A method and system for intelligent recommendation of multi-source integrated news content

Through the intelligent recommendation method of multi-source integrated news content, using user historical browsing records and TF-IDF models, the problem of insufficient comprehensive news content in the existing technology is solved, and personalized content modification of user interests is realized and expand reading scope.

CN118626731BActive Publication Date: 2024-11-15SHENZHEN CHACE NETWORK INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411117151.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-15
Estimated Expiration
2044-08-15

AI Technical Summary

Technical Problem

The existing news recommendation methods cannot adapt to customizing news content, resulting in the news content being insufficiently comprehensive and unable to meet users' diverse reading needs.

Method used

Using the intelligent recommendation method of multi-source fusion news content, build a user database by obtaining user history browsing history, filtering news collections, extracting news keywords using TF-IDF model, obtaining multi-source extended information, and integrating them into the news to be pushed. Form multi-source integrated news.

Benefits of technology

It realizes personalized content modification to user interests, expands the user's reading range, avoids information cocoons, and improves the fit and user satisfaction of news content.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118626731B_ABST
    Figure CN118626731B_ABST
Patent Text Reader

Abstract

The present invention is applicable to the field of news recommendation technology, and in particular, relates to a method and system for intelligent recommendation of multi-source integrated news content, the method comprising: obtaining user historical browsing records, and constructing a user database based on the user historical browsing records; obtaining a news collection, and screening the news collection based on the user database to screen out the news to be pushed; using the TF-IDF model to extract keywords from the news to be pushed, and obtaining news keywords of each news to be pushed, and collecting extended information based on the news keywords to obtain multi-source extended information; screening the multi-source extended information based on the user database, and integrating the screened multi-source extended information into the news to be pushed to obtain multi-source integrated news. The present invention rewrites the news to be pushed according to the extended information collected from multiple data sources to obtain multi-source integrated news, thereby realizing adaptive content modification for users, so that the news content can better fit the reading habits of users.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of news recommendation, and in particular relates to a method and system for intelligently recommending multi-source fusion news content. Background Art

[0002] Intelligent news content recommendation is a service that uses advanced algorithms and technologies to automatically filter and present the most relevant and attractive news information to users based on personalized data such as user reading history, interest preferences, and interactive behavior. This recommendation system can continuously learn user preferences and improve the accuracy of recommendations over time, thereby helping users efficiently discover content of interest, while also effectively improving user engagement and satisfaction on media platforms.

[0003] Existing news recommendation methods mainly screen news based on user portraits and directly push corresponding news to users. The content of the news cannot be adaptively customized and the news content is not comprehensive enough. Summary of the invention

[0004] The purpose of the present invention is to provide a multi-source fusion news content intelligent recommendation method, aiming to solve the problem that the corresponding news is directly pushed to users, the content of the news cannot be adaptively customized, and the news content is not comprehensive enough.

[0005] The present invention is implemented as follows: a method for intelligently recommending multi-source fusion news content, the method comprising:

[0006] Acquire user browsing history records, and build a user database based on the user browsing history records, wherein the user database includes a preferred keyword library, a neutral keyword library, and an excluded keyword library;

[0007] Acquire a news collection, filter the news collection based on a user database, and filter out news to be pushed, where the news to be pushed includes at least preferred news and unclassified news;

[0008] The TF-IDF model is used to extract keywords from the news to be pushed, and the news keywords of each news to be pushed are obtained. Extended information is collected based on the news keywords to obtain multi-source extended information.

[0009] The multi-source extended information is screened based on the user database, and the screened multi-source extended information is integrated into the news to be pushed to obtain the multi-source integrated news.

[0010] Preferably, the step of obtaining a news collection, screening the news collection based on a user database to screen out news to be pushed, wherein the news to be pushed includes at least preferred news and unclassified news, specifically includes:

[0011] Determine the preference contribution of each keyword in the user database based on the frequency of keyword occurrence, and classify news that does not contain keywords in the preference keyword library and the exclusion keyword library as unclassified news;

[0012] Read each news item in the news collection word by word to determine the comprehensive contribution dot matrix corresponding to the news item;

[0013] A contribution curve is constructed based on the comprehensive contribution dot matrix, and the slope of each point on the contribution curve is calculated. The news collection is classified based on the slope to obtain preferred news. The news whose slope on the contribution curve is lower than the preset slope and whose proportion is less than the preset value is preferred news. Unclassified news and preferred news are both news to be pushed.

[0014] Preferably, the step of extracting keywords from the news to be pushed using the TF-IDF model to obtain news keywords of each news to be pushed, collecting extended information based on the news keywords, and obtaining multi-source extended information specifically includes:

[0015] Build a TF-IDF model, import the news to be pushed into the TF-IDF model one by one, and extract the news keywords corresponding to each news to be pushed;

[0016] Divide news keywords based on the components of news to obtain multiple groups of component keywords;

[0017] Based on the element keywords, an online query is performed to obtain multi-source extended information, and each group of element keywords corresponds to a group of multi-source extended information.

[0018] Preferably, the step of screening the multi-source extended information based on the user database and integrating the screened multi-source extended information into the news to be pushed to obtain the multi-source integrated news specifically includes:

[0019] Extract keywords from multi-source extended information to obtain extended information keywords;

[0020] Calculate the extension contribution of the multi-source extended information corresponding to each element according to the extended information keywords, and screen the multi-source extended information based on the extension contribution to obtain the multi-source information to be fused;

[0021] Identify each keyword in the news to be pushed, add the corresponding multi-source information to be integrated to the corresponding position, and obtain multi-source integrated news.

[0022] Preferably, in the step of integrating the filtered multi-source extended information into the news to be pushed, statistics are taken on the user's historical browsing records, the time values ​​of the users browsing different news elements in the news are counted, and a browsing time coordinate is constructed, wherein the horizontal axis of the browsing time coordinate is the number of times the news is read, and the vertical axis is the time value of browsing a certain news element in the news, and an element browsing prediction function is constructed based on the browsing time coordinate, and a predicted browsing time of each news element is generated according to the browsing prediction function, and the ratio of the multi-source extended information corresponding to each news element to be integrated into the news to be pushed is determined based on the ratio of the predicted browsing time, and the multi-source extended information is integrated.

[0023] Another object of the present invention is to provide a multi-source fusion news content intelligent recommendation system, the system comprising:

[0024] A user behavior analysis module, used to obtain user browsing history records and build a user database based on the user browsing history records, wherein the user database includes a preferred keyword library, a neutral keyword library and an exclusion keyword library;

[0025] A news classification module is used to obtain a news collection, filter the news collection based on the user database, and filter out news to be pushed, wherein the news to be pushed includes at least preferred news and unclassified news;

[0026] The information extension module is used to extract keywords from the news to be pushed using the TF-IDF model, obtain news keywords of each news to be pushed, collect extended information based on the news keywords, and obtain multi-source extended information;

[0027] The news fusion module is used to filter the multi-source extended information based on the user database, and fuse the filtered multi-source extended information into the news to be pushed to obtain multi-source fused news.

[0028] Preferably, the news classification module includes:

[0029] A keyword classification unit, used to determine the preference contribution of each keyword in the user database based on the frequency of occurrence of the keyword, and classify news that does not contain keywords in the preference keyword library and the exclusion keyword library as unclassified news;

[0030] A contribution analysis unit is used to read each news in the news collection word by word and determine the comprehensive contribution dot matrix corresponding to the news;

[0031] The news classification unit is used to construct a contribution curve based on the comprehensive contribution dot matrix, calculate the slope of each point on the contribution curve, classify the news collection based on the slope, and obtain preferred news. The news whose slope on the contribution curve is lower than the preset slope and the proportion of the curve is less than the preset value is the preferred news. Unclassified news and preferred news are both news to be pushed.

[0032] Preferably, the information expansion module includes:

[0033] The data extraction unit is used to construct a TF-IDF model, import the news to be pushed into the TF-IDF model one by one, and extract the news keywords corresponding to each news to be pushed;

[0034] An element classification unit is used to classify news keywords based on the constituent elements of the news to obtain multiple groups of element keywords;

[0035] The information query unit is used to perform online query based on element keywords to obtain multi-source extended information, and each group of element keywords corresponds to a group of multi-source extended information.

[0036] Preferably, the news fusion module includes:

[0037] An extended data extraction unit, used to extract keywords from multi-source extended information to obtain extended information keywords;

[0038] An information screening unit, used to calculate the extended contribution of the multi-source extended information corresponding to each element according to the extended information keyword, and screen the multi-source extended information based on the extended contribution to obtain the multi-source information to be fused;

[0039] The multi-source fusion unit is used to identify each keyword in the news to be pushed, add the corresponding multi-source information to be fused to the corresponding position, and obtain the multi-source fused news.

[0040] Preferably, in the step of integrating the filtered multi-source extended information into the news to be pushed, statistics are taken on the user's historical browsing records, the time values ​​of the users browsing different news elements in the news are counted, and a browsing time coordinate is constructed, wherein the horizontal axis of the browsing time coordinate is the number of times the news is read, and the vertical axis is the time value of browsing a certain news element in the news, and an element browsing prediction function is constructed based on the browsing time coordinate, and a predicted browsing time of each news element is generated according to the browsing prediction function, and the ratio of the multi-source extended information corresponding to each news element to be integrated into the news to be pushed is determined based on the ratio of the predicted browsing time, and the multi-source extended information is integrated.

[0041] The multi-source integrated news content intelligent recommendation method provided by the present invention analyzes the news to be pushed, thereby obtaining information related to the news through online query, and rewrites the news to be pushed according to the user's preferences and the extended information collected from multiple data sources to obtain multi-source integrated news, thereby realizing adaptive content modification for users and making the news content more in line with the reading habits of users. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1A flowchart of a multi-source fusion news content intelligent recommendation method provided by an embodiment of the present invention;

[0043] Figure 2 A flowchart of the steps of obtaining a news collection, filtering the news collection based on a user database, and filtering out news to be pushed provided in an embodiment of the present invention;

[0044] Figure 3 A flowchart of the steps of using the TF-IDF model to extract keywords from news to be pushed, obtaining news keywords for each news to be pushed, collecting extended information based on the news keywords, and obtaining multi-source extended information provided by an embodiment of the present invention;

[0045] Figure 4 A flowchart of the steps of filtering multi-source extended information based on a user database, integrating the filtered multi-source extended information into the news to be pushed, and obtaining multi-source integrated news provided by an embodiment of the present invention;

[0046] Figure 5 An architecture diagram of a multi-source fusion news content intelligent recommendation system provided by an embodiment of the present invention;

[0047] Figure 6 An architecture diagram of a news classification module provided by an embodiment of the present invention;

[0048] Figure 7 An architecture diagram of an information expansion module provided by an embodiment of the present invention;

[0049] Figure 8 An architectural diagram of a news fusion module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0051] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, without departing from the scope of this application, a first xx script may be referred to as a second xx script, and similarly, a second xx script may be referred to as a first xx script.

[0052] like Figure 1 FIG. 1 is a flowchart of a method for intelligently recommending multi-source integrated news content provided by an embodiment of the present invention, wherein the method comprises:

[0053] S100, obtaining a user's historical browsing records, and constructing a user database based on the user's historical browsing records, wherein the user database includes a preferred keyword library, a neutral keyword library, and an excluded keyword library.

[0054] In this step, the user's historical browsing records are obtained, the news data that the user has browsed is read, and the TF-IDF model is used to extract the keywords contained in the news that the user has browsed. In the user's historical browsing records, the user's interest in the news is determined according to the user's browsing time. Based on the preset reading speed, the standard reading time for each group of news is determined, such as 250 words per minute. The standard reading time is determined according to the number of words contained in the news. For example, if the number of words contained in the news is 500 words, the standard reading time is two minutes. Whether the user is interested is determined based on the ratio of the user's actual reading time to the standard reading time. Specifically, different ratio ranges are set, so as to determine the user's interest in the news according to the range in which the ratio of the actual reading time to the standard reading time falls. Whether news is preferred or rejected, determine the types of news that users have read, including news of interest and news of rejection, assign values ​​to the two different types of news, set the measurement value of keywords appearing in news of interest to 1, and set the measurement value of keywords appearing in news of rejection to -1, and perform statistics on the measurement value of each keyword. For example, if keyword A appears 10 times and 5 times in news of interest and news of rejection respectively, the sum of the measurement values ​​is 10+ (-5) = 5. Based on this, determine the sum of the measurement values ​​of each keyword, and divide each keyword according to the preset measurement value range, and divide it into preferred keywords, neutral keywords and rejected keywords, thereby constructing a preferred keyword library, a neutral keyword library and a rejected keyword library.

[0055] S200, obtaining a news collection, filtering the news collection based on a user database, filtering out news to be pushed, wherein the news to be pushed includes at least preferred news and unclassified news.

[0056] In this step, a news collection is obtained. When it is necessary to push news, all the news generated at this moment are retrieved, that is, the above news can all be pushed, which is a news collection. The news collection contains a large amount of real-time news. In order to cater to the user's reading preferences, each news in the news collection is analyzed based on the user database to determine the degree of user interest, so as to select the news that the user is interested in, which is the preferred news. In the existing push process, recommendations are only made based on the historical records of interest that the user has browsed. Therefore, the types of news pushed are exactly the same, which can easily cause reading fatigue to the user, thereby causing aversion to this type of news. However, the pushed news are all generated based on historical records, and the scope of new news cannot be expanded. In the present invention, a preset proportion of unclassified news is selected from the news collection. Unclassified news is a type of news that the user has not been exposed to before, or the user has not expressed his like or rejection. It is also pushed as news to be pushed, thereby actively expanding the user's news reading type, avoiding the user from entering the information cocoon, and improving the user's reading range.

[0057] S300, using the TF-IDF model to extract keywords from the news to be pushed, obtaining news keywords of each news to be pushed, collecting extended information based on the news keywords, and obtaining multi-source extended information.

[0058] In this step, the TF-IDF model is used to extract keywords from the news to be pushed. The TF-IDF model is used to evaluate the importance of a word to a document set or a single document in a corpus. The TF-IDF model can be used to extract keywords contained in each news to be pushed, and avoid the influence of common keywords, so as to expand information based on the extracted keywords, that is, to collect supplementary information based on various elements of the news. News elements can be divided into people involved in the news, event types, time of the event, location of the event, cause of the event and course of the event. Information collection is carried out from the above six aspects, including the collection of person information, introduction of event types and online query of similar events. In addition, for the same event, there will be multiple media reporting on it, so multiple news versions corresponding to the same event can be collected, that is, multi-source extended information is obtained.

[0059] S400, filtering the multi-source extended information based on the user database, and integrating the filtered multi-source extended information into the news to be pushed to obtain the multi-source integrated news.

[0060] In this step, the multi-source extended information is screened based on the user database. Similarly, after obtaining the multi-source extended information, the multi-source extended information obtained by the online query is screened again using the user database to extract the part that the user is interested in, and then the part that the user is interested in is added to the news to be pushed to obtain multi-source integrated news, that is, the multi-source integrated news contains extended content, such as an introduction to news figures, an introduction to event types, an introduction to event locations, an introduction to the relationship between news figures, and the connection between events, etc. The obtained multi-source integrated news contains both topics that the user is interested in and topics that the user has never touched upon. Therefore, it can not only expand the user's reading range, but also modify the content of the news based on the user's interests.

[0061] like Figure 2 As shown, as a preferred embodiment of the present invention, the step of obtaining a news collection, filtering the news collection based on a user database, filtering out news to be pushed, wherein the news to be pushed includes at least preferred news and unclassified news, specifically includes:

[0062] S201, determining the preference contribution of each keyword in the user database based on the frequency of occurrence of the keyword, and classifying news that does not contain the keywords in the preference keyword library and the exclusion keyword library as unclassified news.

[0063] In this step, the preference contribution of each keyword in the user database is determined based on the frequency of keyword appearance. In the process of building the user database, the number of times each keyword in the user database appears is counted. Then in the preference keyword library, the number of occurrences Pi corresponding to the i-th keyword is known, and the proportion of keyword appearance frequency P=Pi / P0 is calculated, where P0 is the total number of times all keywords appear, and P is used as the preference contribution of the keyword. Similarly, the keywords in the exclusion keyword library are also processed in the same way, with -P as the preference contribution value of each keyword in the exclusion keyword library. News that does not contain keywords in the preference keyword library and the exclusion keyword library is classified as unclassified news. This type of news is a type of news that users have never touched upon, and is used to expand the scope of active reading and further clarify user portraits.

[0064] S202, reading each piece of news in the news collection word by word, and determining a comprehensive contribution dot matrix corresponding to the piece of news.

[0065] In this step, each news item in the news collection is read word by word, each word in the news collection is numbered consecutively, and is read consecutively in the order of the numbers. The sum of the preference contributions of the keywords contained in the read words is determined to obtain a comprehensive contribution, and a two-dimensional coordinate system is constructed, with the word number as the horizontal axis and the comprehensive contribution when the word is read as the vertical axis. Multiple points are marked in the two-dimensional coordinate system to obtain a comprehensive contribution dot matrix based on the two-dimensional coordinate system.

[0066] S203, constructing a contribution curve based on the comprehensive contribution dot matrix, calculating the slope of each point on the contribution curve, classifying the news collection based on the slope, and obtaining preferred news. The news whose slope on the contribution curve is lower than the preset slope and whose proportion is less than the preset value is the preferred news. Both unclassified news and preferred news are news to be pushed.

[0067] In this step, a contribution curve is constructed based on the comprehensive contribution dot matrix, and each point in the comprehensive contribution dot matrix is ​​connected with a smooth curve to obtain a continuous curve. The slope of each point on the contribution curve is calculated. When the slope is a negative value, it means that the user's reading interest at this position is decreasing. Therefore, the news in which the user's reading interest is gradually decreasing is screened out and discarded to obtain preliminary screened news. Subsequently, the average value of the contribution curve of the preliminary screened news is calculated. If the average value is higher than the set value, it is classified as preferred news.

[0068] like Figure 3 As shown, as a preferred embodiment of the present invention, the steps of extracting keywords of the news to be pushed using the TF-IDF model to obtain news keywords of each news to be pushed, collecting extended information based on the news keywords, and obtaining multi-source extended information specifically include:

[0069] S301, constructing a TF-IDF model, importing the news to be pushed into the TF-IDF model one by one, and extracting news keywords corresponding to each news to be pushed.

[0070] In this step, a TF-IDF model is constructed. TF-IDF consists of two parts: term frequency (TF) measures the frequency of a word appearing in a document; inverse document frequency (IDF) reduces the weight of common words by dividing the total number of documents by the number of documents containing the word. By multiplying these two values, TF-IDF can effectively highlight the key information in the document while reducing the influence of common words. The TF-IDF model can accurately extract the corresponding news keywords from each news to be pushed.

[0071] S302, dividing news keywords based on constituent elements of the news to obtain multiple groups of element keywords.

[0072] In this step, news keywords are divided based on the constituent elements of news. The constituent elements of news are generally divided into six parts. Dividing the news keywords into the above six parts will result in six groups of element keywords, such as the element keywords corresponding to news figures, the element keywords before and after the news event, etc.

[0073] S303, performing online query based on the element keywords to obtain multi-source extended information, where each group of element keywords corresponds to a group of multi-source extended information.

[0074] In this step, an online query is performed based on the element keywords. By performing the online query, a large amount of information related to the above-mentioned element keywords can be obtained from the network, such as news released by various official media, information content released by social media, and information published on websites, etc., to obtain multi-source extended information. Each group of element keywords corresponds to a group of multi-source extended information.

[0075] like Figure 4 As shown, as a preferred embodiment of the present invention, the step of filtering multi-source extended information based on the user database, integrating the filtered multi-source extended information into the news to be pushed, and obtaining multi-source integrated news specifically includes:

[0076] S401, extract keywords from multi-source extended information to obtain extended information keywords.

[0077] In this step, keyword extraction is performed on the multi-source extended information. Based on the same extraction method, the same TF-IDF model and the same extraction rules are used to extract keywords. Then, each multi-source extended information corresponds to a set of extended information keywords.

[0078] S402, calculating the extension contribution of the multi-source extended information corresponding to each element according to the extended information keyword, and screening the multi-source extended information based on the extension contribution to obtain the multi-source information to be fused.

[0079] In this step, the extended contribution of the multi-source extended information corresponding to each element is calculated based on the extended information keywords. For example, if the multi-source extended information is the character relationship introduction information of the event characters in the news, after keyword extraction, the above character relationship introduction information can be characterized by keywords. According to the user database, it can be determined which keyword library the above extended information keywords belong to. Similarly, based on the same method as the comprehensive contribution, the extended contribution is calculated. The extended contribution can indicate whether the user is interested in the current multi-source extended information. The multi-source extended information is screened based on the preset extended contribution threshold to obtain the multi-source information to be integrated.

[0080] S403, identifying each keyword in the news to be pushed, adding the corresponding multi-source information to be integrated to the corresponding position, and obtaining the multi-source integrated news.

[0081] In this step, we identify the keywords in the news to be pushed, count the user's historical browsing records, and count the time values ​​of users browsing different news elements in the news. For example, when a user reads the first news, it takes 45 seconds to read the information related to the event and person. When the user reads the second news, it takes 73 seconds to read the information related to the event and person. When the user reads the third news, it takes 89 seconds to read the information related to the event and person. We then construct a browsing time coordinate, where the horizontal axis of the browsing time coordinate is the number of times the news is read, and the vertical axis is the time value of browsing a certain news element in the news, i.e. (1, 45), (2, 73), (3, 89). Based on the browsing time coordinate, we construct an element browsing prediction function, and obtain the fitting function by fitting:

[0082] The predicted browsing time of each news element is generated according to the browsing prediction function. There are six groups of news elements, that is, there are six groups of fitting functions. The independent variables of the six groups of fitting functions are the number of news read by users. By bringing the number of this news into the fitting function, the reading time ratio of the user for the six groups of news elements in this reading process can be predicted, and six groups of time calculation values ​​are obtained, namely F1, F2, F3, F4, F5 and F6. Based on the proportion of predicted browsing time, the proportion of multi-source extended information corresponding to each news element integrated into the news to be pushed is determined. The proportion of multi-source extended information corresponding to each news element integrated into the news to be pushed is Pi=Fi / (F1+F2+F3+F4+F5+F6), Fi is the i-th time calculation value, according to the corresponding integration ratio Pi of each news element, the corresponding number of words of text are selected from the multi-source information to be integrated and added to the corresponding position to obtain multi-source integrated news.

[0083] like Figure 5 As shown, 6. A multi-source integrated news content intelligent recommendation system provided by an embodiment of the present invention, the system includes:

[0084] The user behavior analysis module 100 is used to obtain the user's historical browsing records and build a user database based on the user's historical browsing records. The user database includes a preference keyword library, a neutral keyword library and an exclusion keyword library.

[0085] In this system, the user behavior analysis module 100 obtains the user's historical browsing records, reads the news data that the user has browsed, and uses the TF-IDF model to extract the keywords contained in the news that the user has browsed. In the user's historical browsing records, the user's interest in the news is determined according to the user's browsing time. Based on the preset reading speed, the standard reading time for each group of news is determined, such as 250 words per minute. The standard reading time is determined according to the number of words contained in the news. For example, if the number of words contained in the news is 500 words, the standard reading time is two minutes. Whether the user is interested is determined according to the ratio of the user's actual reading time to the standard reading time. Specifically, different ratio ranges are set, so as to determine the range within which the ratio of the actual reading time to the standard reading time falls. Determine whether the user prefers or rejects the news, determine the types of news the user has read, including news of interest and news of rejection, assign values ​​to the two different types of news, set the measurement value of the keyword appearing in the news of interest to 1, and set the measurement value of the keyword appearing in the rejected news to -1, and perform statistics on the measurement value of each keyword. For example, if keyword A appears 10 times and 5 times in the news of interest and the rejected news respectively, the sum of the measurement values ​​is 10+(-5)=5. Based on this, the sum of the measurement values ​​of each keyword is determined, and each keyword is divided according to the preset measurement value range, and is divided into preferred keywords, neutral keywords and rejected keywords, thereby constructing a preferred keyword library, a neutral keyword library and a rejected keyword library.

[0086] The news classification module 200 is used to obtain a news collection, filter the news collection based on the user database, and filter out news to be pushed, where the news to be pushed includes at least preferred news and unclassified news.

[0087] In this system, the news classification module 200 obtains a news collection. When it is necessary to push news, all the news generated at this moment is retrieved, that is, the above news can be pushed, that is, a news collection. The news collection contains a large amount of real-time news. In order to cater to the user's reading preferences, each news in the news collection is analyzed based on the user database to determine the degree of user interest, so as to select the news that the user is interested in, that is, the preferred news. In the existing push process, only recommendations are made based on the historical records of interest that the user has browsed. Therefore, the types of news pushed are exactly the same, which can easily cause reading fatigue to the user, thereby causing aversion to such news. However, the pushed news are all generated based on historical records, and the new news range cannot be expanded. In the present invention, a preset proportion of unclassified news is selected from the news collection. The unclassified news is a type of news that the user has not been exposed to before, or the user has not expressed his likes or rejections. It is also pushed as news to be pushed, thereby actively expanding the user's news reading type, avoiding the user from entering the information cocoon, and improving the user's reading range.

[0088] The information extension module 300 is used to extract keywords from the news to be pushed using the TF-IDF model to obtain news keywords of each news to be pushed, collect extended information based on the news keywords, and obtain multi-source extended information.

[0089] In this system, the information extension module 300 uses the TF-IDF model to extract keywords from the news to be pushed. The TF-IDF model is used to evaluate the importance of a word to a document set or a single document in a corpus. The TF-IDF model can be used to extract keywords contained in each news to be pushed, and avoid the influence of common keywords, so as to expand information based on the extracted keywords, that is, to supplement and collect information based on various elements of the news. News elements can be divided into people involved in the news, event types, time of the event, location of the event, cause of the event and course of the event. Information collection is performed from the above six aspects, including the collection of person information, introduction of event types and online query of information such as similar events. In addition, for the same event, there will be multiple media reporting on it, so multiple news versions corresponding to the same event can be collected, that is, multi-source extended information can be obtained.

[0090] The news fusion module 400 is used to filter the multi-source extended information based on the user database, and fuse the filtered multi-source extended information into the news to be pushed to obtain multi-source fused news.

[0091] In this system, the news fusion module 400 filters the multi-source extended information based on the user database. Similarly, after obtaining the multi-source extended information, the user database is used to filter the multi-source extended information obtained by the online query again to extract the part that the user is interested in, and then add the part that the user is interested in to the news to be pushed to obtain multi-source fusion news, that is, the multi-source fusion news contains extended content, such as an introduction to the news figures, an introduction to the event type, an introduction to the event location, an introduction to the relationship between the news figures, and the connection between the events, etc. The obtained multi-source fusion news contains both topics that the user is interested in and topics that the user has never touched upon. Therefore, it can not only expand the user's reading range, but also modify the content of the news based on the user's interests.

[0092] like Figure 6 As shown, as a preferred embodiment of the present invention, the news classification module 200 includes:

[0093] The keyword classification unit 201 is used to determine the preference contribution of each keyword in the user database based on the frequency of occurrence of the keyword, and classify the news that does not contain the keywords in the preference keyword library and the exclusion keyword library as unclassified news.

[0094] In this module, the keyword classification unit 201 determines the preference contribution of each keyword in the user database based on the frequency of keyword appearance. In the process of constructing the user database, the number of times each keyword in the user database appears is counted. Then in the preference keyword library, the number of occurrences Pi corresponding to the i-th keyword is known, and the proportion of keyword appearance frequency P=Pi / P0 is calculated, where P0 is the total number of times all keywords appear, and P is used as the preference contribution of the keyword. Similarly, the keywords in the exclusion keyword library are also processed in the same way, with -P as the preference contribution value of each keyword in the exclusion keyword library. News that does not contain keywords in the preference keyword library and the exclusion keyword library is classified as unclassified news. This type of news is a type of news that users have never touched upon, and is used to actively expand the scope of reading and further clarify user portraits.

[0095] The contribution analysis unit 202 is used to read each news in the news collection word by word to determine the comprehensive contribution dot matrix corresponding to the news.

[0096] In this module, the contribution analysis unit 202 reads each news in the news collection word by word, consecutively numbers each word in the news collection, reads continuously in the order of numbers, determines the sum of the preference contributions of the keywords contained in the read words, obtains the comprehensive contribution, constructs a two-dimensional coordinate system, uses the word number as the horizontal axis, and uses the comprehensive contribution when the word is read as the vertical axis, marks multiple points in the two-dimensional coordinate system, and obtains a comprehensive contribution dot matrix based on the two-dimensional coordinate system.

[0097] The news classification unit 203 is used to construct a contribution curve based on the comprehensive contribution dot matrix, calculate the slope of each point on the contribution curve, classify the news collection based on the slope, and obtain preferred news. The news whose slope on the contribution curve is lower than the preset slope and the proportion of the curve is less than the preset value is the preferred news. Unclassified news and preferred news are both news to be pushed.

[0098] In this module, the news classification unit 203 constructs a contribution curve based on the comprehensive contribution dot matrix, connects each point in the comprehensive contribution dot matrix with a smooth curve to obtain a continuous curve, and calculates the slope of each point on the contribution curve. When the slope is a negative value, it means that the user's reading interest at this position is decreasing. Therefore, the news in which the user's reading interest is gradually decreasing is screened out and discarded to obtain preliminary screened news. Subsequently, the average value of the contribution curve of the preliminary screened news is calculated. If the average value is higher than the set value, it is classified as preferred news.

[0099] like Figure 7 As shown, as a preferred embodiment of the present invention, the information expansion module 300 includes:

[0100] The data extraction unit 301 is used to construct a TF-IDF model, import the news to be pushed into the TF-IDF model one by one, and extract news keywords corresponding to each news to be pushed.

[0101] In this module, the data extraction unit 301 constructs a TF-IDF model. TF-IDF consists of two parts: the term frequency (TF) measures the frequency of a word appearing in a document; the inverse document frequency (IDF) reduces the weight of common words by dividing the total number of documents by the number of documents containing the word. By multiplying these two values, TF-IDF can effectively highlight the key information in the document while reducing the influence of common words. The TF-IDF model can accurately extract the corresponding news keywords from each news to be pushed.

[0102] The element classification unit 302 is used to classify news keywords based on the constituent elements of the news to obtain multiple groups of element keywords.

[0103] In this module, the element classification unit 302 divides news keywords based on the constituent elements of news. The constituent elements of news are generally divided into six parts. The news keywords are divided into the above six parts, that is, six groups of element keywords are obtained, such as the element keywords corresponding to news figures, the element keywords before and after the news event, etc.

[0104] The information query unit 303 is used to perform online query based on the element keywords to obtain multi-source extended information, where each group of element keywords corresponds to a group of multi-source extended information.

[0105] In this module, the information query unit 303 performs an online query based on the element keywords. By performing the online query, a large amount of information related to the above-mentioned element keywords can be obtained from the network, such as news released by various official media, information content released by social media, and information published on websites, etc., to obtain multi-source extended information, and each group of element keywords corresponds to a group of multi-source extended information.

[0106] like Figure 8 As shown, as a preferred embodiment of the present invention, the news fusion module 400 includes:

[0107] The extended data extraction unit 401 is used to extract keywords from multi-source extended information to obtain extended information keywords.

[0108] In this module, the extended data extraction unit 401 performs keyword extraction on multi-source extended information. Based on the same extraction method, the same TF-IDF model and the same extraction rules are used to extract keywords. Then, each multi-source extended information corresponds to a set of extended information keywords.

[0109] The information screening unit 402 is used to calculate the extended contribution of the multi-source extended information corresponding to each element according to the extended information keyword, and screen the multi-source extended information based on the extended contribution to obtain the multi-source information to be fused.

[0110] In this module, the information screening unit 402 calculates the extended contribution of the multi-source extended information corresponding to each element according to the extended information keywords. For example, if the multi-source extended information is the character relationship introduction information of the event characters in the news, after keyword extraction, the above character relationship introduction information can be characterized by keywords. According to the user database, it can be determined which keyword library the above extended information keywords belong to. Similarly, based on the same method as the comprehensive contribution, the extended contribution is calculated. The extended contribution can indicate whether the user is interested in the current multi-source extended information. The multi-source extended information is screened based on the preset extended contribution threshold to obtain the multi-source information to be fused.

[0111] The multi-source fusion unit 403 is used to identify each keyword in the news to be pushed, add the corresponding multi-source information to be fused to the corresponding position, and obtain the multi-source fused news.

[0112] In this module, the multi-source fusion unit 403 identifies each keyword in the news to be pushed, counts the user's historical browsing records, and counts the time values ​​of users browsing different news elements in the news. For example, when the user reads the first news, the time for reading the relevant information of the event and the person is 45 seconds, when the user reads the second news, the time for reading the relevant information of the event and the person is 73 seconds, and when the user reads the third news, the time for reading the relevant information of the event and the person is 89 seconds. The browsing time coordinate is constructed, and the horizontal axis of the browsing time coordinate is the number of times the news is read, and the vertical axis is the time value of browsing a certain news element in the news, that is, (1, 45), (2, 73), (3, 89). The element browsing prediction function is constructed based on the browsing time coordinate, and the fitting function is obtained by fitting:

[0113] The predicted browsing time of each news element is generated according to the browsing prediction function. There are six groups of news elements, that is, there are six groups of fitting functions. The independent variables of the six groups of fitting functions are the number of news read by users. By bringing the number of this news into the fitting function, the reading time ratio of the user for the six groups of news elements in this reading process can be predicted, and six groups of time calculation values ​​are obtained, namely F1, F2, F3, F4, F5 and F6. Based on the proportion of predicted browsing time, the proportion of multi-source extended information corresponding to each news element integrated into the news to be pushed is determined. The proportion of multi-source extended information corresponding to each news element integrated into the news to be pushed is Pi=Fi / (F1+F2+F3+F4+F5+F6), Fi is the i-th time calculation value, according to the corresponding integration ratio Pi of each news element, the corresponding number of words of text are selected from the multi-source information to be integrated and added to the corresponding position to obtain multi-source integrated news.

[0114] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0115] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for intelligent recommendation of multi-source fusion news content, characterized in that: The method comprises: Acquire user browsing history records, and build a user database based on the user browsing history records, wherein the user database includes a preferred keyword library, a neutral keyword library, and an excluded keyword library; Acquire a news collection, filter the news collection based on a user database, and filter out news to be pushed, where the news to be pushed includes at least preferred news and unclassified news; The TF-IDF model is used to extract keywords from the news to be pushed, and the news keywords of each news to be pushed are obtained. Extended information is collected based on the news keywords to obtain multi-source extended information. The multi-source extended information is screened based on the user database, and the screened multi-source extended information is integrated into the news to be pushed to obtain the multi-source integrated news; The steps of obtaining a news collection, filtering the news collection based on a user database, and filtering out news to be pushed, wherein the news to be pushed includes at least preferred news and unclassified news, specifically include: Determine the preference contribution of each keyword in the user database based on the frequency of keyword occurrence, and classify news that does not contain keywords in the preference keyword library and the exclusion keyword library as unclassified news; Read each news item in the news collection word by word to determine the comprehensive contribution dot matrix corresponding to the news item; Construct a contribution curve based on the comprehensive contribution dot matrix, calculate the slope of each point on the contribution curve, classify the news collection based on the slope, and obtain preferred news. The news whose slope on the contribution curve is lower than the preset slope and whose proportion is less than the preset value is the preferred news. Unclassified news and preferred news are both news to be pushed; The step of filtering the multi-source extended information based on the user database and integrating the filtered multi-source extended information into the news to be pushed to obtain the multi-source integrated news specifically includes: Extract keywords from multi-source extended information to obtain extended information keywords; Calculate the extension contribution of the multi-source extended information corresponding to each element according to the extended information keywords, and screen the multi-source extended information based on the extension contribution to obtain the multi-source information to be fused; Identify each keyword in the news to be pushed, add the corresponding multi-source information to be integrated to the corresponding position, and obtain multi-source integrated news; In the step of integrating the filtered multi-source extended information into the news to be pushed, statistics are collected on the user's historical browsing records, the time values ​​of the user browsing different news elements in the news are counted, and a browsing time coordinate is constructed. The horizontal axis of the browsing time coordinate is the number of times the news is read, and the vertical axis is the time value of browsing a certain news element in the news. An element browsing prediction function is constructed based on the browsing time coordinate, and a predicted browsing time of each news element is generated according to the browsing prediction function. Based on the ratio of the predicted browsing time, the ratio of the multi-source extended information corresponding to each news element to be integrated into the news to be pushed is determined, and the multi-source extended information is integrated.

2. The multi-source integrated news content intelligent recommendation method according to claim 1 is characterized in that: The steps of extracting keywords from the news to be pushed using the TF-IDF model to obtain news keywords for each news to be pushed, collecting extended information based on the news keywords, and obtaining multi-source extended information specifically include: Build a TF-IDF model, import the news to be pushed into the TF-IDF model one by one, and extract the news keywords corresponding to each news to be pushed; Divide news keywords based on the components of news to obtain multiple groups of component keywords; Based on the element keywords, an online query is performed to obtain multi-source extended information, and each group of element keywords corresponds to a group of multi-source extended information.

3. A multi-source integrated news content intelligent recommendation system, characterized in that: The system comprises: A user behavior analysis module, used to obtain user browsing history records and build a user database based on the user browsing history records, wherein the user database includes a preferred keyword library, a neutral keyword library and an exclusion keyword library; A news classification module is used to obtain a news collection, filter the news collection based on the user database, and filter out news to be pushed, wherein the news to be pushed includes at least preferred news and unclassified news; The information extension module is used to extract keywords from the news to be pushed using the TF-IDF model, obtain news keywords of each news to be pushed, collect extended information based on the news keywords, and obtain multi-source extended information; A news fusion module is used to filter multi-source extended information based on a user database, and fuse the filtered multi-source extended information into the news to be pushed to obtain multi-source fused news; The news classification module includes: A keyword classification unit, used to determine the preference contribution of each keyword in the user database based on the frequency of occurrence of the keyword, and classify news that does not contain keywords in the preference keyword library and the exclusion keyword library as unclassified news; A contribution analysis unit is used to read each news in the news collection word by word and determine the comprehensive contribution dot matrix corresponding to the news; A news classification unit is used to construct a contribution curve based on the comprehensive contribution dot matrix, calculate the slope of each point on the contribution curve, classify the news collection based on the slope, and obtain preferred news. The news whose slope on the contribution curve is lower than a preset slope and whose proportion is less than a preset value is the preferred news. Unclassified news and preferred news are both news to be pushed; The news fusion module includes: An extended data extraction unit, used to extract keywords from multi-source extended information to obtain extended information keywords; An information screening unit, used to calculate the extended contribution of the multi-source extended information corresponding to each element according to the extended information keyword, and screen the multi-source extended information based on the extended contribution to obtain the multi-source information to be fused; The multi-source fusion unit is used to identify each keyword in the news to be pushed, add the corresponding multi-source information to be fused to the corresponding position, and obtain the multi-source fused news; In the step of integrating the filtered multi-source extended information into the news to be pushed, statistics are collected on the user's historical browsing records, the time values ​​of the user browsing different news elements in the news are counted, and a browsing time coordinate is constructed. The horizontal axis of the browsing time coordinate is the number of times the news is read, and the vertical axis is the time value of browsing a certain news element in the news. An element browsing prediction function is constructed based on the browsing time coordinate, and a predicted browsing time of each news element is generated according to the browsing prediction function. Based on the ratio of the predicted browsing time, the ratio of the multi-source extended information corresponding to each news element to be integrated into the news to be pushed is determined, and the multi-source extended information is integrated.

4. The multi-source integrated news content intelligent recommendation system according to claim 3 is characterized in that: The information expansion module includes: The data extraction unit is used to build a TF-IDF model, import the news to be pushed into the TF-IDF model one by one, and extract the news keywords corresponding to each news to be pushed; An element classification unit is used to classify news keywords based on the constituent elements of the news to obtain multiple groups of element keywords; The information query unit is used to perform online query based on element keywords to obtain multi-source extended information, and each group of element keywords corresponds to a group of multi-source extended information.