributes and professional attributes of the target user indicated by the character portrait feature model; use the static attribute tag and the multiple time series attribute tags to construct a character label group of the target user, and perform vector conversion on the character label group to obtain the character feature vector of the target user.

[0147] In a specific application scenario, the push module 303 is used to retrieve multiple candidate document intelligence in the document vector database based on the character feature vector; set a first intermediate value for each candidate document intelligence according to the historical browsing situation of the target user, wherein, for each candidate document intelligence, when the historical browsing situation indicates that the candidate document intelligence has been browsed, the first intermediate value corresponding to the candidate document intelligence is set to 0, and when the historical browsing situation indicates that the candidate document intelligence has not been browsed, the first intermediate value corresponding to the candidate document intelligence is set to 1; obtain a preset gravity coefficient, and use the browsing volume, release time and preset gravity coefficient of each candidate document intelligence to calculate a second intermediate value for each candidate document intelligence; calculate a sorting weight for each candidate document intelligence based on the first intermediate value and the second intermediate value corresponding to each candidate document intelligence; sort the multiple candidate document intelligence according to the sorting weight corresponding to each candidate document intelligence to obtain a sorting result; determine a preset push quantity, extract the multiple target document intelligence from the sorting result according to the preset push quantity, and push the multiple target document intelligence to the target user.

[0148] The device provided in the embodiment of the present application determines the target user to whom enterprise intelligence is to be pushed, constructs a character feature vector of the target user, and obtains a document vector database. The document vector database is constructed using multiple document intelligence labeled with explicit labels and implicit labels. The explicit label labeled with each document intelligence is labeled according to the subject word of the document intelligence, and the implicit label labeled with each document intelligence is labeled after the document intelligence is classified according to the multi-dimensional classification system of the enterprise. The character feature vector is used to perform a search operation in the document vector database, and multiple retrieved target document intelligence is pushed to the target user. The document intelligence is pushed in combination with multi-dimensional labels, so that knowledge recommendation in business scenarios is more accurate. While ensuring that intelligence outside the technical field is not recommended, it can also avoid missing valuable document intelligence, achieve targeted document intelligence recommendation, and improve recommendation efficiency and accuracy.

[0149] It should be noted that for other corresponding descriptions of the functional units involved in the enterprise intelligence push device based on a multi-dimensional tag vector provided in the embodiment of the present application, reference can be made to Figure 1 and FIG. 2A to FIG. 2B The corresponding description in will not be repeated here.

[0150] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0151] The above embodiments and the technical features in the embodiments can be arbitrarily combined with each other. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0152] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

[0153] In an exemplary embodiment, see Figure 4, and also provides a computer device, which includes a bus, a processor, a memory and a communication interface, and may also include an input and output interface and a display device, wherein each functional unit can communicate with each other through the bus. The memory stores a computer program, and the processor is used to execute the program stored in the memory and execute the enterprise intelligence push method based on multi-dimensional tag vectors in the above embodiment.

[0154] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the enterprise intelligence push method based on multi-dimensional tag vectors.

[0155] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by hardware, or by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each implementation scenario of the present application.

[0156] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily required for implementing the present application.

[0157] Those skilled in the art will appreciate that the modules in the device in the implementation scenario can be distributed in the device in the implementation scenario according to the implementation scenario description, or can be changed accordingly and located in one or more devices different from the implementation scenario. The modules in the above implementation scenario can be combined into one module, or can be further split into multiple sub-modules.

[0158] The above application serial numbers are for description only and do not represent the advantages or disadvantages of the implementation scenarios.

[0159] The above disclosure only discloses several specific implementation scenarios of the present application. However, the present application is not limited thereto, and any changes that can be conceived by technicians in this field should fall within the scope of protection of the present application.

Claims

1. A method for pushing enterprise intelligence based on multi-dimensional tag vectors, characterized in that: include: Determine the target user to whom enterprise intelligence is to be pushed, and construct a character feature vector of the target user; Acquire a document vector database, wherein the document vector database is constructed using a plurality of document intelligences annotated with explicit tags and implicit tags, wherein each explicit tag annotated with the document intelligence is annotated according to a subject word of the document intelligence, and each implicit tag annotated with the document intelligence is annotated after the document intelligence is classified according to a multi-dimensional classification system of the enterprise; The character feature vector is used to perform a search operation in the document vector database, and the retrieved multiple target document information is pushed to the target user.

2. The method according to claim 1, characterized in that Before determining the target user to be pushed enterprise intelligence and constructing the character feature vector of the target user, the method further includes: Collecting a plurality of corpus data and a plurality of core documents related to the enterprise to which the target user belongs, and creating a subject word library for the enterprise based on the plurality of corpus data and the keywords and subject words corresponding to each of the core documents; The subject word library is used as an explicit tag library, the explicit tag library is used to annotate the multiple core documents, and the annotated multiple core documents are used to train an explicit tagging model; Creating an implicit label library according to the multi-dimensional classification system associated with the enterprise, using the implicit label library to label the multiple core documents, and using the labeled multiple core documents to train an implicit label labeling model; Acquire the plurality of document information under the enterprise, determine the keywords and subject terms corresponding to each of the document information, and input each of the document information and the corresponding keywords and subject terms into the explicit label marking model, so that the explicit label marking model marks an explicit label for each of the document information; Determining abnormal explicit tags that are not related to the enterprise among the explicit tags annotated for each of the document information based on the subject word library, and replacing the abnormal explicit tags with default tags; Inputting each of the document information into the implicit label marking model, so that the implicit label marking model marks an implicit label for each of the document information; The explicit labels annotated with each of the document intelligence are combined to form a label phrase corresponding to each of the document intelligence, and the label phrase corresponding to each of the document intelligence is converted into a label vector. The document vector database is constructed using the converted label vector and the annotated multiple document intelligence.

3. The method according to claim 2, characterized in that The step of creating a subject word library for the enterprise based on the plurality of corpus data and the keywords and subject words corresponding to each of the core documents includes: Using the plurality of corpus data, construct a seed subject word library for the enterprise; The TF-IDF word frequency method is used to analyze the content of each core document and determine the corresponding keywords for each core document; Using the pre-trained BERT model, determine the corresponding keyword for each core document; Filter the keywords and subject terms corresponding to each of the core documents, delete the keywords or subject terms that are not relevant to the enterprise, and integrate the keywords and subject terms retained in each of the core documents to form an initial word library; Comparing the words included in the initial word library with the words included in the seed theme word library semantically, determining the words to be supplemented in the initial word library, the words to be supplemented being words whose semantic similarity between the words in the initial word library and the words in the seed theme word library is greater than a similarity threshold; The words to be supplemented are added to the seed subject word library, and the supplemented seed subject word library is used as the subject word library.

4. The method according to claim 3, characterized in that: The pre-trained BERT model is used to determine corresponding keywords for each core document, including: For each of the core documents, encoding the core document using the pre-trained BERT model to extract a semantic vector representation of the core document; Performing word segmentation processing on the core document to obtain a plurality of word segmentation phrases, and extracting a word vector corresponding to each of the word segmentation phrases; The cosine similarity between the semantic vector representation and the word vector corresponding to each of the segmentation phrases is calculated, and the segmentation phrase with the largest cosine similarity among the multiple segmentation phrases is used as the subject word corresponding to the core document.

5. The method according to claim 2, characterized in that: The step of using the implicit tag library to tag the plurality of core documents, and using the annotated plurality of core documents to train an implicit tag tagging model includes: Using the implicit tag library to annotate the plurality of core documents to obtain the annotated plurality of core documents; Preprocessing the original texts of the annotated core documents, wherein the preprocessing includes word segmentation, stop word removal, and text format conversion; Calculating the word frequency vectors of the processed original texts of the plurality of core documents, and generating a text feature vector of each of the core documents based on the word frequency vectors; A shallow neural network is used to learn the correspondence between the text feature vector of each core document and the annotated implicit label to obtain the implicit label annotation model.

6. The method according to claim 1, characterized in that The step of constructing the character feature vector of the target user includes: Constructing a character portrait feature model of the target user based on the basic user information and historical interaction information of the target user; Referring to the character portrait feature model, determining at least one designated document having an interactive relationship with the target user, obtaining a plurality of explicit tags annotated by the at least one designated document, and using the plurality of explicit tags as a plurality of temporal attribute tags; Calculating a time sharp reduction coefficient for each of the time series attribute labels, and calculating a label weight for each of the time series attribute labels using the time sharp reduction coefficient corresponding to each of the time series attribute labels; According to the calculated label weights, the plurality of time series attribute labels are divided into a preset number of label groups; Using a random function, extracting a plurality of target time series attribute labels whose number meets a quantity threshold corresponding to the label group in each label group; Extracting a static attribute label of the target user from the character attribute and professional attribute of the target user indicated by the character portrait feature model; The static attribute label and the multiple time series attribute labels are used to construct a character label group of the target user, and the character label group is vector-converted to obtain a character feature vector of the target user.

7. The method according to claim 1, characterized in that The step of utilizing the character feature vector to perform a search operation in the document vector database and pushing the retrieved multiple target document information to the target user includes: Based on the character feature vector, searching for multiple candidate document information in the document vector database; According to the historical browsing situation of the target user, a first intermediate value is set for each candidate document information, wherein, for each candidate document information, when the historical browsing situation indicates that the candidate document information has been browsed, the first intermediate value corresponding to the candidate document information is set to 0, and when the historical browsing situation indicates that the candidate document information has not been browsed, the first intermediate value corresponding to the candidate document information is set to 1; Obtaining a preset gravity coefficient, and using the page views, publishing time, and the preset gravity coefficient of each candidate document information to calculate a second intermediate value for each candidate document information; Calculate a ranking weight for each candidate document information based on the first intermediate value and the second intermediate value corresponding to each candidate document information; Sorting the plurality of candidate document information according to the ranking weight corresponding to each candidate document information to obtain a ranking result; A preset push quantity is determined, the plurality of target document information is extracted from the sorting result according to the preset push quantity, and the plurality of target document information is pushed to the target user.

8. An enterprise intelligence push device based on multi-dimensional tag vectors, characterized in that: include: A construction module, used to determine the target user to be pushed enterprise intelligence and construct a character feature vector of the target user; An acquisition module is used to acquire a document vector database, wherein the document vector database is constructed using a plurality of document intelligences annotated with explicit tags and implicit tags, wherein each explicit tag annotated with the document intelligence is annotated according to a subject word of the document intelligence, and each implicit tag annotated with the document intelligence is annotated after the document intelligence is classified according to a multi-dimensional classification system of an enterprise; The push module is used to use the character feature vector to perform a search operation in the document vector database and push the retrieved multiple target document information to the target user.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images