A method and system for automatically extracting and recommending scientific and technological information

Abstract

The application discloses a kind of scientific and technological information automatic extraction and recommendation method and system, it is related to information data analysis field.The scientific and technological information automatic extraction and recommendation method, the text information of several scientific articles is acquired in real time, and is input to the classification model of pre-training and is extracted and analyzed, the correlation index of each scientific article is obtained, the historical browsing behavior time series data of each scientific topic of user is simultaneously acquired, and is comprehensively analyzed, obtains the comprehensive interest preference index of each historical period of each scientific topic of user, and is carried out moving weighted average processing, obtains the comprehensive interest preference change index of each scientific topic of user, and is combined with correlation index and is comprehensively analyzed, obtains the recommendation index of each scientific article, the present application is judged and analyzed by recommended index and preset recommended index threshold value, and is pushed based on judgment analysis result, to accurately analyze the interest change of user, to dynamically adjust recommendation strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information data analysis, in particular to a scientific and technological information automatic extraction and recommendation method and system. BACKGROUND

[0002] With the continuous development of information technology, the Internet and various information platforms have generated a large amount of scientific and technological information. At present, the traditional information retrieval method mostly relies on keyword search, which can provide certain relevant information, but when facing massive information, the retrieval efficiency is low and it is difficult to meet the personalized needs. The automatic information extraction and recommendation technology can quickly extract key information from a large information database and make personalized recommendations according to user needs through natural language processing, machine learning and data mining technology. With the continuous development of big data and artificial intelligence technology, automatic information extraction and recommendation systems are rapidly evolving.

[0003] The prior art such as the patent application with the publication number CN117743564B discloses a scientific and technological policy information automatic extraction and recommendation method and system. Based on a preset crawler strategy, data is collected from a target website to obtain scientific and technological policy source data. Key words are extracted from the scientific and technological policy source data to form feature word data. Based on a scientific and technological policy database, text conversion and semantic analysis are performed to extract entities, attributes and relationship data in the scientific and technological policy and construct a knowledge graph. According to the feature word data and user feature keywords, based on a collaborative recommendation algorithm, recommended data is searched from the knowledge graph to obtain first recommended policy data. According to user real-time website browsing data, it is judged whether the recommended data needs to be updated through the knowledge graph. If yes, based on the user real-time data and the knowledge graph, secondary data retrieval is performed to obtain second recommended policy data. Through the present application, the update time of the recommended data can be effectively located and the corresponding recommended data can be effectively inferred, thereby improving the efficiency of user analysis of scientific and technological policies.

[0004] Based on the above-mentioned scheme, the limitations of the prior art at least include the following problems. First, the prior art does not fully consider the dynamic change of the scientific interest of the user to be recommended in different time periods, which easily leads to a lag in the recommended results or a mismatch with the current actual needs of the user, thereby making it difficult to reflect the interest points of the user in real time. Second, the analysis of interactive behavior is ignored, which easily leads to neglecting the deep needs and interests of the user, thereby affecting the individualization and accuracy of the recommended results, and failing to provide scientific and technological information that best meets the needs of the user. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a scientific and technological information automatic extraction and recommendation method and system, which solves the problems that the prior art does not fully consider the dynamic change of the scientific interest of the user to be recommended in different time periods and ignores the analysis of interactive behavior.

[0006] To achieve the above object, the present application is implemented by the following technical scheme: a scientific and technological information automatic extraction and recommendation method, comprising the following steps: acquiring text information of a plurality of scientific articles to be extracted in real time and performing pretreatment; inputting the text information of each scientific article to be extracted after pretreatment into a pre-trained classification model for extraction analysis to obtain the relevance index of each scientific article to be extracted and each scientific and technological topic; simultaneously acquiring historical browsing behavior time series data of each scientific and technological topic of the user to be recommended, the historical reading behavior time series data including historical reading time series data, historical interaction time series data, and performing data analysis respectively to obtain the historical reading preference index, the historical interaction preference index of each historical period of each scientific and technological topic of the user to be recommended, and performing comprehensive analysis to obtain the comprehensive interest preference index of each historical period of each scientific and technological topic of the user to be recommended; performing moving weighted average processing on the comprehensive interest preference index of each historical period of each scientific and technological topic of the user to be recommended to obtain the comprehensive interest preference change index of each scientific and technological topic of the user to be recommended, and combining the relevance index of each scientific article to be extracted and each scientific and technological topic to perform comprehensive analysis to obtain the recommendation index of each scientific article to be extracted; judging and analyzing the recommendation index of each scientific article to be extracted with the preset recommendation index threshold value respectively, and pushing the user to be recommended based on the judgment and analysis result.

[0007] Further, the historical reading time series data includes page scroll depth value, page focus dwell time value, page scroll rate value, page pause duration value, reading duration value, reading frequency value, reading response duration value of each historical period, and the historical interaction time series data includes interaction conversion rate value, interaction frequency value, interaction response speed value.

[0008] Further, the classification model is specifically a bidirectional encoder representation model, the bidirectional encoder representation model comprises an input embedding layer, an encoder, a classification layer, an output layer, and the specific process of obtaining the relevance index of each scientific article to be extracted and each scientific topic is: in the input embedding layer of the bidirectional encoder representation model, the text information of each scientific article to be extracted is converted to obtain the initial feature representation vector of each scientific article to be extracted; in the encoder of the bidirectional encoder representation model, the initial feature representation vector of each scientific article to be extracted is processed in depth to obtain the representation vector set of each scientific article to be extracted; in the classification layer of the bidirectional encoder representation model, the context-related representation vector set of each scientific article to be extracted is classified to obtain the relevance probability of each scientific article to be extracted and each scientific topic; in the output layer of the bidirectional encoder representation model, the relevance probability of each scientific article to be extracted and each scientific topic is respectively judged and analyzed with the preset relevance probability threshold to obtain the relevance index of each scientific article to be extracted and each scientific topic.

[0009] Further, the pre-training process of the classification model is: obtaining a plurality of groups of training text information, and dividing them into a text information training set and a text information validation set; dividing the text information training set into a plurality of batch training sets, and each batch training set contains a plurality of text information; performing forward propagation processing on each batch training set to obtain the representation vector set of each text information in each batch training set, and calculating the text loss function; simultaneously performing iterative training, evaluating the training result of the classification model each time based on the back propagation algorithm and the text information validation set, and adjusting the model parameters according to the validation result until the model meets the expected standard.

[0010] Further, the specific steps of obtaining the historical reading preference index of each historical period of each science and technology topic of the user to be recommended are as follows: normalizing the page scroll depth value, the page focus dwell time value, the page scroll rate value, the page pause duration value, the reading duration value, the reading frequency value, and the reading response duration value of each historical period of each science and technology topic of the user to be recommended; comprehensively analyzing the normalized page scroll depth value, the page focus dwell time value, the page scroll rate value, and the page pause duration value of each historical period of each science and technology topic of the user to be recommended to obtain the historical attraction index of each historical period of each science and technology topic of the user to be recommended; comprehensively analyzing the normalized reading duration value, the reading frequency value, and the reading response duration value of each historical period of each science and technology topic of the user to be recommended to obtain the historical reading input index of each historical period of each science and technology topic of the user to be recommended; and comprehensively analyzing the historical attraction index and the historical reading input index of each historical period of each science and technology topic of the user to be recommended to obtain the historical reading preference index of each historical period of each science and technology topic of the user to be recommended.

[0011] Further, the specific formula for calculating the historical attraction index, the historical reading input index, and the historical reading preference index of each historical period of each science and technology topic of the user to be recommended is as follows:

[0012] the historical attraction index of the jth historical period of the ith science and technology topic of the user to be recommended, YgS′ ij YmS′ is the normalized page scroll depth value of the jth historical period of the ith science and technology topic of the user to be recommended, α1 is the depth coefficient stored in the database, YmS′ ij YdS′ is the normalized page focus dwell time value of the jth historical period of the ith science and technology topic of the user to be recommended, α2 is the focus coefficient stored in the database, YdS′ ij YtS′ is the normalized page scroll rate value of the jth historical period of the ith science and technology topic of the user to be recommended, α3 is the rate coefficient stored in the database, YtS′ ij TyD is the normalized page pause duration value of the jth historical period of the ith science and technology topic of the user to be recommended, α4 is the pause coefficient stored in the database, α1+α2+α3+α4=1, e is a natural constant, TyD ij YdS′ is the historical reading input index of the jth historical period of the ith science and technology topic of the user to be recommended, YdS′ ij YdS′ is the normalized reading duration value of the jth historical period of the ith science and technology topic of the user to be recommended, β1 is the reading duration coefficient stored in the database, YdS′ ijYdS is the normalized reading frequency value of the jth historical period of the ith scientific topic of the user to be recommended, β2 is the reading frequency coefficient stored in the database, YdX ij YdR is the normalized reading response duration value of the jth historical period of the ith scientific topic of the user to be recommended, β3 is the response duration coefficient stored in the database, β1+β2+β3=1, YdX ij YdH is the historical reading preference index of the jth historical period of the ith scientific topic of the user to be recommended, δ1 is the attraction coefficient stored in the database, δ2 is the input coefficient stored in the database, δ1+δ2=1, i=1, 2, 3, …, i0, i0 is the number of scientific topic categories, j=1, 2, 3, …, j0, j0 is the number of historical periods.

[0013] Further, the specific steps of obtaining the historical interaction preference index of each historical period of each scientific topic of the user to be recommended are as follows: standardizing the interaction conversion rate value, interaction frequency value and interaction response speed value of each historical period of each scientific topic of the user to be recommended; comprehensively analyzing the standardized interaction conversion rate value, interaction frequency value and interaction response speed value of each historical period of each scientific topic of the user to be recommended to obtain the historical interaction preference index of each historical period of each scientific topic of the user to be recommended.

[0014] Further, the formula for calculating the comprehensive interest preference index of each historical period of each scientific topic of the user to be recommended is as follows: wherein, ZhQ ij YdX is the comprehensive interest preference index of the jth historical period of the ith scientific topic of the user to be recommended, YdH ij YdH is the historical reading preference index of the jth historical period of the ith scientific topic of the user to be recommended, LhD is the reading coefficient stored in the database, ij YdH is the historical interaction preference index of the jth historical period of the ith scientific topic of the user to be recommended, LhI is the interaction coefficient stored in the database, i=1, 2, 3, …, i0, i0 is the number of scientific topic categories, j=1, 2, 3, …, j0, j0 is the number of historical periods, e is a natural constant.

[0015] Further, the specific formula for calculating the recommendation index of each scientific article to be extracted is as follows: wherein, TqZ a BhZ is the recommendation index of the ath scientific article to be extracted, ZtG i ZtG is the comprehensive interest preference change index of the ith scientific topic of the user to be recommended, μ1 is the comprehensive interest coefficient stored in the database, aiThe relevance index of the a-th scientific article to be extracted and the i-th scientific topic is denoted as mu2, mu1+mu2=1, a=1, 2, 3, …, a0, a0 is the number of scientific articles, i=1, 2, 3, …, i0, i0 is the number of scientific topic categories.

[0016] An automatic extraction and recommendation system of scientific information comprises a data acquisition module, a data extraction module, a data analysis module, a recommendation analysis module and a user recommendation module.

[0017] The automatic extraction and recommendation method of scientific information has the following beneficial effects:

[0018] (1) The automatic extraction and recommendation method of scientific information extracts the relevance of scientific articles and each scientific topic in real time, and combines the change of user historical preferences, so that the recommendation is more accurate, uses the recommendation index as the filtering standard, effectively screens out the scientific articles that best meet the user's needs for pushing, and then helps the user quickly obtain the scientific information of interest, improves the exposure rate and reading rate of the article, and then enhances the information dissemination effect and influence.

[0019] (2), the scientific and technological information automatic extraction and recommendation method, through the analysis of the user's browsing behavior time series data, thereby accurately analyzing the user's interest change for each scientific and technological theme, thereby using the comprehensive interest preference index, thereby dynamically adjusting the recommendation strategy, thereby ensuring that the recommended content is highly consistent with the user's latest interest, and enhancing the user's stickiness, thereby enabling the recommendation system to continuously provide high-quality scientific and technological content that meets the user's needs, and improving the user's reading experience.

[0020] (3), the scientific and technological information automatic extraction and recommendation system, through the cooperative work of each module, realizes efficient information extraction and intelligent recommendation, the data acquisition and extraction module guarantees the real-time and accuracy, and the data analysis module analyzes the user's historical behavior time series data, thereby ensuring that the recommended content follows the change of user demand, the weighted average processing of the recommendation analysis module and the comprehensive analysis of the correlation index make the recommendation not only accurate, but also dynamically adapt to the change of user interest, thereby significantly improving the efficiency and intelligence of the recommendation, and being able to adjust the recommendation strategy in real time, thereby providing a continuously optimized recommendation service experience.

[0021] Of course, implementing any product of the present application does not necessarily require all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A scientific and technological information automatic extraction and recommendation method flow chart of the present application.

[0023] Figure 2 A scientific and technological information automatic extraction and recommendation system block diagram of the present application. DETAILED DESCRIPTION

[0024] The problem in the embodiment of the present application is as follows:

[0025] First, the text information of the scientific articles to be extracted is obtained in real time, then input into the pre-trained classification model for extraction analysis, and the correlation index of each scientific article to be extracted and each scientific theme is obtained, the historical reading preference index and the historical interaction preference index of each historical period of each scientific theme of the user are obtained by analyzing the user's historical behavior time series data, and the comprehensive interest preference index of each scientific theme is obtained by comprehensive analysis, then the comprehensive interest preference change index is obtained by weighted average processing of the comprehensive interest preference index of each historical period of the user to be recommended, and finally the recommendation index is obtained by combining the correlation index of each scientific article and comprehensive analysis, and the scientific articles are pushed to the user according to the set recommendation index threshold.

[0026] Please refer to Figure 1The embodiment of the application provides a technical scheme: a scientific and technological information automatic extraction and recommendation method, comprising the following steps: acquiring text information of a plurality of scientific and technological articles to be extracted in real time, and performing pretreatment; inputting the text information of each scientific and technological article to be extracted after pretreatment into a pre-trained classification model for extraction analysis, to obtain a correlation index of each scientific and technological article to be extracted and each scientific and technological topic; simultaneously acquiring historical browsing behavior time sequence data of each scientific and technological topic of a user to be recommended, the historical reading behavior time sequence data comprising historical reading time sequence data and historical interaction time sequence data, and respectively performing data analysis to obtain a historical reading preference index and a historical interaction preference index of each historical period of each scientific and technological topic of the user to be recommended, and performing comprehensive analysis to obtain a comprehensive interest preference index of each historical period of each scientific and technological topic of the user to be recommended; performing moving weighted average processing on the comprehensive interest preference index of each historical period of each scientific and technological topic of the user to be recommended, to obtain a comprehensive interest preference change index of each scientific and technological topic of the user to be recommended, and combining the correlation index of each scientific and technological article to be extracted and each scientific and technological topic to perform comprehensive analysis, to obtain a recommendation index of each scientific and technological article to be extracted; respectively judging and analyzing the recommendation index of each scientific and technological article to be extracted and a preset recommendation index threshold value, and pushing the user to be recommended based on the judgment and analysis result (if the recommendation index of the scientific and technological article to be extracted is lower than or equal to the preset recommendation index threshold value, the user to be recommended is not pushed; if the recommendation index of the scientific and technological article to be extracted is higher than the preset recommendation index threshold value, the user to be recommended is pushed).

[0027] The historical reading time sequence data comprises a page scroll depth value, a page focus dwell time value, a page scroll rate value, a page pause duration value, a reading duration value, a reading times value and a reading response duration value of each historical period.

[0028] The page scroll depth value is the average value of the proportion of the page content from the top to the bottom of each article browsed by the user, expressed as a percentage of the page length, and can be obtained by a user behavior analysis tool (such as Hotjar).

[0029] The page focus dwell time value is the average value of the time of the user staying in each article, and the focus is on the dwell time of a specific area, such as the article title, picture, paragraph, etc., which can be obtained by a user behavior analysis tool (such as Hotjar).

[0030] The page scroll rate value is the average value of the scroll speed of the user in browsing each article, and a slow scroll speed means that the user reads carefully, and a fast scroll speed means that the user quickly browses the content, which can be obtained by a user behavior analysis tool (such as Hotjar).

[0031] Page dwell time value is the average of the time length that the user stays in each article without any scrolling or interaction, indicating that the user may be thinking, reading or viewing the content of a certain area, which can be obtained through user behavior analysis tools (such as Hotjar).

[0032] Reading response time value is the average of the time that the page fully responds after the user clicks or scrolls the page while browsing each article, which can be obtained through user behavior analysis tools (such as Hotjar).

[0033] Interaction conversion rate value is the proportion of users who perform certain interactive behaviors (such as likes, comments, shares, click links, etc.) after browsing the content in a period of time, that is, interaction conversion rate value = actual interaction behavior times / page view times, and actual interaction behavior times and page view times can be obtained through user behavior analysis tools (such as Hotjar).

[0034] Interaction frequency value is the sum of the number of interactions (such as collection, like, share) of the user on the theme or content in the period, and the number of each interaction can be obtained through user behavior analysis tools (such as Hotjar).

[0035] Interaction response speed value is the average of the time interval from the user initiating a certain interactive behavior (such as clicking, sliding, scrolling, inputting, etc.) to the response or feedback of the interactive behavior, which can be obtained through user behavior analysis tools (such as Hotjar).

[0036] Specifically, the classification model is specifically a bidirectional encoder representation model (BERT), which comprises an input embedding layer, an encoder, a classification layer, and an output layer, and the specific process of obtaining the relevance index of each scientific article to be extracted and each scientific topic is as follows: in the input embedding layer of the bidirectional encoder representation model, the text information of each scientific article to be extracted is converted (i.e., the input text is converted into a vector representation that the model can understand), to obtain an initial feature representation vector of each scientific article to be extracted (including a word semantic vector, a position representation vector, and a sentence distinction vector); in the encoder of the bidirectional encoder representation model, the initial feature representation vector of each scientific article to be extracted is subjected to deep processing (i.e., a multi-head self-attention mechanism, which calculates the dependency between each word and all other words in the article, thereby capturing global semantic and contextual information; a feedforward neural network, which performs a nonlinear transformation on the representation of each word to enhance its feature expression capability; a residual connection, which retains the input feature information and alleviates the gradient vanishing problem; and layer normalization, which standardizes the output of each layer to ensure the stability of the training), to obtain a (context-dependent) representation vector set of each scientific article to be extracted (i.e., the semantic and contextual information of the words in the initial feature representation vector in the context of the article); in the classification layer of the bidirectional encoder representation model, the context-dependent representation vector set of each scientific article to be extracted is subjected to a classification operation, to obtain the relevance probability of each scientific article to be extracted to each scientific topic; and in the output layer of the bidirectional encoder representation model, the relevance probability of each scientific article to be extracted to each scientific topic is respectively subjected to a judgment analysis with a preset relevance probability threshold (i.e., if it is lower than or equal to the preset relevance probability threshold, it is discarded, otherwise, it is retained and marked as a relevance index), to obtain the relevance index of each scientific article to be extracted to each scientific topic.

[0037] wherein the specific steps of the classification operation are as follows: the context-dependent representation vector set is subjected to pooling (such as average pooling or maximum pooling) to generate an article-level representation (which is a high-dimensional vector), and a fully connected layer is used to reduce the article-level representation to a feature space required for the classification task, and then an activation function is applied to convert the score into an interpretable probability value, i.e., a relevance index.

[0038] In the embodiment, the bidirectional encoder representation (BERT) can fully mine the deep semantic information in the article text, and through the multiple processing of the input embedding layer, the encoder and the classification layer, the BERT can capture the dependency between each word and other words through the multi-head self-attention mechanism, thereby better understanding the context information in the article, comprehensively understanding the semantics and context of the text, and improving the depth analysis and accurate recognition of the article content. In combination with the context-related representation vector of the BERT model and the correlation probability output by the classification layer, the highly accurate matching of the scientific article and each scientific topic can be realized, and the article-level high-dimensional vector is generated through the pooling operation, which is further converted into a correlation index by using the full connection layer and the activation function, so as to ensure that the recommendation of each article is highly consistent with the user's interest.

[0039] Specifically, the pre-training process of the classification model is as follows: a plurality of sets of training text information are obtained, and are divided into a text information training set and a text information validation set; the text information training set is divided into a plurality of batch training sets, and each batch training set contains a plurality of text information; each batch training set is processed by forward propagation to obtain a representation vector set of each text information in each batch training set, and a text loss function is calculated; iterative training is performed, the classification model is evaluated based on the back propagation algorithm and the text information validation set, and the model parameters are adjusted according to the validation result until the model meets the expected standard.

[0040] In the embodiment, the training data is divided into a plurality of batch training sets for forward propagation processing, and the model can gradually optimize its learning process, thereby avoiding the overfitting problem that may occur in a single training process, and the representation vector set of each batch training set can effectively transmit feature information, so that the classification model can better understand the semantic features of different texts. In addition, batch training can effectively improve the training efficiency and reduce the consumption of computing resources. At the same time, through iterative training using the back propagation algorithm and the text information validation set, the classification model can adjust the parameters based on the validation result after each training, so that the model can adaptively optimize its parameters and continuously improve the classification performance, thereby ensuring that the performance of the final model in actual application meets the expected standard, and the classification model can continuously adapt to new data features, thereby enhancing its prediction ability for unknown data.

[0041] Specifically, the specific steps of obtaining the historical reading preference index of each historical period of each technology topic of the user to be recommended are as follows: normalizing (i.e., removing units) the page scroll depth value, the page focus dwell time value, the page scroll rate value, the page dwell duration value, the reading duration value, the reading frequency value, and the reading response duration value of each historical period of each technology topic of the user to be recommended; comprehensively analyzing the normalized page scroll depth value, the page focus dwell time value, the page scroll rate value, and the page dwell duration value of each historical period of each technology topic of the user to be recommended to obtain the historical attraction index of each historical period of each technology topic of the user to be recommended; comprehensively analyzing the normalized reading duration value, the reading frequency value, and the reading response duration value of each historical period of each technology topic of the user to be recommended to obtain the historical reading input index of each historical period of each technology topic of the user to be recommended; and comprehensively analyzing the historical attraction index and the historical reading input index of each historical period of each technology topic of the user to be recommended to obtain the historical reading preference index of each historical period of each technology topic of the user to be recommended.

[0042] The specific formula for calculating the historical attraction index, the historical reading input index, and the historical reading preference index of each historical period of each technology topic of the user to be recommended is as follows: wherein, XyZ ij is the historical attraction index of the jth historical period of the ith technology topic of the user to be recommended, YgS′ ij is the normalized page scroll depth value of the jth historical period of the ith technology topic of the user to be recommended, and α1 is the depth coefficient stored in the database, YmS′ ij is the normalized page focus dwell time value of the jth historical period of the ith technology topic of the user to be recommended, and α2 is the focus coefficient stored in the database, YdS′ ij is the normalized page scroll rate value of the jth historical period of the ith technology topic of the user to be recommended, and α3 is the rate coefficient stored in the database, YtS′ ij is the normalized page dwell duration value of the jth historical period of the ith technology topic of the user to be recommended, and α4 is the dwell coefficient stored in the database, α1+α2+α3+α4=1, e is a natural constant, and in this embodiment example, e is 2.71, TyD ij is the historical reading input index of the jth historical period of the ith technology topic of the user to be recommended, YdS′ ij is the normalized reading duration value of the jth historical period of the ith technology topic of the user to be recommended, and β1 is the reading duration coefficient stored in the database, YdS′ ijYdS′ represents the normalized reading count of the i-th technology topic for the user to be recommended during the j-th historical period, β2 is the reading count coefficient stored in the database, and YdS′ represents the reading count coefficient. ij Let β1 be the normalized reading response time value of the i-th technology topic for the j-th historical time period of the user to be recommended, and let β3 be the response time coefficient stored in the database, β1+β2+β3=1, YdX ij Let δ1 be the historical reading preference index of the i-th technology topic for the j-th historical period of the user to be recommended, δ2 be the attraction coefficient stored in the database, δ1+δ2=1, i=1,2,3,…,i0, i0 be the number of technology topics, j=1,2,3,…,j0, j0 be the number of historical periods.

[0043] It needs to be explained that the specific acquisition process of α1, α2, α3, and α4 is as follows: read the page scroll depth, page focus dwell time, page scroll rate, and page pause duration values ​​for each historical period of each technology topic of the user to be recommended after normalization, perform mean analysis, and perform summation analysis based on the mean analysis results to obtain the attraction sum value. Then, perform a ratio analysis between the mean analysis results and the attraction sum value, and use the ratio analysis results as the corresponding coefficients.

[0044] The specific acquisition process of β1, β2, and β3 is as follows: Read the reading duration, reading frequency, and reading response duration values ​​of each historical period for each technology topic of the user to be recommended after normalization, perform mean analysis, and perform summation analysis based on the mean analysis results to obtain the input sum value. Then, perform a ratio analysis between the mean analysis results and the input sum value, and use the ratio analysis results as the corresponding coefficients.

[0045] The specific acquisition process of δ1 and δ2 is as follows: For each historical period of each technology topic of the user to be recommended, the historical attraction index and historical reading input index (both historical attraction index and historical reading input index are dimensionless indices and can be calculated) are subjected to mean analysis, and the summation analysis is performed based on the mean analysis results to obtain the preference sum value. The mean analysis results are then compared with the preference sum value to obtain the proportion analysis results, and the proportion analysis results are used as the corresponding coefficients.

[0046] In this embodiment, by analyzing the behavior data of each historical period of each technology topic of the user, the user's interests and preferences are accurately captured, which helps to more personalized recommend technology topic content that meets the user's taste, thereby improving the accuracy of the recommendation. Secondly, the historical attraction index and the historical reading input index are evaluated by multiple indicators, which avoids the limitation of relying on a single data point, thereby comprehensively reflecting the user's attention and participation in different technology topics. Through normalization processing and coefficient analysis, different behavior indicators can be converted into unified dimensionless indexes, which helps to compare and optimize uniformly in the recommendation process. In addition, the coefficient proportion analysis can provide more detailed preference index, further improve the accuracy of the recommendation. Finally, the comprehensive analysis of the historical reading preference index can recommend content that meets the user's interests and needs, improve the user's satisfaction and participation, and reduce the interference of irrelevant content.

[0047] Specifically, the specific steps of obtaining the historical interaction preference index of each historical period of each technology topic of the user to be recommended are as follows: standardizing (i.e. unit processing) the interaction conversion rate value, interaction frequency value and interaction response speed value of each historical period of each technology topic of the user to be recommended; comprehensively analyzing the standardized interaction conversion rate value, interaction frequency value and interaction response speed value of each historical period of each technology topic of the user to be recommended to obtain the historical interaction preference index of each historical period of each technology topic of the user to be recommended; wherein the specific formula for calculating the historical interaction preference index of each historical period of each technology topic of the user to be recommended is as follows: Wherein, LhD ij is the historical interaction preference index of the jth historical period of the ith technology topic of the user to be recommended, HzP ij is the standardized interaction conversion rate value of the jth historical period of the ith technology topic of the user to be recommended, φ1 is the conversion coefficient stored in the database, HdP ij is the standardized interaction frequency value of the jth historical period of the ith technology topic of the user to be recommended, φ2 is the interaction frequency coefficient stored in the database, HsD′ ij is the standardized interaction response speed value of the jth historical period of the ith technology topic of the user to be recommended, φ3 is the response speed coefficient stored in the database, φ1+φ2+φ3=1, i=1, 2, 3, …, i0, i0 is the number of technology topics, j=1, 2, 3, …, j0, j0 is the number of historical periods, e is the natural constant, and in this embodiment, the value is 2.71.

[0048] It needs to be explained that the specific acquisition process of φ1, φ2, φ3 is: reading the interaction conversion rate value, interaction frequency value, and interaction response speed value of each historical period of each technology topic of the user to be recommended after standardization processing, and performing mean value analysis, and performing summation analysis based on the mean value analysis result to obtain the interaction preference sum value, and performing proportion analysis of the mean value analysis result and the interaction preference sum value respectively, and taking the proportion analysis result as the corresponding coefficient.

[0049] In the embodiment, by standardizing the interaction conversion rate, interaction frequency, and interaction response speed and the like, different interaction behaviors can be compared on the same scale, thereby helping to accurately measure the interaction preference of the user on different historical periods and technology topics, and further improving the personalized accuracy of the recommendation. Secondly, the multiple interaction indexes are comprehensively analyzed, thereby being able to understand the interaction tendency of the user from multiple angles, which helps to reflect the interest and interaction depth of the user on a specific technology topic. Finally, according to the interaction data of the historical period and as time goes on, the recommendation strategy is dynamically adjusted based on the interaction feedback of the user, thereby enhancing the adaptability of the recommendation.

[0050] Specifically, the formula for calculating the comprehensive interest preference index of each historical period of each technology topic of the user to be recommended is as follows: wherein, ZhQ ij is the comprehensive interest preference index of the jth historical period of the ith technology topic of the user to be recommended, YdX ij is the historical reading preference index of the jth historical period of the ith technology topic of the user to be recommended, is the reading coefficient stored in the database, LhD ij is the historical interaction preference index of the jth historical period of the ith technology topic of the user to be recommended, is the interaction coefficient stored in the database, i = 1, 2, 3, …, i0, i0 is the number of technology topic categories, j = 1, 2, 3, …, j0, j0 is the number of historical periods, e is a natural constant, and in the embodiment, the value is 2.71.

[0051] It needs to be explained that, the specific acquisition process of is: reading the historical reading preference index and the historical interaction preference index of each historical period of each technology topic of the user to be recommended, and performing mean value analysis, and performing summation analysis based on the mean value analysis result to obtain the comprehensive interest preference sum value, and performing proportion analysis of the mean value analysis result and the comprehensive interest preference respectively, and taking the proportion analysis result as the corresponding coefficient.

[0052] In the embodiment, the historical reading behavior and the historical interaction behavior of the user are combined, so that the comprehensive interest preference index more comprehensively and accurately reflects the interest and preference of the user. In addition, the mean analysis and the proportion analysis are used to adjust the interest preference index according to the historical behavior and different weights of the user, thereby avoiding the limitation of a single index, and the recommendation is flexibly adapted to the change of the user behavior. The mean analysis and the proportion analysis enable the multi-dimensional data to be reasonably integrated, so as to evaluate the interest of the user from multiple angles and multiple behavior levels, and further improve the recommendation effect.

[0053] Specifically, the specific formula for calculating the recommendation index of each science and technology article to be extracted is as follows: Wherein, TqZ a is the recommendation index of the a-th science and technology article to be extracted, BhZ i is the comprehensive interest preference change index of the i-th science and technology topic of the user to be recommended, μ1 is the comprehensive interest coefficient stored in the database, ZtG ai is the relevance index of the a-th science and technology article to be extracted and the i-th science and technology topic, μ2 is the correlation coefficient stored in the database, μ1+μ2=1, a=1, 2, 3, …, a0, a0 is the number of science and technology articles, i=1, 2, 3, …, i0, i0 is the number of science and technology topic categories.

[0054] It should be explained that the specific acquisition process of μ1 and μ2 is as follows: the comprehensive interest preference change index of each science and technology topic of the user to be recommended and the relevance index of each science and technology article to be extracted and each science and technology topic are read, mean analysis is performed, and summation analysis is performed based on the mean analysis result to obtain a recommendation sum, and the mean analysis result is respectively subjected to proportion analysis, and the proportion analysis result is taken as the corresponding coefficient.

[0055] In the embodiment, the interest preference change of the user is combined with the relevance of the article, so as to ensure that the recommended science and technology article not only meets the interest change trend of the user, but also is highly relevant to the science and technology topic concerned by the user, thereby helping to improve the accuracy of the recommendation and making the recommendation more meet the dynamic needs of the user. In addition, by introducing the “comprehensive interest preference change index”, the current interest of the user is considered, and the change trend of the interest of the user is also reflected, so that the recommendation can recognize and adapt to the change of the interest of the user, and make more flexible adjustment. Then, the summation and proportion analysis based on the mean analysis are used to give different weights to the interest change of the user and the relevance of the article, so that the recommendation strategy can be adjusted in real time according to the change of the data, thereby improving the timeliness and accuracy of the recommendation.

[0056] Please refer to Figure 2This invention provides a technical solution: an automatic extraction and recommendation system for scientific and technological information, comprising a data acquisition module, a data extraction module, a data analysis module, a recommendation analysis module, and a user recommendation module. The data acquisition module is used to acquire the text information of several scientific and technological articles to be extracted in real time and perform preprocessing. The data extraction module is used to input the preprocessed text information of each scientific and technological article to be extracted into a pre-trained classification model for extraction and analysis, obtaining the relevance index between each scientific and technological article and each scientific and technological topic. The data analysis module is used to simultaneously acquire the historical browsing behavior time-series data of the user to be recommended for each scientific and technological topic, including historical reading behavior time-series data and historical interaction time-series data, and perform data analysis on each to obtain the user to be recommended. The system collects and analyzes the historical reading preference index and historical interaction preference index of each user for each technology topic in each historical period to obtain the comprehensive interest preference index of the user to be recommended for each technology topic in each historical period. The recommendation analysis module performs a moving weighted average on the comprehensive interest preference index of each user for each technology topic in each historical period to obtain the comprehensive interest preference change index of the user to be recommended for each technology topic. It also performs a comprehensive analysis on the relevance index between each technology article to be extracted and each technology topic to obtain the recommendation index of each technology article to be extracted. The user recommendation module compares the recommendation index of each technology article to be extracted with a preset recommendation index threshold and pushes recommendations to the user to be recommended based on the results of the comparison analysis.

[0057] In summary, this application has at least the following effects:

[0058] By extracting the relevance of technology articles to each technology topic in real time and combining it with changes in users' historical preferences, the recommendations become more accurate. Using a recommendation index as a filtering standard, the system effectively selects and pushes technology articles that best meet users' needs, thereby helping users quickly obtain technology information that interests them and increasing the exposure and readership of articles, thus enhancing the dissemination effect and influence of information.

[0059] By analyzing users' browsing behavior data over time, the system can accurately determine changes in users' interest in each technology topic. By utilizing a comprehensive interest preference index, the recommendation strategy can be dynamically adjusted to ensure that recommended content closely matches users' latest interests, thereby enhancing user engagement. This allows the recommendation system to continuously provide high-quality technology content that meets user needs, thus improving the user's reading experience.

[0060] Through the cooperative work of each module, efficient information extraction and intelligent recommendation are realized, the data acquisition and extraction module guarantees real-time and accuracy, and the data analysis module ensures that the recommended content closely follows the change of user demand by analyzing the historical behavior time series data of the user, the weighted average processing of the recommendation analysis module and the comprehensive analysis of the correlation index make the recommendation not only accurate, but also dynamically adapt to the change of user interest, thereby significantly improving the efficiency and intelligence of the recommendation, and being able to adjust the recommendation strategy in real time, and thus providing a continuously optimized recommendation service experience.

[0061] Although preferred embodiments of the application have been described herein, changes and modifications can be suggested to one skilled in the art, and it is intended that the application encompass such changes and modifications as fall within the scope of the appended claims.

[0062] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A method for automatically extracting and recommending scientific and technological information, characterized in that, Includes the following steps: The text information of several scientific articles to be extracted is acquired in real time and preprocessed. The preprocessed text information of each scientific article to be extracted is input into a pre-trained classification model for extraction and analysis, so as to obtain the correlation index between each scientific article to be extracted and each scientific topic. The classification model is specifically a bidirectional encoder representation model, which includes an input embedding layer, an encoder, a classification layer, and an output layer. The specific process for obtaining the relevance index between each scientific article and each scientific topic is as follows: In the input embedding layer of the bidirectional encoder representation model, the text information of each scientific article to be extracted is transformed to obtain the initial feature representation vector of each scientific article to be extracted. In the encoder of the bidirectional encoder representation model, the initial feature representation vector of each scientific article to be extracted is processed in depth to obtain the representation vector set of each scientific article to be extracted. In the classification layer of the bidirectional encoder representation model, the context-related representation vector set of each scientific article to be extracted is classified to obtain the relevance probability of each scientific article to each scientific topic. In the output layer of the bidirectional encoder representation model, the correlation probability of each scientific article to be extracted with each scientific topic is judged and analyzed against the preset correlation probability threshold to obtain the correlation index of each scientific article to be extracted with each scientific topic. Simultaneously, historical browsing behavior time-series data of each technology topic of the user to be recommended is obtained. The historical browsing behavior time-series data includes historical reading time-series data and historical interaction time-series data. Data analysis is performed separately to obtain the historical reading preference index and historical interaction preference index of each technology topic of the user to be recommended for each historical period. Comprehensive analysis is then performed to obtain the comprehensive interest preference index of each technology topic of the user to be recommended for each historical period. The comprehensive interest preference index for each technology topic of the user to be recommended is processed by moving weighted average for each historical period to obtain the comprehensive interest preference change index for each technology topic of the user to be recommended. This is then combined with the correlation index between each technology article to be extracted and each technology topic for comprehensive analysis to obtain the recommendation index for each technology article to be extracted. The specific formula is as follows: ； in, For the first to be extracted The recommendation index of this science and technology article The first user to be recommended A comprehensive interest preference change index for various technology-related topics. The comprehensive interest coefficients stored in the database. For the first to be extracted The first science article and the first The relevance index of various technology themes The correlation coefficients are stored in the database. , 1, 2, 3, ... , For the number of science and technology articles, 1, 2, 3, ... , Number of science and technology themes; The recommendation index of each science and technology article to be extracted is compared with the preset recommendation index threshold for analysis, and recommendations are pushed to users based on the analysis results.

2. The method for automatic extraction and recommendation of scientific and technological information according to claim 1, characterized in that, The historical reading time series data includes page scrolling depth, page focus dwell time, page scrolling rate, page pause duration, reading duration, number of readings, and reading response time for each historical period. The historical interaction time series data includes interaction conversion rate, interaction frequency, and interaction response speed.

3. The method for automatic extraction and recommendation of scientific and technological information according to claim 1, characterized in that, The pre-training process for the classification model is as follows: Acquire several sets of training text information and divide them into a text information training set and a text information validation set; The text information training set is divided into several batch training sets, and each batch training set contains several text information sets. Forward propagation is performed on each batch of training sets to obtain the representation vector set of each text information in each batch of training sets, and the text loss function is calculated. Simultaneously, iterative training is performed, and the classification model's training results are evaluated based on the backpropagation algorithm and the text information validation set. The model parameters are then adjusted according to the validation results until the model meets the expected standards.

4. The method for automatic extraction and recommendation of scientific and technological information according to claim 2, characterized in that, The specific steps to obtain the historical reading preference index for each historical period of each technology topic for the user to be recommended are as follows: For each historical period of each technology topic recommended to users, the page scroll depth, page focus dwell time, page scroll rate, page pause duration, reading duration, number of reads, and reading response time are normalized. By comprehensively analyzing the normalized page scroll depth, page focus dwell time, page scroll rate, and page pause duration values ​​for each historical period of each technology topic for the user to be recommended, the historical attraction index of each historical period for each technology topic for the user to be recommended is obtained. Furthermore, a comprehensive analysis was conducted on the reading duration, number of reads, and reading response time of each historical period for each technology topic of the users to be recommended after normalization, to obtain the historical reading input index of each historical period for each technology topic of the users to be recommended. Furthermore, a comprehensive analysis was conducted on the historical attraction index and historical reading engagement index of each technology topic for each historical period of the user to be recommended, to obtain the historical reading preference index of each technology topic for each historical period of the user to be recommended.

5. The method for automatic extraction and recommendation of scientific and technological information according to claim 4, characterized in that, The specific formulas for calculating the historical attraction index, historical reading engagement index, and historical reading preference index for each historical period of each technology topic for the user to be recommended are as follows: ； in, The first user to be recommended The first of its kind in the technology theme Historical attraction index for each historical period The first normalized version of the user to be recommended The first of its kind in the technology theme Page scroll depth value for a historical time period, The depth coefficient is stored in the database. The first normalized version of the user to be recommended The first of its kind in the technology theme Page focus dwell time value for each historical period, The focus coefficients are stored in the database. The first normalized version of the user to be recommended The first of its kind in the technology theme Page scroll rate values ​​for each historical time period. The rate coefficients are stored in the database. The first normalized version of the user to be recommended The first of its kind in the technology theme Page pause duration values ​​for each historical period, This refers to the pause coefficient stored in the database. , It is a natural constant. The first user to be recommended The first of its kind in the technology theme Historical reading engagement index for each historical period The first normalized version of the user to be recommended The first of its kind in the technology theme Reading time value for each historical period, This refers to the reading time coefficient stored in the database. The first normalized version of the user to be recommended The first of its kind in the technology theme The number of reads for a historical period. The coefficient of the number of reads stored in the database. The first normalized version of the user to be recommended The first of its kind in the technology theme Reading response time value for each historical period, The response time coefficient is stored in the database. , The first user to be recommended The first of its kind in the technology theme Historical reading preference index for each historical period The attraction coefficient is stored in the database. The input coefficients are stored in the database. , 1, 2, 3, ... , For the number of science and technology themes, 1, 2, 3, ... , This represents the number of historical periods.

6. The method for automatic extraction and recommendation of scientific and technological information according to claim 2, characterized in that, The specific steps to obtain the historical interaction preference index for each historical period of each technology topic for the user to be recommended are as follows: The interaction conversion rate, interaction frequency, and interaction response speed values ​​for each historical period of each technology topic for recommended users are standardized. By comprehensively analyzing the interaction conversion rate, interaction frequency, and interaction response speed of each technology topic for each historical period of the users to be recommended after standardization, the historical interaction preference index of each technology topic for each historical period of the users to be recommended is obtained.

7. The method for automatic extraction and recommendation of scientific and technological information according to claim 1, characterized in that, The formula for calculating the comprehensive interest preference index for each technology topic of the user to be recommended for each historical period is as follows: ； in, The first user to be recommended The first of its kind in the technology theme The comprehensive interest preference index for each historical period The first user to be recommended The first of its kind in the technology theme Historical reading preference index for each historical period The reading coefficients are stored in the database. The first user to be recommended The first of its kind in the technology theme Historical interaction preference index for each historical period The interaction coefficients are stored in the database. , 1, 2, 3, ... , For the number of science and technology themes, 1, 2, 3, ... , The number of historical periods. It is a natural constant.

8. A system for automatically extracting and recommending scientific and technological information, using the method for automatically extracting and recommending scientific and technological information as described in any one of claims 1-7, characterized in that, include: Data acquisition module, data extraction module, data analysis module, recommendation analysis module, user recommendation module; The data acquisition module is used to acquire the text information of several scientific articles to be extracted in real time and perform preprocessing. The data extraction module is used to input the preprocessed text information of each scientific article to be extracted into a pre-trained classification model for extraction and analysis, so as to obtain the correlation index between each scientific article to be extracted and each scientific topic. The data analysis module is used to simultaneously acquire the historical browsing behavior time-series data of each technology topic of the user to be recommended. The historical reading behavior time-series data includes historical reading time-series data and historical interaction time-series data. The module performs data analysis on each of these data to obtain the historical reading preference index and historical interaction preference index of each technology topic of the user to be recommended for each historical period. The module then performs comprehensive analysis to obtain the comprehensive interest preference index of each technology topic of the user to be recommended for each historical period. The recommendation analysis module is used to perform a moving weighted average of the comprehensive interest preference index of each technology topic of the user to be recommended for each historical period, so as to obtain the comprehensive interest preference change index of each technology topic of the user to be recommended, and to perform a comprehensive analysis in combination with the correlation index between each technology article to be extracted and each technology topic, so as to obtain the recommendation index of each technology article to be extracted. The user recommendation module is used to compare the recommendation index of each scientific article to be extracted with a preset recommendation index threshold, and push recommendations to users based on the results of the comparison and analysis.

Images