CN110888990A - Text recommending methods, devices, equipment and media - Google Patents

Abstract

The invention discloses a text recommendation method, a text recommendation device, text recommendation equipment and a text recommendation medium, which relate to the technical field of financial science and technology, and comprise the following steps: monitoring the operation behavior of a target user to determine keywords associated with the target user; more than one updated text containing at least one keyword is retrieved from a preset text database set to serve as a first candidate text; calling a preset affair atlas, selecting an updated text with the total association degree with the first candidate text not less than a preset association threshold value from a preset text database set according to the affair atlas, wherein the affair atlas contains the association relation between the text and the text, and each association relation has the corresponding association degree; and screening out a selected text from the first candidate text and the second candidate text according to the operation behavior, and recommending the selected text to the target user. The invention solves the technical problems of single recommendation and low recommendation accuracy in the existing recommendation process.

Description

Text recommending methods, devices, equipment and media

technical field

The present invention relates to the technical field of financial technology (Fintech), and in particular, to a text recommendation method, apparatus, device and medium.

Background technique

With the development of computer technology, more and more technologies are applied in the financial field, and the traditional financial industry is gradually transforming to Finteh, and content recommendation technology is no exception. It also puts forward higher requirements for content recommendation technology. At present, content recommendation technology completely relies on user-configured keywords for recommendation, and completely relying on user-configured keywords for recommendation will narrow the user's preference, and then there is a push to The user's news data and other content are too simplistic and have technical problems such as low accuracy. For example, the user's monitoring target is long-term rental apartments, and the key words are related to long-term rental apartments. Due to the explosion of P2P lightning, the long-term rental apartment will be exploded indirectly. In the absence of P2P-related keywords, users cannot receive such news, and they cannot predict the risk that long-term rental apartments may burst out in advance.

SUMMARY OF THE INVENTION

The main purpose of the present invention is to provide a text recommendation method, device, equipment and medium, which aims to solve the technical problems of excessively single recommendation and low recommendation accuracy in the existing content recommendation process based on keywords.

To achieve the above object, an embodiment of the present invention provides a text recommendation method, where the text recommendation method includes:

Monitoring the operation behavior of the target user, and determining the keywords associated with the target user according to the operation behavior;

Retrieve one or more updated texts containing at least one of the keywords from a preset text database set as the first candidate text;

Retrieve the preset affair atlas, and select, from the preset text database set according to the preset affair atlas, the updated text whose total degree of association with the first candidate text is not less than the preset association threshold, as the first Two candidate texts, the event graph includes the association between texts, and each association has its corresponding degree of association;

According to the operation behavior, the selected text is screened from the first candidate text and the second candidate text, and the selected text is recommended to the target user.

Optionally, the retrieval of a preset event map, according to the preset event map, is selected from the preset text database set and the total degree of association with the first candidate text is not less than a preset association threshold. The Update Text step before includes:

Collect the text to be processed from the preset text database collection every preset time period;

Perform html tag filtering, symbol filtering and clause processing on the to-be-processed text through a preset regular expression, to obtain a preprocessed text formed by a clause list;

The preset event graph is generated according to the preprocessed text.

Optionally, the step of generating the preset event map according to the preprocessed text includes:

Identify multiple preset text associations for each clause in the clause list, and obtain node texts to be processed, wherein the preset text associations include but are not limited to succession, causality, condition, and juxtaposition relation;

Perform word segmentation processing on the node text to be processed by using a preset word segmentation tool, and obtain the word vector of each word segmentation, and obtain the node vector of each node text to be processed based on the word vector of each word segmentation;

Calculate the first node distance between each to-be-processed node text and other to-be-processed node texts according to the node vector of each to-be-processed node text;

Perform iterative grafting processing on two to-be-processed node texts whose node distances are less than the first preset distance, until each of the to-be-processed node texts is in a convergent state in which the relationship between the node texts does not change, wherein the The text of each pending node of the state is set to the text of the convergent node;

The preset event graph is generated based on the node text relationship edges between the convergent node text and the convergent node text.

Optionally, the retrieval of a preset event map, according to the preset event map, is selected from the preset text database set and the total degree of association with the first candidate text is not less than a preset association threshold. The updated text of , as the second candidate text step includes:

Retrieve a preset event map, and determine whether there is a convergence node text containing the keyword in the corresponding clause in the event map;

If there is, set the convergence node text containing the keyword in the corresponding clause as the user's focus node text, and select the first candidate from the text updated in the preset text database within the preset time period For the third candidate text outside the text, a word segmentation process is performed on the title of each text in the third candidate text by using a preset word segmentation tool to obtain the title vector of each text in the third candidate text;

Calculate the second node distance between the title vector and the node vector of each convergent node text in the event graph;

Select the second node distance from the third candidate text that is less than the second preset distance, and the convergent node text corresponding to the less than the second preset distance is the first target text of the node text that the user pays attention to, or from the first target text of the node text. Among the three candidate texts, the distance of the second node selected is less than the second preset distance, and there is a second target text of the node text that the user pays attention to within the preset screening logic depth range of the convergent node text corresponding to the less than the second preset distance, wherein , the screening logic depth is determined according to the correlation degree of each correlation in the event map;

The first target text and the second target text are set as the second candidate text.

Optionally, the step of filtering out the selected text from the first candidate text and the second candidate text according to the operation behavior, and recommending the selected text to the target user includes:

Obtain the dissemination volume of each text in the first candidate text and the second candidate text, and obtain the correlation between each text and the target user, and obtain the preference of the target user according to the operation behavior;

According to the spread, the relevance and the preference, the selected text is screened out from the first candidate text and the second candidate text, and the selected text is recommended to the target user .

Optionally, the step of obtaining the relevance between each text and the target user includes:

Obtain the number of times that the keyword appears in each text of the first candidate text, and set the number of times as the number of words;

Obtain the position where the keyword appears in each text of the first candidate text, set the position as a word position, and obtain a preset position weight corresponding to the word position, wherein the word position is different, the position The weights are different, and the word positions include the position of the first sentence of the first paragraph of the text, the position of the first sentence of the last paragraph of the text, the position of the non-first sentence of the first paragraph of the text, the position of the non-first sentence of the last paragraph of the text, the position of the first sentence of the non-first paragraph and the first sentence of the non-final paragraph. Location;

Obtain the ratio of the number of sentences between the positions where the keyword appears for the first time and the last time in each text of the first candidate text to the total number of sentences in the full text, and set the ratio as the word span;

Obtain the target text between the positions where the keyword appears for the first time and the last time in each text of the first candidate text, and obtain the target text that contains the keyword in the average number of preset sentences The number of words contained in the average number of sentences per preset sentence is set as the word density;

Obtain the first relevance of each text in the first candidate text according to the word count, the preset position weight corresponding to the word position, the word span and the word density;

The screening logic depth of each text in the second candidate text is acquired, and the second relevancy degree of each text in the second candidate text is determined according to the screening logic depth.

Optionally, the step of obtaining the preference degree of the target user according to the operation behavior includes:

Acquire the historical browsing text of the target user from the operation behavior, acquire the first document vector of each text in the historical browsing text, and acquire each of the first candidate text and the second candidate text the second document vector of the text;

The first Pearson correlation coefficient between the second document vector and the first document vector is obtained, and the preference of the target user is obtained according to the first Pearson correlation coefficient.

Optionally, the step of obtaining the first document vector of each text in the historical browsing text includes:

Obtain a first probability matrix in which each text in the historical browsing text is classified under a first preset category according to a preset clustering algorithm;

Obtain the word segmentation of each text in the historical browsing text according to a preset word segmentation algorithm, and obtain a second probability matrix in which the word segmentation is divided into a second preset category;

Obtain each optimized word vector corresponding to each text in the historical browsing text according to the first probability matrix and the second probability matrix;

The first document vector of each text in the historical browsing text is obtained according to the optimized word vector.

Optionally, the step of obtaining the first Pearson correlation coefficient between the second document vector and the first document vector, and obtaining the preference degree of the target user according to the first Pearson correlation coefficient includes:

Obtain the historical browsing time between the time when each text in the historical browsing text is clicked to browse to the current moment;

Obtain the first Pearson correlation coefficient between the second document vector and the first document vector, and perform interest downweight processing on the first Pearson correlation coefficient according to the historical browsing time to obtain a second Pearson correlation coefficient correlation coefficient;

The preference of the target user is acquired according to the second Pearson correlation coefficient.

Optionally, the selected text is screened out from the first candidate text and the second candidate text according to the spread, the relevancy and the preference, and the selected text is The steps of recommending to the target user include:

Calculate the value score of each text in the first candidate text and the second candidate text according to the spread, the first relevancy, the second relevancy and the preference;

According to the value score, a preset number of texts are sequentially selected as selected texts, and the selected texts are recommended to the target user.

The present invention also provides a text recommendation device, the text recommendation device includes:

a monitoring module for monitoring the operation behavior of the target user, and determining the keywords associated with the target user according to the operation behavior;

a retrieval module, used for retrieving more than one updated text containing at least one of the keywords from a preset text database set, as the first candidate text;

The selection module is used for retrieving a preset affair atlas, and according to the preset affair atlas, selects from the preset text database set the ones whose total degree of association with the first candidate text is not less than a preset association threshold. Update the text, as the second candidate text, the event graph includes the association relationship between the text and the text, and each association relationship has its corresponding association degree;

The screening module is configured to screen out the selected text from the first candidate text and the second candidate text according to the operation behavior, and recommend the selected text to the target user.

Optionally, the text recommendation device further includes:

a collection module, used for collecting texts to be processed from the preset text database set at every preset time period;

a preprocessing module, configured to perform html tag filtering, symbol filtering and clause processing on the text to be processed by using a preset regular expression to obtain preprocessed text formed by a clause list;

A generating module, configured to generate the preset event graph according to the preprocessed text.

Optionally, the generation module includes:

An identification unit, configured to identify multiple preset text associations for each clause in the clause list, and obtain node texts to be processed, wherein the preset text associations include but are not limited to Shuncheng, causality, condition and juxtaposition;

a first obtaining unit, configured to perform word segmentation processing on the node text to be processed by using a preset word segmentation tool, and obtain the word vector of each word segmentation, and obtain the node vector of each node text to be processed based on the word vector of each word segmentation;

a first computing unit, configured to calculate the first node distance between each to-be-processed node text and other to-be-processed node texts according to the node vector of each to-be-processed node text;

The grafting processing unit is used for iterative grafting processing of two to-be-processed node texts whose node distances are less than the first preset distance, until each of the to-be-processed node texts is in a convergent state where the relationship edges of the node texts no longer change, wherein , setting the text of each node to be processed in the convergent state as the text of the convergent node;

A generating unit, configured to generate the preset event graph based on the node text relationship edges between the convergent node text and the convergent node text.

Optionally, the selection module includes:

A retrieving unit, used for retrieving a preset event map, and judging whether there is a convergence node text that includes the keyword in the corresponding clause in the event map;

The first setting unit is configured to set the convergence node text containing the keyword in the corresponding clause as the user's attention node text, if it exists, and update it from the preset text database within a preset time period In the text, select the third candidate text outside the first candidate text, carry out word segmentation processing to the title of each text in the third candidate text by the preset word segmentation tool, and obtain the title vector of each text in the third candidate text;

a second computing unit for computing the second node distance between the title vector and the node vector of each convergent node text in the event graph;

a selection unit, configured to select from the third candidate texts that the second node distance is less than the second preset distance, and the convergent node text corresponding to the less than the second preset distance is the first target text of the node text that the user pays attention to, Or select the second node distance from the third candidate text that is less than the second preset distance, and the preset screening logic depth range of the convergence node text corresponding to the less than the second preset distance exists. Two target texts, wherein the screening logic depth is determined according to the degree of association of the associations in the event map;

A second setting unit, configured to set the first target text and the second target text as the second candidate text.

Optionally, the screening module includes:

The second obtaining unit is configured to obtain the dissemination volume of each text in the first candidate text and the second candidate text, obtain the correlation between each text and the target user, and obtain the information according to the operation behavior. Describe the preferences of target users;

a recommending unit, configured to filter out the selected text from the first candidate text and the second candidate text according to the spread, the relevance and the preference, and recommend the selected text to the target user.

Optionally, the second obtaining unit includes:

A first acquisition subunit, for acquiring the number of times the keyword appears in each text of the first candidate text, and setting the number of times as the number of words;

The second obtaining subunit is configured to obtain the position where the keyword appears in each text of the first candidate text, set the position as a word position, and obtain a preset position weight corresponding to the word position , wherein the word position is different, the position weight is different, and the word position includes the position of the first sentence of the first paragraph of the text, the position of the first sentence of the last paragraph of the text, the position of the non-first sentence of the first paragraph of the text, the position of the non-first sentence of the last paragraph of the text, and the first sentence of the non-first paragraph. Sentence position and non-final first sentence position;

The third obtaining subunit is used to obtain the ratio of the number of sentences between the positions where the keyword appears for the first time and the last time in each text of the first candidate text to the total number of sentences in the full text, The ratio is set to the word span;

The fourth obtaining subunit is used to obtain the target text between the positions where the keyword appears for the first time and the last time in each text of the first candidate text, and obtain the average per preset value in the target text. The number of the keywords contained in the number of sentences, and the number of the keywords contained in the average number of sentences per preset is set as the word density;

The fifth obtaining subunit is used to obtain the first correlation degree of each text in the first candidate text according to the word count, the preset position weight corresponding to the word position, the word span and the word density ;

The sixth obtaining subunit is configured to obtain the screening logic depth of each text in the second candidate text, and determine the second relevancy degree of each text in the second candidate text according to the screening logic depth.

Optionally, the second obtaining unit includes:

The seventh obtaining subunit is used to obtain the historical browsing text of the target user from the operation behavior, obtain the first document vector of each text in the historical browsing text, and obtain the first candidate text and all the texts. Describe the second document vector of each text in the second candidate text;

The eighth obtaining subunit is configured to obtain the first Pearson correlation coefficient between the second document vector and the first document vector, and obtain the preference degree of the target user according to the first Pearson correlation coefficient.

Optionally, the seventh acquisition subunit is used to realize:

Obtain a first probability matrix in which each text in the historical browsing text is classified under a first preset category according to a preset clustering algorithm;

Obtain the word segmentation of each text in the historical browsing text according to a preset word segmentation algorithm, and obtain a second probability matrix in which the word segmentation is divided into a second preset category;

Obtain each optimized word vector corresponding to each text in the historical browsing text according to the first probability matrix and the second probability matrix;

The first document vector of each text in the historical browsing text is obtained according to the optimized word vector.

Optionally, the eighth acquisition subunit is used to implement:

Obtain the historical browsing time between the time when each text in the historical browsing text is clicked to browse to the current moment;

Obtain the first Pearson correlation coefficient between the second document vector and the first document vector, and perform interest downweight processing on the first Pearson correlation coefficient according to the historical browsing time to obtain a second Pearson correlation coefficient correlation coefficient;

The preference of the target user is acquired according to the second Pearson correlation coefficient.

Optionally, the screening module includes:

a third calculation unit, configured to calculate each text in the first candidate text and the second candidate text according to the spread, the first degree of relevancy, the second degree of relevancy and the preference degree value score;

A screening unit, configured to sequentially select a preset number of texts as selected texts according to the value score from high to low, and recommend the selected texts to the target user.

The present invention also provides a medium on which a text recommendation program is stored, and when the text recommendation program is executed by a processor, implements the steps of the above text recommendation method.

The present invention monitors the operation behavior of the target user, and determines the keywords associated with the target user according to the operation behavior; after acquiring the keywords, retrieves one or more documents containing at least one of the keywords from a preset text database set The updated text is regarded as the first candidate text; after obtaining the first candidate text, the preset event map is called, and the preset event map is selected from the preset text database set according to the preset event map. The updated text whose total correlation degree of candidate text is not less than the preset correlation threshold is regarded as the second candidate text. The acquisition of the second candidate text expands the selection category of the candidate text in the recommendation process. Each association has its corresponding degree of association; according to the operation behavior, the selected text is screened out from the first candidate text and the second candidate text, and the selected text is recommended to the target user. That is, in this application, the selected text is not only selected from the first candidate text searched according to the keywords, but the second candidate text and the first candidate text obtained according to the preset event map, etc. The selected text is selected from the collection, thus avoiding the simplification of content recommendation, and because the association relationship between the reference text and the text in this application is used to recommend content instead of just recommending based on keywords alone, this application can Improve recommendation accuracy.

Description of drawings

1 is a schematic flowchart of a first embodiment of a text recommendation method according to the present invention;

2 is a second embodiment of the text recommendation method of the present invention based on retrieving a preset event map, and selecting the total content with the first candidate text from the preset text database set according to the preset event map. A schematic diagram of the refinement process before the step of updating the text with the degree of association not less than the preset association threshold;

3 is a schematic diagram of a device structure of a hardware operating environment involved in a method according to an embodiment of the present invention;

4 is a schematic diagram of a first scenario involved in the text recommendation method of the present invention;

5 is a schematic diagram of a second scenario involved in the text recommendation method of the present invention;

6 is a schematic diagram of a third scenario involved in the text recommendation method of the present invention;

7 is a schematic diagram of a fourth scenario involved in the text recommendation method of the present invention;

8 is a schematic diagram of a fifth scenario involved in the text recommendation method of the present invention;

9 is a schematic diagram of a sixth scenario involved in the text recommendation method of the present invention;

10 is a schematic diagram of a seventh scenario involved in the text recommendation method of the present invention;

11 is a schematic diagram of an eighth scenario involved in the text recommendation method of the present invention;

12 is a schematic diagram of a ninth scenario involved in the text recommendation method of the present invention;

FIG. 13 is a schematic flowchart of the text recommendation method of the present invention.

The realization, functional characteristics and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments.

Detailed ways

It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

The present invention provides a text recommendation method. In an embodiment of the text recommendation method, referring to FIG. 1 , the text recommendation method includes:

Step S10, monitoring the operation behavior of the target user, and determining the keywords associated with the target user according to the operation behavior;

Step S20, retrieve more than one updated text containing at least one of the keywords from a preset text database set, as the first candidate text;

Step S30, fetching a preset affair atlas, and selecting, according to the preset affair atlas, from the preset text database set an updated text whose total degree of association with the first candidate text is not less than a preset association threshold , as the second candidate text, the event graph includes the association relationship between texts, and each association relationship has its corresponding association degree;

Step S40, according to the operation behavior, screen out the selected text from the first candidate text and the second candidate text, and recommend the selected text to the target user.

Specific steps are as follows:

Step S10, monitoring the operation behavior of the target user, and determining the keywords associated with the target user according to the operation behavior;

At present, more and more public opinion systems have appeared on the market to meet the needs of enterprises to monitor online public opinions and individual thematic tracking of hot events. Specifically, public opinion systems can help enterprises to listen to the ideas of target audiences and analyze industry trends. , manage brand reputation and carry out crisis warning and other functions, the current public opinion system generally achieves the above functions through the following processes: 1. Data collection: collect all information sources on the entire network, including news media, forums, blogs, microblogs 2. Data screening: filter news data according to the monitoring task keywords configured on the public opinion system. For example, if the text of a certain news contains keywords configured by the user, the article will be kept. News is used for subsequent processing; 3. Data processing: For all news containing keywords, sequentially calculate the emotional tendency of the text, the amount of news dissemination, the degree of correlation between news text and keywords, etc.; 4. Data push: comprehensively consider news The sentiment, the amount of communication, the degree of relevance, and the user’s click preference on the historically pushed news data on the public opinion system, sort the news data processed in steps 1-3, and select multiple news articles that users are most likely to be interested in to push, and also That is, in the prior art, the recommendation of the target content is carried out completely relying on the keywords configured by the user and the user's click preference on the historically pushed news data, and in the process of pushing the target content, the user preference is analyzed based on word2vec (one The word vector model, which can describe the semantic similarity of Chinese words according to the distance between the word vectors) is processed based on the word vector of the word2vec type, and is processed based on the word2vec type word vector when analyzing user preferences. The preference is narrow, which leads to the simplification of the pushed content. Specifically, for example, a user has clicked on a piece of news such as "Enterprise A and Bank B have reached a cooperation" multiple times in a certain period of time. After processing the word2vec word vector, the public opinion system learns that the user prefers the relationship between Enterprise A and Bank B. news in between. If there is another piece of news "Enterprise A and Bank C invest hundreds of millions in a university to jointly build a laboratory", the public opinion system will probably not think that this is the news that the user prefers or likes, but the actual preference of the user is that Enterprise A is in The layout of the financial field, and only pushing the news between enterprise A and bank B obviously leads to the simplification of the content of the push. In addition, in the process of pushing the target content, the recommendation of the target content completely relies on the keywords configured by the user and the user's click preference on the historically pushed news data, which will lead to the recommendation of the news data and other content recommended to the user is too single and the recommendation is accurate. For technical problems with low rates, such as the user's monitoring target is long-term rental apartments, the keywords are all related to long-term rental apartments. Since the explosion of P2P mines will indirectly lead to the explosion of long-term rental apartments, if P2P-related keywords are not configured, users cannot receive such news, and they cannot predict the risk of explosion of long-term rental apartments in advance. Therefore, it is completely dependent on the data screening of keywords, and the recommendation accuracy is low in many scenarios.

In order to solve the above-mentioned technical problems, in this embodiment, the operation behavior of the target user is monitored, and the keywords associated with the target user are determined according to the operation behavior, and the operation behavior includes the search behavior triggered by sliding behavior or inputting keywords. is a sliding behavior, and determining the keyword associated with the target user according to the operation behavior may be: extracting a pre-stored keyword associated with the target user, and the pre-stored keyword associated with the target user is obtained based on the historical browsing text of the target user. , if the operation behavior is a search behavior of inputting a keyword, the keyword associated with the target user may be the input keyword, or may be a combination of the input keyword and a pre-stored keyword associated with the target user.

Step S20, retrieve more than one updated text containing at least one of the keywords from a preset text database set, as the first candidate text;

The preset text database collection includes database collections composed of news network media, forums, blogs, Weibo and other various information clients.

One or more updated texts containing at least one of the keywords are retrieved from the preset text database set, and as the first candidate text, one or more updated texts containing at least one of the keywords are retrieved from the preset text database set The updated text, as the first candidate text, can be: retrieve more than one update text containing at least one of the keywords from the preset text database collection in real time, as the first candidate text (for real-time recommendation), or every At intervals of a certain period of time, one or more updated texts containing at least one of the keywords are retrieved from the preset text database collection in real time, as the first candidate text (for timing recommendation), or only this time from the preset text. One or more updated texts containing at least one of the keywords are retrieved from the database set as the first candidate texts (to facilitate the target user to make recommendations when searching), and the like.

Step S30, fetching a preset affair atlas, and selecting, according to the preset affair atlas, from the preset text database set an updated text whose total degree of association with the first candidate text is not less than a preset association threshold , as the second candidate text, the event graph includes the association relationship between texts, and each association relationship has its corresponding association degree;

Among them, the preset affair map has been generated and updated in real time or regularly, and the total relevance can be determined by the logical depth between words or the distance between words, for example, the total relevance is not less than the preset relevance threshold. It can be that the logical depth is not less than 2 depth units or the distance between words is not less than 10 preset distance units, etc., and the relationship between texts can be causality, inheritance and other related relationships. The correlation degree of the relationship is different. Retrieve the preset affair atlas, and select, from the preset text database set according to the preset affair atlas, the updated text whose total degree of association with the first candidate text is not less than the preset association threshold, as the first Two candidate texts, the second candidate text improves the recommendation range in the content recommendation process.

Wherein, as shown in FIG. 2 , when the preset event map is retrieved, the total correlation degree with the first candidate text is selected from the preset text database set according to the preset event map and is not less than The update text step for preset association thresholds includes:

Step A1, collecting the text to be processed from the preset text database set at every preset time period;

Step A2, performing html tag filtering, symbol filtering and clause processing on the text to be processed by a preset regular expression, to obtain a preprocessed text formed by a clause list;

Step A3, generating the preset event map according to the preprocessed text.

In this embodiment, texts to be processed are collected from the preset text database set every preset time period (which may include real-time), wherein, since the amount of texts to be processed collected every day is on the order of tens of millions, it is possible to use The preset collection model, such as the preset spark streaming model, is used to complete the collection. After the text to be processed is obtained, html tag filtering, symbol filtering and clause processing are performed on the text to be processed by using a preset regular expression, so as to obtain a preprocessed text formed by a clause list, optionally, every preset time interval The segment collects the text to be processed from the preset text database set, and after obtaining the text to be processed, performs html tag filtering, symbol filtering and sentence segmentation on each text in the text to be processed through a preset regular expression After processing, the preprocessed text composed of the clause list is obtained. Specifically, the html tags in each text body of the text to be processed are filtered out through the following four regular expressions, the first one: '//<! \[CDATA\[[^>]*//\]\]>', the second line: '<\s*script[^>]*>[^<]*<\s*/\s*script\ s*>', the third item: '<\s*style[^>]*>[^<]*<\s*/\s*style\s*>', the fourth item: '<! --[^>]*-->', after filtering out the html tags in the text through the above 4 regular expressions, filter out the emojis in each text of the text to be processed through the following 4 regular expressions, the first One: "\U0001F600-\U0001F64F", the second: "\U0001F300-\U0001F5FF", the third: "\U0001F680-\U0001F6FF", the fourth: "\U0001F1E0-\U0001F1FF", after filtering , the text to be processed, such as the main text of each text, is divided into sentences: according to the punctuation marks ".", "?", "?", "!", "!", etc., the main text is divided into sentence lists. In this embodiment, although Sentence processing is performed separately for each text, but the sentences for each text in the sentence list can be mixed rather than differentiated on a per-text basis.

The step of generating the preset event map according to the preprocessed text includes:

Step A31: Identify multiple preset text associations for each clause in the clause list, and obtain node texts to be processed, wherein the preset text associations include but are not limited to succession, causality, conditions and juxtapositions;

Identify multiple preset text associations for each clause in the clause list. The preset text associations include but are not limited to types such as inheritance, causality, condition, and parallel relationship. After multiple preset text associations are identified for each clause in the list, the node text to be processed is obtained.

Specifically, two events (two event phrases) representing succession/causality/condition/parallel relationship can be identified from each text sentence of each text of the preprocessed text according to the preset event word segmentation tool. The phrase can be set as the node text to be processed in the event graph, and the event map uses a directed edge (pointed line) to connect the two node texts to be processed, for example, from "The central bank's interest rate cut will make the loan cost lower" The node text to be processed as shown in Fig. 4 can be obtained from the text sentence of . In the present embodiment, the table can also be identified from each text sentence of each text of the preprocessed text based on the preset text association extraction model. /Cause/Condition/Parallel relationship of node texts to be processed, wherein the preset word combinations in the preset text association relationship are: (first, second), (first, then), (on the one hand, on the other hand ), (first, then), (first, then), (first, then), etc., if a sentence contains both the above-mentioned words and expressions to combine two words in a certain phrase, and the two words are in the sentence The order of appearance is consistent with the order defined in the phrase, and the two clauses guided by these two words are extracted by the preset guiding clause model in the preset text association extraction model, and all punctuation marks in the clauses are removed ( default), modal particles (default), auxiliary words (default) and stop words (default), etc., as the two pending node texts in the event graph, and at the same time use a table to follow the To edge (the edge of the logical relationship) connects the two pending node texts. For example, the sentence "First A, then B" is processed as shown in Figure 5.

Similarly, the presupposed word combinations that express cause and effect are: (because, so), (because, cause), (because, cause), (because, therefore), (because, so), (because, cause) , (because, make), (because, therefore), (since, then), (since, then), (once, then), (because, therefore), (because, so), (because, cause) , (due to, therefore), (due to, made), (due to, therefore), (_, therefore), (_, therefore), (_, caused), (_, thus), (_, made), ( _, therefore) etc. The underscore "_" in the phrase represents an empty word, and the matching of the empty word can be ignored in the subsequent matching process. If a sentence contains the above causal words to combine two words in a certain phrase, and the order of appearance of the two words in the sentence is the same as the order defined in the phrase, the two words are divided by the preset guiding clause model. The two clauses guided by each word are extracted, and all punctuation marks (preset), modal particles (preset), auxiliary words (preset) and stop words (preset) in the clauses are removed. As two to-be-processed node texts in the graph, a causal directed edge (logical relationship edge) is used to connect the two to-be-processed node texts. For example, the sentence "because of A, so B" is processed to obtain Figure 6.

Similarly, the word combinations of the preset table conditions are: (if, then), (if, then), (if, then), (if, then), (if, then), (if, then), ( If, then), (if, then), (once, then), (as long as, then), (if, then), (only, only) and so on. If a sentence contains the words of the above table conditions at the same time to combine two words in a certain phrase, and the appearance order of the two words in the sentence is the same as the order defined in the phrase, the two words are divided by the preset guiding clause model. The two clauses guided by each word are extracted, and all punctuation marks (preset), modal particles (preset), auxiliary words (preset) and stop words (preset) in the clauses are removed. As two node texts to be processed in the graph, a directed edge (the edge of a logical relationship) of a table condition is used to connect the two node texts to be processed, such as the sentence "If A, then B", after processing, we get Figure 7.

Likewise, the presupposed list of juxtaposed word combinations are: (not only, and), (not only, and), (not only, also), (not only, also), (not only, and), (not only, and), ( not only, also), (not only, also), (not only, and), (not only, and), (not only, also), (not only, also), (not only, but), (not only, and), (not only, also), (not only, also), (either, or), (or, or), (or, or), etc. If a sentence contains the parallel words in the above table to combine two words in a certain phrase, and the order of appearance of the two words in the sentence is the same as the order defined in the phrase, the two words are divided by the preset guiding clause model. The two clauses guided by each word are extracted, and all punctuation marks (preset), modal particles (preset), auxiliary words (preset) and stop words (preset) in the clauses are removed. As two to-be-processed node texts in the graph, the two to-be-processed node texts are connected by a table-parallel directed edge (edge of logical relationship). For example, the sentence "Not only A, but also B" is processed to obtain Figure 8.

It should be noted that if the preprocessed text contains more than a preset number of text data, such as 100,000 pieces of text data, a preset bidirectional extraction network model can be used to extract multiple preset text associations in each text sentence in the preprocessed text. Relational event phrases.

Step A32, perform word segmentation processing on the node text to be processed by a preset word segmentation tool, and obtain the word vector of each word segmentation, and obtain the node vector of each node text to be processed based on the word vector of each word segmentation;

The word segmentation process is performed on the node text to be processed by using a preset word segmentation tool, such as the preset stammer word segmentation tool (an open source Chinese word segmentation tool, which can perform word segmentation and part-of-speech tagging on the input Chinese text). word2vec (a word vector model that maps each Chinese word into a high-dimensional vector (a 200-dimensional vector can be taken), assuming that 5 high-dimensional vectors are obtained, abcde, the 5 word vectors are divided into corresponding dimensions and dimension weights Add up to get the word vector of the segmented word) to get the word vector of each segmented word, and get the node vector of each node text to be processed based on the word vector of each segmented word.

Step A33, calculating the first node distance between each to-be-processed node text and other to-be-processed node texts according to the node vector of each to-be-processed node text;

For any two Chinese words, the closer they are in terms of semantics, the closer the vector distance is after mapping. Therefore, the semantic similarity of Chinese words can be described according to the distance between word vectors.

In this embodiment, the first node distance between each to-be-processed node text and other to-be-processed node texts is calculated according to the node vector of each to-be-processed node text, and specifically, the preset node text Pearson correlation is obtained. The calculation formula of the coefficient is to calculate the node text Pearson correlation coefficient between the two to-be-processed node text vectors according to the calculation formula of the node vector of each node text to be processed and the Pearson correlation coefficient of the preset node text, using θ represents the node text Pearson correlation coefficient between two to-be-processed node text vectors, then the first node distance between the two to-be-processed node texts can be expressed as 1-(θ+1)/2, for each to-be-processed node text The node text is processed, and the distances between the to-be-processed node text and all other to-be-processed node texts are calculated in turn.

Step A34, performing iterative grafting processing on two to-be-processed node texts whose node distances are less than the first preset distance, until each of the to-be-processed node texts is in a convergent state in which the node-text relationship edges no longer change, wherein all The text of each pending node in the convergent state is set as the text of the convergent node;

Step A35: Generate the preset event graph based on the node text relationship edge between the convergent node text and the convergent node text.

Perform iterative grafting processing on two to-be-processed node texts whose node distances are less than the first preset distance, until each of the to-be-processed node texts is in a convergent state in which the relationship between the node texts does not change, wherein the Each node text to be processed in the state is set as the convergence node text. Specifically, for example, if it is found that the distance between the node text A to be processed and a certain node B is less than the first preset distance, such as less than 0.3, the node text to be processed is set. All the relationships of A are grafted onto the node text B to be processed, and the node text A to be processed is deleted at the same time, as shown in Figure 9, if the distance between the node text A to be processed and the node text C to be processed is smaller than the first preset distance as is less than 0.3, the relationship between the texts of the nodes to be processed as shown in Figure 10 is obtained, and the two texts of the nodes to be processed whose node distance is less than the first preset distance are grafted, until each of the texts to be processed is The processing node text is in a convergent state, that is, iteratively executes the calculation process of grafting two to-be-processed node texts whose node distance is less than the first preset distance, until each to-be-processed node text of the to-be-processed text is in Convergence state, so the relation edge (relation boundary) of each node text to be processed will not change any more, so an event graph of directed edges is generated, that is, it is considered that the graph has reached a convergent state. It should be noted that, in the fusion In the process, it is also necessary to deal with the relationship edge of the "parallel" table in the event graph. For example, suppose there is the following Figure 11, the tail node D of the "parallel" directed edge of the table is grafted to the head node B of the directed edge. On the parent to-be-processed node text A, create a new directed edge that is the same as that between A and B to connect A and D, and obtain Figure 12.

Step S40, according to the operation behavior, screen out the selected text from the first candidate text and the second candidate text, and recommend the selected text to the target user.

After the second candidate text and the first candidate text are obtained, the selected text is comprehensively screened from the first candidate text and the second candidate text, and the selected text is recommended to the target user without Only the selected text is filtered out from the first candidate text, and the selected text is recommended to the target user.

The present invention monitors the operation behavior of the target user, and determines the keywords associated with the target user according to the operation behavior; after acquiring the keywords, retrieves one or more documents containing at least one of the keywords from a preset text database set The updated text is regarded as the first candidate text; after obtaining the first candidate text, the preset event map is called, and the preset event map is selected from the preset text database set according to the preset event map. The updated text whose total correlation degree of candidate text is not less than the preset correlation threshold is regarded as the second candidate text. The acquisition of the second candidate text expands the selection category of the candidate text in the recommendation process. Each association has its corresponding degree of association; according to the operation behavior, the selected text is screened out from the first candidate text and the second candidate text, and the selected text is recommended to the target user. That is, in this application, the selected text is not only selected from the first candidate text searched according to the keywords, but the second candidate text and the first candidate text obtained according to the preset event map, etc. The selected text is selected from the collection, thus avoiding the simplification of content recommendation, and because the association relationship between the reference text and the text in this application is used to recommend content instead of just recommending based on keywords alone, this application can Improve recommendation accuracy.

Further, on the basis of the first embodiment, the present invention provides another embodiment of the text recommendation method. The preset event map is retrieved, and the preset event map is retrieved from the preset text database according to the preset event map. Selecting an update text whose total degree of association with the first candidate text is not less than a preset association threshold in the set, as the second candidate text, the steps include:

Step S31, fetching a preset event map, and determining whether there is a convergence node text containing the keyword in the corresponding clause in the event map;

Step S32, if there is, set the convergence node text containing the keyword in the corresponding clause as the user's attention node text, and select from the text updated in the preset text database within the preset time period. For the third candidate text other than the first candidate text, a word segmentation process is performed on the title of each text in the third candidate text by using a preset word segmentation tool to obtain the title vector of each text in the third candidate text;

Retrieve a preset event graph, and determine whether there is a convergence node text that contains the keyword in the corresponding clause, wherein the corresponding clause may contain the keyword or may not contain the key word, if there is no convergent node text containing the keyword in the corresponding clause in the event graph, no subsequent processing is performed, and the selected text can be directly selected from the first candidate text for recommendation. If there is a convergence node text that contains the keyword in the corresponding clause, the node text to be processed is marked as "user concerned node text", and the text updated in the preset time period from the preset text database Select the third candidate text outside the first candidate text, perform word segmentation processing on the title of each text in the third candidate text by a preset word segmentation tool, and obtain the title vector of each text in the third candidate text, specifically, Perform preset stammer segmentation for the title of each text in the third candidate text, and obtain the title vector of each text in the third candidate text with the help of the preset word2vec tool. The purpose of obtaining the title vector is to calculate the second node distance.

Step S33, calculating the second node distance between the title vector and the node vector of each convergent node text in the event map;

Step S34, selecting from the third candidate text that the second node distance is less than the second preset distance, and the convergence node text corresponding to the less than the second preset distance is the first target text of the node text that the user pays attention to, or from In the third candidate text, the distance between the selected second node is less than the second preset distance, and there is a second target of the node text that the user pays attention to within the preset screening logic depth range of the convergent node text corresponding to the less than the second preset distance The text, wherein the screening logic depth is determined according to the degree of association of the associations in the event map;

Calculate the second node distance between the title vector and the node vector of each convergent node text in the event graph, if the distance is less than the second preset distance, such as less than 0.4, and the node text to be processed is "Users focus on node text" ”, then it is determined that the text corresponding to the title vector is the first target text, if the distance is less than the second preset distance, such as less than 0.4, and within the preset preset filtering logical depth range of the node text node to be processed, such as the logical depth is 2 If there are other nodes marked as "user-focused node text" in the range, the text corresponding to the title vector is reserved as the second target text, wherein the screening logic depth is based on the relationship between the relationships in the event map. The correlation degree is determined, that is, the logical depth of screening can be defined as follows: the logical depth of the edge of the table "parallel" logical relationship is recorded as 0.5, the logical depth of the edge of the table "sequential" logical relationship is recorded as 0.7, the table "causal" logical relationship and The logical depth of the edge of the logical relationship in the table "condition" is denoted as 1, and the logical depth between the texts of two nodes to be processed is the sum of the logical depths of all edges between the nodes. The fastest relationship between texts C can be achieved through two edges, and the two edges represent the “causal” logical relationship and the “successive” logical relationship respectively, then the screening logic depth of the node text to be processed in B or the corresponding title vector of the text of the text. The logical depth is 1.7, which is within the logical depth of 2.

Step S35, set the first target text and the second target text as the second candidate text.

After the first target text and the second target text are obtained, the first target text and the second target text are set as the second candidate text.

In this embodiment, by retrieving a preset event graph, it is judged whether there is a convergence node text that contains the keyword in the corresponding clause; The convergence node text of the keyword is set as the user's attention node text, and the third candidate text except the first candidate text is selected from the texts updated in the preset text database within the preset time period, and the preset word segmentation tool is used. Perform word segmentation processing on the title of each text in the third candidate text to obtain the title vector of each text in the third candidate text; calculate the relationship between the title vector and the node vector of each convergent node text in the event map The second node distance is selected from the third candidate text, and the second node distance is less than the second preset distance, and the convergence node text corresponding to the less than the second preset distance is the first target text of the node text that the user pays attention to. , or select from the third candidate text that the second node distance is less than the second preset distance, and the preset screening logic depth range of the convergent node text corresponding to the less than the second preset distance contains the node text that the user is concerned about. The second target text, wherein the screening logic depth is determined according to the correlation degree of the respective associations in the matter graph; the first target text and the second target text are set as the second candidate texts . This embodiment realizes the accurate acquisition of the second candidate text, and lays a foundation for realizing accurate text recommendation.

Further, on the basis of the above embodiment, the present invention provides another embodiment of the text recommendation method. In this embodiment, according to the operation behavior, the first candidate text and the second candidate text are The steps of filtering out the selected text from the text and recommending the selected text to the target user include:

Step S41, obtaining the dissemination volume of each text in the first candidate text and the second candidate text, and obtaining the correlation between each text and the target user, and obtaining the target user's information according to the operation behavior. preference;

Retrieve the title of each text in the first candidate text and the second candidate text in a preset search engine to obtain the spread of each text, wherein the spread of each text reflects the text In this embodiment, it is considered that: two news with the same title belong to two reposts of the same news, and the calculation steps of the spread can be as follows: first delete the title of the first candidate text and the second candidate text. All punctuation marks (due to the process of text collection, some punctuation marks may be modified from half-width to full-width, in addition, some media will also modify some punctuation from half-width to full-width, or from full-width when forwarding text Modified to half-width, so when calculating the spread here, the difference in punctuation in the title is not considered), and then use the title after removing all punctuation to retrieve a preset number of texts from the preset search engine, such as 1000 texts (usually In this case, the maximum forwarding volume of a text is on the order of 100, and it will not exceed 1000), delete the punctuation marks in the titles of the retrieved 1000 texts, and count the current first candidate text and the second candidate text. The number of titles that are completely consistent with the candidate text titles is used as the spread of the current news α.

Wherein, the step of obtaining the relevance between each text and the target user includes:

Step S41, obtain the number of times that the keyword appears in each text of the first candidate text, and set the number of times as the number of words;

Step S42, obtain the position where the keyword appears in each text of the first candidate text, set the position as a word position, and obtain a preset position weight corresponding to the word position, wherein the word position Different, the position weight is different, and the word position includes the position of the first sentence of the first paragraph of the text, the position of the first sentence of the last paragraph of the text, the position of the non-first sentence of the first paragraph of the text, the position of the non-first sentence of the last paragraph of the text, the position of the first sentence of the non-first paragraph and the non-tail position. the position of the first sentence of the paragraph;

Step S43: Obtain the ratio of the number of sentences between the positions where the keyword appears for the first time and the last time in each text of the first candidate text to the total number of sentences in the full text, and set the ratio as a word. span;

Step S44, obtaining the target text between the positions where the keyword appears for the first time and the last time in each text of the first candidate text, and obtaining the average number of sentences in the target text that contains the The number of stated keywords, and the number of said keywords contained in the average number of preset sentences is set as the word density;

Step S45, obtaining the first relevance of each text in the first candidate text according to the word count, the preset position weight corresponding to the word position, the word span and the word density;

The correlation calculation of the first candidate text and the second candidate text is different, as shown in FIG. 13 .

The first correlation degree of the first candidate text is calculated as follows: obtaining the word count, word position, word span, word density, etc. of the keyword appearing in each text of the first candidate text, and according to the word count, word position, etc. , word span, word density to obtain the first relevance of each text of the first candidate text, specifically, the number of words a: the total number of keywords appearing in the text body; word position b: the initial time b The value is 0. If the keyword appears in the first sentence of the first paragraph or the first sentence of the last paragraph of the text, add 2 to b; if the keyword appears in the first sentence or the last sentence of the text, add 2 b plus 1; if the keyword appears in the first sentence of the rest of the paragraphs except the first and last paragraphs, add 0.5 to b; word span c: the keyword is between the first and last occurrences of the keyword in the text The ratio of the number of sentences in the interval to the total number of sentences in the full text; word density d: intercept the text between the first and last occurrences of keywords in the text. In this part of the text, the average number of sentences per preset magnitude is If the number of keywords contained in every 10 sentences is defined as the word density d, the correlation calculation formula is: β=0.3a+b+0.1c*d.

Step S46: Obtain the screening logic depth of each text in the second candidate text, and determine the second relevance of each text in the second candidate text according to the screening logic depth.

The second candidate text does not contain keywords, therefore, the second degree of relevancy of each text of the second candidate text is determined according to the depth of the screening logic. Specifically, the second degree of relevancy is defined as β=6*0.8 ^l , where l Indicates the logical depth of screening, specifically, the logical depth of screening refers to the logical depth that exists between the initial node text of the political graph to the corresponding user-focused node text or the depth of the least logical edges it contains in the process of screening the second candidate text Wait.

The step of acquiring the preference degree of the target user according to the operation behavior includes:

Step S47, obtaining the historical browsing text of the target user from the operation behavior, obtaining the first document vector of each text in the historical browsing text, and obtaining the first candidate text and the second candidate text the second document vector for each text in ;

Step S48: Obtain a first Pearson correlation coefficient between the second document vector and the first document vector, and obtain the preference of the target user according to the first Pearson correlation coefficient.

In this embodiment, the historical browsing text of the target user text is obtained, and the historical browsing text may be the historical browsing text in the past month, the first document vector of each text in the historical browsing text is obtained, and the the second document vector of each text in the first candidate text and the second candidate text, and obtain the first Pearson correlation coefficient between the second document vector and the first document vector, according to the first The Pearson correlation coefficient is used to obtain the preference of the target user, and the user preference is obtained according to the historical browsing text, the first candidate text and the second candidate text.

Step S42: Screen out the selected text from the first candidate text and the second candidate text according to the communication amount, the correlation degree and the preference degree, and recommend the selected text to the selected text. describe the target users.

In this embodiment, the selected text is screened out from the first candidate text and the second candidate text, and the selected text is recommended to the target user.

In this embodiment, by acquiring the dissemination volume of each text in the first candidate text and the second candidate text, and acquiring the relevance of each text and the target user, the information is acquired according to the operation behavior. the preference of the target user; according to the spread, the relevancy and the preference, screen out the selected text from the first candidate text and the second candidate text, and select the selected text from the first candidate text and the second candidate text. The text is recommended to the target user. In this embodiment, three factors are considered to screen the selected text, thereby improving the accuracy of the recommendation.

Further, on the basis of the above embodiment, the present invention provides another embodiment of a text recommendation method. In this embodiment, the step of obtaining the first document vector of each text in the historical browsing text includes:

Step B1, obtaining, according to a preset clustering algorithm, a first probability matrix in which each text in the historical browsing text is divided into a first preset category;

Step B2, obtaining the word segmentation of each text in the historical browsing text according to a preset word segmentation algorithm, and obtaining a second probability matrix in which the word segmentation is divided into a second preset category;

Step B3, obtaining each optimized word vector corresponding to each text in the historical browsing text according to the first probability matrix and the second probability matrix;

Step B4, obtaining the first document vector of each text in the historical browsing text according to the optimized word vector.

Obtain a first probability matrix in which each text in the historical browsing text is divided into a first preset category (including the number of 200 text subcategories) according to a preset clustering algorithm. Specifically, LDA (Latent DirichletAllocation, Implicit Dirichlet distribution) algorithm performs unsupervised clustering of historical browsing texts (the number of clusters can be set to 200) to obtain the first probability matrix p that each text is divided into the first preset category, Obtain the word segmentation of each text in the historical browsing text according to the preset word segmentation algorithm, and obtain the second probability matrix q in which the word segmentation is divided into the second preset category (including the number of sub-categories of 200 words), according to The first probability matrix and the second probability matrix obtain each optimized word vector W corresponding to each text in the historical browsing text, W=0.6p+0.4q, and obtain the historical browsing according to the optimized word vector The first document vector for each text in the text. Specifically, after adding the optimized word vectors of all the segmented words, the document vector of the corresponding news is obtained.

obtaining the first Pearson correlation coefficient between the second document vector and the first document vector, and obtaining the preference of the target user according to the first Pearson correlation coefficient,

Obtain the first Pearson correlation coefficient between the second document vector and the first document vector, and obtain the preference of the target user according to the first Pearson correlation coefficient, first from a preset text database collection Retrieve the set of texts (N ₁ , N2, ..... Nk) that the current target user has clicked and browsed in history, with a total of k texts. Then, the first candidate text, the second candidate text and the retrieved k texts such as news are respectively subjected to preset stuttering word segmentation processing, and the word vectors of all word segmentations are added to obtain the document vector V of the corresponding text. Let (V ₁ ,V2,.....Vk) represent the document vector of each text in the historically browsed text set (N ₁ ,N2,.....Nk), and let Vv represent the current first candidate text and the document vector of each text in the second candidate text, use ρ(a, b) to represent the first Pearson correlation coefficient between the two document vectors (a, b), then the first candidate text and the The user preference of the second candidate text can be expressed as

Wherein, Vv in the formula represents the document vector of each text in the first candidate text and the second candidate text, and _Vj represents the document vector of each text in the historically browsed text set.

The step of obtaining the first Pearson correlation coefficient between the second document vector and the first document vector, and obtaining the preference degree of the target user according to the first Pearson correlation coefficient includes:

Step C1, obtaining the historical browsing time between the time when each text in the historical browsing text is clicked to browse to the current moment;

Step C2: Obtain the first Pearson correlation coefficient between the second document vector and the first document vector, and perform interest reduction processing on the first Pearson correlation coefficient according to the historical browsing time, to obtain the first Pearson correlation coefficient. Two Pearson correlation coefficient;

Step C3, obtaining the preference of the target user according to the second Pearson correlation coefficient.

Take into account that the preferences of target users may shift significantly over time. For example, a week ago, the text that an operator cared about most was the company's new product launch conference, and the text that he cared most about now was the public's evaluation of the company's new products. Therefore, it is necessary to perform time interest reduction processing on the texts that the user has clicked and browsed in the history. First, obtain the historical browsing time between the time when each text in the historically browsed text is clicked and browsed to the current moment. Specifically, use t _k indicates that the text N _k was clicked and browsed before t _k days, that is, the historical browsing time is t _k , and the first Pearson correlation coefficient between the second document vector and the first document vector is obtained, according to the The historical browsing time performs interest reduction processing on the first Pearson correlation coefficient to obtain a second Pearson correlation coefficient, and obtains the preference of the target user according to the second Pearson correlation coefficient, then the first candidate text and The user preference of the second candidate text is finally expressed as

In this embodiment, by obtaining the historical browsing time between the time when each piece of historical browsing text is clicked and browsing to the current moment; the first document vector between the second document vector and the first document vector is obtained. Pearson correlation coefficient, performing interest downweight processing on the first Pearson correlation coefficient according to the historical browsing time to obtain a second Pearson correlation coefficient; obtaining the preference of the target user according to the second Pearson correlation coefficient Spend. In this embodiment, the preference of the target user can be accurately obtained, which lays a foundation for accurate recommendation.

Further, on the basis of the above-mentioned embodiment, the present invention provides another embodiment of the text recommendation method. In this embodiment, the The steps of filtering out the selected text from the first candidate text and the second candidate text, and recommending the selected text to the target user include:

Step D1, calculating the value score of each text in the first candidate text and the second candidate text according to the spread, the first correlation, the second correlation and the preference;

Step D2, selecting a preset number of texts in order from high to low according to the value score as selected texts, and recommending the selected texts to the target user.

The selected text is selected from the first candidate text and the second candidate text according to the spread, the first relevancy, the second relevancy and the user preference, and the selected text is The selected text is pushed to the target user. Specifically, the value score of each text in the first candidate text and the second candidate text is obtained through a preset calculation formula such as s=0.7αβ+0.3γ. A preset number of texts are selected in sequence at the lowest level as selected texts, and the selected texts are recommended to the target user. For example, 10 news articles with the highest scores are selected as the target content to be pushed to the target user, which can be pushed once a day.

In this embodiment, each piece of text in the first candidate text and the second candidate text is calculated according to the spread, the first correlation degree, the second correlation degree, and the preference degree value score; according to the value score, a preset number of texts are sequentially selected as selected texts from high to low, and the selected texts are recommended to the target user. In this embodiment, accurate recommendation of text is performed according to the value score.

Referring to FIG. 3 , FIG. 3 is a schematic diagram of a device structure of a hardware operating environment involved in an embodiment of the present invention.

The text recommendation device in the embodiment of the present invention may be a PC, and may also be a terminal device such as a smart phone, a tablet computer, and a portable computer.

As shown in FIG. 3 , the text recommendation device may include: a processor 1001 , such as a CPU, a memory 1005 , and a communication bus 1002 . Among them, the communication bus 1002 is used to realize the connection communication between the processor 1001 and the memory 1005 . The memory 1005 may be high-speed RAM memory, or may be non-volatile memory, such as disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001 .

Optionally, the text recommendation device may further include a target user interface, a network interface, a camera, an RF (Radio Frequency, radio frequency) circuit, a sensor, an audio circuit, a WiFi module, and the like. The target user interface may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional target user interface may also include a standard wired interface and a wireless interface. Optional network interfaces may include standard wired interfaces and wireless interfaces (eg, WI-FI interfaces).

Those skilled in the art can understand that the structure of the text recommendation device shown in FIG. 3 does not constitute a limitation on the text recommendation device, and may include more or less components than the one shown, or combine some components, or different components layout.

As shown in FIG. 3 , the memory 1005 as a computer storage medium may include an operating system, a network communication module and a text recommendation program. The operating system is a program that manages and controls the hardware and software resources of the text recommendation device, and supports the operation of the text recommendation program and other software and/or programs. The network communication module is used to realize the communication between various components in the memory 1005, as well as the communication with other hardware and software in the text recommendation device.

In the text recommendation device shown in FIG. 3 , the processor 1001 is configured to execute the text recommendation program stored in the memory 1005 to implement the steps of any one of the text recommendation methods described above.

The specific implementation manner of the text recommendation device of the present invention is basically the same as the above-mentioned embodiments of the text recommendation method, and will not be repeated here.

In addition, an embodiment of the present invention also provides a text recommendation device, where the text recommendation device includes:

a monitoring module for monitoring the operation behavior of the target user, and determining the keywords associated with the target user according to the operation behavior;

a retrieval module, used for retrieving more than one updated text containing at least one of the keywords from a preset text database set, as the first candidate text;

The selection module is used for retrieving a preset affair atlas, and according to the preset affair atlas, selects from the preset text database set the ones whose total degree of association with the first candidate text is not less than a preset association threshold. Update the text, as the second candidate text, the event graph includes the association relationship between the text and the text, and each association relationship has its corresponding association degree;

The screening module is configured to screen out the selected text from the first candidate text and the second candidate text according to the operation behavior, and recommend the selected text to the target user.

Optionally, the text recommendation device further includes:

a collection module, used for collecting texts to be processed from the preset text database set at every preset time period;

a preprocessing module, configured to perform html tag filtering, symbol filtering and clause processing on the text to be processed by using a preset regular expression to obtain preprocessed text formed by a clause list;

A generating module, configured to generate the preset event graph according to the preprocessed text.

Optionally, the generation module includes:

An identification unit, configured to identify multiple preset text associations for each clause in the clause list, and obtain node texts to be processed, wherein the preset text associations include but are not limited to Shuncheng, causality, condition and juxtaposition;

a first obtaining unit, configured to perform word segmentation processing on the node text to be processed by using a preset word segmentation tool, and obtain the word vector of each word segmentation, and obtain the node vector of each node text to be processed based on the word vector of each word segmentation;

a first computing unit, configured to calculate the first node distance between each to-be-processed node text and other to-be-processed node texts according to the node vector of each to-be-processed node text;

The grafting processing unit is used for iterative grafting processing of two to-be-processed node texts whose node distances are less than the first preset distance, until each of the to-be-processed node texts is in a convergent state where the relationship edges of the node texts no longer change, wherein , setting the text of each node to be processed in the convergent state as the text of the convergent node;

A generating unit, configured to generate the preset event graph based on the node text relationship edges between the convergent node text and the convergent node text.

Optionally, the selection module includes:

A retrieving unit, used for retrieving a preset event map, and judging whether there is a convergence node text that includes the keyword in the corresponding clause in the event map;

The first setting unit is configured to set the convergence node text containing the keyword in the corresponding clause as the user's attention node text, if it exists, and update it from the preset text database within a preset time period In the text, select the third candidate text outside the first candidate text, carry out word segmentation processing to the title of each text in the third candidate text by the preset word segmentation tool, and obtain the title vector of each text in the third candidate text;

a second computing unit for computing the second node distance between the title vector and the node vector of each convergent node text in the event graph;

a selection unit, configured to select from the third candidate texts that the second node distance is less than the second preset distance, and the convergent node text corresponding to the less than the second preset distance is the first target text of the node text that the user pays attention to, Or select the second node distance from the third candidate text that is less than the second preset distance, and the preset screening logic depth range of the convergence node text corresponding to the less than the second preset distance exists. Two target texts, wherein the screening logic depth is determined according to the degree of association of the associations in the event map;

A second setting unit, configured to set the first target text and the second target text as the second candidate text.

Optionally, the screening module includes:

The second obtaining unit is configured to obtain the dissemination volume of each text in the first candidate text and the second candidate text, obtain the correlation between each text and the target user, and obtain the information according to the operation behavior. Describe the preferences of target users;

a recommending unit, configured to filter out the selected text from the first candidate text and the second candidate text according to the spread, the relevance and the preference, and recommend the selected text to the target user.

Optionally, the second obtaining unit includes:

A first acquisition subunit, for acquiring the number of times the keyword appears in each text of the first candidate text, and setting the number of times as the number of words;

The second obtaining subunit is configured to obtain the position where the keyword appears in each text of the first candidate text, set the position as a word position, and obtain a preset position weight corresponding to the word position , wherein the word position is different, the position weight is different, and the word position includes the position of the first sentence of the first paragraph of the text, the position of the first sentence of the last paragraph of the text, the position of the non-first sentence of the first paragraph of the text, the position of the non-first sentence of the last paragraph of the text, and the first sentence of the non-first paragraph. Sentence position and non-final first sentence position;

The third obtaining subunit is used to obtain the ratio of the number of sentences between the positions where the keyword appears for the first time and the last time in each text of the first candidate text to the total number of sentences in the full text, The ratio is set to the word span;

The fourth obtaining subunit is used to obtain the target text between the positions where the keyword appears for the first time and the last time in each text of the first candidate text, and obtain the average per preset value in the target text. The number of the keywords contained in the number of sentences, and the number of the keywords contained in the average number of sentences per preset is set as the word density;

The fifth obtaining subunit is used to obtain the first correlation degree of each text in the first candidate text according to the word count, the preset position weight corresponding to the word position, the word span and the word density ;

The sixth obtaining subunit is configured to obtain the screening logic depth of each text in the second candidate text, and determine the second relevancy degree of each text in the second candidate text according to the screening logic depth.

Optionally, the second obtaining unit includes:

The seventh obtaining subunit is used to obtain the historical browsing text of the target user from the operation behavior, obtain the first document vector of each text in the historical browsing text, and obtain the first candidate text and all the texts. Describe the second document vector of each text in the second candidate text;

The eighth obtaining subunit is configured to obtain the first Pearson correlation coefficient between the second document vector and the first document vector, and obtain the preference degree of the target user according to the first Pearson correlation coefficient.

Optionally, the seventh acquisition subunit is used to realize:

Obtain a first probability matrix in which each text in the historical browsing text is classified under a first preset category according to a preset clustering algorithm;

Obtain the word segmentation of each text in the historical browsing text according to a preset word segmentation algorithm, and obtain a second probability matrix in which the word segmentation is divided into a second preset category;

Obtain each optimized word vector corresponding to each text in the historical browsing text according to the first probability matrix and the second probability matrix;

The first document vector of each text in the historical browsing text is obtained according to the optimized word vector.

Optionally, the eighth acquisition subunit is used to implement:

Obtain the historical browsing time between the time when each text in the historical browsing text is clicked to browse to the current moment;

Obtain the first Pearson correlation coefficient between the second document vector and the first document vector, and perform interest downweight processing on the first Pearson correlation coefficient according to the historical browsing time to obtain a second Pearson correlation coefficient correlation coefficient;

The preference of the target user is acquired according to the second Pearson correlation coefficient.

Optionally, the screening module includes:

a third calculation unit, configured to calculate each text in the first candidate text and the second candidate text according to the spread, the first degree of relevancy, the second degree of relevancy and the preference degree value score;

A screening unit, configured to sequentially select a preset number of texts as selected texts according to the value score from high to low, and recommend the selected texts to the target user. The specific implementation manner of the text recommendation apparatus is basically the same as that of the above-mentioned text recommendation method embodiments, and will not be repeated here.

In addition, an embodiment of the present invention also provides a text recommendation device, the device includes: a memory 109 , a processor 110 , and a text recommendation program stored in the memory 109 and executable on the processor 110 , and the text recommendation program is executed by the processor 110 At the same time, the steps of each embodiment of the above text recommendation method are implemented.

In addition, the present invention also provides a computer medium, the computer medium stores one or more programs, and the one or more programs can also be executed by one or more processors to implement the above text recommendation method Steps of each example.

The expanded contents of the specific implementation manner of the device and medium (ie, computer medium) of the present invention are basically the same as those of the above-mentioned embodiments of the text recommendation method, and are not repeated here.

It should be noted that, in the text, the terms "comprising", "comprising" or any other variation thereof are intended to encompass non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements, It also includes other elements not expressly listed or inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

The above-mentioned serial numbers of the embodiments of the present invention are only for description, and do not represent the advantages or disadvantages of the embodiments.

From the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general hardware platform, and of course hardware can also be used, but in many cases the former is better implementation. Based on this understanding, the technical solutions of the present invention can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products are stored in a storage medium (such as ROM/RAM, magnetic disk, CD-ROM), including several instructions to make a terminal (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in the various embodiments of the present invention.

The embodiments of the present invention have been described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments, which are merely illustrative rather than restrictive. Under the inspiration of the present invention, without departing from the scope of protection of the present invention and the claims, many forms can be made, which all belong to the protection of the present invention.

Claims (13)

1. A text recommendation method, characterized in that the text recommendation method comprises:

monitoring the operation behavior of a target user, and determining keywords associated with the target user according to the operation behavior;

more than one updated text containing at least one keyword is retrieved from a preset text database set to serve as a first candidate text;

calling a preset case map, selecting an updated text with the total association degree with the first candidate text not less than a preset association threshold value from the preset text database set according to the preset case map as a second candidate text, wherein the case map comprises association relations between the texts, and each association relation has a corresponding association degree;

and according to the operation behaviors, screening out a selected text from the first candidate text and the second candidate text, and recommending the selected text to the target user.

2. The method for recommending texts according to claim 1, wherein the step of retrieving a preset case map and selecting updated texts from the preset text database set according to the preset case map, wherein the total association degree with the first candidate text is not less than a preset association threshold value, comprises:

collecting texts to be processed from the preset text database set at intervals of a preset time period;

performing html label filtering, symbol filtering and clause processing on the text to be processed through a preset regular expression to obtain a preprocessed text formed by a clause list;

and generating the preset affair map according to the preprocessed text.

3. The text recommendation method of claim 2, wherein the step of generating the preset case map from the preprocessed text comprises:

identifying a plurality of preset text association relations of each clause in the clause list to obtain a node text to be processed, wherein the preset text association relations include but are not limited to sequential bearing, cause and effect, conditions and parallel relations;

performing word segmentation on the node text to be processed through a preset word segmentation tool, acquiring a word vector of each word segmentation, and acquiring a node vector of each node text to be processed based on the word vector of each word segmentation;

calculating a first node distance between each node text to be processed and other node texts to be processed according to the node vector of each node text to be processed;

performing iterative grafting processing on two node texts to be processed, of which the node distance is smaller than a first preset distance, until each node text to be processed is in a convergence state in which the node text relation side is not changed any more, wherein each node text to be processed in the convergence state is set as a convergence node text;

and generating the preset case map based on the convergent node text and the node text relation edge between the convergent node texts.

4. The text recommendation method according to claim 3, wherein the step of calling a preset case map, and selecting an updated text with a total association degree with the first candidate text not less than a preset association threshold from the preset text database set according to the preset case map as the second candidate text comprises:

calling a preset case map, and judging whether a convergence node text containing the keywords in the corresponding clause exists in the case map;

if yes, setting the convergent node text containing the keywords in the corresponding clause as a user attention node text, selecting a third candidate text except the first candidate text from the texts updated in a preset time period in the preset text database, and performing word segmentation on the title of each text in the third candidate text through a preset word segmentation tool to obtain a title vector of each text in the third candidate text;

calculating a second node distance between the title vector and a node vector of each convergence node text in the matter graph;

selecting a second node distance from the third candidate text which is smaller than a second preset distance, wherein the convergence node text corresponding to the second node distance smaller than the second preset distance is a first target text of the user attention node text, or selecting a second node distance from the third candidate text which is smaller than the second preset distance, wherein a second target text of the user attention node text exists in a preset screening logic depth range of the convergence node text corresponding to the second preset distance, wherein the screening logic depth is determined according to the association degree of each association relation in the case map;

setting the first target text and the second target text as the second candidate text.

5. The text recommendation method of any of claims 1-4, wherein said step of screening out selected text from the first candidate text and the second candidate text and recommending the selected text to the target user according to the operation behavior comprises:

acquiring the transmission quantity of each text in the first candidate text and the second candidate text, acquiring the correlation degree of each text and the target user, and acquiring the preference degree of the target user according to the operation behavior;

and screening out a selected text from the first candidate text and the second candidate text according to the propagation amount, the relevance and the preference, and recommending the selected text to the target user.

6. The text recommendation method of claim 5, wherein the step of obtaining the relevance of each text to the target user comprises:

acquiring the occurrence times of the keywords in each text of the first candidate text, and setting the occurrence times as word times;

acquiring the position of the keyword appearing in each text of the first candidate text, setting the position as a word position, and acquiring the position weight corresponding to the word position, wherein the word positions are different and the position weights are different, and the word positions comprise a first sentence position of a first text segment, a first sentence position of a last text segment, a non-first sentence position of the first text segment, a non-first sentence position of the last text segment, a non-first sentence position of the non-first text segment and a non-first sentence position of the non-last text segment;

acquiring the ratio of the number of sentences spaced between the first and last appearing positions of the keyword in each text of the first candidate text to the total number of full-text sentences, and setting the ratio as a word span;

acquiring a target text of the keywords between positions where the keywords appear for the first time and the last time in each text of the first candidate text, acquiring the number of the keywords contained in each preset sentence number in the target text, and setting the number of the keywords contained in each preset sentence number as word density;

acquiring a first relevance of each text in the first candidate text according to the word times, the position weight corresponding to the word position and preset position weight, the word span and the word density;

and obtaining the screening logic depth of each text in the second candidate texts, and determining the second relevance of each text in the second candidate texts according to the screening logic depth.

7. The text recommendation method of claim 5, wherein the step of obtaining the preference degree of the target user according to the operation behavior comprises:

obtaining historical browsing texts of the target user from the operation behaviors, obtaining a first document vector of each text in the historical browsing texts, and obtaining a second document vector of each text in the first candidate text and the second candidate text;

and acquiring a first Pearson correlation coefficient between the second document vector and the first document vector, and acquiring the preference degree of the target user according to the first Pearson correlation coefficient.

8. The text recommendation method of claim 7, wherein the step of obtaining a first document vector for each of the historically viewed texts comprises:

acquiring a first probability matrix of each text in the historical browsing texts, which is divided under a first preset category, according to a preset clustering algorithm;

obtaining word segmentation words of each text in the historical browsing text according to a preset word segmentation algorithm, and obtaining a second probability matrix of the word segmentation words under a second preset category;

obtaining each optimized word vector corresponding to each text in the historical browsing text according to the first probability matrix and the second probability matrix;

and acquiring a first document vector of each text in the historical browsing text according to the optimized word vector.

9. The text recommendation method of claim 7, wherein the obtaining a first pearson correlation coefficient between the second document vector and the first document vector, and the obtaining the preference of the target user according to the first pearson correlation coefficient comprises:

acquiring historical browsing time from the time when each text in the historical browsing texts is clicked and browsed to the current time;

acquiring a first Pearson correlation coefficient between the second document vector and the first document vector, and performing interest weight reduction processing on the first Pearson correlation coefficient according to the historical browsing time to obtain a second Pearson correlation coefficient;

and acquiring the preference degree of the target user according to the second Pearson correlation coefficient.

10. The text recommendation method of claim 6, wherein the step of selecting a selected text from the first candidate text and the second candidate text according to the propagation amount, the degree of correlation, and the degree of preference, and recommending the selected text to the target user comprises:

calculating a value score of each text in the first candidate text and the second candidate text according to the propagation amount, the first correlation degree, the second correlation degree and the preference degree;

and sequentially selecting a preset number of texts from high to low according to the value scores as selected texts, and recommending the selected texts to the target user.

11. A text recommendation apparatus, characterized in that the text recommendation apparatus comprises:

the monitoring module is used for monitoring the operation behavior of a target user and determining keywords related to the target user according to the operation behavior;

the retrieval module is used for retrieving more than one updated text containing at least one keyword from a preset text database set to serve as a first candidate text;

the selecting module is used for calling a preset case map, selecting an updated text with the total association degree with the first candidate text not less than a preset association threshold value from a preset text database set according to the preset case map as a second candidate text, wherein the case map comprises association relations between the texts, and each association relation has the corresponding association degree;

and the screening module is used for screening out a selected text from the first candidate text and the second candidate text according to the operation behavior and recommending the selected text to the target user.

12. A text recommendation apparatus, characterized in that the apparatus comprises: a memory, a processor and a text recommender stored on the memory and operable on the processor, the text recommender when executed by the processor implementing the steps of the text recommendation method as claimed in any one of claims 1 to 10.

13. A medium having stored thereon a text recommendation program which when executed by a processor performs the steps of the text recommendation method of any one of claims 1 to 10.

Images