Personalized recommendation method and system based on natural language processing technology

Abstract

The present invention belongs to the field of artificial intelligence information technology. The present invention provides a personalized recommendation method and system based on natural language processing technology. The collected user data is pre-processed to obtain cleaned text data. Based on the chat information and search records in the cleaned user text data, the user's points of interest are predicted and evaluated, so as to avoid the situation of information cocoons appearing in the process of the model predicting the user's points of interest. The prediction results are evaluated again, which is also helpful to improve the user experience and increase the user's usage frequency and duration. The accuracy of the model performance is calculated, and the accuracy value is used as the representative value of the model performance index. On the basis of the representative value of the model performance index, a representative value of the model performance index fluctuation is obtained. The representative value of the model performance index fluctuation is used as an indicator for judging whether the model performance is affected, so as to provide a performance early warning mechanism for users when using the model.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence information technology, and specifically, a personalized recommendation method and system based on natural language processing technology. Background Art

[0002] With the rapid development of the Internet and the explosive growth of information, users are faced with a huge amount of information to choose from. Traditional recommendation systems mainly make recommendations based on users' historical behavior records, but this approach has the problem of information isolation. Natural language processing technology makes recommendations based on multiple aspects of user information.

[0003] A Chinese patent application with publication number CN112417835A discloses an intelligent review method and system for procurement documents based on natural language processing technology, including: solidifying online document templates for technical specifications and feasibility studies and estimates using web technology and frameworks; exporting the core field data of the work item parts of the solidified technical specifications and feasibility studies and estimates, and performing data preprocessing; using a similarity algorithm to analyze the processed core field data of the technical specifications and the core field data of the feasibility studies and estimates to obtain a review report.

[0004] In the existing technology, only the repetitive and tedious work in manual review is reduced, and detailed errors caused by high-load manual review are avoided. The review report obtained after the data analysis is not further analyzed, so there is an inaccurate review report, and the fluctuation value of the system during operation is not obtained, and the fluctuation of the system during operation is not analyzed, and the effect of optimizing and upgrading the system in the later stage is not achieved. Summary of the invention

[0005] In order to make up for the deficiencies of the prior art, at least one technical problem raised in the background technology is solved.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a personalized recommendation method and system based on natural language processing technology, comprising the following steps:

[0007] Step 1: Collect user text data and pre-process the text data;

[0008] In step 1, the text data is preprocessed. The preprocessing steps include deduplication, missing value processing, and outlier detection and processing. The specific processing process is as follows:

[0009] A1, performing deduplication processing on the user text data to obtain deduplicated text data;

[0010] A2, based on the deduplicated text data, judge and process the outliers in the text data;

[0011] A3, detect and process missing values ​​in text data to obtain cleaned text data;

[0012] Step 2: Processing based on user text data to establish a user recommendation model;

[0013] B1, processing the user's text data to obtain the user's personal interest information, the text data includes chat information and search records;

[0014] B2, train the model based on the user's personal interest information, predict the user's points of interest and make recommendations;

[0015] B3, evaluate the conclusion of the model predicting the user's interest points to determine whether the model prediction is accurate;

[0016] Step three: collect feedback information from users using the user recommendation model and judge the performance indicators of the user recommendation model.

[0017] The present invention further illustrates that in A2, by converting the deduplicated text data into a data collection, establishing a three-dimensional coordinate system, and converting the data values ​​in the data collection into a box-type graph, the specific processing process is as follows:

[0018] A201, based on the data values ​​in the data collection, respectively obtain the minimum value, the first quartile, the median, the third quartile and the maximum value in the data collection;

[0019] A202, connect the first quartile and the third quartile to get the box outline, connect the minimum value and the maximum value so that the upper and lower edges of the box are connected, and the median is located in the center of the box;

[0020] A203, putting the data in the data collection into the three-dimensional box, and obtaining the data in the data collection that exists in the three-dimensional box and the data that does not exist in the three-dimensional box;

[0021] The data in the data collection that exist in the three-dimensional box are recorded as normal values;

[0022] The data in the data set that does not exist in the three-dimensional box is recorded as an outlier.

[0023] The present invention further illustrates that in A3, based on the three-dimensional box figure in A2, the data in the data collection that does not exist in the three-dimensional box is processed to determine whether it is a missing value. The specific process is as follows:

[0024] A301, based on the data not existing in the three-dimensional box, the data value is subtracted from the minimum value, the first quartile, the median, the third quartile and the maximum value, the difference is summed up to get the average value, and the average value is used as the representative value for determination;

[0025] A302, compare the determination representative value with the preset determination threshold value, the specific comparison process is as follows:

[0026] If the representative value is greater than or equal to the preset threshold, the data value is determined to be a missing value;

[0027] If the representative value is determined to be less than or equal to the preset determination threshold, the data value is determined to be not a missing value.

[0028] The present invention further illustrates: In B1, the specific process is as follows:

[0029] B101, by recording the search record contents in the user period, the frequency of the search record in the user period is obtained;

[0030] B102, extracting keywords from the chat information and recording the number of times the keywords appear to obtain the frequency of keyword appearance;

[0031] B103, based on the frequency of search records in the user period and the frequency of keywords in the chat information, a representative value of the user's personal interest information is obtained;

[0032] The representative value E of the user's personal interest information is obtained through the formula: E=(P+L)×α, where P represents the frequency of search records within the user period, L represents the frequency of keywords appearing in chat information, and α represents the preset proportional coefficient.

[0033] The present invention further illustrates: in B2, based on the representative value of the user's personal interest information, the user's personal interest areas are divided, and the representative value of the user's personal interest areas is obtained. The click rate and comment rate of the user's personal interest areas are recorded to obtain the user's personal interest intensity representative value. Based on the user's personal interest intensity representative value, the user's interest points are predicted. The specific process is as follows:

[0034] B201, dividing the representative values ​​of several user personal interest information into fields, recording the click rate of the user in each interest field, and summing up the click rate values ​​of multiple fields to obtain an average value, and using the average value as a threshold for determining the interest field;

[0035] B202, compare the click-through rate values ​​of multiple fields with the threshold for selecting the field of interest, and the comparison process is as follows:

[0036] If there is a field with a click rate greater than or equal to the threshold for determining an interest field among multiple fields, the field with the larger click rate is determined as the interest field;

[0037] If there is a field with a click-through rate value lower than the threshold for determining an interest field among multiple fields, then the field is determined to be the interest field with the smaller click-through rate value;

[0038] B203, based on the interest areas with larger click-through rate values, records the comment rate of the interest areas;

[0039] B204, by formula: The user's personal interest intensity representative value W is obtained, R represents the comment rate, Q represents the click rate, and β represents the preset proportional coefficient.

[0040] The present invention further illustrates: In B3, the specific process is as follows:

[0041] B301, obtaining user purchase records, matching keywords in the user purchase records with keywords in the user search records within a period, summing the user purchase record frequency with the user search record frequency within a period, and obtaining an evaluation threshold;

[0042] B302, subtract the user's personal interest intensity representative value from the evaluation threshold, and use the difference as the judgment evaluation value;

[0043] B303, compare the judgment evaluation value with the preset judgment evaluation threshold, the specific comparison process is as follows:

[0044] If the judgment evaluation value is greater than or equal to the judgment evaluation threshold, the judgment model prediction is inaccurate;

[0045] If the judgment evaluation value is less than the judgment evaluation threshold, the judgment model prediction is accurate.

[0046] The present invention further illustrates that in step three, the performance index of the model is obtained based on the feedback information and analysis and processing, and based on the fluctuation trend of the performance index, it is identified whether the fluctuation trend affects the model performance index. The specific process is as follows:

[0047] C1, analyze and process the feedback information to obtain the representative value of the model performance index;

[0048] C2, based on the representative value of the model performance index within the period, obtains the representative value of the model performance index fluctuation, and determines whether the fluctuation trend affects the model performance index.

[0049] The present invention further illustrates that in C1, by calculating the accuracy of the model performance, the accuracy value is used as the representative value of the model performance index. The specific process is as follows:

[0050] C101, obtain the frequency of user complaint rates within a period, sum and average the frequency of user complaint rates, and use the average as a representative value of the user complaint rate within the period;

[0051] C102, by the formula: The precision value F is obtained, M represents the representative value of the user complaint rate within the period, N represents the evaluation threshold, and θ represents the preset proportional coefficient.

[0052] The present invention further illustrates that: in C2, by statistically recording the representative values ​​of the model performance indicators in each cycle, a two-dimensional coordinate system is established, and the representative values ​​of the model performance indicators in each cycle are punctuated and connected to obtain a model performance indicator fluctuation curve diagram;

[0053] Based on the representative values ​​of the model performance indicators of two adjacent periods on the model performance indicator fluctuation curve, the representative values ​​of the model performance indicators of every two adjacent periods are subtracted, and the difference is processed to obtain the square difference value, and the square difference value is used as the representative value of the model performance indicator fluctuation;

[0054] Compare the representative value of the model performance index fluctuation with the preset fluctuation range. The specific comparison process is as follows:

[0055] If the representative value of the model performance index fluctuation exists within the preset fluctuation range, it is determined that the fluctuation trend has no impact on the model performance index;

[0056] If the representative value of the model performance index fluctuation does not exist within the preset fluctuation range, it is determined that the fluctuation trend has an impact on the model performance index.

[0057] The present invention further illustrates: a data collection and processing module collects user text data and pre-processes the text data;

[0058] By preprocessing the text data, the preprocessing steps include deduplication, missing value processing, and outlier detection and processing. The specific processing process is as follows:

[0059] A1, performing deduplication processing on the user text data to obtain deduplicated text data;

[0060] A2, based on the deduplicated text data, judge and process the outliers in the text data;

[0061] A3, detect and process missing values ​​in text data to obtain cleaned text data;

[0062] The data analysis and judgment module processes user text data and establishes a user recommendation model;

[0063] B1, processing the user's text data to obtain the user's personal interest information, the text data includes chat information and search records;

[0064] B2, train the model based on the user's personal interest information, predict the user's points of interest and make recommendations;

[0065] B3, evaluate the conclusion of the model predicting the user's interest points to determine whether the model prediction is accurate;

[0066] The self-detection module judges the performance indicators of the user recommendation model based on the feedback information of users using the user recommendation model.

[0067] The beneficial effects of the present invention are as follows:

[0068] 1. The personalized recommendation method and system based on natural language processing technology described in the present invention obtains cleaned text data by preprocessing the collected user data, and predicts and evaluates the user's points of interest based on the chat information and search records in the cleaned user text data, thereby avoiding the situation of information cocoons appearing in the process of the model predicting the user's points of interest. The evaluation of the prediction results again also helps to improve the user experience and increase the user's usage frequency and duration.

[0069] 2. The personalized recommendation method and system based on natural language processing technology described in the present invention calculates the accuracy of model performance and takes the accuracy value as the representative value of model performance index. Then, by statistically recording the representative value of model performance index in each period, a two-dimensional coordinate system is established, and the representative value of model performance index in each period is punctuated and connected to obtain a model performance index fluctuation curve. Based on the representative value of model performance index of two adjacent periods on the model performance index fluctuation curve, the representative value of model performance index of every two adjacent periods is subtracted, and the difference is processed to obtain the square difference. The square difference is used as the representative value of model performance index fluctuation. The representative value of model performance index fluctuation is compared with a preset fluctuation range to identify whether the fluctuation trend affects the model performance index. A performance warning mechanism is provided for users when using it, and it is helpful to quickly locate model performance problems, and also provides historical data for later optimization and adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The present invention will be further described below in conjunction with the accompanying drawings.

[0071] Figure 1 is a flow chart of Embodiment 1 of the present invention;

[0072] Figure 2 is a flow chart of Embodiment 2 of the present invention;

[0073] Figure 3 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0074] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0075] Embodiment 1

[0076] like Figure 1 As shown, the personalized recommendation method based on natural language processing technology described in the embodiment of the present invention includes the following steps:

[0077] Step 1: Collect user text data and pre-process the text data;

[0078] In step 1, the text data is preprocessed. The preprocessing steps include deduplication, missing value processing, and outlier detection and processing. The specific processing process is as follows:

[0079] A1, performing deduplication processing on the user text data to obtain deduplicated text data;

[0080] A2, based on the deduplicated text data, judge and process the outliers in the text data;

[0081] In A2, the deduplicated text data is converted into a data collection, a three-dimensional coordinate system is established, and the data values ​​in the data collection are converted into box-shaped graphs. The specific processing process is as follows:

[0082] A201, based on the data values ​​in the data collection, respectively obtain the minimum value, the first quartile, the median, the third quartile and the maximum value in the data collection;

[0083] A202, connect the first quartile and the third quartile to get the box outline, connect the minimum value and the maximum value so that the upper and lower edges of the box are connected, and the median is located in the center of the box;

[0084] A203, putting the data in the data collection into the three-dimensional box, and obtaining the data in the data collection that exists in the three-dimensional box and the data that does not exist in the three-dimensional box;

[0085] The data in the data collection that exist in the three-dimensional box are recorded as normal values;

[0086] The data in the data collection that does not exist in the three-dimensional box is recorded as an outlier;

[0087] A3, detect and process missing values ​​in text data to obtain cleaned text data;

[0088] In A3, based on the 3D box graph in A2, the data in the data collection that does not exist in the 3D box is processed to determine whether it is a missing value. The specific process is as follows:

[0089] A301, based on the data not existing in the three-dimensional box, the data value is subtracted from the minimum value, the first quartile, the median, the third quartile and the maximum value, the difference is summed up to get the average value, and the average value is used as the representative value for determination;

[0090] A302, compare the determination representative value with the preset determination threshold value, the specific comparison process is as follows:

[0091] If the representative value is greater than or equal to the preset threshold, the data value is determined to be a missing value;

[0092] If the representative value is less than or equal to the preset judgment threshold, the data value is judged not to be a missing value;

[0093] Step 2: Processing based on user text data to establish a user recommendation model;

[0094] In step 2, a model is built based on the user's historical behavior and interests to predict the user's future interests and evaluate the model's conclusions. The specific process is as follows:

[0095] B1, processing the user's text data to obtain the user's personal interest information, the text data includes chat information and search records;

[0096] The specific process is as follows:

[0097] B101, by recording the search record contents in the user period, the frequency of the search record in the user period is obtained;

[0098] B102, extracting keywords from the chat information and recording the number of times the keywords appear to obtain the frequency of keyword appearance;

[0099] B103, based on the frequency of search records in the user period and the frequency of keywords in the chat information, a representative value of the user's personal interest information is obtained;

[0100] The representative value E of the user's personal interest information is obtained by the formula: E = (P + L) × α, where P represents the frequency of search records in the user's period, L represents the frequency of keywords in the chat information, and α represents the preset proportional coefficient;

[0101] B2, train the model based on the user's personal interest information, predict the user's points of interest and make recommendations;

[0102] In B2, based on the representative value of the user's personal interest information, the user's personal interest areas are divided, and the representative value of the user's personal interest areas is obtained. The click rate and comment rate of the user's personal interest areas are recorded to obtain the user's personal interest intensity representative value. Based on the user's personal interest intensity representative value, the user's interest points are predicted. The specific process is as follows:

[0103] B201, dividing the representative values ​​of several user personal interest information into fields, recording the click rate of the user in each interest field, and summing up the click rate values ​​of multiple fields to obtain an average value, and using the average value as a threshold for determining the interest field;

[0104] B202, compare the click-through rate values ​​of multiple fields with the threshold for selecting the field of interest, and the comparison process is as follows:

[0105] If there is a field with a click rate greater than or equal to the threshold for determining an interest field among multiple fields, the field with the larger click rate is determined as the interest field;

[0106] If there is a field with a click-through rate value lower than the threshold for determining an interest field among multiple fields, then the field is determined to be the interest field with the smaller click-through rate value;

[0107] B203, based on the interest areas with larger click-through rate values, records the comment rate of the interest areas;

[0108] B204, by formula: Get the user's personal interest strength representative value W, R represents the comment rate, Q represents the click rate, and β represents the preset proportional coefficient;

[0109] B3, evaluate the conclusion of the model predicting the user's interest points to determine whether the model prediction is accurate. The specific process is as follows:

[0110] B301, obtaining user purchase records, matching keywords in the user purchase records with keywords in the user search records within a period, summing the user purchase record frequency with the user search record frequency within a period, and obtaining an evaluation threshold;

[0111] B302, subtract the user's personal interest intensity representative value from the evaluation threshold, and use the difference as the judgment evaluation value;

[0112] B303, compare the judgment evaluation value with the preset judgment evaluation threshold, the specific comparison process is as follows:

[0113] If the judgment evaluation value is greater than or equal to the judgment evaluation threshold, the judgment model prediction is inaccurate;

[0114] If the judgment evaluation value is less than the judgment evaluation threshold, the judgment model prediction is accurate;

[0115] The technical solution of the first embodiment of the present invention is as follows: by pre-processing the collected user data to obtain cleaned text data, based on the chat information and search records in the cleaned user text data, the user's points of interest are predicted and evaluated, so as to avoid the situation of information cocoons appearing in the process of the model predicting the user's points of interest, and once again evaluating the prediction results also helps to improve the user experience and increase the user's usage frequency and duration;

[0116] Embodiment 2

[0117] Step 3: Collect feedback information from users using the user recommendation model and judge the performance indicators of the user recommendation model;

[0118] like Figure 2 As shown, in step three, the performance indicators of the model are obtained based on the feedback information and analysis and processing. Based on the fluctuation trend of the performance indicators, it is identified whether the fluctuation trend affects the model performance indicators, and adjustments and optimizations are performed. The specific process is as follows:

[0119] C1, analyze and process the feedback information to obtain the representative value of the model performance index;

[0120] In C1, the accuracy of the model performance is calculated and the accuracy value is used as the representative value of the model performance indicator. The specific process is as follows:

[0121] C101, obtain the frequency of user complaint rates within a period, sum and average the frequency of user complaint rates, and use the average as a representative value of the user complaint rate within the period;

[0122] C102, by the formula: The precision value F is obtained, where M represents the representative value of the user complaint rate within the period, N represents the evaluation threshold, and θ represents the preset proportional coefficient;

[0123] C2, based on the representative value of the model performance index within the period, obtain the representative value of the model performance index fluctuation, and determine whether the fluctuation trend affects the model performance index;

[0124] In C2, by statistically recording the representative values ​​of the model performance indicators in each cycle, a two-dimensional coordinate system is established, and the representative values ​​of the model performance indicators in each cycle are punctuated and connected to obtain the model performance indicator fluctuation curve;

[0125] Based on the representative values ​​of the model performance indicators of two adjacent periods on the model performance indicator fluctuation curve, the representative values ​​of the model performance indicators of every two adjacent periods are subtracted, and the difference is processed to obtain the square difference value, and the square difference value is used as the representative value of the model performance indicator fluctuation;

[0126] Compare the representative value of the model performance index fluctuation with the preset fluctuation range. The specific comparison process is as follows:

[0127] If the representative value of the model performance index fluctuation exists within the preset fluctuation range, it is determined that the fluctuation trend has no impact on the model performance index;

[0128] If the representative value of the model performance index fluctuation does not exist within the preset fluctuation range, it is determined that the fluctuation trend has an impact on the model performance index;

[0129] The technical solution in the second embodiment of the present invention is as follows: by calculating the accuracy of the model performance, taking the accuracy value as the representative value of the model performance index, then by statistically recording the representative value of the model performance index in each period, establishing a two-dimensional coordinate system, punctuating and connecting the representative value of the model performance index in each period, and obtaining a model performance index fluctuation curve diagram, based on the model performance index representative value of two adjacent periods on the model performance index fluctuation curve diagram, the model performance index representative value of each two adjacent periods is subtracted, and the difference is processed to obtain a square difference value, and the square difference value is used as the representative value of the model performance index fluctuation, and the model performance index fluctuation representative value is compared with a preset fluctuation range to identify whether the fluctuation trend affects the model performance index, so as to provide a performance warning mechanism for users when using it, and help to quickly locate the problem of model performance, and also provide historical data for later optimization and adjustment;

[0130] Embodiment 3

[0131] like Figure 3 As shown in the figure, the personalized recommendation system based on natural language processing technology includes the following modules:

[0132] The data collection and processing module collects user text data and pre-processes the text data;

[0133] By preprocessing the text data, the preprocessing steps include deduplication, missing value processing, and outlier detection and processing. The specific processing process is as follows:

[0134] A1, performing deduplication processing on the user text data to obtain deduplicated text data;

[0135] A2, based on the deduplicated text data, judge and process the outliers in the text data;

[0136] A3, detect and process missing values ​​in text data to obtain cleaned text data;

[0137] The data analysis and judgment module processes user text data and establishes a user recommendation model;

[0138] B1, processing the user's text data to obtain the user's personal interest information, the text data includes chat information and search records;

[0139] B2, train the model based on the user's personal interest information, predict the user's points of interest and make recommendations;

[0140] B3, evaluate the conclusion of the model predicting the user's interest points to determine whether the model prediction is accurate;

[0141] The self-detection module judges the performance indicators of the user recommendation model based on the feedback information of users using the user recommendation model.

[0142] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A personalized recommendation method based on natural language processing technology, characterized by: The following steps are involved: Step 1: Collect user text data and pre-process the text data; In step 1, the text data is preprocessed. The preprocessing steps include deduplication, missing value processing, and outlier detection and processing. The specific processing process is as follows: A1, performing deduplication processing on the user text data to obtain deduplicated text data; A2, based on the deduplicated text data, judge and process the outliers in the text data; A3, detect and process missing values ​​in text data to obtain cleaned text data; Step 2: Process the user text data and establish a user recommendation model. The specific processing process is as follows; B1, processing the user's text data to obtain the user's personal interest information, the text data includes chat information and search records; B2, train the model based on the user's personal interest information, predict the user's points of interest and make recommendations; B3, evaluate the conclusion of the model predicting the user's interest points to determine whether the model prediction is accurate; Step 3: Collect feedback information from users using the user recommendation model and judge the performance indicators of the user recommendation model; In B2, based on the representative value of the user's personal interest information, the user's personal interest areas are divided, and the representative value of the user's personal interest areas is obtained. The click rate and comment rate of the user's personal interest areas are recorded to obtain the user's personal interest intensity representative value. Based on the user's personal interest intensity representative value, the user's interest points are predicted. The specific process is as follows: B201, dividing the representative values ​​of several user personal interest information into fields, recording the click rate of the user in each interest field, and summing up the click rate values ​​of multiple fields to obtain an average value, and using the average value as a threshold for determining the interest field; B202, compare the click-through rate values ​​of multiple fields with the threshold for selecting the field of interest, and the comparison process is as follows: If there is a field with a click rate greater than or equal to the threshold for determining an interest field among multiple fields, the field with the larger click rate is determined as the interest field; If there is a field with a click-through rate value lower than the threshold for determining an interest field among multiple fields, the field is determined to be the interest field with the smaller click-through rate value; B203, based on the interest areas with larger click-through rate values, records the comment rate of the interest areas; B204, by formula: Get the user's personal interest strength representative value W, R represents the comment rate, Q represents the click rate, and β represents the preset proportional coefficient; In B3, the specific process is as follows: B301, obtaining user purchase records, matching keywords in the user purchase records with keywords in the user search records within a period, summing the user purchase record frequency with the user search record frequency within a period, and obtaining an evaluation threshold; B302, subtract the user's personal interest intensity representative value from the evaluation threshold, and use the difference as the judgment evaluation value; B303, compare the judgment evaluation value with the preset judgment evaluation threshold, the specific comparison process is as follows: If the judgment evaluation value is greater than or equal to the judgment evaluation threshold, the judgment model prediction is inaccurate; If the judgment evaluation value is less than the judgment evaluation threshold, the judgment model prediction is accurate; In step three, the performance indicators of the model are obtained based on the feedback information and analyzed and processed. Based on the fluctuation trend of the performance indicators, it is identified whether the fluctuation trend affects the model performance indicators. The specific process is as follows: C1, analyze and process the feedback information to obtain the representative value of the model performance index; C2, based on the representative value of the model performance index within the period, obtain the representative value of the model performance index fluctuation, and determine whether the fluctuation trend affects the model performance index; In C1, the accuracy of the model performance is calculated and the accuracy value is used as the representative value of the model performance indicator. The specific process is as follows: C101, obtain the frequency of user complaint rates within a period, sum and average the frequency of user complaint rates, and use the average as a representative value of the user complaint rate within the period; C102, by the formula: The precision value F is obtained, where M represents the representative value of the user complaint rate within the period, N represents the evaluation threshold, and θ represents the preset proportional coefficient; In C2, by statistically recording the representative values ​​of the model performance indicators in each cycle, a two-dimensional coordinate system is established, and the representative values ​​of the model performance indicators in each cycle are punctuated and connected to obtain the model performance indicator fluctuation curve; Based on the representative values ​​of the model performance indicators of two adjacent periods on the model performance indicator fluctuation curve, the representative values ​​of the model performance indicators of every two adjacent periods are subtracted, and the difference is processed to obtain the square difference value, and the square difference value is used as the representative value of the model performance indicator fluctuation; Compare the representative value of the model performance index fluctuation with the preset fluctuation range. The specific comparison process is as follows: If the representative value of the model performance index fluctuation exists within the preset fluctuation range, it is determined that the fluctuation trend has no impact on the model performance index; If the representative value of the model performance index fluctuation does not exist within the preset fluctuation range, it is determined that the fluctuation trend has an impact on the model performance index.

2. The personalized recommendation method based on natural language processing technology according to claim 1, characterized in that: In A2, the deduplicated text data is converted into a data collection, a three-dimensional coordinate system is established, and the data values ​​in the data collection are converted into box-shaped graphs. The specific processing process is as follows: A201, based on the data values ​​in the data collection, respectively obtain the minimum value, the first quartile, the median, the third quartile and the maximum value in the data collection; A202, connect the first quartile and the third quartile to get the box outline, connect the minimum value and the maximum value so that the upper and lower edges of the box are connected, and the median is located in the center of the box; A203, putting the data in the data collection into the three-dimensional box, and obtaining the data in the data collection that exists in the three-dimensional box and the data that does not exist in the three-dimensional box; The data in the data collection that exist in the three-dimensional box are recorded as normal values; The data in the data set that does not exist in the three-dimensional box is recorded as an outlier.

3. The personalized recommendation method based on natural language processing technology according to claim 1, characterized in that: In A3, based on the 3D box graph in A2, the data in the data collection that does not exist in the 3D box is processed to determine whether it is a missing value. The specific process is as follows: A301, based on the data not existing in the three-dimensional box, the data value is subtracted from the minimum value, the first quartile, the median, the third quartile and the maximum value, the difference is summed up to get the average value, and the average value is used as the representative value for determination; A302, compare the determination representative value with the preset determination threshold value, the specific comparison process is as follows: If the representative value is greater than or equal to the preset threshold, the data value is determined to be a missing value; If the representative value is determined to be less than or equal to the preset determination threshold, the data value is determined to be not a missing value.

4. The personalized recommendation method based on natural language processing technology according to claim 1, characterized in that: In B1, the specific process is as follows: B101, by recording the search record contents in the user period, the frequency of the search record in the user period is obtained; B102, extracting keywords from the chat information and recording the number of times the keywords appear to obtain the frequency of keyword appearance; B103, based on the frequency of search records in the user period and the frequency of keywords in the chat information, a representative value of the user's personal interest information is obtained; The representative value E of the user's personal interest information is obtained through the formula: E=(P+L)×α, where P represents the frequency of search records within the user period, L represents the frequency of keywords appearing in chat information, and α represents the preset proportional coefficient.

5. A personalized recommendation system based on natural language processing technology, the system is based on the method according to any one of claims 1 to 4, characterized in that: include: The data collection and processing module collects user text data and pre-processes the text data; By preprocessing the text data, the preprocessing steps include deduplication, missing value processing, and outlier detection and processing. The specific processing process is as follows: A1, performing deduplication processing on the user text data to obtain deduplicated text data; A2, based on the deduplicated text data, judge and process the outliers in the text data; A3, detect and process missing values ​​in text data to obtain cleaned text data; The data analysis and judgment module processes user text data and establishes a user recommendation model; B1, processing the user's text data to obtain the user's personal interest information, the text data includes chat information and search records; B2, train the model based on the user's personal interest information, predict the user's points of interest and make recommendations; B3, evaluate the conclusion of the model predicting the user's interest points to determine whether the model prediction is accurate; The self-detection module judges the performance indicators of the user recommendation model based on the feedback information of users using the user recommendation model.

Images