A method for recommending target knowledge features for enterprise-related archive management
By acquiring and vectorizing enterprise archive data, combining deep learning and Transformer models, generating user preference sets, the limitations and adaptability of feature recommendations in the enterprise-related archive management system are solved, and efficient and personalized archive recommendations and manage.
Patent Information
- Application Number
- CN202410134410.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-31
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-01-31
AI Technical Summary
The existing enterprise-related archive management system has different indexing purposes in the process of recommending knowledge characteristics of specific targets, resulting in limitations in representation requirements, lack of active learning and adaptive adjustment capabilities, and it is difficult to automatically adjust the index content according to user needs.
By obtaining archival graph structure data and text structure data, using text embedding technology to convert it into vector representation, and building a deep learning classification model, combining the Transformer model to learn user preferences, generate a personal exclusive preference collection, and provide personalized recommendations.
It improves the efficiency of file management and the accuracy of recommendations, improves the personalization level of enterprise recommendation systems, can adapt to complex and diverse file management needs, and enhances user experience.
Smart Images

Figure CN118035440B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of archive management, and in particular to a method for recommending target knowledge features of enterprise-related archive management. Background Art
[0002] Enterprise-related file management is to achieve unified management of enterprise employee information by establishing and maintaining a personal relationship network between various departments and positions of the enterprise. Its main functions include: setting up standardized classifications for various departments and positions within the enterprise, establishing employee affiliation, collecting and integrating employee personal information, such as name, contact information, education experience, and resume, etc., tracking employee job changes and job transfers within the enterprise, updating files in a timely manner, and supporting file query and access management under different authorization levels. Enterprise-related file management can help enterprises fully and dynamically understand the internal personnel situation, provide support for subsequent personnel evaluation and career planning, and also facilitate coordination and cooperation between different departments;
[0003] However, in the existing enterprise-related archive management process, there are still some shortcomings in the recommendation process of specific target knowledge features, such as:
[0004] 1. Different managers or information organizers have different indexing purposes. Therefore, they have different representation requirements for different index targets. It is difficult to automatically provide corresponding numerical representations according to identity changes, so it has limitations;
[0005] 2. Lack of active learning capabilities. In the long-term indexing and recommendation process, it is difficult to automatically adjust the index content according to the needs of the indexer. When the indexer's needs change in a certain period of time, it is difficult to make adaptive adjustments. Summary of the invention
[0006] 1. Technical issues to be solved
[0007] In view of the above-mentioned shortcomings of the prior art, the present invention provides a method for recommending target knowledge features of enterprise-related archive management, which can effectively solve the problems of the prior art.
[0008] (II) Technical solution
[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions:
[0010] The present invention discloses a method for recommending target knowledge features of enterprise-related archive management, comprising the following steps:
[0011] Step 1: Intervene in the archive management system to obtain the enterprise's archive data, including archive graph structure data and text structure data;
[0012] Step 2: Convert the archive graph structure data and text structure data into vector representations through text embedding technology. Convert the text data into a vector of fixed length, fuse the vector representation of the archive graph structure data and the vector representation of the text structure data, and integrate different types of information into the same feature space.
[0013] Step 3: Based on the vectorized and fused data, build and train a deep learning classification model, input the data into the classification model for classification, divide the archival data into different categories, and generate corresponding classification labels;
[0014] Step 4: Based on the classification label definition of the data in step 3, obtain the recommended data of the historical index and the corresponding index content, and determine the archive data related to the user's interests by analyzing the existing index data and its corresponding archive content, and use it as the recommended data;
[0015] Step 5: By analyzing the final cited data, compare the difference between the corresponding index content and the cited data;
[0016] Step 6: Use the Transformer model to learn the fused feature vector to capture the relationship and contextual information between features. By training the Transformer model on the fused feature vector, combined with the recommendation data obtained in step 4 and the difference representation obtained in step 5 for comprehensive analysis, the learned user preferences are converted into data representation as the final preference reference value;
[0017] Step 7: Under the classification label, integrate the final preference reference values obtained into a personal preference set and provide a modification port;
[0018] Step 8: After the user verifies the identity information, the user enters personal information and a short text of the subject to be queried on the decoding end. According to the user's personal preference set, the corresponding content is captured under different classification tags, and the archival data related to the user's query subject is filtered out, and the content of interest is recommended to the user.
[0019] Furthermore, the archive graph structure data in step 1 is used to describe the relationship and hierarchy between different types of archives, and basic attributes include: folder structure and file association.
[0020] Furthermore, the text embedding technology in step 2 includes word embedding and paragraph embedding.
[0021] Furthermore, the text embedding technology in step 2 generates vectorized representations of each entity and relationship, vectorizes different types of nodes and relationships in a unified feature space, and performs embedding training. Assuming that there is a relationship r∈R (relationship set) between entities (h, t)∈E (entity set), h, t, r are used as vectorized representations, (h, t, r) constitute a triple, and given constraints The calculation formula of the loss function of its embedding training is:
[0022] L=∑ a∈S ∑ b∈D [y+f r (h,t)-f v (k,p)] + ;
[0023] Where L represents the coefficient of the loss function, a∈S represents the triple (h, t, r) in the knowledge graph, y represents the hyperparameter of the edge distance, b∈D represents the negative sample triple (k, p, v) randomly extracted during training, and f r (h, t) represents the triple performance, f v (k,p) represents the performance of negative sample triples, [y+f r (h,t)-f v (k,p)] represents x, [y+f r (h,t)-f v (k,p)] + represents max(0,x);
[0024] Among them, f r The calculation formula for (h,t) is:
[0025]
[0026] Furthermore, the classification tags in step 3 identify and organize the archive data as reference data for subsequent recommendations and analysis.
[0027] Furthermore, the difference representation in step 5 is used as a representation of the user's historical behavior characteristics, reflecting the user's behavior and preference in the archive management system.
[0028] Furthermore, the Transformer model in step 6 obtains and calculates the recommendation score of each type of data, performs normalization operation and classifies it by a Softmax layer in the fully connected layer, sorts the output predicted probability distribution to obtain the top n recommendations, and calculates the final recommendation value through binary cross entropy as a representation of the final preference reference value.
[0029] Furthermore, the personal preference set in step 7 includes profile data related to user interests. After being enabled, related recommendations can be made based on the user's historical behavior characteristics and index content corresponding to the preferences obtained through feature learning.
[0030] Furthermore, after the user with modification authority intervenes, the modification port in step 7 can be used by the user to trigger the operation of customizing and modifying personal preference characteristics. After the modification is completed, the modification result is fed back.
[0031] (III) Beneficial effects
[0032] Compared with the known prior art, the technical solution provided by the present invention has the following beneficial effects:
[0033] 1. By obtaining archival graph structure data and text structure data, and then vectorizing and fusing them in the same feature space, the data processing process is greatly simplified and the efficiency of archival management is improved. Large amounts of archival data are managed in an organized manner through classification labels, features are learned from the corpus, and valuable information is extracted, thereby improving the validity and accuracy of archival retrieval and recommendation.
[0034] 2. By learning the historical behavior characteristics of users and building personal preference sets, the personalization level of the enterprise recommendation system is further improved. Users only need to enter their personal information and query topics to capture and recommend content of interest under different classification tags based on their exclusive preference sets. This can adapt to more complex and diverse file management needs, and has important application value for improving the information management capabilities of enterprises and further enhancing user experience. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0036] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION
[0037] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0038] The present invention will be further described below in conjunction with the embodiments.
[0039] Example 1
[0040] A method for recommending target knowledge features of enterprise-related archive management in this embodiment is as follows: Figure 1 As shown, the following steps are included:
[0041] Step 1: Access the archive management system to obtain the enterprise's archive data, including archive graph structure data and text structure data. The archive graph structure data is used to describe the relationship and hierarchy between different types of archives. The basic attributes include: folder structure and file association;
[0042] Step 2: Convert the archive graph structure data and text structure data into vector representations through text embedding technology. Convert the text data into a fixed-length vector, fuse the vector representation of the archive graph structure data and the vector representation of the text structure data, and integrate different types of information into the same feature space. Text embedding technology includes word embedding and paragraph embedding.
[0043] Step 3: Based on the vectorized and fused data, a deep learning classification model is constructed and trained. The data is input into the classification model for classification, the archival data is divided into different categories, and corresponding classification labels are generated. The classification labels identify and organize the archival data as reference data for subsequent recommendations and analysis;
[0044] Step 4: Based on the classification label definition of the data in step 3, obtain the recommended data of the historical index and the corresponding index content, and determine the archive data related to the user's interests by analyzing the existing index data and its corresponding archive content, and use it as the recommended data;
[0045] Step 5: By analyzing the final cited data, the difference representation between the corresponding index content and the cited data is compared. The difference representation is used as the user's historical behavior feature representation, reflecting the user's behavior and preference in the archive management system;
[0046] Step 6: Use the Transformer model to learn the fused feature vector to capture the relationship between features and contextual information. Train the Transformer model on the fused feature vector, combine the recommendation data obtained in step 4 and the difference representation obtained in step 5 for comprehensive analysis, and convert the learned user preferences into data representation as the final preference reference value. The Transformer model obtains and calculates the recommendation score of each type of data, performs normalization, and classifies it through a Softmax layer in the fully connected layer. The output predicted probability distribution is sorted to obtain the top n recommendations, and the final recommendation value is calculated through binary cross entropy as the representation of the final preference reference value.
[0047] Step 7: Under the classification label, integrate the final preference reference values obtained into a personal preference set and provide a modification port;
[0048] Step 8: After the user verifies the identity information, the user enters personal information and a short text of the subject to be queried on the decoding end. According to the user's personal preference set, the corresponding content is captured under different classification tags, and the archival data related to the user's query subject is filtered out, and the content of interest is recommended to the user.
[0049] Compared with the existing technology, by obtaining archival graph structure data and text structure data, and then vectorizing and fusing them in the same feature space, the data processing process is greatly simplified, the efficiency of archival management is improved, and a large amount of archival data is organized through classification labels. Features are learned from the corpus and valuable information is extracted, thereby improving the validity and accuracy of archival retrieval and recommendation.
[0050] Example 2
[0051] At other levels, this embodiment also provides a method for calculating a loss function. The text embedding technology in step 2 generates vectorized representations of each entity and relationship, vectorizes different types of nodes and relationships in a unified feature space, and performs embedding training. It is assumed that there is a relationship r∈R (relationship set) between entities (h, t)∈E (entity set), h, t, r are used as vectorized representations, (h, t, r) constitute a triple, and given constraints The calculation formula of the loss function of its embedding training is:
[0052] L=∑ a∈S ∑ b∈D [y+f r (h,t)-f v (k,p)] + ;
[0053] Where L represents the coefficient of the loss function, a∈S represents the triple (h, t, r) in the knowledge graph, y represents the hyperparameter of the edge distance, b∈D represents the negative sample triple (k, p, v) randomly extracted during training, and f r (h, t) represents the triple performance, f v (k,p) represents the performance of negative sample triples, [y+f r (h,t)-f v (k,p)] represents x, [y+f r (h,t)-f v (k,p)] + represents max(0,x);
[0054] Among them, f r The calculation formula for (h,t) is:
[0055]
[0056] Example 3
[0057] In this embodiment, the personal preference set includes archival data related to the user's interests. After being enabled, related recommendations can be made based on the user's historical behavior characteristics and the index content corresponding to the preferences obtained by feature learning. After the user with modification authority intervenes, the modification port can be used by the user to trigger the operation of customizing and modifying personal preference characteristics. After the modification is completed, the modification results will be fed back, and the content topics that the user has recently queried, browsed, and referenced will be compared with past records to see if there are obvious new interest points. The frequency of use of user interest point classification tags is counted to see if there is any adjustment, whether the tags that were less used before but are now beginning to increase reflect new interests, and whether the new personal preference features submitted by the user contain previously unavailable topic keywords to prove that the user's interests have been extended.
[0058] Working principle: By acquiring archival graph structure data and text structure data, the present invention performs vectorized expression and fusion in the same feature space, significantly simplifying the data processing process and improving the efficiency of archival management. In addition, a large amount of archival data is managed in an organized manner through classification labels, features are learned from the corpus, and valuable information is extracted, thereby improving the validity and accuracy of archival retrieval and recommendation;
[0059] In order to better meet the personalized needs of users, the present invention also builds personal preference sets by learning the historical behavior characteristics of users, further improving the personalization level of the enterprise recommendation system. Users only need to enter their personal information and query topics, and the content of interest can be captured and recommended under different classification tags according to the user's exclusive preference set;
[0060] In addition, the present invention also uses big data technology to analyze users' long-term behavior patterns, extract users' implicit topic preferences, and provide data-driven personalized services to provide differentiated experiences for different people. At the same time, active learning is used to continuously optimize the recognition of personal preferences and improve the quality of recommendations and questions and answers.
[0061] In summary, the present invention improves file management efficiency, optimizes user experience, and meets personalized needs through various means, and has important application value.
[0062] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for recommending target knowledge features of enterprise-related archive management, characterized in that: The following steps are involved: Step 1: Intervene in the archive management system to obtain the enterprise's archive data, including archive graph structure data and text structure data; Step 2: Convert the archive graph structure data and text structure data into vector representations through text embedding technology, convert the text data into a vector of fixed length, fuse the vector representation of the archive graph structure data and the vector representation of the text structure data, and integrate different types of information into the same feature space; Step 3: Based on the vectorized and fused data, build and train a deep learning classification model, input the data into the classification model for classification, divide the archival data into different categories, and generate corresponding classification labels; Step 4: Based on the classification label definition of the data in step 3, obtain the recommended data of the historical index and the corresponding index content, and determine the archive data related to the user's interests by analyzing the existing index data and its corresponding archive content, and use it as the recommended data; Step 5: By analyzing the final cited data, compare the difference between the corresponding index content and the cited data; Step 6: Use the Transformer model to learn the fused feature vector to capture the relationship and contextual information between features. By training the Transformer model on the fused feature vector, combined with the recommendation data obtained in step 4 and the difference representation obtained in step 5 for comprehensive analysis, the learned user preferences are converted into data representation as the final preference reference value; Step 7: Under the classification label, integrate the final preference reference values obtained into a personal preference set and provide a modification port; Step 8: After the user verifies the identity information, the user enters personal information and a short text of the subject to be queried on the decoding end. According to the user's personal preference set, the corresponding content is captured under different classification tags, and the archival data related to the user's query subject is filtered out, and the content of interest is recommended to the user.
2. The method for recommending target knowledge features of enterprise-related archive management according to claim 1, characterized in that: The archive graph structure data in step 1 is used to describe the relationship and hierarchy between different types of archives, and basic attributes include: folder structure and file association.
3. The method for recommending target knowledge features of enterprise-related archive management according to claim 1, characterized in that: The text embedding technology in step 2 includes word embedding and paragraph embedding.
4. The method for recommending target knowledge features of enterprise-related archive management according to claim 1, characterized in that: The text embedding technology in step 2 generates vectorized representations of each entity and relationship, vectorizes different types of nodes and relationships in a unified feature space, and performs embedding training. Assuming that there is a relationship r∈R (relationship set) between entities (h, t)∈E (entity set), h, t, r are used as vectorized representations, (h, t, r) constitutes a triple, and given constraints , , , the calculation formula of the loss function of its embedding training is: ; In the formula, L represents the loss function coefficient, represents the triple (h, t, r) in the knowledge graph, y represents the hyperparameter of the edge distance, represents the negative sample triple (k, p, v) randomly extracted during training, represents the triple performance, represents the performance of negative sample triples, represents x, represents max(0,x); in, The calculation formula is: 。 5. The method for recommending target knowledge features of enterprise-related archive management according to claim 1, characterized in that: The classification labels in step 3 identify and organize the archival data as reference data for subsequent recommendations and analysis.
6. The method for recommending target knowledge features of enterprise-related archive management according to claim 1, characterized in that: The difference representation in step 5 is used as a representation of the user's historical behavior characteristics, reflecting the user's behavior and preference in the archive management system.
7. The method for recommending target knowledge features of enterprise-related archive management according to claim 1, characterized in that: The Transformer model in step 6 obtains and calculates the recommendation score of each type of data, performs normalization operation and classifies it by a Softmax layer in the fully connected layer, sorts the output predicted probability distribution to obtain the top n recommendations, and calculates the final recommendation value through binary cross entropy as a representation of the final preference reference value.
8. The method for recommending target knowledge features of enterprise-related archive management according to claim 1, characterized in that: The personal preference set in step 7 includes profile data related to user interests. After being enabled, it can make associated recommendations based on the user's historical behavior characteristics and the index content corresponding to the preferences obtained through feature learning.
9. The method for recommending target knowledge features of enterprise-related archive management according to claim 1, characterized in that: The modification port in step 7 can be used by the user to trigger the operation of customizing and modifying personal preference characteristics after the user with modification authority intervenes. After the modification is completed, the modification result will be fed back.
