An article recognition method and device, a computer device, and a storage medium

Abstract

The application provides an article identification method and device, computer equipment and a storage medium, relates to the technical field of artificial intelligence, which can obtain article information of a target article; input the article information of the target article into an article evaluation model to predict first information and second information of the target article, the first information representing the probability that the exposure click rate of the target article is higher than the average exposure click rate of the market, and the second information representing the probability that the single-visit reading time of the target article is longer than the average single-visit reading time of the market; and then, based on the first information and the second information of the target article, an identification result of the target article can be generated. In the application, the generation of the article identification result does not depend on the subjective experience of the staff for screening, is not affected by artificial subjective experience, reduces the distance between the article sample annotation and the real interest points that the actual users like, and improves the accuracy of the identification of the articles that attract users.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an article recognition method, apparatus, computer device, and storage medium. Background Technology

[0002] In news feed scenarios, a large number of articles are added to the database daily, with varying quality. The goal of news feed recommendation systems is to filter out and maximize the exposure of newly added high-quality articles, thereby increasing click-through rate (CTR) and improving the efficiency of the recommendation system.

[0003] Existing technologies rely on the subjective experience of staff to select articles that attract users. However, there is a significant gap between the subjective experience of staff and the actual interests of users. Therefore, there is usually a problem of low accuracy in identifying articles that attract users. Summary of the Invention

[0004] In view of this, to solve the above problems, the present invention provides an article recognition method, apparatus, computer device, and storage medium to improve the accuracy of recognizing articles that attract users. The technical solution is as follows:

[0005] An article recognition method includes:

[0006] Obtain article information of the target article, wherein the article information includes any one or more of the tags that characterize the article content of the target article and the title of the target article;

[0007] The article evaluation model predicts first and second information of the target article based on the article information. The first information represents the probability that the exposure click-through rate of the target article is higher than the average exposure click-through rate of the overall market, and the second information represents the probability that the single access reading time of the target article is higher than the average single access reading time of the overall market.

[0008] Based on the first and second information of the target article, an identification result of the target article is generated, wherein the identification result indicates whether the target article attracts users or does not attract users;

[0009] The article evaluation model is based on an article sample. It is obtained by determining the first and second information predicted by the article evaluation model based on the article information of the article sample, and training the article evaluation model with the goal of the predicted first information being close to the target first information carried by the article sample and the predicted second information being close to the target second information carried by the article sample.

[0010] An article recognition device includes:

[0011] An article information acquisition unit is used to acquire article information of a target article, wherein the article information includes any one or more of the tags characterizing the content of the target article and the title of the target article;

[0012] The information prediction unit is used to predict the first information and the second information of the target article based on the article information through the article evaluation model. The first information represents the probability that the exposure click-through rate of the target article is higher than the average exposure click-through rate of the overall market, and the second information represents the probability that the single access reading time of the target article is higher than the average single access reading time of the overall market.

[0013] The identification result generation unit is used to generate an identification result of the target article based on the first information and the second information of the target article, wherein the identification result indicates whether the target article attracts users or does not attract users;

[0014] The article evaluation model is based on an article sample. It is obtained by determining the first and second information predicted by the article evaluation model based on the article information of the article sample, and training the article evaluation model with the goal of the predicted first information being close to the target first information carried by the article sample and the predicted second information being close to the target second information carried by the article sample.

[0015] A computer device includes: a processor and a memory, the processor and the memory being connected via a communication bus; wherein the processor is used to call and execute a program stored in the memory; the memory is used to store the program, the program being used to implement the article recognition method.

[0016] A computer-readable storage medium having a computer program stored thereon, the computer program being loaded and executed by a processor to implement the steps of the article recognition method.

[0017] This application provides an article recognition method, apparatus, computer device, and storage medium. The method involves acquiring article information of a target article, including tags representing the article content and any one or more of the article's title. The article information is input into an article evaluation model to predict first and second information of the target article. The first information represents the probability that the target article's exposure-to-click rate is higher than the average exposure-to-click rate of the overall market, and the second information represents the probability that the target article's single-visit reading time is higher than the average single-visit reading time of the overall market. Based on the first and second information, an article recognition result can be generated. The article recognition result generated in this application does not rely on the subjective experience of staff for screening, is unaffected by human subjective experience, narrows the gap between the article sample annotations and the actual interests of users, and improves the accuracy of identifying articles that attract users. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 A flowchart of an article determination method provided in this application embodiment;

[0020] Figure 2 A flowchart illustrating an article evaluation model generation method provided in this application embodiment;

[0021] Figure 3 A flowchart illustrating a method for determining the first target information carried by an article sample, as provided in this application embodiment;

[0022] Figure 4 A flowchart illustrating a method for determining target second information carried by an article sample, provided in an embodiment of this application;

[0023] Figure 5 This is a schematic diagram of an article recognition model generation method provided in an embodiment of this application;

[0024] Figure 6 A flowchart of an article evaluation model optimization method provided in this application embodiment;

[0025] Figure 7 This is a schematic diagram of an article evaluation model optimization method provided in an embodiment of this application;

[0026] Figure 8This is a schematic diagram of the structure of an article determining device provided in an embodiment of this application;

[0027] Figure 9 This is a hardware structure block diagram of a computer device to which an article determination method provided in this application is applicable. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] In news feed scenarios, a large number of articles are added to the database daily, with varying quality. The goal of news feed recommendation systems is to filter out and maximize the exposure of newly added high-quality articles, thereby increasing click-through rate (CTR) and improving the efficiency of the recommendation system.

[0030] Existing technologies rely on the subjective experience of staff to select articles that attract users. However, there is a significant gap between the subjective experience of staff and the actual interests of users. Therefore, there is usually a problem of low accuracy in identifying articles that attract users.

[0031] In order to improve the accuracy of identifying articles that attract users, the inventors of this application first proposed a high-quality article identification scheme based on article titles. This scheme is based on manually formulated attractive title standards; under the guidance of prior knowledge from humans, samples are labeled and a classification model is constructed.

[0032] However, the inventors discovered through research that this solution for attracting articles also struggles to grasp the points of user interest in the titles. Standards typically derived from human experience are insufficient to select titles that truly appeal to users, often falling short of their genuine interests. Furthermore, in actual content production, user interests shift with current trending events, and the style of titles preferred by users changes to some extent as headline development and expression trends evolve. This shift in user interests and the evolution of headline styles can easily lead to the degradation of classification models, and current methods of manually selecting titles are unlikely to improve metrics such as click-through rates and viewing time on the business side.

[0033] Therefore, the inventors of this application further propose an article identification method to screen articles with high-quality titles from a massive article database. The key technical points of this method are: 1) an article evaluation model that integrates high-conversion keywords is proposed, which can combine multiple posterior indicators to meet multiple posterior constraints; 2) a complete system for constructing training samples, training models, and automatically updating and iterating models is proposed.

[0034] The article identification method mainly involves the following two points: 1) proposing a method for automatically filtering attractive articles based on an article evaluation model; 2) by designing an automatic update and iteration process for the article evaluation model, it can be ensured that the article evaluation model does not degenerate over time and can still learn the article content that users like.

[0035] In information flow platforms, this application specifically serves the scenario of pushing high-quality articles. For example, in information flow scenarios such as QQ Kandian, Tiantian Kuaibao, and browsers, a number of high-quality articles are pushed to specific users daily. These articles are often the latest and most relevant. How to filter articles with high engagement time and high click-through rates from a massive amount of content is a key to improving recommendation efficiency. This application constructs a multi-objective model for filtering high-quality articles and adds an automatic model update process, enabling the model to continuously and efficiently filter out high-quality titles.

[0036] In this embodiment, high-quality articles can be considered as articles that attract users, and articles that attract users can be considered as articles with high engagement time and high click-through rates. This embodiment provides an article identification method that relies on a pre-trained article evaluation model to identify whether an article is attractive to users.

[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] Figure 1 This is a flowchart of an article recognition method provided in an embodiment of this application.

[0039] like Figure 1 As shown, the method includes:

[0040] S101. Obtain the article information of the target article, which includes any one or more of the tags that characterize the content of the target article and the title of the target article.

[0041] In this embodiment of the application, for ease of distinction, the article to be determined as a high-quality article can be temporarily referred to as the target article. Accordingly, when determining whether a target article is a high-quality article (i.e., when determining whether a target article is an article that attracts users), it is necessary to obtain the article information of the target article. The article information of the target article includes any one or more of the tags that characterize the content of the target article and the title of the target article.

[0042] Taking the article information of the target article as an example, which includes tags that characterize the content of the target article, the process of obtaining the article information of the target article includes: determining a pre-set tag whitelist, and determining tags that match the content of the target article from the tag whitelist.

[0043] In this embodiment of the application, the tag whitelist includes at least one tag; for each of the at least one tag, it is determined whether the tag matches the article content of the target article; if the tag matches the article content of the target article, the tag is determined as a tag representing the article content of the target article.

[0044] For example, at least one tag is related to an entity in the article content. For instance, at least one tag may include the name of a variety show, guests, authors, etc. The above is merely a preferred content of at least one tag provided in the embodiments of this application. Those skilled in the art can set the specific content of at least one tag according to their own needs, and no limitation is made here.

[0045] S102. Using the article evaluation model, the first and second information of the target article are predicted based on the article information. The first information represents the probability that the exposure click-through rate of the target article is higher than the average exposure click-through rate of the overall market, and the second information represents the probability that the single access reading time of the target article is higher than the average single access reading time of the overall market.

[0046] In this embodiment of the application, the article evaluation model is based on an article sample. It is obtained by determining the first information and the second information predicted by the article evaluation model to be trained based on the article information of the article sample, with the training objective being that the predicted first information is close to the target first information carried by the article sample and the predicted second information is close to the target second information carried by the article sample.

[0047] For details on the generation process of the article evaluation model, please refer to [link / reference]. Figure 2 A detailed description of it is omitted here.

[0048] S103. Based on the first and second information of the target article, generate the identification result of the target article. The identification result indicates whether the target article attracts users or does not attract users.

[0049] In this embodiment of the application, an identification result of the target article can be generated based on the first information and the second information of the target article. The identification result of the target article indicates whether the target article attracts users or not. If the target article attracts users, it means that the target article is an article that attracts users. If the target article does not attract users, it means that the target article is an article that does not attract users.

[0050] If the first information of the target article exceeds a preset first threshold and the second information of the target article exceeds a preset second threshold, then the identification result of the target article indicates that the target article is attractive to users; if the first information of the target article does not exceed the preset first threshold or the second information of the target article does not exceed the preset second threshold, then the identification result of the target article indicates that the target article is not attractive to users.

[0051] One implementation method for generating the identification result of the target article based on the first and second information of the target article can be as follows: determine whether the first information exceeds a preset first threshold; if the first information exceeds the first threshold, determine whether the second information exceeds a preset second threshold; if the second information exceeds the second threshold, determine that the identification result of the target article indicates that the target article is attractive to users; if the first information does not exceed the first threshold or the second information does not exceed the second threshold, determine that the identification result of the target article indicates that the target article is not attractive to users.

[0052] For example, the first threshold can be 0.5 and the second threshold can be 0.5. The above is just an example of the first threshold and the second threshold provided in the embodiments of this application. Those skilled in the art can set the specific content of the first threshold and the second threshold according to their own needs, and there is no limitation here.

[0053] Figure 2 This is a flowchart illustrating a method for generating an article evaluation model, as provided in an embodiment of this application. Figure 2 As shown, the method includes:

[0054] S201. Obtain article data from the information flow platform within the first historical time period. The article data indicates at least one article pushed to the user by the information flow platform within the first historical time period, and each article in the at least one article is an article sample.

[0055] S202. Determine the article information, target primary information, and target secondary information of the article sample;

[0056] S203. Input the article information of the article sample into the article evaluation model to be trained, determine the first information and second information of the article sample predicted by the article evaluation model based on the article information of the article sample, and take the prediction of the first information of the article sample being close to the target first information of the article sample and the prediction of the second information of the article sample being close to the target second information of the article sample as the training target, and train the article evaluation model to obtain the article evaluation model.

[0057] For example, the process of training the article evaluation model to generate the article evaluation model can be as follows: obtain article samples, determine the article information, target first information, and target second information of the article samples; input the article information of the article samples into the article evaluation model to be trained, determine the first information and second information predicted by the article evaluation model to be trained based on the article information of the article samples, and take the predicted first information approaching the target first information of the article samples and the predicted second information approaching the target second information of the article samples as the training target, and train the article evaluation model to obtain the article evaluation model.

[0058] The article sample can be obtained by: acquiring article data from the information flow platform within the first historical time period. The article data indicates at least one article pushed to the user by the information flow platform within the first historical time period, and each article in the at least one article can be considered as an article sample.

[0059] Taking an article sample as an example, the method for determining the primary target information carried by the article sample is as follows: Figure 3 As shown. See also Figure 3 The methods for determining the target primary information of an article sample can be:

[0060] S301. Determine the exposure click-through rate of the article sample;

[0061] S302. Determine the average exposure click-through rate of the entire market based on the exposure click-through rate of all articles in the market.

[0062] Taking an article sample as an example, the target primary information of the article sample can be determined as follows: calculate the exposure click-through rate of the article sample; identify all articles in the overall market and calculate the exposure click-through rate of each article in the overall market; determine the average exposure click-through rate of the overall market based on the exposure click-through rates of each article in the overall market; and then determine the target primary information of the article sample based on the exposure click-through rate of the article sample and the average exposure click-through rate of the overall market.

[0063] For example, the average exposure click-through rate (CTR) of the entire market can be determined by calculating the sum of the CTRs of all articles in the market (for ease of distinction, the sum of the CTRs of all articles in the market can be called the first value), determining the total number of articles in the market (for ease of distinction, the total number of articles in the market can be called the second value), and dividing the first value by the second value to obtain the average exposure click-through rate of all articles in the market. The average exposure click-through rate of all articles in the market can be considered as the average exposure click-through rate of the entire market, i.e., the average exposure click-through rate of the entire market.

[0064] S303. Determine whether the exposure click-through rate of the article sample is not lower than the average exposure click-through rate of the overall market; if the exposure click-through rate of the article sample is not lower than the average exposure click-through rate of the overall market, proceed to step S304; if the exposure click-through rate of the article sample is lower than the average exposure click-through rate of the overall market, proceed to step S305.

[0065] S304. Determine the target primary information of the article sample as the first preset value;

[0066] S305. Determine the target first information of the article sample as the second preset value.

[0067] For example, the first preset value can be 1 and the second preset value can be 0. The above are merely preferred contents of the first and second preset values ​​provided in the embodiments of this application. Those skilled in the art can set the specific contents of the first and second preset values ​​according to their own needs, and there is no limitation here.

[0068] Taking an article sample as an example, the method for determining the target second information carried by the article sample is as follows: Figure 4 As shown. See also Figure 4 The methods for determining the target secondary information of an article sample can be:

[0069] S401. Determine the duration of a single access to an article sample;

[0070] S402. Determine the average single access reading time of the market based on the single access reading time of all articles in the market.

[0071] Taking an article sample as an example, the target second information of the article sample can be determined as follows: calculate the single access reading time of the article sample; determine all articles in the market and calculate the single access reading time of each article in the market; determine the average single access reading time of the market based on the single access reading time of each article in the market; and then determine the target second information of the article sample based on the single access reading time of the article sample and the average single access reading time of the market.

[0072] For example, the average single-access reading time of the market can be determined based on the single-access reading time of each article in the market as follows: calculate the sum of the single-access reading times of all articles in the market (for ease of distinction, the sum of the single-access reading times of all articles in the market can be called the third value), determine the total number of articles in the market (for ease of distinction, the total number of articles in the market can be called the second value), and divide the third value by the second value to obtain the average single-access reading time of all articles in the market; the average single-access reading time of all articles in the market can be considered as the average single-access reading time of the market, that is, the average single-access reading time of the market.

[0073] S403. Determine whether the single access reading time of the article sample is not less than the average single access reading time of the overall market; if the single access reading time of the article sample is not less than the average single access reading time of the overall market, proceed to step S404; if the single access reading time of the article sample is less than the average single access reading time of the overall market, proceed to step S405.

[0074] S404. Determine the target second information of the article sample as the first preset value;

[0075] S405. Determine the target second information of the article sample as the second preset value.

[0076] For example, the first preset value can be 1 and the second preset value can be 0. The above are merely preferred contents of the first and second preset values ​​provided in the embodiments of this application. Those skilled in the art can set the specific contents of the first and second preset values ​​according to their own needs, and there is no limitation here.

[0077] In this embodiment of the application, when training the article evaluation model, each article among at least one article pushed to the user by the information flow platform within a first historical time period is taken as an article sample, and the article sample carries two pieces of information: one is target first information, and the other is target second information.

[0078] Because the primary information about an article sample depends on the relationship between its exposure click-through rate (CTR) and the overall average CTR, the article evaluation model trained on this primary information can predict the relationship between the target article's CTR and the overall average CTR when predicting the target article's performance. Specifically, it predicts the probability that the target article's CTR is higher than the overall average CTR. Conversely, the secondary information about an article sample depends on the relationship between its single-access reading time and the overall average single-access reading time. The article evaluation model trained on this secondary information can predict the relationship between the target article's single-access reading time and the overall average single-access reading time when predicting the target article's performance. Specifically, it predicts the probability that the target article's single-access reading time is higher than the overall average single-access reading time.

[0079] One approach is to consider at least one article pushed to a user by the information flow platform within the first historical time period as the overall picture.

[0080] For example, when training the article evaluation model, at least one article pushed to users by the information flow platform within a first historical time period is identified, and each article is considered an article sample. The target first information of the article sample is determined based on the relationship between the exposure click-through rate (CTR) of the article sample and the average CTR of the at least one article. The target second information of the article sample is determined based on the relationship between the single-access reading time of the article sample and the average single-access reading time of the at least one article. Then, the article evaluation model is trained using the article samples carrying the target first and target second information to obtain the article evaluation model. Thus, when using the article evaluation model for prediction, inputting the article information of the target article into the model allows prediction of the target article's first and second information. The first information represents the probability that the target article's CTR is higher than the average CTR of the overall platform, and the second information represents the probability that the target article's single-access reading time is higher than the average single-access reading time of the overall platform.

[0081] Another approach is to consider the articles pushed to users by the information flow platform during the fourth historical time period as a whole, with the end time of the first historical time period earlier than the end time of the fourth historical time period, and the end time of the fourth historical time period earlier than the start time of training and generating the article evaluation model based on the article samples.

[0082] For example, when training the article evaluation model, at least one article pushed to the user by the information flow platform within a first historical time period is identified, and each of these articles is considered an article sample. Articles pushed to the user by the information flow platform within a fourth historical time period are considered as the overall data. The end time of the fourth historical time period is earlier than the start time of training the article evaluation model based on the article samples, and the end time of the fourth historical time period is later than the end time of the first historical time period. This ensures that the articles in the overall data are closer to the training and generation time of the article evaluation model, and more closely reflect the user's latest article preferences, thereby improving the accuracy of the article evaluation model's article recognition results.

[0083] The following is combined with Figure 5 The schematic diagram shown illustrates the article evaluation model generation method provided in this application embodiment, which provides a detailed explanation of the article evaluation model generation method.

[0084] For example, this application uses a multi-task model to classify articles based on their attractiveness, as follows: Figure 5 As shown, the titles and corresponding labels of the article samples are word2vec encoded and input into the article evaluation model to be trained, which is built based on the Recurrent Neural Network (RNN)-Attention mechanism. Multi-task modeling is performed, and the article samples are labeled using two posterior data dimensions, Label_ctr and label_read_dur (which can be regarded as user reading behavior data).

[0085] Figure 5 The Label_ctr is a sample label constructed based on the CTR. The label is constructed as follows: calculate the average CTR of all article samples. If the CTR of an article sample is not lower than the average CTR, the label_ctr corresponding to that article sample is marked as 1. If the CTR of an article sample is lower than the average CTR, the label_ctr corresponding to that article sample is marked as 0. The label_ctr corresponding to an article sample can be regarded as the target primary information of the article sample.

[0086] Figure 5The label_read_dur is a sample label constructed based on the single access reading time. The label is constructed as follows: calculate the single access reading time of each article sample and the average single access reading time of all article samples. If the single access reading time of an article sample is not less than the average single access reading time, the label_read_dur of the article sample is 1. If the single access reading time of an article sample is less than the average single access reading time, the label_read_dur of the article sample is marked as 0. The label_read_dur of an article sample can be regarded as the target secondary information of the article sample.

[0087] For example, the reading time of an article sample can be divided by the total number of visits to the article sample to obtain the reading time per visit.

[0088] In this embodiment of the application, each article sample corresponds to two labels, namely the first target information and the second target information. Since the goal is to select high-quality articles from the two label directions, a high-quality article selection model (the article selection model can also be called the article evaluation model) is constructed using the binary cross-entropy (BCE) loss function.

[0089] Accordingly, during the model prediction process, it is necessary to combine the scores of the two labels, label_ctr and label_read_dur. When the score of label_ctr of the target article is higher than the first threshold and the score of label_read_dur of the target article is higher than the second threshold, the target article can be considered as a high-quality article.

[0090] Furthermore, the article recognition method provided in this application embodiment can also optimize the article evaluation model; the optimization method is described in [reference needed]. Figure 6 .

[0091] like Figure 6 As shown, the method includes:

[0092] S601. Using the article data from the information flow platform in the second historical time period, train the article evaluation model to be trained to obtain the target article evaluation model, wherein the second historical time period is later than the first historical time period.

[0093] S602. Using an article evaluation model, predict the identification results of each article pushed to users by the information flow platform in the third historical time period to obtain the first identification result set; the third historical time period is later than the second historical time period.

[0094] For example, an information flow platform can be a content push platform. The above is merely a preferred embodiment of the information flow platform provided in this application. Those skilled in the art can set the specific content of the information flow platform according to their own needs, and it is not limited here.

[0095] In this embodiment of the application, an article platform model is used to predict the identification results of each article pushed to the user by the information flow platform in a third historical time period, and the identification results of each article constitute the first identification result set.

[0096] S603. Using the target article evaluation model, predict the identification result of each article pushed to the user by the information flow platform in the third historical time period, and obtain the second identification result set.

[0097] In this embodiment of the application, the target article platform model is used to predict the identification result of each article pushed to the user by the information flow platform in the third historical time period, and the identification results of each article constitute the second identification result set.

[0098] S604. Determine the standard identification results of each article pushed to users by the information flow platform in the third historical time period, and obtain the third identification result set;

[0099] For example, when a news feed platform pushes articles to users in a third historical time period, it can determine the standard identification result of the article based on the user's feedback. For instance, if the user's feedback indicates dislike, it can be determined that the standard identification result indicates that the article is not attractive to the user; if the user's feedback indicates liking, it can be determined that the standard identification result indicates that the article is attractive to the user.

[0100] The above is merely a preferred method for determining the standard identification result of an article provided in the embodiments of this application. The specific method for determining the standard identification result of an article can be set by those skilled in the art according to their own needs, and is not limited here.

[0101] The information flow platform pushes the standard recognition results of each article to users within the third historical time period, which constitute the third recognition result set.

[0102] S605. If the second identification result set is closer to the third identification result set than the first identification result set, update the article evaluation model to the target article evaluation model.

[0103] Because the third historical time period is later than the second historical time period, and the second historical time period is later than the first historical time period, the target article evaluation model uses slightly more up-to-date article samples during training compared to the currently used article evaluation model. By predicting the articles pushed to users by the information flow platform within the third historical time period using both the current article evaluation model and the target article evaluation model, it can be determined whether the current article evaluation model or the target article evaluation model is more suitable for the current situation. If the target article evaluation model is more suitable, the currently used article evaluation model is updated to the target article evaluation model. This reduces the problem of decreased accuracy in article recognition due to changes in time, trending events, and user preferences.

[0104] The automatic update process of the article evaluation model is as follows: As time goes by, trending events change, and users' preferred article content also changes. To prevent the model from degrading when new content emerges, a hot update process is used to update and iterate the model, ensuring that its performance remains at a relatively good level. The specific process of automatic model update is as follows: Figure 7 As shown. When training the new model, data from the past month is used to extract posterior data such as title, tags, click-through rate, and viewing time. Data from the first 28 days of the past month is selected as training data, according to... Figure 5 The data is constructed using a data labeling method. The model for evaluating articles is trained using data from the previous 28 days to generate a new article evaluation model. Furthermore, the currently used online article evaluation model is considered the old model. Data from the last two days of the past month is used as the test set. The old and new article evaluation models are tested on this test set, and metrics such as AUC, recall, and precision are compared. If the new article evaluation model outperforms the old model in all these metrics, the old model is automatically replaced by the new one.

[0105] This application provides an article recognition method that uses a multi-task approach and posterior data for modeling. Compared to traditional methods that rely on prior knowledge of operational products, this method builds data that is more closely aligned with user preferences and reduces reliance on manually labeled datasets to some extent. Furthermore, the automatic model update and iteration system allows the model to continuously update and optimize, consistently learning from trending and classic content favored by platform users, preventing model degradation, and continuously improving the model.

[0106] Experiments have shown that the article recognition method provided in this application, when fully deployed in information flow scenarios such as Kandian and Kandian Daily, improves both the main feed click-through rate and the user conversion rate when applied to the high-quality content pool in the Kandian scenario. In the Kandian Daily push scenario, both the first-time conversion rate and the user conversion rate show significant improvements.

[0107] Figure 8 This is a schematic diagram of the structure of an article recognition device provided in an embodiment of this application.

[0108] like Figure 8 As shown, the device includes:

[0109] The article information acquisition unit 801 is used to acquire the article information of the target article, which includes any one or more of the tags that characterize the content of the target article and the title of the target article.

[0110] The information prediction unit 802 is used to predict the first information and the second information of the target article based on the article information through the article evaluation model. The first information represents the probability that the exposure click-through rate of the target article is higher than the average exposure click-through rate of the whole market, and the second information represents the probability that the single access reading time of the target article is higher than the average single access reading time of the whole market.

[0111] The recognition result generation unit 803 is used to generate the recognition result of the target article based on the first information and the second information of the target article. The recognition result indicates whether the target article attracts users or does not attract users.

[0112] The article evaluation model is based on article samples. It is trained by determining the first and second information predicted by the article evaluation model based on the article information of the article samples, with the training objective being that the predicted first information is close to the target first information carried by the article samples and the predicted second information is close to the target second information carried by the article samples.

[0113] In this embodiment of the application, preferably, the article information of the target article includes tags that characterize the article content of the target article; correspondingly, the article information acquisition unit is used to determine a pre-set tag whitelist and determine tags that match the article content of the target article from the tag whitelist.

[0114] In this embodiment of the application, preferably, the recognition result generation unit includes:

[0115] The first generation unit is used to determine that the recognition result of the target article represents that the target article attracts users if the first information of the target article exceeds a preset first threshold and the second information of the target article exceeds a preset second threshold.

[0116] The second generation unit is used to determine that the identification result of the target article indicates that the target article is not attractive to users if the first information of the target article does not exceed the first threshold or the second information of the target article does not exceed the second threshold.

[0117] Furthermore, the article recognition device provided in this application embodiment also includes a model generation unit, which includes:

[0118] The article data acquisition unit is used to acquire article data of the information flow platform within the first historical time period. The article data indicates at least one article pushed to the user by the information flow platform within the first historical time period, and each article in the at least one article is an article sample.

[0119] The information determination unit is used to determine the article information, target first information, and target second information of the article sample.

[0120] The training unit is used to input the article information of the article sample into the article evaluation model to be trained, determine the first information and the second information of the article sample predicted by the article evaluation model based on the article information of the article sample, and use the prediction of the first information of the article sample approaching the target first information of the article sample and the prediction of the second information of the article sample approaching the target second information of the article sample as the training objective to train the article evaluation model to obtain the article evaluation model.

[0121] In this embodiment of the application, the preferred information determining unit for determining the target first information of the article sample includes:

[0122] The first determining unit is used to determine the exposure click-through rate of the article sample and the average exposure click-through rate of all articles in the overall market;

[0123] The second determining unit is used to determine the target first information of the article sample as the first preset value if the exposure click rate of the article sample is not lower than the average exposure click rate.

[0124] The third determining unit is used to determine the target first information of the article sample as the second preset value if the exposure click-through rate of the article sample is lower than the average exposure click-through rate.

[0125] In this embodiment of the application, preferably, the information determining unit for determining the target second information of the article sample includes:

[0126] The fourth determination unit is used to determine the single access reading time of the article sample and the average single access reading time of all articles in the overall market;

[0127] The fifth determining unit is used to determine the target second information of the article sample as the first preset value if the single access reading time of the article sample is not less than the average single access reading time.

[0128] The sixth determining unit is used to determine the target second information of the article sample as the second preset value if the single access reading time of the article sample is lower than the average single access reading time.

[0129] Furthermore, the article determination apparatus provided in this application embodiment further includes a model optimization unit, which includes:

[0130] The target article evaluation model generation unit is used to train the article evaluation model to be trained using article data from the information flow platform in the second historical time period, and obtain the target article evaluation model. The second historical time period is later than the first historical time period.

[0131] The first prediction unit is used to use the article evaluation model to predict the identification result of each article pushed to the user by the information flow platform in the third historical time period, and obtain the first identification result set; the third historical time period is later than the second historical time period.

[0132] The second prediction unit is used to use the target article evaluation model to predict the identification result of each article pushed to the user by the information flow platform in the third historical time period, and obtain the second identification result set.

[0133] The standard identification result determination unit is used to determine the standard identification result of each article pushed to the user by the information flow platform in the third historical time period, and obtain the third identification result set.

[0134] The model optimization subunit is used to update the article evaluation model to the target article evaluation model if the second identification result set is closer to the third identification result set than the first identification result set.

[0135] like Figure 9 The diagram shown is a structural diagram of one implementation of a computer device provided in this application. The computer device includes:

[0136] Memory 901 is used to store programs;

[0137] Processor 902 is used to execute programs, specifically for:

[0138] Obtain the article information of the target article, which includes tags that characterize the content of the target article and any one or more items from the title of the target article;

[0139] The article evaluation model predicts the first and second information of the target article based on the article information. The first information represents the probability that the exposure click-through rate of the target article is higher than the average exposure click-through rate of the overall market, and the second information represents the probability that the single access reading time of the target article is higher than the average single access reading time of the overall market.

[0140] Based on the first and second information of the target article, the identification result of the target article is generated, which indicates whether the target article attracts users or not.

[0141] The article evaluation model is based on article samples. It is trained by determining the first and second information predicted by the article evaluation model based on the article information of the article samples, with the training objective being that the predicted first information is close to the target first information carried by the article samples and the predicted second information is close to the target second information carried by the article samples.

[0142] Processor 902 could be a central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0143] The control device may also include a communication interface 903 and a communication bus 904, wherein the memory 901, the processor 902 and the communication interface 903 communicate with each other through the communication bus 904.

[0144] This application also provides a readable storage medium on which a computer program is stored. The computer program is loaded and executed by a processor to implement the steps of the above-described article recognition method. The specific implementation process can be referred to the description of the corresponding part of the above embodiments, and will not be repeated in this embodiment.

[0145] This application also proposes a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations of the above-described article recognition method or article recognition device. Specific implementation processes can be referred to the descriptions of the corresponding embodiments above, and will not be repeated here.

[0146] This application provides an article recognition method, apparatus, computer device, and storage medium. It acquires article information of a target article and inputs this information into a pre-trained article evaluation model to generate the recognition result of the target article. The article evaluation model in this application does not rely on prior human knowledge to label article samples as attractive to users. Instead, it labels article samples based on click-through rate and single-visit reading time, thus avoiding the influence of subjective human experience. This reduces the distance between the labeled article samples and the actual interests of users, improving the accuracy of the recognition results for articles that are attractive to users. Furthermore, this multi-task modeling approach, relying on click-through rate and single-visit reading time, further improves the accuracy of article recognition results compared to the single-task modeling approach of existing technologies that depends on whether an article is attractive to users.

[0147] The foregoing has provided a detailed description of the article recognition method, apparatus, computer device, and storage medium provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

[0148] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0149] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that elements inherent to a process, method, article, or apparatus that comprises a list of elements, or elements inherent to such processes, methods, articles, or apparatus, are also included. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0150] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An article recognition method, characterized in that, include: Obtain article information of the target article, wherein the article information includes any one or more of the tags that characterize the article content of the target article and the title of the target article; The article evaluation model predicts first and second information about the target article based on the article information. The first information represents the probability that the target article's exposure click-through rate is higher than the average exposure click-through rate of the overall market, and the second information represents the probability that the target article's single-visit reading time is higher than the average single-visit reading time of the overall market. The average exposure click-through rate of the overall market is determined based on the exposure click-through rate of all articles in the overall market, and the average single-visit reading time of the overall market is determined based on the single-visit reading time of all articles in the overall market. Based on the first and second information of the target article, an identification result of the target article is generated, wherein the identification result indicates whether the target article attracts users or does not attract users; The article evaluation model is based on an article sample. It is obtained by determining the first and second information predicted by the article evaluation model based on the article information of the article sample, and training the article evaluation model with the goal of the predicted first information being close to the target first information carried by the article sample and the predicted second information being close to the target second information carried by the article sample.

2. The method according to claim 1, characterized in that, The article information of the target article includes tags that characterize the article content of the target article, and obtaining the article information of the target article includes: A pre-set tag whitelist is determined, and tags that match the content of the target article are identified from the tag whitelist.

3. The method according to claim 1, characterized in that, The step of generating the identification result of the target article based on the first and second information of the target article includes: If the first information of the target article exceeds a preset first threshold, and the second information of the target article exceeds a preset second threshold, the identification result of the target article is determined to indicate that the target article attracts users; If the first information of the target article does not exceed the first threshold, or the second information of the target article does not exceed the second threshold, the identification result of the target article indicates that the target article is not attractive to users.

4. The method according to claim 1, characterized in that, Also includes: Obtain article data from an information flow platform within a first historical time period, wherein the article data indicates at least one article pushed to a user by the information flow platform within the first historical time period, and each of the at least one articles is an article sample; Determine the article information, target first information, and target second information of the article sample; The article information of the article sample is input into the article evaluation model to be trained. The first information and the second information of the article sample predicted by the article evaluation model based on the article information of the article sample are determined. The training objective is to make the predicted first information of the article sample approach the target first information of the article sample and the predicted second information of the article sample approach the target second information of the article sample. The article evaluation model is then trained to obtain the article evaluation model.

5. The method according to claim 4, characterized in that, Determining the target first information of the article sample includes: Determine the exposure click-through rate of the article sample; The average exposure and click-through rate of the entire market is determined based on the exposure and click-through rate of all articles in the market. If the exposure click-through rate of the article sample is not lower than the average exposure click-through rate of the overall market, the target first information of the article sample is determined as the first preset value; If the exposure click-through rate of the article sample is lower than the average exposure click-through rate of the overall market, the target first information of the article sample is determined to be the second preset value.

6. The method according to claim 4, characterized in that, Determining the target second information of the article sample includes: Determine the duration of a single access to the article sample; The average single access reading time of the overall market is determined based on the single access reading time of all articles in the overall market. If the single access reading time of the article sample is not less than the average single access reading time of the overall market, the target second information of the article sample is determined as the first preset value; If the single access reading time of the article sample is lower than the average single access reading time of the overall market, the target second information of the article sample is determined to be the second preset value.

7. The method according to claim 1, characterized in that, Also includes: By using article data from the information flow platform within the second historical time period, the article evaluation model to be trained is obtained, and the target article evaluation model is obtained. The second historical time period is later than the first historical time period. Using the article evaluation model, the identification result of each article pushed to the user by the information flow platform in the third historical time period is predicted to obtain a first identification result set; the third historical time period is later than the second historical time period. Using the target article evaluation model, the identification result of each article pushed to the user by the information flow platform within the third historical time period is predicted, and a second identification result set is obtained; Determine the standard identification result of each article pushed to the user by the information flow platform within the third historical time period to obtain the third identification result set; If the second identification result set is closer to the third identification result set than the first identification result set, the article evaluation model is updated to the target article evaluation model.

8. An article recognition device, characterized in that, include: An article information acquisition unit is used to acquire article information of a target article, wherein the article information includes any one or more of the tags characterizing the content of the target article and the title of the target article; The information prediction unit is used to predict, based on the article information, the first information and the second information of the target article using an article evaluation model. The first information represents the probability that the exposure click-through rate of the target article is higher than the average exposure click-through rate of the overall market, and the second information represents the probability that the single-access reading time of the target article is higher than the average single-access reading time of the overall market. The average exposure click-through rate of the overall market is determined based on the exposure click-through rate of all articles in the overall market, and the average single-access reading time of the overall market is determined based on the single-access reading time of all articles in the overall market. The identification result generation unit is used to generate an identification result of the target article based on the first information and the second information of the target article, wherein the identification result indicates whether the target article attracts users or does not attract users; The article evaluation model is based on an article sample. It is obtained by determining the first and second information predicted by the article evaluation model based on the article information of the article sample, and training the article evaluation model with the goal of the predicted first information being close to the target first information carried by the article sample and the predicted second information being close to the target second information carried by the article sample.

9. The apparatus according to claim 8, characterized in that, The article information of the target article includes tags that characterize the article content of the target article; correspondingly, the article information acquisition unit is used to determine a pre-set tag whitelist and determine tags that match the article content of the target article from the tag whitelist.

10. The apparatus according to claim 8, characterized in that, The recognition result generation unit includes: The first generation unit is configured to determine that the identification result of the target article represents that the target article attracts users if the first information of the target article exceeds a preset first threshold and the second information of the target article exceeds a preset second threshold. The second generation unit is configured to determine, if the first information of the target article does not exceed the first threshold, or the second information of the target article does not exceed the second threshold, that the identification result of the target article indicates that the target article is not attractive to users.

11. The apparatus according to claim 8, characterized in that, The article recognition device further includes a model generation unit, which includes: The article data acquisition unit is used to acquire article data of the information flow platform within a first historical time period. The article data indicates at least one article pushed to the user by the information flow platform within the first historical time period, and each of the at least one articles is an article sample. An information determination unit is used to determine the article information, target first information, and target second information of the article sample. The training unit is used to input the article information of the article sample into the article evaluation model to be trained, determine the first information and the second information of the article sample predicted by the article evaluation model to be trained based on the article information of the article sample, and train the article evaluation model to be trained with the goal that the predicted first information of the article sample is close to the target first information of the article sample and the predicted second information of the article sample is close to the target second information of the article sample, so as to obtain the article evaluation model.

12. The apparatus according to claim 11, characterized in that, The information determination unit for determining the target primary information of the article sample includes: The first determining unit is used to determine the exposure click-through rate of the article sample; and to determine the average exposure click-through rate of the entire market based on the exposure click-through rates of all articles in the market. The second determining unit is used to determine the target first information of the article sample as a first preset value if the exposure click-through rate of the article sample is not lower than the average exposure click-through rate of the overall market. The third determining unit is used to determine the target first information of the article sample as the second preset value if the exposure click-through rate of the article sample is lower than the average exposure click-through rate of the overall market.

13. The apparatus according to claim 11, characterized in that, The information determination unit for determining the target second information of the article sample includes: The fourth determining unit is used to determine the single access reading time of the article sample; and to determine the average single access reading time of the entire market based on the single access reading time of all articles in the market. The fifth determining unit is used to determine the target second information of the article sample as a first preset value if the single access reading time of the article sample is not less than the average single access reading time of the overall market. The sixth determining unit is used to determine the target second information of the article sample as the second preset value if the single access reading time of the article sample is lower than the average single access reading time of the overall market.

14. The apparatus according to claim 8, characterized in that, The article determination device further includes a model optimization unit, which includes: The target article evaluation model generation unit is used to train the article evaluation model to be trained using article data from the information flow platform in the second historical time period, and obtain the target article evaluation model, wherein the second historical time period is later than the first historical time period; The first prediction unit is used to use the article evaluation model to predict the identification result of each article pushed to the user by the information flow platform in the third historical time period, and obtain a first identification result set; the third historical time period is later than the second historical time period. The second prediction unit is used to use the target article evaluation model to predict the identification result of each article pushed to the user by the information flow platform in the third historical time period, and obtain a second identification result set. The standard identification result determination unit is used to determine the standard identification result of each article pushed to the user by the information flow platform in the third historical time period, and to obtain the third identification result set. The model optimization subunit is used to update the article evaluation model to the target article evaluation model if the second identification result set is closer to the third identification result set than the first identification result set.

15. A computer device, characterized in that, include: A processor and a memory are connected via a communication bus; wherein the processor is used to call and execute a program stored in the memory; The memory is used to store a program for implementing the article recognition method as described in any one of claims 1-7.

16. A computer-readable storage medium, characterized in that, It stores a computer program, which is loaded and executed by a processor to implement the steps of the article recognition method as described in any one of claims 1-7.

17. A computer program product, characterized in that, The computer program product includes instructions that, when executed on a computer device, cause the computer device to perform the steps of the article recognition method as described in any one of claims 1-7.

Images