Intelligent office assistant system based on natural language processing

Abstract

The invention discloses an intelligent office assistant system based on natural language processing, and relates to the technical field of natural language processing. When the system runs, a user intention vector I is generated by acquiring query behavior data of a user, generating a user behavior feature vector B and analyzing the real demand of the user by using a natural language technology; the method comprises the following steps of: performing similarity calculation with an entry vector R of a rule and regulation inventory, generating a matching result sim (I, R), generating a personalized recommendation list TC, comparing the personalized recommendation list TC with a pushing execution threshold TZ, accurately controlling the pushing content, adaptively adjusting the pushing execution threshold TZ by using an operation record after a user receives the pushing content, and realizing dynamic response to user demand change. According to the technical scheme, a comprehensive and efficient solution is provided, the problem of inaccurate matching caused by dependence on keyword search is solved, the phenomenon that the pushed content is redundant or does not accord with requirements is effectively avoided, and the information matching efficiency and the user satisfaction degree are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to an intelligent office assistant system based on natural language processing. Background Art

[0002] In the invention of Chinese patent application number CN202410709352.8, an efficient office system for power plants based on artificial intelligence is disclosed, including a data acquisition module, a data management module, a data analysis and processing module, a decision support module, and a data display module. The present invention integrates artificial intelligence technologies such as natural language processing, machine learning, and data analysis to realize an intelligent power plant office management system. It provides intelligent data management and classification functions to facilitate employees to quickly obtain and share information. It uses machine learning and data analysis technology to provide accurate data support and decision-making suggestions to improve the accuracy and efficiency of decision-making. It optimizes work processes and provides intelligent assistant support to reduce work duplication and human errors, and improve work efficiency and accuracy.

[0003] It can be seen that at present, most office assistant systems are mainly based on "passive query", and users must explicitly enter questions to obtain corresponding answers. This model has significant shortcomings. The first is the lag in information access, and users cannot actively query when they are not aware of certain new regulations or important policies; the second is the low interaction efficiency, and frequent input and retrieval increase the user's operating costs; finally, the existing system is difficult to conduct in-depth analysis and mining of user behavior data, and fails to make full use of existing interaction records to optimize the service experience. For example, if a user searches for a specific rule or regulation many times, it may reflect changes in business process needs or management blind spots, but existing technologies cannot capture this potential information, resulting in insufficient intelligence. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides an intelligent office assistant system based on natural language processing, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: an intelligent office assistant system based on natural language processing, including a behavior data collection module, a user intention analysis module, a rule matching module, a recommendation generation module, a push strategy execution module and a feedback optimization module;

[0006] The behavior data collection module collects the user's query behavior data, processes and analyzes it, and obtains the user behavior feature vector B;

[0007] The user intention analysis module analyzes the intention of the user query using natural language technology according to the acquired user behavior feature vector B to generate a user intention vector I;

[0008] The rule matching module calculates the similarity between the user intention vector I and the item vectors R in the preset rules and regulations database, and generates a matching result sim(I, R);

[0009] The recommendation generation module combines the user behavior feature vector B and the user intention vector I based on the matching result sim(I, R) to generate a personalized recommendation list TC;

[0010] The push strategy execution module compares the recommendation list TC with a preset push execution threshold TZ to determine whether the content in the recommendation list C is pushed to the user;

[0011] The feedback optimization module adaptively adjusts the push execution threshold TZ according to the user's operation record after the user receives the push content.

[0012] Preferably, the behavior data collection module includes a query data collection unit and a behavior feature processing unit;

[0013] The query data collection unit collects data in real time from the user's query interaction, including the query keyword input by the user, the query frequency and the query timestamp;

[0014] The query keywords obtain the text content input by the user by using onkeyup and onchange events to monitor the real-time input of the user in the input box on the interactive front-end page, and use natural language processing technology to extract the key words in the text content and mark them as query keywords, and then form a query keyword list of {Q1, Q2, Q3, ..., Qn};

[0015] After receiving the user's query request through the backend server, the query frequency will be recorded and stored in the database, including recording the keyword, user ID and timestamp request data of each query, and then querying the query frequency of the user's query keyword obtained in real time through SQL statements, and then forming a query frequency list of each query keyword {F1, F2, F3, ..., Fn};

[0016] The timestamp records the time of each keyword query to form a query timestamp list of the time series {T1, T2, T3, ..., Tn}.

[0017] Preferably, the behavior feature processing unit extracts and processes the query keyword list {Q1, Q2, Q3, ..., Qn}, the query frequency list {F1, F2, F3, ..., Fn} of each query keyword, and the query timestamp list of the time series {T1, T2, T3, ..., Tn}, obtains the keyword Qk of the kth query, the query frequency Fk of the kth query keyword Qk, and the timestamp Tk of the kth query keyword Qk, and performs feature processing to obtain the user behavior feature vector B of the user i;

[0018] The user behavior feature vector B is obtained by the following calculation formula:

[0019]

[0020] Where Bi represents the user behavior feature vector B of user i, n represents the total number of user queries, log represents the logarithmic function, e represents the exponential decay function, and △Tk represents the time interval of the query timestamp Tk.

[0021] Preferably, the user intention analysis module includes an intention classification modeling unit and an intention vector generation unit;

[0022] The intent classification modeling unit is used to define and vectorize the model, including using the BGE model and the M3E model for definition and training, and obtaining the intent category matrix E, wherein the intent category matrix E dimension is specifically C*d dimension, C represents the number of intent categories, and d represents the dimension of each category feature vector;

[0023] The intention category matrix E is specifically:

[0024] Among them, the number of intent categories C and the dimension d of each category feature vector are obtained by extracting the company's rules and regulations text.

[0025] Preferably, the intention vector generation unit uses the intention category matrix E, combined with the user behavior feature vector B of user i, to generate the intention vector I of user i through vector calculation, and performs a probability distribution reflecting the intention of the current query of user i;

[0026] The intention vector I of user i is specifically obtained by projecting the user behavior feature vector B of user i onto the intention category matrix E obtained by pre-training, calculating the correlation of the intention category C, obtaining the correlation score z matrix, and then transforming the correlation score z matrix using the Softmax normalized exponential function to obtain the intention vector I matrix of user i of C*d dimension;

[0027] The relevance score z is obtained by the following calculation formula:

[0028]

[0029] where zq represents the relevance score of the qth intent category in intent category C, σ represents a nonlinear activation function, including Sigmoid and ReLU, d represents the dimension of the feature vector of each category, specifically the total number of feature vectors of the qth intent category in intent category C, k represents the kth feature in dimension d of the feature vector of each category, wk represents the global weight value of the kth feature, eqk represents the kth feature of the qth intent category in intent category C in the intent category matrix E, specifically the weight value of the kth feature dimension of the qth intent category in intent category C, Bi,k represents the kth feature in the user behavior feature vector B of user i, and bq represents the bias term of the qth intent category in intent category C.

[0030] Preferably, the rule matching module includes an entry vector generating unit and a similarity calculating unit;

[0031] The item vector generating unit converts each item in the rules and regulations database into a characteristic vector form by using natural language processing technology to obtain each item vector R;

[0032] The item vectors R are specifically R={R1, R2, R3, ..., Rr}.

[0033] Preferably, the similarity calculation unit calculates the matching result by obtaining all the intention categories in the user intention vector I of the user i and all the vectors in the item vector R, obtains the similarity score sim(i, p, q) between the qth intention category in the intention category C in the intention vector of the user i and the pth item vector in the item vector R, and then performs aggregation processing to generate a matching result sim(I, R);

[0034] The similarity score sim(i, p, q) is obtained by the following calculation formula:

[0035]

[0036] Wherein, sim(i, p, q) represents the similarity score between the qth intent category in the intent category C in the intent vector of user i and the pth item vector in the item vector R, Ii,q represents the component of the qth intent category in the intent vector of user i, specifically represents the correlation score of the qth intent category in the intent category C of user i, and Rp,q represents the correlation score between the pth item vector in the item vector R and the qth intent category;

[0037] The matching result sim(I, R) is obtained by The calculation formula is obtained, where sim(Ii, Rp) represents the matching result of user i regarding the p-th entry vector in the entry vector R, aq represents the preset weight value of the q-th intent category in the intent category C, and q represents the q-th intent category in the intent category C.

[0038] Preferably, the recommendation generation module includes a personalized recommendation generation unit;

[0039] The personalized recommendation generation unit combines the user behavior feature vector B and the user intention vector I based on the matching result sim(I, R), obtains the comprehensive score Fp of the p-th item vector in each item vector R, and generates a personalized recommendation list TC according to the comprehensive score Fp;

[0040] The comprehensive score Fp is obtained by the following calculation formula:

[0041] Fp=f1*sim(Ii,Rp)+f2*f(Bi,Bp)+f3*g(Ii,Rp);

[0042] Where f1, f2 and f3 represent the weight values ​​of the matching result, user behavior feature score and user intention score respectively, and f1+f2+f3=1. The specific value is set by the user. sim represents the similarity function, f represents the behavior correlation function, and g represents the deep matching function, which is specifically used to amplify the matching degree between user intention and items in key categories and highlight the highly correlated categories.

[0043] The recommendation list TC is associated and marked by the comprehensive score Fp of each item vector in the item vector R, and then bubble sorted to reorganize into the recommendation list TC = {(R1, F1), (R2, F2), (R3, F3), ..., (Rr, Fp)}.

[0044] Preferably, the recommendation list TC is compared with the preset push execution threshold TZ in the following manner:

[0045] When the comprehensive score Fp ≥ the push execution threshold TZ, the comprehensive score Fp and the associated item vectors Rr in the recommendation list TC are obtained and pushed to the user;

[0046] When the comprehensive score Fp is less than the push execution threshold TZ, the comprehensive score Fp and the associated item vectors Rr in the recommendation list TC are obtained and not pushed to the user.

[0047] Preferably, the user's operation record is obtained by extracting the front-end log record to obtain the user's non-interactive behavior time period information on the recommended content, including the pop-up time TC and the closing time GB, and the stay time TL is calculated, and the push execution threshold TZ is proportionally adjusted according to the calculated stay time TL to obtain the secondary adjustment push execution threshold ETZ;

[0048] The length of stay TL is obtained by the calculation formula TL=TC-GB;

[0049] The secondary adjustment push execution threshold ETZ is passed Where β represents the learning rate, and Tmax represents the preset maximum duration of a user's use of recommended content.

[0050] The present invention provides an intelligent office assistant system based on natural language processing, which has the following beneficial effects:

[0051] (1) When the system is running, the user behavior feature vector B is generated by obtaining the user's query behavior data, and the user's real needs are analyzed using natural language technology to generate the user intention vector I. Subsequently, the similarity is calculated with the entry vector R of the rules and regulations inventory to generate the matching result sim(I, R), and then the personalized recommendation list TC is generated. It is compared with the push execution threshold TZ to accurately control the push content. The user's operation record after receiving the push content is used to adaptively adjust the push execution threshold TZ to achieve dynamic response to changes in user needs, providing a comprehensive and efficient solution. On the one hand, natural language technology is used to analyze user needs and generate user intentions, overcoming the problem of inaccurate matching caused by relying on keyword search; on the other hand, the user behavior characteristics and push execution threshold TZ are combined for recommendation and dynamic optimization, effectively avoiding the redundancy of push content or the phenomenon of inconsistency with needs. The accurate push and personalized recommendation of rules and regulations information are achieved, greatly improving the information matching efficiency and user satisfaction.

[0052] (2) By using natural language processing technology to convert the items in the rules and regulations database into item vectors R, and by matching the user intention vector I with the item vector R, the similarity score sim(i, p, C) is calculated by category, thus achieving deep matching between user behavior and rules and regulations items, and fully exploring the semantic relevance of rules and regulations items under multi-category features. By combining the user intention vector I with the intention category matrix E, it is possible to provide a multi-dimensional interpretation of complex user query behaviors, and through the sub-item aggregation strategy, dynamically weigh the importance of different intent categories to ensure that the matching result sim(I, R) is more in line with the user's current actual needs. Compared with the previous single-dimensional matching method, this system greatly improves the matching accuracy and relevance, and optimizes the query efficiency and user experience of rules and regulations.

[0053] (3) Based on the matching result sim(I, R), combined with the user behavior feature vector B and the user intention vector I, and the scoring result is associated with the item, the personalized recommendation list TC is generated by bubble sorting to further filter the items that meet the user's needs. For the user's non-interactive behavior time period, the push execution threshold TZ is proportionally adjusted by calculating the ratio of the stay time TL to the preset maximum stay time Tmax, and the secondary adjustment push execution threshold ETZ is generated. The push strategy is adaptively optimized, which reflects user needs at a deeper level compared with the traditional static scoring and fixed push rules. By adjusting the push execution threshold ETZ twice, it can adapt to the dynamic changes of user interests in real time, ensuring that the recommended content is neither redundant nor omitted, further optimizing the push effect and user satisfaction, and comprehensively improving the intelligence level of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 The present invention is a schematic diagram of a block diagram of an intelligent office assistant system based on natural language processing. DETAILED DESCRIPTION

[0055] The technical solutions 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 only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0056] Example 1

[0057] The present invention provides an intelligent office assistant system based on natural language processing, please refer to Figure 1 , including behavior data collection module, user intention analysis module, rule matching module, recommendation generation module, push strategy execution module and feedback optimization module;

[0058] The behavior data collection module collects the user's query behavior data, processes and analyzes it, and obtains the user behavior feature vector B;

[0059] The user intention analysis module analyzes the intention of the user query using natural language technology according to the acquired user behavior feature vector B to generate a user intention vector I;

[0060] The rule matching module calculates the similarity between the user intention vector I and the item vectors R in the preset rules and regulations database, and generates a matching result sim(I, R);

[0061] The recommendation generation module combines the user behavior feature vector B and the user intention vector I based on the matching result sim(I, R) to generate a personalized recommendation list TC;

[0062] The push strategy execution module compares the recommendation list TC with a preset push execution threshold TZ to determine whether the content in the recommendation list C is pushed to the user;

[0063] The feedback optimization module adaptively adjusts the push execution threshold TZ according to the user's operation record after the user receives the push content.

[0064] In this embodiment, by obtaining the user's query behavior data, a user behavior feature vector B is generated, and the user's real needs are analyzed using natural language technology to generate a user intention vector I. Subsequently, similarity calculation is performed with the entry vector R of the rules and regulations inventory to generate a matching result sim(I, R), and then a personalized recommendation list TC is generated. By comparing it with the push execution threshold TZ, the push content is accurately controlled, and the user's operation record after receiving the push content is used to adaptively adjust the push execution threshold TZ to achieve dynamic response to changes in user needs, providing a comprehensive and efficient solution. On the one hand, natural language technology is used to analyze user needs and generate user intentions, overcoming the problem of inaccurate matching caused by relying on keyword search; on the other hand, the user behavior characteristics and the push execution threshold TZ are combined for recommendation and dynamic optimization, effectively avoiding the redundancy of the push content or the phenomenon of inconsistency with the needs. Accurate push and personalized recommendation of rules and regulations information are achieved, greatly improving information matching efficiency and user satisfaction.

[0065] Example 2

[0066] This embodiment is explained in Example 1, please refer to Figure 1 ,Specifically: the behavior data collection module includes a query data collection unit and a behavior feature processing unit;

[0067] The query data collection unit collects data in real time from the user's query interaction, including the query keyword input by the user, the query frequency and the query timestamp;

[0068] The query keywords obtain the text content input by the user by using onkeyup and onchange events to monitor the real-time input of the user in the input box on the interactive front-end page, and use natural language processing technology to extract the key words in the text content and mark them as query keywords, and then form a query keyword list of {Q1, Q2, Q3, ..., Qn};

[0069] After receiving the user's query request through the backend server, the query frequency will be recorded and stored in the database, including recording the keyword, user ID and timestamp request data of each query, and then querying the query frequency of the user's query keyword obtained in real time through SQL statements, and then forming a query frequency list of each query keyword {F1, F2, F3, ..., Fn};

[0070] The timestamp records the time of each keyword query to form a query timestamp list of the time series {T1, T2, T3, ..., Tn}.

[0071] The behavior feature processing unit extracts and processes the query keyword list {Q1, Q2, Q3, ..., Qn}, the query frequency list {F1, F2, F3, ..., Fn} of each query keyword, and the query timestamp list of the time series {T1, T2, T3, ..., Tn}, obtains the keyword Qk of the kth query, the query frequency Fk of the kth query keyword Qk, and the timestamp Tk of the kth query keyword Qk, and performs feature processing to obtain the user behavior feature vector B of user i;

[0072] The user behavior feature vector B is obtained by the following calculation formula:

[0073]

[0074] Where Bi represents the user behavior feature vector B of user i, which is specifically used to describe the query behavior pattern of user i, including keyword preference, query frequency and time interval characteristics, n represents the total number of user queries, specifically represents the total number in the query keyword list {Q1, Q2, Q3, …, Qn}, log represents the logarithmic function, e represents the exponential decay function, and △Tk represents the time interval of the query timestamp Tk.

[0075] In this embodiment, the text content input by the user is obtained in real time by using the event monitoring technology of the interactive front-end page, and the query keywords are extracted and a query keyword list is generated by combining the natural language processing technology; at the same time, the request data and timestamp information of the user query are recorded by the back-end server, and the query frequency list and query timestamp list of each query keyword are counted in real time to form a high-precision behavior data record, and further based on the combination of query keywords, query frequency and query timestamp, the user's behavior pattern is feature extracted, and finally the user behavior feature vector B of user i is generated by formula calculation, which effectively quantifies the user's query behavior characteristics, ensures the real-time and integrity of the user query behavior data, and also realizes the pattern analysis of the user's query behavior, making up for the lack of the traditional query system's ability to capture dynamic changes in behavior. By generating the user behavior feature vector B, the module can accurately describe the user's keyword preference, query frequency change and time interval characteristics, thereby providing richer input data. Compared with the traditional method that relies only on keyword matching, the dynamic characteristics of behavior are integrated into the user portrait, which significantly enhances the system's ability to understand the user's needs globally. This transition from a single query behavior to a comprehensive behavior pattern lays a more scientific foundation for subsequent intent analysis and personalized recommendations.

[0076] Example 3

[0077] This embodiment is explained in Example 2. Please refer to Figure 1 ,Specifically: the user intention analysis module includes an intention classification modeling unit and an intention vector generation unit;

[0078] The intent classification modeling unit is used to define and vectorize the model, including using the BGE model and the M3E model for definition and training, and obtaining the intent category matrix E, wherein the intent category matrix E dimension is specifically C*d dimension, C represents the number of intent categories, and d represents the dimension of each category feature vector;

[0079] The intention category matrix E is specifically:

[0080] Among them, the number of intent categories C and the dimension d of each category feature vector are obtained by extracting the company's rules and regulations text.

[0081] The intention vector generation unit uses the intention category matrix E and the user behavior feature vector B of user i to generate the intention vector I of user i through vector calculation to reflect the intention probability distribution of the current query of user i;

[0082] The intention vector I of user i is specifically obtained by projecting the user behavior feature vector B of user i onto the intention category matrix E obtained by pre-training, calculating the correlation of the intention category C, obtaining the correlation score z matrix, and then transforming the correlation score z matrix using the Softmax normalized exponential function to obtain the intention vector I matrix of user i of C*d dimension;

[0083] The relevance score z is obtained by the following calculation formula:

[0084]

[0085] Wherein, zq represents the relevance score of the qth intent category in intent category C, σ represents a nonlinear activation function, including Sigmoid and ReLU, d represents the dimension of each category feature vector, specifically the total number of feature vectors of the qth intent category in intent category C, k represents the kth feature in dimension d of each category feature vector, wk represents the global weight value of the kth feature, eqk represents the kth feature of the qth intent category in intent category C in the intent category matrix E, specifically represents the weight value of the kth feature dimension of the qth intent category in intent category C, Bi,k represents the kth feature in the user behavior feature vector B of user i, and bq represents the bias term of the qth intent category in intent category C;

[0086] The relevance score zC of the intent category C is specifically

[0087] By integrating the relevance scores zC of the intent categories C, a relevance score z matrix of the C*d dimension is obtained. The relevance score z matrix is ​​specifically:

[0088] The intention vector I matrix of user i is specifically:

[0089] The rule matching module includes an entry vector generating unit and a similarity calculating unit;

[0090] The item vector generating unit converts each item in the rules and regulations database into a characteristic vector form by using natural language processing technology to obtain each item vector R;

[0091] The item vectors R are specifically R={R1, R2, R3, ..., Rr}.

[0092] The similarity calculation unit obtains all the intention categories in the user intention vector I of user i and all the vectors in the item vector R to perform matching result calculation, obtains the similarity score sim(i, p, q) between the qth intention category in the intention category C in the intention vector of user i and the pth item vector in the item vector R, and then performs aggregation processing to generate a matching result sim(I, R);

[0093] The similarity score sim(i, p, q) is obtained by the following calculation formula:

[0094]

[0095] Wherein, sim(i, p, q) represents the similarity score between the qth intent category in the intent category C in the intent vector of user i and the pth item vector in the item vector R, Ii,q represents the component of the qth intent category in the intent vector of user i, specifically represents the correlation score of the qth intent category in the intent category C of user i, and Rp,q represents the correlation score between the pth item vector in the item vector R and the qth intent category;

[0096] The matching result sim(I, R) is obtained by The calculation formula is obtained, where sim(Ii, Rp) represents the matching result of user i on the p-th item vector in the item vector R, aq represents the preset weight value of the q-th intent category in the intent category C, and q represents the q-th intent category in the intent category C;

[0097] The matching result sim(I, R) is specifically sim(I, R)={sim(Ii, R1), sim(Ii, R2), sim(Ii, R3), ..., sim(Ii, Rr)}.

[0098] In this embodiment, through accurate intention analysis and efficient item matching, the relevance between rules and regulations and user needs is comprehensively improved. Combined with the user behavior feature vector B of user i, based on the pre-trained intention category matrix E, the intention vector I of user i is generated through nonlinear mapping and Softmax normalized exponential function, and the user's behavior pattern is converted into a multi-category intention probability distribution. At the same time, natural language processing technology is used to convert the items in the rules and regulations database into item vectors R. By matching the user intention vector I with the item vector R, the similarity score sim(i, p, C) by category is calculated, and the deep matching of user behavior and rules and regulations items is achieved, and the semantic relevance of rules and regulations items under multi-category features is fully explored. Through the combination of user intention vector I and intention category matrix E, a multi-dimensional interpretation can be provided for complex user query behaviors, and the importance of different intention categories can be dynamically weighed through the sub-item aggregation strategy to ensure that the matching result sim(I, R) is more in line with the current actual needs of users. Compared with the previous single-dimensional matching method, this system greatly improves the accuracy and relevance of matching, and optimizes the query efficiency and user experience of rules and regulations.

[0099] Example 4

[0100] This embodiment is explained in Example 3, please refer to Figure 1 ,Specifically: the recommendation generation module includes a personalized recommendation generation unit;

[0101] The personalized recommendation generation unit combines the user behavior feature vector B and the user intention vector I based on the matching result sim(I, R), obtains the comprehensive score Fp of the p-th item vector in each item vector R, and generates a personalized recommendation list TC according to the comprehensive score Fp;

[0102] The comprehensive score Fp is obtained by the following calculation formula:

[0103] Fp=f1*sim(Ii,Rp)+f2*f(Bi,Bp)+f3*g(Ii,Rp);

[0104] Where, f1, f2 and f3 represent the weight values ​​of the matching result, user behavior feature score and user intention score respectively, and f1+f2+f3=1. The specific value is set by the user. sim represents the similarity function, which is used to provide a global matching score and measure the overall fit between user intention and item content. f represents the behavior correlation function, which is used to reflect the matching degree between user historical behavior and item content and strengthen the behavioral basis for recommendation. g represents the deep matching function, which is used to amplify the matching degree between user intention and item in key categories and highlight high correlation categories.

[0105] The recommendation list TC is associated and marked by the comprehensive score Fp of each item vector in the item vector R, and then bubble sorted to reorganize into the recommendation list TC = {(R1, F1), (R2, F2), (R3, F3), ..., (Rr, Fp)}.

[0106] The method of comparing the recommendation list TC with the preset push execution threshold TZ is as follows:

[0107] When the comprehensive score Fp ≥ the push execution threshold TZ, the comprehensive score Fp and the associated item vectors Rr in the recommendation list TC are obtained and pushed to the user;

[0108] When the comprehensive score Fp is less than the push execution threshold TZ, the comprehensive score Fp and the associated item vectors Rr in the recommendation list TC are obtained and not pushed to the user.

[0109] The user's operation record is obtained by extracting the front-end log record to obtain the user's non-interactive behavior time period information on the recommended content, including the pop-up time TC and the closing time GB, and the stay time TL is calculated. The push execution threshold TZ is proportionally adjusted according to the calculated stay time TL to obtain the secondary adjustment push execution threshold ETZ;

[0110] The length of stay TL is obtained by the calculation formula TL=TC-GB;

[0111] The secondary adjustment push execution threshold ETZ is passed Where β represents the learning rate, which is used to control the amplitude of the proportional adjustment, and Tmax represents the preset maximum stay time of users using the recommended content.

[0112] In this embodiment, based on the matching result sim(I, R), combined with the user behavior feature vector B and the user intention vector I, and the scoring result and the item are associated and marked, a personalized recommendation list TC is generated by bubble sorting to further filter the items that meet the user's needs. For the user's non-interactive behavior time period, the push execution threshold TZ is proportionally adjusted by calculating the ratio of the stay time TL to the preset maximum stay time Tmax, and the secondary adjustment push execution threshold ETZ is generated. The push strategy is adaptively optimized, which reflects user needs at a deeper level compared to the traditional static scoring and fixed push rules. By adjusting the push execution threshold ETZ twice, it is possible to adapt to the dynamic changes of user interests in real time, ensuring that the recommended content is neither redundant nor omitted, further optimizing the push effect and user satisfaction, and comprehensively improving the intelligence level of the recommendation system.

[0113] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent office assistant system based on natural language processing, characterized in that: It includes behavior data collection module, user intention analysis module, rule matching module, recommendation generation module, push strategy execution module and feedback optimization module; The behavior data collection module collects the user's query behavior data, processes and analyzes it, and obtains the user behavior feature vector B; The user intention analysis module analyzes the intention of the user query using natural language technology according to the acquired user behavior feature vector B to generate a user intention vector I; The rule matching module calculates the similarity between the user intention vector I and the item vectors R in the preset rules and regulations database, and generates a matching result sim(I, R); The recommendation generation module combines the user behavior feature vector B and the user intention vector I based on the matching result sim(I, R) to generate a personalized recommendation list TC; The push strategy execution module compares the recommendation list TC with a preset push execution threshold TZ to determine whether the content in the recommendation list C is pushed to the user; The feedback optimization module adaptively adjusts the push execution threshold TZ according to the user's operation record after the user receives the push content.

2. The intelligent office assistant system based on natural language processing according to claim 1, characterized in that: The behavior data collection module includes a query data collection unit and a behavior feature processing unit; The query data collection unit collects data in real time from the user's query interaction, including the query keyword input by the user, the query frequency and the query timestamp; The query keywords obtain the text content input by the user by using onkeyup and onchange events to monitor the real-time input of the user in the input box on the interactive front-end page, and use natural language processing technology to extract the key words in the text content and mark them as query keywords, and then form a query keyword list of {Q1, Q2, Q3, ..., Qn}; After receiving the user's query request through the backend server, the query frequency will be recorded and stored in the database, including recording the keyword, user ID and timestamp request data of each query, and then querying the query frequency of the user's query keyword obtained in real time through SQL statements, and then forming a query frequency list of each query keyword {F1, F2, F3, ..., Fn}; The timestamp records the time of each keyword query to form a query timestamp list of the time series {T1, T2, T3, ..., Tn}.

3. The intelligent office assistant system based on natural language processing according to claim 2, characterized in that: The behavior feature processing unit extracts and processes the query keyword list {Q1, Q2, Q3, ..., Qn}, the query frequency list {F1, F2, F3, ..., Fn} of each query keyword, and the query timestamp list of the time series {T1, T2, T3, ..., Tn}, obtains the keyword Qk of the kth query, the query frequency Fk of the kth query keyword Qk, and the timestamp Tk of the kth query keyword Qk, and performs feature processing to obtain the user behavior feature vector B of user i; The user behavior feature vector B is obtained by the following calculation formula: Where Bi represents the user behavior feature vector B of user i, n represents the total number of user queries, log represents the logarithmic function, e represents the exponential decay function, and △Tk represents the time interval of the query timestamp Tk.

4. The intelligent office assistant system based on natural language processing according to claim 1, characterized in that: The user intention analysis module includes an intention classification modeling unit and an intention vector generation unit; The intent classification modeling unit is used to define and vectorize the model, including using the BGE model and the M3E model for definition and training, and obtaining the intent category matrix E, wherein the intent category matrix E dimension is specifically C*d dimension, C represents the number of intent categories, and d represents the dimension of each category feature vector; The intention category matrix E is specifically: Among them, the number of intent categories C and the dimension d of each category feature vector are obtained by extracting the company's rules and regulations text.

5. The intelligent office assistant system based on natural language processing according to claim 4, characterized in that: The intention vector generation unit uses the intention category matrix E and the user behavior feature vector B of user i to generate the intention vector I of user i through vector calculation to reflect the intention probability distribution of the current query of user i; The intention vector I of user i is specifically obtained by projecting the user behavior feature vector B of user i onto the intention category matrix E obtained by pre-training, calculating the correlation of the qth intention category in the intention category C, obtaining the correlation score z matrix, and then transforming the correlation score z matrix using the Softmax normalized exponential function to obtain the intention vector I matrix of user i with C*d dimensions; The relevance score z is obtained by the following calculation formula: where zq represents the relevance score of the qth intent category in intent category C, σ represents a nonlinear activation function, including Sigmoid and ReLU, d represents the dimension of the feature vector of each category, specifically the total number of feature vectors of the qth intent category in intent category C, k represents the kth feature in dimension d of the feature vector of each category, wk represents the global weight value of the kth feature, eqk represents the kth feature of the qth intent category in intent category C in the intent category matrix E, specifically the weight value of the kth feature dimension of the qth intent category in intent category C, Bi,k represents the kth feature in the user behavior feature vector B of user i, and bq represents the bias term of the qth intent category in intent category C.

6. The intelligent office assistant system based on natural language processing according to claim 1, characterized in that: The rule matching module includes an entry vector generating unit and a similarity calculating unit; The item vector generating unit converts each item in the rules and regulations database into a characteristic vector form by using natural language processing technology to obtain each item vector R; The item vectors R are specifically R={R1, R2, R3, ..., Rr}.

7. The intelligent office assistant system based on natural language processing according to claim 6, characterized in that: The similarity calculation unit obtains all the intention categories in the user intention vector I of user i and all the vectors in the item vector R to perform matching result calculation, obtains the similarity score sim(i, p, q) between the qth intention category in the intention category C in the intention vector of user i and the pth item vector in the item vector R, and then performs aggregation processing to generate a matching result sim(I, R); The similarity score sim(i, p, q) is obtained by the following calculation formula: Wherein, sim(i, p, q) represents the similarity score between the qth intent category in the intent category C in the intent vector of user i and the pth item vector in the item vector R, Ii,q represents the component of the qth intent category in the intent vector of user i, specifically represents the correlation score of the qth intent category in the intent category C of user i, and Rp,q represents the correlation score between the pth item vector in the item vector R and the qth intent category; The matching result sim(I, R) is obtained by The calculation formula is obtained, where sim(Ii, Rp) represents the matching result of user i regarding the p-th entry vector in the entry vector R, aq represents the preset weight value of the q-th intent category in the intent category C, and q represents the q-th intent category in the intent category C.

8. The intelligent office assistant system based on natural language processing according to claim 7, characterized in that: The recommendation generation module includes a personalized recommendation generation unit; The personalized recommendation generation unit combines the user behavior feature vector B and the user intention vector I based on the matching result sim(I, R), obtains the comprehensive score Fp of the p-th item vector in each item vector R, and generates a personalized recommendation list TC according to the comprehensive score Fp; The comprehensive score Fp is obtained by the following calculation formula: Fp=f1*sim(Ii,Rp)+f2*f(Bi,Bp)+f3*g(Ii,Rp); Where f1, f2 and f3 represent the weight values ​​of the matching result, user behavior feature score and user intention score respectively, and f1+f2+f3=1. The specific value is set by the user. sim represents the similarity function, f represents the behavior correlation function, and g represents the deep matching function, which is specifically used to amplify the matching degree between user intention and items in key categories and highlight the highly correlated categories. The recommendation list TC is associated and marked by the comprehensive score Fp of each item vector in the item vector R, and then bubble sorted to reorganize into the recommendation list TC = {(R1, F1), (R2, F2), (R3, F3), ..., (Rr, Fp)}.

9. The intelligent office assistant system based on natural language processing according to claim 8, characterized in that: The method of comparing the recommendation list TC with the preset push execution threshold TZ is as follows: When the comprehensive score Fp ≥ the push execution threshold TZ, the comprehensive score Fp and the associated item vectors Rr in the recommendation list TC are obtained and pushed to the user; When the comprehensive score Fp is less than the push execution threshold TZ, the comprehensive score Fp and the associated item vectors Rr in the recommendation list TC are obtained and not pushed to the user.

10. The intelligent office assistant system based on natural language processing according to claim 1, characterized in that: The user's operation record is obtained by extracting the front-end log record to obtain the user's non-interactive behavior time period information on the recommended content, including the pop-up time TC and the closing time GB, and the stay time TL is calculated. The push execution threshold TZ is proportionally adjusted according to the calculated stay time TL to obtain the secondary adjustment push execution threshold ETZ; The length of stay TL is obtained by the calculation formula TL=TC-GB; The secondary adjustment push execution threshold ETZ is passed Where β represents the learning rate, and Tmax represents the preset maximum duration of a user's use of recommended content.

Images