Personalized enterprise knowledge recommendation system and method for intelligent office platform

Abstract

The invention relates to a personalized enterprise knowledge recommendation system and method for a smart office platform, and belongs to the technical field of data mining and knowledge recommendation. Comprising a user behavior perception type multi-source information embedded representation learning module which is used for calculating a personalized business portrait representation vector based on business platform data, user role data and user personalized data of a current user; the document attribute perception type enterprise document representation learning module is used for calculating a comprehensive document embedding representation vector of each knowledge document based on the enterprise knowledge documents and the document attributes; the knowledge recommendation module is used for recommending an enterprise knowledge document to the user based on the set similarity of the personalized business portrait representation vector and each comprehensive document embedded representation vector; and the training module is used for training and adjusting related parameters based on the recommended enterprise knowledge document. According to the system and the method, accurate and reasonable documents can be provided for users according to specific service information in an enterprise office platform.

Description

Technical Field

[0001] The present invention belongs to the technical field of data mining and knowledge recommendation, and in particular, relates to a personalized enterprise knowledge recommendation system and method for a smart office platform. Background Art

[0002] As corporate rules and regulations become more and more complete, business scope gradually expands, and industry standards are frequently added and updated, the amount of data in the corporate knowledge base is increasing day by day, which makes it time-consuming and laborious for staff who handle related business based on standards and specifications to search for relevant documents, greatly affecting their work efficiency. Therefore, it is increasingly important to provide accurate and reasonable document recommendations for business staff based on specific business information.

[0003] The existing enterprise knowledge recommendation system has the following shortcomings: the recommendation of enterprise knowledge is based on keyword association matching. Since the keyword-based search method has a small amount of information and repeated information, it cannot fully represent the enterprise document knowledge to provide more comprehensive knowledge representation information. In addition, the existing recommendation system lacks consideration and analysis of user characteristics and business characteristics. The enterprise knowledge content recommended by the system may be far from the actual needs and has a low accuracy rate. Summary of the invention

[0004] In view of the above analysis, the present invention aims to provide a personalized enterprise knowledge recommendation system for a smart office platform, which analyzes and mines the multi-dimensional correlations between users and businesses implied in the source data based on the existing data in the office platform and fully considers the influence of document-specific attribute characteristics in the enterprise office platform to provide users with personalized enterprise knowledge recommendations.

[0005] The present invention provides a personalized enterprise knowledge recommendation system for a smart office platform, comprising:

[0006] The user behavior-aware multi-source information embedding representation learning module is used to calculate the personalized business portrait representation vector of the current user based on the business platform data, user role data and user personalized data of the current user on the office platform;

[0007] A document attribute-aware enterprise document representation learning module is used to calculate the comprehensive document embedding representation vector of each knowledge document based on the enterprise knowledge documents and corresponding document attributes of the office platform;

[0008] A knowledge recommendation module, used for recommending enterprise knowledge documents to users based on the set similarity between the personalized business portrait representation vector and each of the comprehensive document embedding representation vectors;

[0009] A training module is used to train the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module based on the enterprise knowledge documents recommended by the knowledge recommendation module and the loss function.

[0010] Furthermore, the user behavior-aware multi-source information embedding representation learning module includes:

[0011] The personalized user embedding representation vector learning layer is used to obtain the personalized user embedding representation vector based on the user personalized data and obtain the user business feature vector based on the current user's business platform data and user role data; wherein the business platform data includes the name of the business of the current user, the business process node, the business process form attributes, and the business process form confidentiality level; the user role data includes the current user's position name, role name, department name, user confidentiality level, the projects in which the user participates, and the functional permissions; the user personalized data includes the name of the file accessed by the user, the number of times the user accessed the file, the summary of the user's like data, and the summary of the user's favorite data;

[0012] The business portrait representation vector layer is used to calculate the user's business portrait representation vector based on the personalized user embedding representation vector and the user's business feature vector;

[0013] The personalized business portrait representation vector learning layer is used to calculate the user's personalized business portrait representation vector based on the personalized user embedding representation vector and the user's business portrait representation vector.

[0014] Furthermore, the business portrait representation vector layer calculates the business portrait representation vector of the user based on the personalized user embedding representation vector and the user business feature vector as follows:

[0015]

[0016] Among them, p t The business portrait representation vector for the user; c ui is the i-th user service feature vector of the current user; α ui Embed the correlation weight between the personalized user representation vector and the user business feature vector; is the normalized correlation weight vector; is the normalized relevance weight; the softmax function is used to convert the calculated weight into a probability distribution so that each element has a value between 0 and 1 and the sum of all elements is 1; u is the personalized user embedding representation vector; W ur For calculating u and vector c ui The attention weight matrix of the correlation between .

[0017] Furthermore, the process of the personalized business portrait representation vector learning layer calculating the personalized business portrait representation vector of the user based on the personalized user embedding representation vector and the business portrait representation vector of the user is expressed as follows:

[0018] p = tanh(W c [p t ;u]);

[0019] Where p is the user's personalized business portrait representation vector; tanh is the hyperbolic tangent function; p t is the user's business portrait representation vector; u is the personalized user embedding representation vector; [p t ;u] is p t The vector obtained by concatenating W and u; c is the weight matrix obtained through training.

[0020] Furthermore, the document attribute-aware enterprise document representation learning module includes:

[0021] A document attribute vector representation learning module is used to calculate a document attribute feature representation vector based on document attributes;

[0022] The word vector extraction layer is used to extract word vectors based on document content;

[0023] A document attribute-aware word vector fusion layer, used to obtain a sentence vector based on the fusion of the word vector and the document attribute feature representation vector;

[0024] A document attribute-aware sentence vector fusion layer, used for calculating a document-level embedding representation vector based on the sentence vector and the document attribute feature representation vector;

[0025] The fully connected output layer is used to calculate a comprehensive document embedding representation vector based on the attribute feature representation vector and the document-level embedding representation vector.

[0026] Furthermore, the calculation process of the document-level embedding representation vector calculated by the document attribute-aware sentence vector fusion layer is expressed as:

[0027]

[0028] in, is the document-level embedding representation vector; i is the i-th sentence vector; α fi is the correlation weight between the i-th sentence vector and the document attribute feature representation vector; is the normalized correlation weight vector; is the normalized relevance weight; the softmax function is used to convert the calculated weight into a probability distribution so that each element has a value between 0 and 1 and the sum of all elements is 1; f is the document attribute feature representation vector; [f; s i ] for f and s i The concatenated vector; With W fs are all vector parameters learned through training.

[0029] Furthermore, the calculation process of the fully connected output layer to obtain a comprehensive document embedding representation vector based on the attribute feature representation vector and the document-level embedding representation vector includes:

[0030]

[0031] Where d is the comprehensive document embedding representation vector; tanh is the hyperbolic tangent function; W d and W f is the vector parameter learned through training.

[0032] Furthermore, the training module trains the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module based on the enterprise knowledge documents recommended by the knowledge recommendation module and the loss function, including:

[0033] Calculate the personalized business portrait representation vector based on the enterprise knowledge documents recommended to the user by the knowledge recommendation module and calculate the corresponding set similarity between each of the recommended enterprise knowledge documents;

[0034] Calculating a loss function based on the corresponding set similarities;

[0035] Based on the loss function, the parameters of the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module are adjusted for iterative training.

[0036] Furthermore, the loss function is expressed as:

[0037]

[0038] in, is the total loss function value; is the i-th personalized business portrait representation vector p i The corresponding loss function value; A collection of embedding representation vectors for all comprehensive documents; for A subset of j is the j-th comprehensive document embedding representation vector, is d j and pi The set similarity, sim{d j ,p i}=σ(d j ⊙p i ), σ is the Sigmoid activation function, and ⊙ represents the dot product operation.

[0039] The present invention also provides a personalized enterprise knowledge recommendation method for a smart office platform, comprising the following steps:

[0040] Based on the business platform data, user role data and user personalized data of the current user of the office platform, a personalized business portrait representation vector of the current user is calculated;

[0041] Calculate the comprehensive document embedding representation vector of each knowledge document based on the enterprise knowledge documents and corresponding document attributes of the office platform;

[0042] Recommending enterprise knowledge documents to users based on the set similarity between the personalized business portrait representation vector and each of the comprehensive document embedding representation vectors;

[0043] During training, the recommended enterprise knowledge documents and the loss function train the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module; when in use, enterprise knowledge documents are recommended to users based on the set similarity between the personalized business portrait representation vector and each of the comprehensive document embedding representation vectors.

[0044] The present invention can achieve at least one of the following beneficial effects:

[0045] By analyzing and mining the implicit multi-dimensional correlations in multi-source data in the enterprise office platform and fully considering the impact of personalized user and document-specific attribute characteristics in the enterprise office platform, using multi-source information embedding representation learning, multi-source information intelligent fusion and other technologies, a personalized knowledge recommendation system and method for the enterprise smart office platform is realized to meet the active, efficient, accurate and intelligent recommendation of knowledge documents, so as to achieve the goal of improving quality and efficiency.

[0046] By considering the user's personalized data to extract feature information that reflects the user's cognitive level and knowledge reserves, and by considering the user's business platform data and user role data to extract features that are business-related to the user, a user personalized business portrait representation vector with rich feature information is constructed; by considering the document's own content and proprietary attribute information to extract features, a comprehensive document embedding representation vector of the knowledge document is constructed; based on the user's personalized features and the comprehensive features of the document, the similarity between the two is calculated, which can accurately provide users with personalized enterprise knowledge recommendations and improve the work efficiency of enterprise staff.

[0047] By fusing word vectors and sentence vectors layer by layer in the process of extracting document vectors, and fusing document attributes layer by layer into comprehensive document embedding representation vectors, the influence of irrelevant information is reduced and the influence of relevant information is enhanced, realizing the transformation of enterprise knowledge documents from raw data to advanced vector representation, ensuring the accuracy and effectiveness of recommended enterprise knowledge documents.

[0048] Other features and advantages of the present invention will be described in the following description, and some advantages may become apparent from the description, or may be understood by practicing the present invention. The purpose and other advantages of the present invention may be realized and obtained through the contents particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0050] Figure 1 It is a schematic diagram of the overall technical solution of the system of the present invention;

[0051] Figure 2 Schematic diagram of user behavior-aware multi-source information embedding representation learning module;

[0052] Figure 3 Schematic diagram of the document attribute-aware enterprise document representation learning module. DETAILED DESCRIPTION

[0053] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0054] System Example

[0055] A specific embodiment of the present invention discloses a personalized enterprise knowledge recommendation system for a smart office platform, including a user behavior-aware multi-source information embedding representation learning module, a document attribute-aware enterprise document representation learning module, a knowledge recommendation module and a training module; Figure 1 The working principle of the personalized enterprise knowledge recommendation system of the present invention is demonstrated;

[0056] Specifically, the user behavior-aware multi-source information embedding representation learning module is used to calculate the personalized business portrait representation vector of the current user based on the business platform data, user role data and user personalized data of the current user on the office platform;

[0057] A document attribute-aware enterprise document representation learning module is used to calculate the comprehensive document embedding representation vector of each knowledge document based on the enterprise knowledge documents and corresponding document attributes of the office platform;

[0058] A knowledge recommendation module, used for recommending enterprise knowledge documents to users based on the set similarity between the personalized business portrait representation vector and each of the comprehensive document embedding representation vectors;

[0059] A training module is used to train the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module based on the enterprise knowledge documents recommended by the knowledge recommendation module and the loss function.

[0060] The present embodiment discloses a personalized enterprise knowledge recommendation system for a smart office platform. By considering the personalized data of the user, feature information reflecting the user's cognitive level and knowledge reserve can be extracted. By considering the user's business platform data and user role data, features with business relevance to the user are extracted to construct a user personalized business portrait representation vector with rich feature information. Features are extracted by considering the content of the document itself and proprietary attribute information to construct a comprehensive document embedding representation vector of the knowledge document. Based on the similarity between the user's personalized features and the comprehensive features of the document, the similarity between the two is considered, and personalized enterprise knowledge recommendations can be accurately provided to the user, thereby improving the work efficiency of the enterprise staff.

[0061] User behavior-aware multi-source information embedding representation learning module

[0062] Specifically, the user behavior-aware multi-source information embedding representation learning module includes a personalized user embedding representation vector learning layer, a business profile representation vector layer, and a personalized business profile representation vector learning layer (such as Figure 2 shown).

[0063] Before describing the user behavior-aware multi-source information embedding representation learning module in detail, the user's business platform data, user role data and user personalized data in this application are first described.

[0064] Specifically, the user's business platform data includes the name of the business the current user is in, business process nodes, business process form attributes, business process form confidentiality levels, etc.

[0065] Furthermore, the business name refers to a specific type of business, such as invoice management business, contract payment approval business, contract change business, contract termination business, etc.; the business process node refers to the step node that needs to be passed to execute a specific business, such as the payment approval business includes the step nodes of contract drafting by the contract clerk, contract approval by the supervisor, contract payment application by the contract clerk, contract payment approval by the supervisor, and payment registration by the financial specialist; the business process form attribute executes a specific business record document, such as the contract payment approval business attributes include: contract code, contract name, project, contract amount, payment code, payee, etc.; the business process form confidentiality level refers to the corresponding confidentiality level of the business, such as non-confidential, secret, confidential, and top secret.

[0066] Specifically, user role data includes the current user's position title, role name, department name, user confidentiality level, projects the user is involved in, user functional permissions, etc.

[0067] Specifically, user personalized data includes the names of files accessed by the user, the content of files accessed by the user, the number of times the user accessed the files, the names of web pages browsed by the user, the number of times the user browsed the web pages, the user's historical search terms, the text content of emails sent and received by the user, the number of times the user sent and received emails, the number of button clicks (sort buttons, query buttons), the number of system logins, the number of device uses, a summary of the user's likes data, and a summary of the user's favorites data, etc.

[0068] Specifically, the personalized user embedding representation vector learning layer is used to obtain a personalized user embedding representation vector based on the user's personalized data and to obtain a user business feature vector based on the current user's business platform data and user role data.

[0069] It should be noted that the user's business platform data, user role data and user personalized data are multi-heterogeneous, but can all be divided into three types, including text type data, text or numerical coded data and numerical data.

[0070] Furthermore, text type data exemplarily includes the names of files accessed by users, the contents of files accessed by users, the text contents of emails sent and received by users, the names of web pages browsed by users, user historical search terms, summaries of user likes data, and summaries of user favorites data, etc.

[0071] Furthermore, text or numeric coded data refers to data in which the attribute value of a field stored in a database is presented as text but the attribute value is stored in numeric code, and examples include the user's job title, role name, department name, user confidentiality level, projects in which the user participates, user's functional permissions, business process form confidentiality level, business process nodes, etc.

[0072] Furthermore, numerical data exemplarily include the number of button clicks (sort button, query button), the number of file accesses, the number of device uses, the number of emails sent and received, the number of web page views, the number of system logins, etc.

[0073] Specifically, for text data, based on pre-trained large language models PLMs such as RoBERTa, the Chinese word segmentation tool is used to segment the text data, and then the corresponding word vectors are obtained through retrieval. The calculation example is as follows:

[0074] [b1,b2...b m ]=RoBERTa(t1,t2...t m ); (Formula 1)

[0075] Among them, t1, t2...t m b1, b2...b m Based on the word vectors retrieved from the pre-trained large language model, the word vectors are further fused to obtain the text type data embedding representation vector.

[0076] Specifically, for text or numerically encoded data, high-dimensional word embedding representation vectors are extracted for fields and attribute values ​​of fields based on the pre-trained large language model RoBERTa, and then a comprehensive feature input vector is obtained through weighted calculation.

[0077] For example, based on the field and attribute value "role: supervisor" (denoted as b r and b s ) respectively get vector t r and t s , after weighted calculation, the comprehensive feature input vector is expressed as:

[0078] x=W r t r +W s t s ;(Formula 2)

[0079] Among them, W r and W s t r and t s The corresponding learnable weight matrix, the weight matrix value continues to change with multiple rounds of training optimization until acceptable recommendation accuracy and performance are achieved.

[0080] Specifically, for numerical data, the field is text and the attribute value of the field is a numeric value. It is necessary to first divide the attribute value of the field into intervals, convert it into category data, and then encode it. Similarly, according to the vector extraction method for text or numeric encoded data, the corresponding comprehensive feature input vector of the numerical data is obtained.

[0081] Furthermore, based on the multiple types of data in the user personalized data, the corresponding multiple text type data embedding representation vectors or comprehensive feature input vectors are obtained respectively, and the above multiple vectors are represented as the user multi-dimensional attribute vector matrix X = [x1, x2..., x N ].

[0082] Furthermore, the matrix X is passed through the self-attention network to obtain the personalized user embedding representation vector u. The calculation process is expressed as:

[0083]

[0084] Where Q = XW Q , K = XW K 、V=XW V are the query vector, key vector, and value vector in the self-attention network calculation, respectively. k is the dimension of the key vector, divided by In order to stabilize the gradient, the softmax function is used to convert the calculated attention weights into a probability distribution so that each element has a value between 0 and 1 and the sum of all elements is 1.

[0085] Furthermore, corresponding text type data embedding representation vectors or comprehensive feature input vectors are obtained based on each type of data in the user's business platform data and user role data, which are collectively referred to as user business feature vectors.

[0086] Specifically, the business portrait representation vector layer is used to calculate the user's business portrait representation vector based on the personalized user embedding representation vector and the user's business feature vector. The calculation process is expressed as follows:

[0087]

[0088] Among them, p t The business portrait representation vector for the user; c ui is the i-th user service feature vector of the current user; α ui Embed the correlation weight between the personalized user representation vector and the user business feature vector; is the normalized correlation weight vector; is the normalized relevance weight; the softmax function is used to convert the calculated weight into a probability distribution so that each element has a value between 0 and 1 and the sum of all elements is 1; u is the personalized user embedding representation vector; W ur For calculating u and vector c ui The attention weight matrix of the correlation between .

[0089] Specifically, the personalized business portrait representation vector learning layer is used to calculate the user's personalized business portrait representation vector based on the personalized user embedding representation vector and the user's business portrait representation vector. The calculation process is expressed as:

[0090] p = tanh(W c [p t ;u]);

[0091] Where p is the user's personalized business portrait representation vector; tanh is the hyperbolic tangent function; p t is the user's business portrait representation vector; u is the personalized user embedding representation vector; [p t ;u] is p t The vector obtained by concatenating W and u; c is the weight matrix obtained through training.

[0092] Document attribute-aware enterprise document representation learning module

[0093] Specifically, the document attribute-aware enterprise document representation learning module (such as Figure 3 shown) include:

[0094] A document attribute vector representation learning module is used to calculate a document attribute feature representation vector based on document attributes;

[0095] The word vector extraction layer is used to extract word vectors based on document content;

[0096] A document attribute-aware word vector fusion layer, used to obtain a sentence vector based on the word vector fusion;

[0097] A document attribute-aware sentence vector fusion layer, used for calculating a document-level embedding representation vector based on the sentence vector and the document attribute feature representation vector;

[0098] The fully connected output layer is used to calculate a comprehensive document embedding representation vector based on the attribute feature representation vector and the document-level embedding representation vector.

[0099] It should be noted that document attributes include document source, document category (for example: standard, process), document validity period, document confidentiality, field classification, etc., and document attribute data types include unstructured text type data, numerically coded data or category data.

[0100] Specifically, when the document attribute vector representation learning module calculates the document attribute feature representation vector based on the document attributes, for unstructured text type data, it is necessary to first convert the unstructured text type data into structured data through word segmentation processing.

[0101] Furthermore, for numerically coded data, categorical data, and converted structured data, the same vector extraction method (Formula 2) for text or numerically coded data in the user behavior-aware multi-source information embedding representation learning module is used to obtain the corresponding comprehensive feature input vectors of document attribute data, which are represented as f1, f2, f3...f n .

[0102] Furthermore, the above multiple vectors are represented as a user multi-dimensional attribute vector matrix F = [f1, f2..., f n ]. Pass the matrix F through the self-attention network (see Formula 3 for the calculation process) to obtain the personalized user embedding representation vector f.

[0103] Specifically, enterprise knowledge documents include internal enterprise standards, rules and regulations, process templates, process templates, process common terms, business classification methods, work processes, custom knowledge bases, etc. The document content mainly includes text-type data, and may also include numerical data, classified data, etc.

[0104] Furthermore, the word vector extraction layer is used to extract word vectors based on the document content; it includes segmenting the file content through a word segmentation tool, and then retrieving the words belonging to the text type data based on the pre-trained language model RoBERTa to obtain the corresponding word vector (Formula 1); for other types of data, the method of Formula 2 is used to obtain the corresponding feature vector.

[0105] Furthermore, the document attribute-aware word vector fusion layer is used to calculate the sentence vector based on the fusion of the word vector and the document attribute feature representation vector as follows:

[0106]

[0107] Among them, s is the sentence vector; c' i is the i-th word vector; β i is the correlation weight between the word vector and the document attribute feature representation vector; is the normalized correlation weight vector; is the normalized relevance weight; f is the document attribute feature representation vector; W fr For calculating f and vector c' i The attention weight matrix of the correlation between .

[0108] Furthermore, the calculation process of the document-level embedding representation vector calculated by the document attribute-aware sentence vector fusion layer is expressed as:

[0109]

[0110] in, is the document-level embedding representation vector; i is the i-th sentence vector; α fi is the correlation weight between the i-th sentence vector and the document attribute feature representation vector; is the normalized correlation weight vector; is the normalized relevance weight; f is the document attribute feature representation vector; [f; s i ] for f and s i The concatenated vector; With W fs are all vector parameters learned through training.

[0111] Furthermore, the document attribute vector representation learning module also includes a word-level encoding layer located between the word vector extraction layer and the document attribute-aware word vector fusion layer, and a sentence-level encoding layer located between the document attribute-aware word vector fusion layer and the document attribute-aware sentence vector fusion layer (e.g., Figure 3 As shown in Figure 1.1, the word-level encoding layer fully captures the contextual relationship between word vectors through the Transformer architecture to ensure that the output vector dimension does not change; the sentence-level encoding layer takes the sentence vector representation obtained by the document attribute-aware word vector fusion layer as input, and fully captures the contextual dependency between vectors through the Transformer architecture to ensure that the vector dimension does not change.

[0112] Furthermore, the calculation process of the fully connected output layer to obtain a comprehensive document embedding representation vector based on the attribute feature representation vector and the document-level embedding representation vector includes:

[0113]

[0114] Where d is the comprehensive document embedding representation vector; tanh is the hyperbolic tangent function; W d and W f is the vector parameter learned through training.

[0115] Furthermore, through batch calculation, we can get the set of all comprehensive document embedding representation vectors. Through document attribute-aware enterprise document representation learning and fusion strategy, the transformation of enterprise knowledge documents from raw data to advanced vector representation is achieved.

[0116] Knowledge recommendation module

[0117] Specifically, the knowledge recommendation module is used to recommend enterprise knowledge documents to users based on the set similarity between the personalized business portrait representation vector and each of the comprehensive document embedding representation vectors. Further, the calculation formula of the set similarity is expressed as:

[0118] sim[d j ,p i}=σ(d j ⊙p i );

[0119] Among them, d j is the j-th comprehensive document embedding representation vector, p i is the representation vector of the i-th personalized business portrait; σ is the Sigmoid activation function, and ⊙ represents the dot multiplication operation.

[0120] Furthermore, when the enterprise knowledge document changes, the knowledge recommendation module calculates the set similarity in real time, and calculates the recommendation probability based on the similarity calculation results (for example, the similarity calculation results can be normalized); further, sorting is performed based on the recommendation probability; further, enterprise knowledge documents are recommended to users based on the sorting results and the set recommendation threshold.

[0121] Training Module

[0122] Specifically, the training module trains the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module based on the enterprise knowledge documents recommended by the knowledge recommendation module and the loss function, including:

[0123] Calculate the personalized business portrait representation vector based on the enterprise knowledge documents recommended to the user by the knowledge recommendation module and calculate the corresponding set similarity between each of the recommended enterprise knowledge documents;

[0124] Calculating a loss function based on the corresponding set similarities;

[0125] The parameters of the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module are adjusted based on the loss function for iterative training, wherein the parameters include a learnable weight matrix W when calculating a comprehensive feature input vector for text or numerical coded data, numerical data, and category data. r , W s , used to calculate the personalized user embedding representation vector u and vector c uiThe attention weight matrix W of the correlation between ur , calculate the weight matrix W of the personalized business portrait representation vector c , calculate the document attribute feature representation vector f and vector c' i The attention weight matrix W of the correlation between fr , calculate the vector parameters of the document-level embedding representation vector With W fs , calculate the vector parameter W of the comprehensive document embedding representation vector d and W f wait.

[0126] Furthermore, the loss function is calculated as:

[0127]

[0128] in, is the total loss function value; is the i-th personalized business portrait representation vector p i The corresponding loss function value; A collection of embedding representation vectors for all comprehensive documents; for , representing multiple enterprise knowledge documents corresponding to the i-th personalized business portrait representation vector; d j is the j-th comprehensive document embedding representation vector, sim{d j ,p i} is d j and p i The set similarity, sim{d j ,p i}=σ(d j ⊙p i ), σ is the Sigmoid activation function, and ⊙ represents the dot product operation.

[0129] It should be noted that when constructing a training data set, data samples also need to be labeled. Exemplary labeling is performed in the following ways: explicit feedback, that is, if the user has clear feedback on the enterprise knowledge document (such as likes, ratings), these feedbacks can be directly used as labels, that is, they are recommended documents; implicit feedback, that is, labels are inferred based on user behavior, such as user clicks, reading time, etc., to determine whether the user is interested in a certain document. In the above way, all data samples can be labeled without being divided according to specific restrictions such as user roles or years of work, thereby ensuring the degree of personalization of knowledge document recommendations. In addition, in the actual training process, when the data samples are divided, the data set is usually divided into a training set, a validation set, and a test set, for example, accounting for 70%, 15%, and 15% respectively.

[0130] Furthermore, in the training process of the present application, L2 regularization, batch normalization and Dropout methods are introduced, wherein L2 regularization is achieved by adding a penalty term of the sum of squares of all weight matrix parameters of the model to the loss function, thereby reducing the complexity of the model and preventing overfitting; batch normalization is usually applied before the activation function to standardize the input data of each layer, reduce internal covariate transfer, and increase the network training speed; and Dropout is applied after the fully connected layer and before the activation function to randomly discard a part of neurons during the training process to enhance the generalization ability of the model.

[0131] Furthermore, the total number of iterative optimizations during training is determined based on whether there is a large fluctuation in the model recommendation accuracy, including precision, recall, F1 score, etc. Precision refers to the proportion of positive samples among all recommended samples. It is calculated as follows: TP represents the number of documents correctly recommended as positive, and FP represents the number of samples incorrectly predicted as positive. Recall refers to the proportion of documents recommended by the model among all documents that should actually be recommended. FN represents the number of documents that are incorrectly not recommended by the model; while the F1 score is the average of the precision and recall, which is used to comprehensively judge the performance of the model. Its calculation formula is: The higher the F1 score, the more balanced the model recommendation results are in terms of precision and recall. In actual training, the dynamic changes of the above indicators can be visualized. When the changes of various indicators tend to be stable, the number of training times required to achieve the optimal model parameters can be determined.

[0132] The present embodiment discloses a personalized enterprise knowledge recommendation system for a smart office platform. By analyzing and mining the implicit multivariate correlations in multi-source data in the enterprise office platform and fully considering the influence of personalized user and document-specific attribute characteristics in the enterprise office platform, multi-source information embedding representation learning, multi-source information intelligent fusion and other technologies are used to realize a personalized knowledge recommendation system and method for an enterprise smart office platform, so as to meet the needs of active, efficient, accurate and intelligent recommendation of knowledge documents, and achieve the purpose of improving quality and efficiency.

[0133] By considering the user's personalized data to extract feature information that reflects the user's cognitive level and knowledge reserves, and by considering the user's business platform data and user role data to extract features that are business-related to the user, a user personalized business portrait representation vector with rich feature information is constructed; by considering the document's own content and proprietary attribute information to extract features, a comprehensive document embedding representation vector of the knowledge document is constructed; based on the user's personalized features and the comprehensive features of the document, the similarity between the two is calculated, which can accurately provide users with personalized enterprise knowledge recommendations and improve the work efficiency of enterprise staff.

[0134] By fusing word vectors and sentence vectors layer by layer in the process of extracting document vectors, and fusing document attributes layer by layer into comprehensive document embedding representation vectors, the influence of irrelevant information is reduced and the influence of relevant information is enhanced, realizing the transformation of enterprise knowledge documents from raw data to advanced vector representation, ensuring the accuracy and effectiveness of recommended enterprise knowledge documents.

[0135] Method Embodiment

[0136] A specific embodiment of the present invention discloses a personalized enterprise knowledge recommendation method for a smart office platform, comprising the following steps:

[0137] Based on the business platform data, user role data and user personalized data of the current user of the office platform, a personalized business portrait representation vector of the current user is calculated;

[0138] Calculate the comprehensive document embedding representation vector of each knowledge document based on the enterprise knowledge documents and corresponding document attributes of the office platform;

[0139] Recommending enterprise knowledge documents to users based on the set similarity between the personalized business portrait representation vector and each of the comprehensive document embedding representation vectors;

[0140] During training, the recommended enterprise knowledge documents and the loss function train the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module; when in use, enterprise knowledge documents are recommended to users based on the set similarity between the personalized business portrait representation vector and each of the comprehensive document embedding representation vectors.

[0141] Compared with the prior art, the personalized enterprise knowledge recommendation method for a smart office platform provided in this embodiment has basically the same beneficial effects as those provided by the system embodiment, which will not be elaborated here.

[0142] It should be noted that the above embodiments are based on the same inventive concept and parts not described repeatedly can be used as reference for each other.

[0143] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A personalized enterprise knowledge recommendation system for a smart office platform, characterized in that: include: The user behavior-aware multi-source information embedding representation learning module is used to calculate the personalized business portrait representation vector of the current user based on the business platform data, user role data and user personalized data of the current user on the office platform; A document attribute-aware enterprise document representation learning module is used to calculate the comprehensive document embedding representation vector of each knowledge document based on the enterprise knowledge documents and corresponding document attributes of the office platform; A knowledge recommendation module, used for recommending enterprise knowledge documents to users based on the set similarity between the personalized business portrait representation vector and each of the comprehensive document embedding representation vectors; A training module is used to train the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module based on the enterprise knowledge documents recommended by the knowledge recommendation module and the loss function.

2. The personalized enterprise knowledge recommendation system according to claim 1, characterized in that: The user behavior-aware multi-source information embedding representation learning module includes: The personalized user embedding representation vector learning layer is used to obtain the personalized user embedding representation vector based on the user personalized data and obtain the user business feature vector based on the current user's business platform data and user role data; wherein the business platform data includes the name of the business of the current user, the business process node, the business process form attributes, and the business process form confidentiality level; the user role data includes the current user's position name, role name, department name, user confidentiality level, the projects in which the user participates, and the functional permissions; the user personalized data includes the name of the file accessed by the user, the number of times the user accessed the file, the summary of the user's like data, and the summary of the user's favorite data; The business portrait representation vector layer is used to calculate the user's business portrait representation vector based on the personalized user embedding representation vector and the user's business feature vector; The personalized business portrait representation vector learning layer is used to calculate the user's personalized business portrait representation vector based on the personalized user embedding representation vector and the user's business portrait representation vector.

3. The personalized enterprise knowledge recommendation system according to claim 2, characterized in that: The calculation process of the business portrait representation vector layer to obtain the user's business portrait representation vector based on the personalized user embedding representation vector and the user's business feature vector is expressed as follows: Among them, p t The business portrait representation vector for the user; c ui is the i-th user service feature vector of the current user; α ui Embed the correlation weight between the personalized user representation vector and the user business feature vector; is the normalized correlation weight vector; is the normalized relevance weight; the softmax function is used to convert the calculated weight into a probability distribution so that each element has a value between 0 and 1 and the sum of all elements is 1; u is the personalized user embedding representation vector; W ur For calculating u and vector c ui The attention weight matrix of the correlation between .

4. The personalized enterprise knowledge recommendation system according to claim 3, characterized in that: The process of the personalized business portrait representation vector learning layer calculating the personalized business portrait representation vector of the user based on the personalized user embedding representation vector and the business portrait representation vector of the user is expressed as follows: p=tanh(W c [p t ;u]); Where p is the user's personalized business portrait representation vector; tanh is the hyperbolic tangent function; p t is the user's business portrait representation vector; u is the personalized user embedding representation vector; [p t ;u] is p t The vector obtained by concatenating W and u; c is the weight matrix learned through training.

5. The personalized enterprise knowledge recommendation system according to any one of claims 1 to 4, characterized in that: The document attribute-aware enterprise document representation learning module includes: A document attribute vector representation learning module is used to calculate a document attribute feature representation vector based on document attributes; The word vector extraction layer is used to extract word vectors based on document content; A document attribute-aware word vector fusion layer, used to obtain a sentence vector based on the fusion of the word vector and the document attribute feature representation vector; A document attribute-aware sentence vector fusion layer, used for calculating a document-level embedding representation vector based on the sentence vector and the document attribute feature representation vector; The fully connected output layer is used to calculate a comprehensive document embedding representation vector based on the attribute feature representation vector and the document-level embedding representation vector.

6. The personalized enterprise knowledge recommendation system according to claim 5, characterized in that: The calculation process of the document-level embedding representation vector calculated by the document attribute-aware sentence vector fusion layer is expressed as follows: in, is the document-level embedding representation vector; i is the i-th sentence vector; α fi is the correlation weight between the i-th sentence vector and the document attribute feature representation vector; is the normalized correlation weight vector; is the normalized relevance weight; the softmax function is used to convert the calculated weight into a probability distribution so that each element has a value between 0 and 1 and the sum of all elements is 1; f is the document attribute feature representation vector; [f; s i ] for f and s i The concatenated vector; With W fs are all vector parameters learned through training.

7. The personalized enterprise knowledge recommendation system according to claim 6, characterized in that: The calculation process of the fully connected output layer to obtain a comprehensive document embedding representation vector based on the attribute feature representation vector and the document-level embedding representation vector includes: Where d is the comprehensive document embedding representation vector; tanh is the hyperbolic tangent function; W d and W f is the vector parameter learned through training.

8. The personalized enterprise knowledge recommendation system according to claim 7, characterized in that: The training module trains the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module based on the enterprise knowledge documents recommended by the knowledge recommendation module and the loss function, including: Calculate the personalized business portrait representation vector based on the enterprise knowledge documents recommended to the user by the knowledge recommendation module and calculate the corresponding set similarity between each of the recommended enterprise knowledge documents; Calculating a loss function based on the corresponding set similarities; Based on the loss function, the parameters of the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module are adjusted for iterative training.

9. The personalized enterprise knowledge recommendation system according to claim 8, characterized in that: The loss function is expressed as: in, is the total loss function value; is the i-th personalized business portrait representation vector p i The corresponding loss function value; A collection of embedding representation vectors for all comprehensive documents; for A subset of j is the j-th comprehensive document embedding representation vector, sim{d j ,p i } is d j and p i The set similarity, sim{d j ,p i }=σ(d j ⊙p i ), σ is the Sigmoid activation function, and ⊙ represents the dot product operation.

10. A personalized enterprise knowledge recommendation method for a smart office platform, characterized in that: The steps include: Based on the business platform data, user role data and user personalized data of the current user of the office platform, a personalized business portrait representation vector of the current user is calculated; Calculate the comprehensive document embedding representation vector of each knowledge document based on the enterprise knowledge documents and corresponding document attributes of the office platform; Recommending enterprise knowledge documents to users based on the set similarity between the personalized business portrait representation vector and each of the comprehensive document embedding representation vectors; During training, the recommended enterprise knowledge documents and the loss function train the user behavior-aware multi-source information embedding representation learning module and the document attribute-aware enterprise document representation learning module; when in use, enterprise knowledge documents are recommended to users based on the set similarity between the personalized business portrait representation vector and each of the comprehensive document embedding representation vectors.

Images