A personalized news push method based on big data

Abstract

The present invention relates to the field of electronic digital data processing technology, and specifically to a personalized news push method based on big data, including: collecting news data and user data, obtaining candidate keywords from the title of the news data, combining the text of the news data, and obtaining the initial keywords of the news data from the candidate keywords; screening out comparative news data with similar content to the current news data from all historical news data according to the initial keywords; screening out the final keywords of the current news data by comparing the initial keywords of the news data, and then combining the final keywords of the current news data and the user's interest tags to push news to the user. The present invention can more accurately describe the news data by extracting keywords, and provide a news push service that better meets the user's needs according to the user's interests and needs, thereby improving the user's satisfaction and usage experience.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a personalized news push method based on big data. Background Art

[0002] In the era of modern information explosion, especially in recent years with the development of self-media, people receive a large amount of disorganized news information every day. The traditional news push method is based on the subjective selection of the editorial team and the user's click history, and only recommends based on general indicators. It cannot take into account the user's specific interests and hobbies, and will push a large amount of information that does not meet their interests to users. It is difficult to meet the personalized needs of users, causing users to feel troubled, thereby reducing the satisfaction and usage frequency of the platform.

[0003] Since news is time-sensitive, when it is difficult to classify and locate news data in a personalized manner for newly acquired news, the news cannot be accurately and timely pushed to users who like the news, thereby reducing the personalized effectiveness of news push. Summary of the invention

[0004] The present invention provides a personalized news push method based on big data to solve the existing problems.

[0005] The invention discloses a personalized news push method based on big data using the following technical solutions:

[0006] An embodiment of the present invention provides a method for personalized news push based on big data, the method comprising the following steps:

[0007] Based on the big data, a number of news data and user data are collected respectively, wherein the news data is divided into current news data and historical news data, including the title and text of the news, and the user data is the user's interest tags;

[0008] Obtain candidate keywords based on the title of the news data; obtain several sentences in the text corresponding to the news data and several phrases contained in each sentence, and construct possible core evaluations of the sentences in combination with the number of candidate keywords in each sentence in the text corresponding to the news data and the similarity between the phrases and the candidate keywords; obtain initial keywords of the news data from the candidate keywords in combination with the occurrence frequency of the phrases by analyzing the similarity between the phrases and the candidate keywords and the possible core evaluations of the sentences;

[0009] Filter out comparative news data with similar content to the current news data from all historical news data based on the initial keywords; filter out the final keywords of the current news data by comparing the initial keywords of the news data;

[0010] Combine the final keywords of the current news data and the user's interest tags to push news to the user.

[0011] Furthermore, the specific method of obtaining candidate keywords based on the title of the news data includes:

[0012] The spaCy software is used to obtain nouns and verbs in the news data titles and recorded as candidate keywords.

[0013] Furthermore, the method of obtaining several sentences in the text corresponding to the news data and several phrases contained in each sentence, and constructing a possible core evaluation of the sentence in combination with the number of candidate keywords in each sentence in the text corresponding to the news data and the similar core association between the phrases and the candidate keywords includes the following specific methods:

[0014] Firstly, the character string between any two adjacent periods in the body of the news data is recorded as a sentence, and Jieba word segmentation is used to segment all the sentences in the body of the news data to obtain several phrases;

[0015] According to the number of times the sth phrase and the candidate keyword appear together in the sentence, the relative core weight of the sentence is obtained;

[0016] Calculate the similarity core association between the phrase and the candidate keyword based on the probability that the phrase and the candidate keyword are in the same sentence in the news data and the relative core weight of the sentence;

[0017] Then, the possible core evaluations of the statements are calculated:

[0018]

[0019] Among them, A α represents the possible core evaluation of the αth sentence, N α represents the number of times the candidate keyword appears in the αth sentence, a α,s N′ represents the similarity between the sth phrase in the αth sentence and the candidate keyword. α represents the total number of phrases contained in the αth sentence, n α It represents the number of phrases excluding candidate keywords in the αth sentence.

[0020] Furthermore, the specific calculation method for calculating the similarity core association between the phrase and the candidate keyword according to the probability of the phrase and the candidate keyword in the same sentence in the news data and the relative core weight of the sentence is:

[0021]

[0022] Among them, a α,srepresents the similarity core relevance between the sth phrase in the αth sentence and the candidate keyword, m α,s represents the total number of sentences in the news data that contain the sth phrase in the αth sentence, m α,s,b represents the number of sentences in the news data that contain both the sth phrase of the αth sentence and the bth candidate keyword, Q α,s It represents the relative core weight of the αth sentence with the sth phrase appearing, max[] represents obtaining the maximum value, and softsign() represents the softsign normalization function.

[0023] Furthermore, the method for obtaining the relative core weight of a sentence according to the number of times the sth phrase and the candidate keyword appear together in the sentence is:

[0024]

[0025] Among them, Y α represents the keyword component similarity of the αth sentence, where the keyword component similarity of the αth sentence is the ratio of the number of candidate keywords appearing in the αth sentence to the number of all candidate keywords, g s,α represents the number of occurrences of the sth phrase in the αth sentence, g α Represents the number of candidate keywords in the αth sentence.

[0026] Furthermore, the method of obtaining the initial keywords of the news data from the candidate keywords by analyzing the similarity core association between the phrase and the candidate keywords, the possible core evaluation of the sentence, and combining the occurrence frequency of the phrase includes the following specific methods:

[0027] Obtain candidate key evaluations for the phrases based on the similar core associations between any phrase and the candidate keywords, the possible core evaluations of the sentences to which the phrases belong, and the number of phrases;

[0028]

[0029] Among them, G α,s represents the candidate key evaluation of the sth phrase in the αth sentence, a α,s A represents the similarity core correlation between the sth phrase in the αth sentence and the candidate keyword, α represents the possible core evaluation of the αth sentence, B α,s represents the number of occurrences of the sth phrase in the αth sentence, m α,s represents the total number of all sentences in the news data that contain the sth phrase in the αth sentence, and maxmin[] represents the maximum and minimum normalization function;

[0030] A threshold T is preset, and if the candidate key evaluation of a phrase is greater than or equal to the threshold T, the phrase is recorded as an initial keyword.

[0031] Furthermore, the method of screening out comparative news data having similar contents to the current news data from all historical news data according to the initial keywords and screening out the final keywords of the current news data by comparing the initial keywords of the news data includes the following specific steps:

[0032] respectively obtaining initial keywords of current news data and historical news data, and selecting preliminary screening news from the historical news data according to the similarity relationship between the initial keywords of the current news data and the historical news data;

[0033] The comparative news data is obtained by screening the preliminary screened news, the initial keywords in the comparative news data and the initial keywords in the current news data are intersected, and the initial keywords in the intersection are screened by a stop word list to obtain the final keywords.

[0034] Furthermore, the specific screening method for screening out preliminary screening news from historical news data is:

[0035] For any initial keyword of current news data, historical news data whose title and text contain the initial keyword of the current news data are obtained and recorded as preliminary screening news.

[0036] Furthermore, the method of obtaining the comparative news data from the preliminary screened news includes:

[0037] The preliminary screened news are sorted according to the number of initial keywords of the current news data appearing in the preliminary screened news, the frequency of each initial keyword appearing, and the candidate key evaluation of the initial keywords in the preliminary screened news, and the preliminary screened news ranked in the top B are selected and marked as comparison news data, where B is a preset parameter.

[0038] Furthermore, the method of pushing news to the user by combining the final keyword of the current news data and the user's interest tag includes the following specific methods:

[0039] Combine the final keywords of the current news data and the user's interest tags, and use the RNN neural network to push news to users;

[0040] The training method of the RNN neural network is as follows: taking the final keyword of any current news data as a sample, manually labeling the interest tag corresponding to the sample as the label of the sample, and forming a data set for training the RNN neural network by a number of labeled samples. Using the data set, according to the cross entropy loss function, the RNN neural network is trained using a stochastic gradient descent algorithm to converge, thereby obtaining a trained RNN neural network for news push.

[0041] The beneficial effect of the technical solution of the present invention is: compared with the traditional news push method, the embodiment of the present invention takes into account that the content of the news is highly summarized in the news title. Therefore, by extracting candidate keywords from the news title, it can more accurately reflect the core content and theme of the news, and further combine the similarity between the phrases in the news text and the candidate keywords, which is helpful to have a deeper understanding of the news content, further refine the key information, and accurately quantify the possible core evaluation of the possibility of reflecting the sentence as the core sentence, so as to screen out the final keywords that can better represent the current news data from the current news data, which is helpful to supplement and verify the content of the current news, improve the comprehensiveness and credibility of the information, and improve the representative ability of keywords for news data, so that keywords can more accurately describe the main points and characteristics of the news, so as to further improve the personalization of the corresponding push effect when the current news is pushed in combination with the user's interest tags. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 The present invention is a flowchart of the steps of a personalized news push method based on big data. DETAILED DESCRIPTION

[0044] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the specific implementation method, structure, features and effects of a personalized news push method based on big data proposed by the present invention in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0045] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0046] The following is a detailed description of a specific scheme of a personalized news push method based on big data provided by the present invention in conjunction with the accompanying drawings.

[0047] See also Figure 1 , which shows a flowchart of a method for personalized news push based on big data provided by an embodiment of the present invention, the method comprising the following steps:

[0048] Step S001: Collect a number of news data and user data based on big data.

[0049] It should be noted that this embodiment intends to extract keywords from news, and compare and screen with some similar news, and finally push the current news to interested users according to the keywords. Therefore, it is necessary to collect news data and user data first.

[0050] Specifically, several news data and user data at different times on the news publishing platform are obtained. The news data is divided into current news data and historical news data. One piece of news data corresponds to one piece of news, including the title, text, and other news published by the publisher of the news in the past period of time (preset to be half a year in view of the changes in publisher habits); the user data refers to the user's interest tags.

[0051] It should be noted that the title and text of the news are used as the basis for extracting keywords, which is also the basis for obtaining similar news. The publisher and publishing platform of the news data are retrieved, which is related to the screening and positioning of news keywords. All news published on the same publishing platform at different times are retrieved to provide data support for screening similar news. The same publishing platform is selected because the news on the same platform has similarities in terms of positioning in large categories. Finally, since the final news is to be pushed to users, it is necessary to retrieve the interest tags of each user on the news platform.

[0052] Step S002: Obtain candidate keywords based on the title of the news data, and construct a possible core evaluation of the sentence based on the number of candidate keywords in each sentence of the corresponding text of the news data and the similarity between the phrase and the candidate keywords; obtain the initial keywords of the news data from the candidate keywords based on the frequency of occurrence of the phrase by analyzing the similarity between the phrase and the candidate keywords and the possible core evaluation of the sentence.

[0053] It should be noted that this embodiment splits the title of the news data into several phrases, and preliminarily determines the candidate keywords of the news based on the similarity and relevance between the title of the news data and each sentence in the text content. Among these possible candidate keywords, the phrase with higher repetition rate has a higher probability of being the final keyword of the news. That is, the more times a sentence is repeated in the text, the higher the possibility that it belongs to the core; the higher the number of times a phrase is repeated in the core sentence, the more core the phrase is; the more times a phrase appears in the full text, the higher the possibility that it belongs to the core keyword. When calculating, the more core words in the core sentence, the higher the possibility that it belongs to the final keyword.

[0054] Specifically, in step 2.1, candidate keywords are obtained based on the titles of news data.

[0055] The spaCy software is used to obtain nouns and verbs in the news data titles and recorded as candidate keywords.

[0056] It should be noted that spaCy is an existing natural language processing library, so it will not be described in detail in this embodiment.

[0057] It should be noted that since news titles are often a sentence containing a lot of information, there is a certain amount of information redundancy. In order to facilitate the summary of the content of the news title, this embodiment chooses to perform preliminary keyword extraction on the news title. The titles of the collected news data are split and divided into multiple phrases according to the properties of nouns, verbs, adjectives, etc. Since the subject, predicate and object in the news title all summarize the main content of the news, and the subject, predicate and object are usually nouns and verbs, this embodiment obtains the nouns and verbs in the news title as a keyword extraction process.

[0058] At this point, the nouns and verbs in the title are filtered out, and these phrases are recorded as candidate keywords.

[0059] Step 2.2, construct the possible core evaluation of the sentence based on the number of candidate keywords in each sentence of the corresponding text of the news data and the similarity between the phrases and the candidate keywords.

[0060] Firstly, the character string between any two adjacent periods in the body of the news data is regarded as a sentence, and all the sentences in the body of the news data are segmented to obtain a number of phrases, and the body of the news data is divided into sentences composed of a plurality of phrases.

[0061] It should be noted that Jieba word segmentation is used here to obtain several phrases. Jieba word segmentation is an existing Chinese word segmentation tool, so it is not described in detail in this embodiment.

[0062] As an example, a specific method for obtaining possible core evaluations of a statement is as follows:

[0063]

[0064] Among them, A α represents the possible core evaluation of the αth sentence, N α represents the number of times the candidate keyword appears in the αth sentence, a α,s N represents the similarity core relevance between the sth phrase in the αth sentence and the candidate keyword, α ′ represents the total number of phrases contained in the αth sentence, n α It represents the number of phrases excluding candidate keywords in the αth sentence.

[0065] It should be noted that the more candidate keywords a sentence contains, the more similar the sentence is to the title, which means that the sentence is more likely to describe the core content of the news data. In addition, some synonyms and antonyms may be used to replace candidate keywords and expand some content during the description. Therefore, the more similar or related phrases there are and the more similar the associations are, the more likely the sentence is to be the core content.

[0066] The method for obtaining the similar core association degree is: calculating the similar core association degree between the phrase and the candidate keyword according to the probability of the phrase and the candidate keyword in the same sentence in the news data and the relative core weight of the sentence.

[0067] As an embodiment, a specific method for obtaining the similarity core relevance between a phrase and a candidate keyword is as follows:

[0068]

[0069] Among them, a α,s represents the similarity core relevance between the sth phrase in the αth sentence and the candidate keyword, m α,s represents the total number of sentences in the news data that contain the sth phrase in the αth sentence, m α,s,b represents the number of sentences in the news data that contain both the sth phrase of the αth sentence and the bth candidate keyword, Q α,s It represents the relative core weight of the αth sentence with the sth phrase appearing, max[] represents obtaining the maximum value, and softsign() represents the softsign normalization function.

[0070] It should be noted that the similarity core correlation between a phrase and a candidate keyword is used to describe the similarity between the phrase and the candidate keyword. The higher the frequency with which the s-th phrase appears together with the candidate keyword, that is, the greater the probability that the two appear in the same sentence, the stronger the correlation between the phrase and the candidate keyword; the greater the relative core weight of the sentence, the stronger the correlation between the phrase and the candidate keyword, that is, the more similar they are.

[0071] The method for obtaining the relative core weight is: obtaining the relative core weight of the sentence according to the number of times the sth phrase and the candidate keyword appear in the sentence.

[0072] As an embodiment, a specific method for obtaining the relative core weight of the sentence in which the phrase is located is:

[0073]

[0074] Among them, Q α,s represents the relative core weight of the αth sentence with the sth phrase appearing, Y α represents the keyword component similarity of the αth sentence, where the keyword component similarity of the αth sentence is the ratio of the number of candidate keywords appearing in the αth sentence to the number of all candidate keywords, g s,α represents the number of occurrences of the sth phrase in the αth sentence, g α represents the number of candidate keywords in the αth sentence, m α,s It represents the total number of all sentences in the news data that contain the sth phrase of the αth sentence.

[0075] It should be noted that the more times the s-th phrase in a sentence appears together with the candidate keywords, the more likely it is that the sentence belongs to the core content; the more similar the sentence is to the title, that is, the more candidate keywords the sentence contains, the more likely it is that the sentence describes the core content of the news data, and the more likely the phrase in the sentence belongs to the final keyword of the news data.

[0076] At this point, the possible core evaluation of the sentence is obtained.

[0077] Step 2.3, analyze the similarity between the phrase and the candidate keywords, possible core evaluation of the sentence, and obtain the initial keywords of the news data from the candidate keywords in combination with the occurrence frequency of the phrase.

[0078] Specifically, all the statements are sorted in descending order according to possible core evaluations, and the first A statements therein are marked as core statements, where A is a preset first parameter.

[0079] It should be noted that this embodiment presets A=30% based on experience, that is, the top 30% of sentences are marked as core sentences, which can be adjusted according to actual conditions and is not specifically limited in this embodiment.

[0080] The initial keywords are obtained based on the number of times the phrase appears in the full text and the possibility of belonging to the core sentence.

[0081] Specifically, first, based on the similarity core association between any phrase and the candidate keyword, the possible core evaluation of the sentence to which the phrase belongs, and the number of phrases, the candidate key evaluation of the phrase belonging to the initial keyword is obtained.

[0082] As an embodiment, a specific method for obtaining the candidate key evaluation of the sth phrase is as follows:

[0083]

[0084] Among them, G α,s represents the candidate key evaluation of the sth phrase in the αth sentence, a α,s A represents the similarity core correlation between the sth phrase in the αth sentence and the candidate keyword, α represents the possible core evaluation of the αth sentence, B α,s represents the number of occurrences of the sth phrase in the αth sentence, m α,s represents the total number of all sentences in the news data that contain the sth phrase in the αth sentence, and maxmin[] represents the maximum and minimum normalization function.

[0085] It should be noted that, in a sentence, the more times a phrase is repeated, the more important the phrase is; in the full text of the news, the more a sentence describes the core content, the more likely the phrases in it are keywords; the more sentences a phrase appears in, the more likely it is that the phrase is a keyword of the news data; the higher the correlation between a phrase and the phrase in the title of the news data, the more likely it is that the phrase is a keyword of the news.

[0086] Then, all phrases appearing in the news are traversed, and a threshold T is set to filter out the initial keywords that meet the requirements. The specific method is: a threshold T is preset, and if the candidate key evaluation of the phrase is greater than or equal to the threshold T, the phrase is recorded as the initial keyword.

[0087] It should be noted that the threshold T=0.6 for screening candidate key evaluations is preset based on experience and can be adjusted according to actual conditions, and is not specifically limited in this embodiment.

[0088] It should be noted that by analyzing the titles of news data and combining the phrases in each sentence of the body text of the news data, the possible core evaluations of the sentences are constructed; by analyzing the relevance between the phrases and the news titles, as well as the possible core evaluations of the sentences, and combining the occurrence frequencies of the phrases, the initial keywords of the news data are obtained.

[0089] Thus, the initial keywords of the news data are obtained through the above method.

[0090] Step S003: Screen out the comparison news data with similar content to the current news data from all historical news data according to the initial keywords; screen out the final keywords of the current news data through the initial keywords of the comparison news data.

[0091] Respectively obtain the initial keywords of the current news data and the historical news data, and screen out the preliminary screening news from the historical news data according to the similarity relationship between the initial keywords of the current news data and the historical news data;

[0092] The specific screening method is as follows: for any initial keyword of the current news data, obtain the historical news data whose title and body text contain these initial keywords of the current news data, and record it as the preliminary screening news;

[0093] Screen out the comparison news data from the preliminary screening news; sort the preliminary screening news according to the number of occurrences of the initial keywords of the current news data in the preliminary screening news, the frequency of occurrence of each initial keyword, and the candidate key evaluations of the initial keywords in the preliminary screening news, and select the news whose rankings of all three are in the top B, and mark it as the comparison news data, where B is a preset parameter.

[0094] It should be noted that in this embodiment, B = 10% is preset according to experience, that is, it is preset according to experience to mark the news in the top 10% as the comparison news data, which can be adjusted according to the actual situation, and this embodiment does not make specific limitations.

[0095] Take the intersection of the initial keywords in the comparison news data and the initial keywords of the current news data, and then compare with the stop word list to delete common and frequently used words that have no substantial meaning for text analysis, such as "de", "le", "shi", etc., and screen out the final keywords of the current news data.

[0096] Step S004: Push news to the user in combination with the final keywords of the current news data and the user's interest tags.

[0097] Specifically, combine the final keywords of the current news data and the user's interest tags, and use the RNN neural network to push news to the user.

[0098] The training method of the RNN neural network is as follows: taking the final keyword of any current news data as a sample, manually labeling the interest tag corresponding to the sample as the label of the sample, and forming a data set for training the RNN neural network by a number of labeled samples. Using the data set, according to the cross entropy loss function, the RNN neural network is trained using a stochastic gradient descent algorithm to converge, thereby obtaining a trained RNN neural network for news push.

[0099] This is the end of the present invention.

[0100] 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, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A personalized news push method based on big data, characterized in that: The method comprises the following steps: Based on the big data, a number of news data and user data are collected respectively, wherein the news data is divided into current news data and historical news data, including the title and text of the news, and the user data is the user's interest tags; Obtain candidate keywords based on the title of the news data; obtain several sentences in the text corresponding to the news data and several phrases contained in each sentence, and construct possible core evaluations of the sentences in combination with the number of candidate keywords in each sentence in the text corresponding to the news data and the similarity between the phrases and the candidate keywords; obtain initial keywords of the news data from the candidate keywords in combination with the occurrence frequency of the phrases by analyzing the similarity between the phrases and the candidate keywords and the possible core evaluations of the sentences; Filter out comparative news data with similar content to the current news data from all historical news data based on the initial keywords; filter out the final keywords of the current news data by comparing the initial keywords of the news data; Combine the final keywords of the current news data and the user's interest tags to push news to the user; The method of obtaining several sentences in the text corresponding to the news data and several phrases contained in each sentence, and constructing possible core evaluation of the sentence in combination with the number of candidate keywords in each sentence in the text corresponding to the news data and the similar core association between the phrase and the candidate keywords includes the following specific methods: First, the character string between any two adjacent periods in the body of the news data is recorded as a sentence, and Jieba word segmentation is used to segment all the sentences in the body of the news data to obtain several phrases; According to the number of times the sth phrase and the candidate keyword appear together in the sentence, the relative core weight of the sentence is obtained; Calculate the similarity core association between the phrase and the candidate keyword based on the probability that the phrase and the candidate keyword are in the same sentence in the news data and the relative core weight of the sentence; Then, the possible core evaluations of the statements are calculated: Among them, A α represents the possible core evaluation of the αth sentence, N α represents the number of times the candidate keyword appears in the αth sentence, a α,s represents the similarity core relevance between the sth phrase in the αth sentence and the candidate keyword, N α ′ represents the total number of phrases contained in the αth sentence, n α It represents the number of phrases excluding candidate keywords in the αth sentence.

2. According to the method of claim 1, the personalized news push method based on big data is characterized in that: The specific method of obtaining candidate keywords based on the title of the news data includes: The spaCy software is used to obtain nouns and verbs in the news data titles and recorded as candidate keywords.

3. According to the method of claim 1, the personalized news push method based on big data is characterized in that: The specific calculation method for calculating the similarity core association between the phrase and the candidate keyword according to the probability of the phrase and the candidate keyword in the same sentence in the news data and the relative core weight of the sentence is: Among them, a α,s represents the similarity core relevance between the sth phrase in the αth sentence and the candidate keyword, m α,s represents the total number of sentences in the news data that contain the sth phrase in the αth sentence, m α,s,b represents the number of sentences in the news data that contain both the sth phrase of the αth sentence and the bth candidate keyword, Q α,s It represents the relative core weight of the αth sentence with the sth phrase appearing, max[] represents obtaining the maximum value, and softsign() represents the softsign normalization function.

4. According to the method of claim 3, the personalized news push method based on big data is characterized in that: The method for obtaining the relative core weight of a sentence according to the number of times the sth phrase and the candidate keyword appear together in the sentence is: Among them, Y α represents the keyword component similarity of the αth sentence, where the keyword component similarity of the αth sentence is the ratio of the number of candidate keywords appearing in the αth sentence to the number of all candidate keywords, g s,α represents the number of occurrences of the sth phrase in the αth sentence, g α Represents the number of candidate keywords in the αth sentence.

5. According to the method of claim 1, the personalized news push method based on big data is characterized in that: The method of obtaining the initial keywords of the news data from the candidate keywords by analyzing the similarity core association between the phrase and the candidate keywords, the possible core evaluation of the sentence, and combining the occurrence frequency of the phrase includes the following specific methods: Obtain candidate key evaluations for the phrases based on the similar core associations between any phrase and the candidate keywords, the possible core evaluations of the sentences to which the phrases belong, and the number of phrases; Among them, G α,s represents the candidate key evaluation of the sth phrase in the αth sentence, a α,s A represents the similarity core correlation between the sth phrase in the αth sentence and the candidate keyword, α represents the possible core evaluation of the αth sentence, B α,s represents the number of occurrences of the sth phrase in the αth sentence, m α,s represents the total number of all sentences in the news data that contain the sth phrase in the αth sentence, and maxmin[] represents the maximum and minimum normalization function; A threshold T is preset, and if the candidate key evaluation of a phrase is greater than or equal to the threshold T, the phrase is recorded as an initial keyword.

6. According to the method of claim 1, the personalized news push method based on big data is characterized in that: The specific steps of screening out comparative news data having similar contents to the current news data from all historical news data according to the initial keywords and screening out the final keywords of the current news data by comparing the initial keywords of the news data include: respectively obtaining initial keywords of current news data and historical news data, and selecting preliminary screening news from the historical news data according to the similarity relationship between the initial keywords of the current news data and the historical news data; Comparative news data were obtained from the preliminary screening news; The initial keywords in the comparison news data and the initial keywords in the current news data are intersected, and the initial keywords in the intersection are screened through a stop word list to obtain final keywords.

7. A personalized news push method based on big data according to claim 6, characterized in that: The specific screening method for screening out preliminary screening news from historical news data is as follows: For any initial keyword of current news data, historical news data whose title and text contain the initial keyword of the current news data are obtained and recorded as preliminary screening news.

8. According to claim 6, a personalized news push method based on big data is characterized in that: The specific method of obtaining the comparative news data from the preliminary screening news is as follows: The preliminary screened news are sorted according to the number of initial keywords of the current news data appearing in the preliminary screened news, the frequency of each initial keyword appearing, and the candidate key evaluation of the initial keywords in the preliminary screened news, and the preliminary screened news ranked in the top B are selected and marked as comparison news data, where B is a preset parameter.

9. The method for personalized news push based on big data according to claim 1, characterized in that: The specific method of pushing news to the user by combining the final keyword of the current news data and the user's interest tag includes: Combine the final keywords of the current news data and the user's interest tags, and use the RNN neural network to push news to users; The training method of the RNN neural network is as follows: taking the final keyword of any current news data as a sample, manually labeling the interest tag corresponding to the sample as the label of the sample, and forming a data set for training the RNN neural network by a number of labeled samples. Using the data set, according to the cross entropy loss function, the RNN neural network is trained using a stochastic gradient descent algorithm to converge, thereby obtaining a trained RNN neural network for news push.

Images