A news processing method and device

Abstract

This application provides a news processing method and apparatus, relating to artificial intelligence technology. The method includes: extracting a portion of training data that meets timeliness requirements from webpage ranking model training data, and using it as training data for a news ranking model; determining timeliness labels for the news ranking model training data based on its timeliness and relevance labels; training a news ranking model based on the news ranking model training data and the corresponding timeliness labels; and ranking multiple news items based on the trained news ranking model. This application enables the efficient training of a news ranking model with excellent timeliness.

Description

Technical Field

[0001] This application relates to artificial intelligence technology and blockchain technology, and in particular to a news processing method and apparatus. Background Technology

[0002] Artificial intelligence (AI) is the theory, methods, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0003] Information recommendation is an important application of artificial intelligence. With the explosive growth of news information on the internet, personalized news search engines have emerged to address this massive volume of information. News search engines rely on news ranking models to appropriately sort news items and present them to users.

[0004] In other words, in related technologies, developing a news ranking model relies on a large amount of manually labeled training data, which is costly and affects training efficiency and the timeliness of news ranking results. Therefore, there is a lack of effective solutions for news ranking models in related technologies. Summary of the Invention

[0005] This application provides a news processing method, apparatus, electronic device, and computer-readable storage medium that can efficiently train a news ranking model with excellent timeliness.

[0006] The technical solution of this application embodiment is implemented as follows:

[0007] This application provides a news processing method, including:

[0008] A portion of the training data that meets the timeliness requirement is extracted from the training data of the web page ranking model and used as the training data for the news ranking model;

[0009] Based on the timeliness of the news ranking model training data and the relevance labels of the news ranking model training data, the timeliness labels of the news ranking model training data are determined.

[0010] The news ranking model is trained based on the training data of the news ranking model and the corresponding timeliness tags.

[0011] The trained news ranking model is used to rank multiple news items.

[0012] This application provides a news processing device, including:

[0013] The extraction module is used to extract a portion of the training data that meets the timeliness requirements from the training data of the web page ranking model, and use it as the training data for the news ranking model.

[0014] The determination module is used to determine the timeliness label of the news ranking model training data based on the timeliness of the news ranking model training data and the relevance label of the news ranking model training data;

[0015] A training model is used to train the news ranking model based on the news ranking model training data and the corresponding timeliness tags.

[0016] The sorting module is used to sort multiple news items based on the trained news sorting model.

[0017] In the above scheme, the determining module is further configured to obtain the search volume of the query string corresponding to the news ranking model training data in multiple time periods; traverse any two adjacent time periods in the multiple time periods, perform verification processing on the search volume of any two adjacent time periods to obtain the two adjacent time periods corresponding to the query string; and take the time interval between the end time of the next time period corresponding to the two adjacent time periods as the timeliness of the news ranking model training data.

[0018] In the above scheme, the determining module is further configured to perform the following processing on any two adjacent time periods: fit the search volume of the two adjacent time periods to obtain a test statistic; map the test statistic to a confidence level based on the chi-square distribution; wherein the confidence level characterizes the correlation between the end time and the increment of the search volume of the two adjacent time periods; and take the two adjacent time periods corresponding to the maximum confidence level as the two adjacent time periods corresponding to the query string.

[0019] In the above scheme, the determining module is further used to obtain the weight value of the timeliness and the weight value of the relevance label; wherein, the weight value of the timeliness is greater than the weight value of the relevance label; based on the weight value of the timeliness and the weight value of the relevance label, the timeliness and the relevance label are weighted and summed, and the weighted summation result is used as the timeliness label of the news ranking model training data.

[0020] In the above scheme, the training module is further configured to determine the proportion of training data corresponding to the timeliness sensitivity of the news ranking model based on the correspondence between different timeliness sensitivities and different proportions of training data in the news ranking model; wherein, the timeliness sensitivity is positively correlated with the amount of target training data, the target training data is the training data of the news ranking model corresponding to the target level of the timeliness label, and the target level includes at least one of the first half of the timeliness label levels in descending order; and adjust the proportion of different types of training data in the news ranking model training data to conform to the proportion of training data corresponding to the timeliness sensitivity of the news ranking model.

[0021] In the above scheme, the extraction module is further configured to obtain the timestamp of the webpage ranking model training data; use the time interval between the timestamp and the current time point as the timeliness of the webpage ranking model training data; and use the portion of the webpage ranking model training data whose timeliness is greater than the timeliness threshold as the portion of training data that meets the timeliness requirement.

[0022] In the above scheme, the extraction module is also used to obtain the titles of various historical news topics; to match the webpage ranking model training data with the titles of the historical news topics to obtain the title matching degree; and to use the portion of the webpage ranking model training data with the title matching degree greater than the matching degree threshold as the portion of training data that meets the timeliness requirements.

[0023] In the above scheme, the training module is further configured to obtain the click-through rate of the webpage ranking model training data, determine the target webpage ranking model training data, wherein the target webpage ranking model training data is the training data in the webpage ranking model training data whose click-through rate is higher than the click-through rate threshold; determine the timeliness label of the target webpage ranking model training data according to the correspondence between different click-through rates and different timeliness labels; and combine the news ranking model training data and the target webpage ranking model training data to obtain the news ranking model training data used to train the news ranking model.

[0024] In the above scheme, the determining module is further configured to obtain the relevance labels of the news ranking model training data by: performing content matching processing on the query string corresponding to the news ranking model training data and the news ranking model training data to obtain the content matching degree; and determining the relevance label corresponding to the content matching degree of the news ranking model training data according to the correspondence between different content matching degrees and different relevance labels.

[0025] This application provides an electronic device, including:

[0026] Memory, used to store executable instructions;

[0027] The processor, when executing executable instructions stored in the memory, implements the news processing method provided in the embodiments of this application.

[0028] This application provides a computer-readable storage medium storing executable instructions for implementing the news processing method provided in this application when executed by a processor.

[0029] The embodiments of this application have the following beneficial effects:

[0030] By extracting a portion of the training data that meets the timeliness requirement from the webpage ranking model training data, and using it as the training data for the news ranking model, and determining the timeliness label of the news ranking model training data based on the timeliness and relevance labels of the news ranking model training data, the training data of the webpage ranking model can be transferred to the training of the news ranking model. This saves the work of relabeling the training data, enables efficient training of the news ranking model, and ensures the timeliness of the news ranking model while satisfying the relevance requirements. Attached Figure Description

[0031] Figure 1A This is a schematic diagram of the architecture of the news processing system 100 provided in an embodiment of this application;

[0032] Figure 1B This is a schematic diagram of the structure of a news processing system 100 based on a blockchain system provided in an embodiment of this application;

[0033] Figure 1C This is a schematic diagram of the block structure provided in the embodiments of this application;

[0034] Figure 2 This is a schematic diagram of the structure of the server 200 provided in the embodiments of this application;

[0035] Figure 3A This is a schematic flowchart of a news processing method provided in an embodiment of this application;

[0036] Figure 3B This is a schematic flowchart of a news processing method provided in an embodiment of this application;

[0037] Figure 3C This is a schematic flowchart of a news processing method provided in an embodiment of this application;

[0038] Figure 4 This is a schematic diagram of the timeliness tag filing provided in the embodiments of this application;

[0039] Figure 5 This is a schematic diagram of the architecture of the news ranking model provided in the embodiments of this application;

[0040] Figure 6 This is a schematic diagram of a four-cell table provided in an embodiment of this application;

[0041] Figure 7 This is a schematic diagram of the chi-square distribution provided in an embodiment of this application;

[0042] Figure 8 This is a schematic flowchart of a news processing method provided in an embodiment of this application. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0044] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0045] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0047] In the implementation of this application, the collection and processing of relevant data should be strictly in accordance with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0048] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0049] 1) Transfer learning: This is generally divided into two categories: data transfer and model transfer. Among them, the main idea of ​​data transfer is to transfer labeled data or knowledge structures from related fields to complete or improve the learning effect in the target field.

[0050] 2) User-generated content (UGC): Content created and uploaded by users.

[0051] 3) Professionally Generated Content (PGC): Content created and uploaded by organizations.

[0052] 4) A search engine is a system that uses specific computer programs and certain strategies to collect information from the Internet, organizes and processes the information, provides search services to users, and displays the relevant search results to users.

[0053] News apps generate a large amount of user-generated content (UGC) and professionally generated content (PGC) news every day. For users who use these apps' built-in news search engines, different queries implicitly reflect different timeliness requirements. For example, when a user searches for "xx financial report" during quarterly financial report releases, the search result indicates the user wants the most up-to-date news about that report. If a web search engine prioritizes older financial report news, it leads to a poor user experience. Conversely, when a user searches for "tomato and egg stir-fry," the search result doesn't necessarily require timeliness. In this case, the web search engine, while ensuring text relevance, should focus more on factors like content authority and quality. Compared to general web search engines, news search engines primarily focus on news information, unlike web search which is often more diverse and less varied. Users primarily use news search engines to find news information. It is evident that webpage ranking models, while satisfying text relevance, prioritize authority and page quality, giving less consideration to timeliness. In contrast, news search ranking models, while also prioritizing text relevance, place greater emphasis on timeliness; the ranking principle is generally: Relevant & Timely > Relevant & Old > Irrelevant. Therefore, when a webpage search ranking team redevelops a news search engine, it needs to re-annotate the training data to meet the requirements of a webpage search engine. This approach relies heavily on manual re-annotation, resulting in high costs.

[0054] To address the aforementioned technical problems, embodiments of this application provide a news processing method, apparatus, electronic device, and computer-readable storage medium, capable of efficiently training a news ranking model with excellent timeliness. The following describes exemplary applications of the news processing method provided by embodiments of this application. This news processing method can be implemented by various electronic devices, such as smartphones, tablets, smart vehicle terminals, smart wearable devices, and other types of terminals, or it can be implemented as a server. The following will describe exemplary applications when the electronic device is implemented as a server.

[0055] See Figure 1A , Figure 1A This is a schematic diagram of the architecture of a news processing system 100 provided in an embodiment of this application. The news processing system 100 includes a server 200, terminals (terminals 400-1 and 400-2 are shown exemplarily) and a network 300. Terminal 400-1 is connected to the server 200 through the network 300, which can be a wide area network (WAN), a local area network (LAN), or a combination of both.

[0056] In some embodiments, server 200 obtains webpage ranking model training data from terminal 400-1, extracts a portion of the training data that meets the timeliness requirements from the webpage ranking model training data, and uses it as news ranking model training data; based on the timeliness and relevance labels of the news ranking model training data, it determines the timeliness labels of the news ranking model training data; based on the news ranking model training data and the corresponding timeliness labels, it trains a news ranking model; based on the trained news ranking model, it sorts multiple news items, performs news recommendation operations based on the sorting results, and sends the news recommendation results to terminal 400-1 so that users can select news of interest based on the news recommendation results. The news ranking model can be integrated into the news search engine running on the news search server, or it can be deployed on a separate news ranking server, which can be called by the news search engine.

[0057] In some embodiments, terminal 400-1 extracts a portion of training data that meets the timeliness requirements from the webpage ranking model training data and uses it as training data for the news ranking model; based on the timeliness of the news ranking model training data and the relevance labels of the news ranking model training data, it determines the timeliness labels of the news ranking model training data; based on the news ranking model training data and the corresponding timeliness labels, it trains the news ranking model; based on the trained news ranking model, it sorts multiple news information, and performs news recommendation operations based on the sorting results, so that users can select news they are interested in according to the recommendation results.

[0058] In some embodiments, server 200 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Terminals and servers can be connected directly or indirectly via wired or wireless communication, which is not limited in this embodiment.

[0059] In some embodiments, see Figure 1B , Figure 1B This is a schematic diagram of the structure of a news processing system 100 based on a blockchain system provided in this application embodiment. Any machine, such as server 200, terminal 400-1, or terminal 400-2 as shown in Figure 1, can join the blockchain network and become a node. The node includes a hardware layer, a middleware layer, an operating system layer, and an application layer. Figure 1B The example shows a server 200 mapped to node 510-2 of blockchain network 500, and a terminal 400-2 mapped to node 510-1 of blockchain network 500. Each node (e.g., node 510-1 to node 510-2) has consensus function and ledger (i.e., maintaining a state database, such as a KV database) function.

[0060] Server 200 runs a webpage ranking model and applies the news processing method provided in this application embodiment to the training data of the webpage ranking model to obtain the ranking results of multiple news items. The ranking results of multiple news items are sent to node 510-2 of the blockchain network by calling a smart contract. Node 510-2 reaches a consensus on the ranking results of multiple news items. After the consensus is passed, the ranking results of multiple news items are returned to terminal 400-2 so that terminal 400-2 can perform news recommendation operations or news ranking result display operations based on the ranking results. In this application embodiment, the reliability of news information ranking and the trustworthiness of the ranking process are guaranteed by the consensus mechanism between nodes of the blockchain network.

[0061] See Figure 1C , Figure 1C This is a schematic diagram of the block structure provided in this application embodiment. Each block includes the hash value of the webpage ranking model training data and relevance tags (the hash value of this block), as well as the hash value of the previous block. The blocks are connected through their hash values ​​to form a blockchain. Additionally, each block may include information such as a timestamp when it was generated. A blockchain, in essence, is a decentralized database, a chain of data blocks linked together using cryptographic methods.

[0062] Next, the structure of the electronic device for implementing the news processing method provided in the embodiments of this application will be described. As mentioned above, the electronic device provided in the embodiments of this application can be the server 200 in FIG1. ​​See also Figure 2 , Figure 2 This is a schematic diagram of the structure of the server 200 provided in the embodiments of this application. Figure 2 The server 200 shown includes at least one processor 210, memory 230, and at least one network interface 220. The various components of server 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to implement communication between these components. In addition to a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 240.

[0063] Processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0064] The memory 230 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 230 may optionally include one or more storage devices physically located away from the processor 210.

[0065] The memory 230 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 230 described in this application embodiment is intended to include any suitable type of memory.

[0066] In some embodiments, memory 230 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0067] Operating system 231 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0068] The network communication module 232 is used to reach other computing devices via one or more (wired or wireless) network interfaces 220, such as Bluetooth, WiFi, and Universal Serial Bus (USB).

[0069] In some embodiments, the news processing apparatus provided in this application can be implemented in software. Figure 2 A news processing device 233 stored in memory 230 is shown. This device can be software in the form of programs and plug-ins, and includes the following software modules: extraction module 2331, determination module 2332, training module 2333, and sorting module 2334. These modules are logically linked and can therefore be arbitrarily combined or further divided according to their implemented functions. The functions of each module will be described below.

[0070] The news processing method provided in this application embodiment can be executed by the server 200 in Figure 1. The following description uses the execution of the news processing method provided in this application embodiment by the server 200 in Figure 1 alone as an example. See also... Figure 3A , Figure 3A This is a flowchart illustrating a news processing method provided in an embodiment of this application, which will be combined with... Figure 3A The steps shown are explained.

[0071] In step 101, a portion of the training data that meets the timeliness requirement is extracted from the webpage ranking model training data and used as the news ranking model training data.

[0072] In some embodiments, extracting a portion of the training data that meets the timeliness requirement from the web page ranking model training data can be achieved in the following ways: obtaining the timestamp of the web page ranking model training data; using the time interval between the timestamp and the current time point as the timeliness of the web page ranking model training data; and using the portion of the web page ranking model training data whose timeliness is greater than the timeliness threshold as the portion of the training data that meets the timeliness requirement.

[0073] In some examples, since the training data for webpage ranking models consists of massive amounts of historical data, an initial screening process based on timeliness is performed when constructing training data for news ranking models from this data. This involves extracting data from the webpage ranking model training data. Because news ranking model training data places greater emphasis on timeliness, the extracted news ranking model training data can be initially screened based on the timeliness of the news. For example, the time of generation of each piece of webpage ranking model training data (e.g., news information) can be obtained from logs; the time interval between the timestamp and the current time point can be used as the timeliness of that news item; thus, the portion of webpage ranking model training data with a timeliness greater than a timeliness threshold is considered as the portion of training data that meets the timeliness requirement.

[0074] In some embodiments, extracting a portion of the training data that meets the timeliness requirement from the web page ranking model training data can be achieved in the following way: obtaining the titles of various historical news topics; matching the web page ranking model training data with the titles of historical news topics to obtain the title matching degree; and using the portion of the web page ranking model training data with a title matching degree greater than the matching degree threshold as the portion of the training data that meets the timeliness requirement.

[0075] In some examples, the training data for the webpage ranking model includes trending news (old news) from a previous period, meaning the timeliness score of the training data is low. Although trending old news is not very timely, users still have a demand for timeliness for certain trending old news and consider this trending old news to be training data that meets the timeliness requirement. Therefore, the extracted news ranking model training data can be initially screened based on historical trending news. For example, obtain the titles of historical news topics corresponding to historical trending news; match the webpage ranking model training data with the titles of historical news topics to obtain the title matching degree; when the title matching degree is greater than the matching degree threshold, it is determined that the webpage ranking model training data belongs to historical trending news, where the matching degree threshold characterizes whether the webpage ranking model training data belongs to historical trending news; thus, the portion of webpage ranking model training data with a title matching degree greater than the matching degree threshold is used as the portion of training data that meets the timeliness requirement.

[0076] In this embodiment of the application, since the requirements of the webpage ranking model and the news ranking model are different, the two aspects of the user's need for timeliness (timeliness and historical hot news) during the training process of the news ranking model are considered. The timeliness of the webpage ranking model training data is initially screened to reduce the invalid training of massive historical webpage ranking model training data, improve training efficiency, and ensure the timeliness of the webpage ranking model training data after the timeliness screening.

[0077] In step 102, the timeliness label of the news ranking model training data is determined based on the timeliness of the news ranking model training data and the relevance label of the news ranking model training data.

[0078] In some embodiments, the timeliness of the news ranking model training data and the relevance labels of the news ranking model training data are fused to obtain the timeliness labels of the news ranking model training data.

[0079] In some examples, the weight values ​​of timeliness and relevance labels are obtained; where the weight value of timeliness is greater than the weight value of relevance labels; based on the weight values ​​of timeliness and relevance labels, the timeliness and relevance labels are weighted and summed, and the weighted sum is used as the timeliness label of the news ranking model training data.

[0080] For example, the timeliness of the news ranking model training data is categorized into three levels: 1-3 days, 3-11 days, and over 11 days, respectively. The relevance labels are also categorized into three levels: 1-3 days, 2-3 days, and 3-3 days, respectively, representing highly relevant, moderately relevant, and irrelevant. Based on the weight values ​​of timeliness and relevance, a weighted sum is applied to the timeliness and relevance labels to obtain a timeliness label adjustment diagram. See [link / reference] Figure 4 , Figure 4 This is a schematic diagram of the timeliness label adjustment provided in the embodiments of this application. It can be seen that the timeliness label of the news ranking model training data with a timeliness of 3 and a relevance of 2 is 4, and the timeliness label of the news ranking model training data with a timeliness of 2 and a relevance of 3 is 3.

[0081] In this embodiment, the relevance labels of the webpage ranking model training data are migrated to timeliness labels that include timeliness and relevance. This allows for the reuse of the original relevance labels and enhances timeliness. In other words, it can be well compatible with the relevance labels of the webpage ranking model training data and adds some features (timeliness) of the news ranking model for news search scenarios to the relevance labels of the webpage ranking model training data.

[0082] In step 103, the news ranking model is trained based on the news ranking model training data and the corresponding timeliness tags.

[0083] In some embodiments, see Figure 5 , Figure 5This is a schematic diagram of the architecture of the news ranking model provided in this application embodiment. The news ranking model includes an input layer, hidden layers, and a prediction layer; the hidden layer includes convolutional layers, pooling layers, and fully connected layers. The input layer of the news ranking model normalizes the training data. The training data is used as samples, and the corresponding timeliness labels are used as sample labels. The training data is forward-propagated through each hidden layer to obtain feature vectors. Then, the obtained feature vectors are forward-propagated through the prediction layer to determine the prediction result. A loss function is initialized, including the sample label and the prediction result for each sample. The error between the prediction result and the sample label for each sample is determined, and the error is backpropagated through the news ranking model according to the loss function to determine the change value of the news ranking model when the loss function reaches its minimum value. The parameters of the news ranking model are updated based on the change value. The prediction layer can be implemented using the logistic regression function softmax. It should be noted that the news ranking model can be integrated into a news search engine or used as an external service for the news search engine to call.

[0084] In step 104, multiple news items are sorted based on the trained news sorting model.

[0085] In some embodiments, sorting multiple news items based on a trained news ranking model can be achieved in the following ways: performing convolution processing on business data to extract feature vectors of news items; performing prediction processing on the feature vectors of news items to map the feature vectors of news items to the probabilities of multiple candidate timeliness labels, and taking the timeliness label with the highest probability as the prediction result; and sorting the news items in descending order based on the prediction result to obtain the ranking result of multiple news items with the prediction result in descending order.

[0086] In some examples, news recommendation operations based on the ranking results of multiple news items can be achieved in at least one of the following ways: directly pushing the ranking results to the user's terminal and presenting the news items in the ranking results in descending order of timeliness tags; or pushing the top-ranked news items in the ranking results to the user's terminal and presenting the news items in the ranking results in descending order of timeliness tags; or performing diversity processing on the ranking results to display news based on a diversity model.

[0087] It should be noted that the ranking results can be applied to various news ranking scenarios. For example, in a news search scenario, based on a trained news ranking model, news items matching the received search keywords (i.e., query strings) are ranked, and the ranking results are displayed in the news search engine. Alternatively, in a news recommendation scenario, based on a trained news ranking model, the news recommendation server obtains the ranking results of multiple news items as the information stream of the news to be pushed, and the client (e.g., a news client) receives the information stream of news pushed by the news recommendation server.

[0088] In this embodiment of the application, the news information is predicted based on the news ranking model, and the ranking results of multiple news information corresponding to the first one in descending order of the news ranking are obtained. The multiple news information are then recommended to the user terminal. The recommended news information not only ensures the relevance to the search query string, but also ensures that the recommended news information can meet the user's news ranking needs.

[0089] In some embodiments, see Figure 3B , Figure 3B This is a schematic flowchart of a news processing method provided in an embodiment of this application, illustrating that in Figure 3A Between steps 101 and 102, steps 105 to 107 can also be performed, which will be explained in detail below.

[0090] In step 105, the search volume of the query string corresponding to the news ranking model training data is obtained in multiple time periods.

[0091] In some examples, the number of searches (search volume) of the query string corresponding to the news ranking model training data in multiple time periods is obtained from the web search logs. The news ranking model training data is extracted from the web ranking model training data, that is, a portion of the training data in the web search model training data. The search volume at each moment recorded in the web search logs is counted, and the search volume for multiple time periods is obtained based on the statistical time interval. For example, if the statistical time interval is selected as 5 seconds, the search volume for 0-5 seconds from 11:20:00 on April 7, 2021, the search volume for 5-10 seconds from 11:20:00 on April 7, 2021, the search volume for 10-15 seconds from 11:20:00 on April 7, 2021, and so on, is obtained to obtain the search volume of the query string corresponding to the news ranking model training data in multiple time periods.

[0092] In other examples, considering the massive search volume of query strings, training data for news ranking models is obtained from web search logs where the search volume of query strings exceeds a certain threshold.

[0093] In step 106, any two adjacent time periods are traversed among multiple time periods, and the search volume of any two adjacent time periods is checked to obtain the two adjacent time periods corresponding to the query string.

[0094] In some examples, the following processing is performed on any two adjacent time periods: the search volume of the adjacent time periods is fitted to obtain a test statistic; the test statistic is mapped to a confidence level based on a chi-square distribution. Here, the confidence level characterizes the correlation between the end time and the increase in search volume between the two adjacent time periods; the two adjacent time periods corresponding to the maximum confidence level are taken as the two adjacent time periods corresponding to the query string.

[0095] For example, a chi-square test can be used for fitting to examine whether the increase in search volume between adjacent time periods is significantly different. Let's assume two adjacent time periods are T. i-1 ~T i T i ~T i+1 Get the time period t0 (T i-1 Time ~T i Time period t1 (T) i Time ~T i+1 The search volume and non-search volume (each occurrence of the query string in the log is counted as one search volume, and the difference between the total search volume and the search volume is the non-search volume) are used to construct a four-cell table based on two adjacent time periods.

[0096] See Figure 6 , Figure 6 This is a schematic diagram of a four-cell table provided in an embodiment of this application. The chi-square value can be obtained based on the four-cell table. The chi-square value is used as the test statistic, which can detect whether there is a sudden burst of fluctuation in the current time period. a is the search volume in time period t0, b is the search volume in time period t1, c is the number of unsearched items in time period t0, d is the number of unsearched items in time period t1, and n is the sum of a, b, c, and d.

[0097] See Figure 7 , Figure 7This is a schematic diagram of the chi-square distribution provided in an embodiment of this application. Assuming a chi-square value of 6.109, consulting the chi-square distribution reveals that 5.024 < 6.109 < 6.635. The probability corresponding to a chi-square value of 5.024 is 0.025, and the probability corresponding to a chi-square value of 6.635 is 0.010. Therefore, it can be determined that the probability range of the confidence level between two adjacent time periods (time period t0 and time period t1) and the increase in search volume corresponding to the query string is between 1 and 0.01 and 1 and 0.025. Multiple confidence levels are calculated for any two adjacent time periods. Based on the comparison of the probability ranges of multiple confidence levels, the two adjacent time periods corresponding to the highest confidence level are taken as the two adjacent time periods corresponding to the query string (i.e., the two adjacent time periods related to the significant change in the increase of the query string).

[0098] In step 107, the time interval between the end time of the second time interval of two adjacent time intervals corresponding to the query string and the current time is used as the timeliness of the news ranking model training data.

[0099] In some examples, the two adjacent time periods corresponding to the query string are the relevant times that caused the significant change in search volume. For example, the two adjacent time periods corresponding to the query string are T. i-1 ~T i T i ~T i+1 It can be determined that the increase or decrease in the search volume of the query string is due to the significant change in the increase or decrease from the later time period T. i ~T i+1 End time T i+1 News about this started happening all the time, and T i+1 The time interval between a given moment and the current moment serves as the timeliness of the training data for the news ranking model.

[0100] In this embodiment of the application, by verifying the search volume of the query string in multiple time periods, the moment when the increase in search volume changes dramatically can be accurately determined, so as to obtain the timeliness of the training data of the news ranking model.

[0101] In some embodiments, see Figure 3C , Figure 3C This is a flowchart illustrating a news processing method provided in an embodiment of this application. It shows that after step 103 in 3A, steps 108 and 109 can also be executed. The following will describe each step in detail.

[0102] In step 108, based on the correspondence between different timeliness sensitivities and different proportions of training data for the news ranking model, the proportion of training data corresponding to the timeliness sensitivity of the news ranking model is determined. Here, timeliness sensitivity is positively correlated with the amount of target training data, which is the training data for the news ranking model corresponding to the target level of the timeliness label. The target level includes at least one of the top half of the timeliness label levels in descending order.

[0103] It's important to note that the number of samples at different levels of timeliness labels in the news ranking model's training data corresponds to different tones (i.e., timeliness sensitivity). The model is trained based on the number of samples corresponding to the timeliness requirements of the news ranking model, ultimately achieving the desired tone. For example, if the timeliness labels in the news ranking model's training data are tiers 1, 2, 3, 4, and 5, and the timeliness sensitivity is increased by sorting the timeliness labels in descending order, the higher the proportion of tiers 3, 4, and 5 in the training data, the higher the timeliness sensitivity. Here, "total number" refers to the total amount of news ranking model training data corresponding to each timeliness label level.

[0104] In step 109, the proportion of different types of training data in the news ranking model training data is adjusted to match the proportion of training data corresponding to the timeliness sensitivity of the news ranking model.

[0105] In some examples, the amount of training data at different levels in the news ranking model's training data is adjusted according to a determined proportion of training data corresponding to the timeliness sensitivity of the news ranking model. This adjustment can be achieved by copying different numbers of data points from a specific level within the news ranking model's training data. For instance, to enhance the news ranking model's sensitivity to timeliness, multiple copies of training data at levels 4-5 can be made to match the proportion of training data corresponding to the news ranking model's timeliness sensitivity. This achieves the goal of synchronously influencing the final news ranking model's bias through the distribution of training data.

[0106] In this embodiment of the application, the timeliness sensitivity of the news ranking model is improved by adjusting the proportion of training data of different levels in the training data of the news ranking model.

[0107] In some embodiments, before training the news ranking model based on the news ranking model training data and corresponding timeliness tags, the training data in the news ranking model can be increased by selectively supplementing the existing news ranking model with high-click data (data with a click-through rate higher than the click-through rate threshold). Specifically, this can be achieved by: obtaining the click-through rate of the webpage ranking model training data; determining the target webpage ranking model training data, which is the training data in the webpage ranking model training data with a click-through rate higher than the click-through rate threshold; determining the timeliness tags of the target webpage ranking model training data based on the correspondence between different click-through rates and different timeliness tags; and combining the news ranking model training data and the target webpage ranking model training data to obtain the news ranking model training data used for training the news ranking model.

[0108] In some embodiments, since the training data for the webpage ranking model is a massive, constantly updated dataset, training data that has already undergone data migration processing (updating the relevance labels of the webpage ranking model training data to the timeliness labels of the news ranking model training data) can be stored in a database. Data migration can then be performed on training data that has not undergone data migration processing. This enables real-time migration, improving the efficiency and speed of synchronizing the webpage ranking model training data to the news ranking model training data. The first training data corresponding to the timeliness labels is stored in the database, and a correspondence is established between the first training data and the second training data. The first training data is the news ranking model training data, and the second training data is the webpage ranking model training data. When the second training data is updated, the timeliness labels of the second training data for which no correspondence has been established are determined.

[0109] In some embodiments, relevance labels can be determined based on the relevance between the news ranking model training data and the query string corresponding to the news ranking model training data. That is, the relevance labels of the news ranking model training data can be obtained by: performing content matching processing between the query string corresponding to the news ranking model training data and the news ranking model training data to obtain the content matching degree; and determining the relevance label corresponding to the content matching degree of the news ranking model training data based on the correspondence between different content matching degrees and different relevance labels.

[0110] In other embodiments, relevance labels can be determined based on the relevance between the news ranking model training data and the corresponding query string, as well as the timeliness of the news ranking model training data. That is, relevance and timeliness can be combined to determine relevance labels, where the weight of relevance is greater than the weight of timeliness. This combination of relevance and timeliness can be achieved by: obtaining the timeliness weight value and the relevance weight value of the news ranking model training data corresponding to the news ranking model training data; wherein the timeliness weight value is less than the relevance label weight value; and performing a weighted summation of the timeliness and relevance labels based on the timeliness weight value and the relevance weight value, using the weighted summation result as the relevance label of the news ranking model training data.

[0111] In this embodiment, considering that some webpage ranking model training data contains timeliness features, the relevance and timeliness are fused to determine the relevance label, which is applicable to all scenarios of webpage ranking model training data, and achieves the purpose of automatically constructing news ranking model training data based on webpage ranking model training data.

[0112] The following describes an exemplary application of the embodiments of this application in a practical application scenario. The news ranking model using the news processing method can be a news search engine integrated into a news search server, or it can be deployed to a separate news ranking server, which can be called by the news search engine. Taking a news search scenario as an example, the news search server extracts a portion of the training data that meets the timeliness requirements from the webpage ranking model training data as the news ranking model training data; based on the timeliness and relevance labels of the news ranking model training data, it determines the timeliness labels of the news ranking model training data; based on the news ranking model training data and the corresponding timeliness labels, it trains the news ranking model; when a search keyword (i.e., query string) is received, it sorts multiple news items based on the news ranking model trained in the news search server, and displays the sorting results in the news search engine. That is, by migrating the webpage ranking model training data to the news ranking model training data, timeliness is introduced, and high-quality news information that combines timeliness and relevance is recommended to the user. See also... Figure 8 , Figure 8 This is a schematic flowchart of a news processing method provided in an embodiment of this application. The following will combine... Figure 8 The news processing method provided in the embodiments of this application will be described in detail.

[0113] In step 801: The news search server obtains the training data for the webpage ranking model.

[0114] In some embodiments, training data of a web page ranking model labeled with relevance tags is obtained from the web page ranking model. Since the web page search scenario does not emphasize timeliness, the relevance tags are mainly based on relevance.

[0115] In some examples, relevance labels can be determined based on the relevance between the news ranking model training data and the corresponding query string. In other examples, relevance labels can be determined based on the relevance between the news ranking model training data and the corresponding query string, as well as the timeliness of the news ranking model training data. That is, relevance and timeliness can be combined to determine relevance labels, where the weight of relevance is greater than the weight of timeliness.

[0116] In step 802: The news search server extracts and processes the training data of the web page ranking model to obtain the training data of the news ranking model.

[0117] In some embodiments, since the training data for the webpage ranking model is historically accumulated labeled data, the news ranking model training data to be extracted (the portion of training data that meets the timeliness requirement) is mainly divided into two categories: one category is webpage ranking model training data belonging to hot news of a certain historical period. Although the timeliness score of webpage ranking model training data belonging to current old news is low, it should still be extracted as news ranking model training data because it involves popular news that occurred in that year. The other category is webpage ranking model training data that always has the requirement of news timeliness. That is, through the timeliness calculation module of the webpage ranking model training data processing system, webpage ranking model training data with a timeliness greater than the timeliness threshold is obtained, and this webpage ranking model training data will also be extracted as news ranking model training data.

[0118] In some examples, the determination of historical trending news in the training data of the first type of timeliness-based web page ranking model can be achieved as follows: collect various historical news topics and other operational data, and perform text matching between the query string corresponding to the web page ranking model training data and the titles of these news topics. For web page ranking model training data with a matching degree higher than the title matching degree, it can be considered as web page ranking model training data related to historical trending news. Thus, the web page ranking model training data determined to belong to historical trending news is used as news ranking model training data (the part of the training data that meets the timeliness requirement). The determination of the second type of web page ranking model training data can be achieved as follows: obtain the timestamp of the web page ranking model training data; use the time interval between the timestamp and the current time point as the timeliness of the web page ranking model training data; and use the web page ranking model training data with a timeliness greater than the timeliness threshold as news ranking model training data (the part of the training data that meets the timeliness requirement).

[0119] In step 803: The news search server constructs training data for a news ranking model based on the idea of ​​transfer learning.

[0120] In some embodiments, constructing training data for a news ranking model based on transfer learning involves updating the original relevance labels of the training data to timeliness labels. This involves obtaining the weight values ​​of timeliness and relevance labels; wherein the weight value of timeliness is greater than the weight value of relevance labels; and then performing a weighted summation of the timeliness and relevance labels based on these weight values, using the weighted sum as the timeliness label for the news ranking model training data.

[0121] For example, the relevance label (level 3) and timeliness label (level 3) are combined to adjust the timeliness label to level 5. For instance, training data for a news ranking model with original labels (relevance labels) of level 3 and a news time lag (immediacy timeliness) of 1 day is identified as sample data with high text relevance and strong timeliness; that is, the adjusted ranking sample for news search is now at level 5. See [link / reference] Figure 4 This allows us to determine the timeliness label of the updated training data for the news ranking model. The timeliness of the training data is categorized into three levels: 1, 2, and 3, representing 1-3 days, 3-11 days, and more than 11 days, respectively. The relevance label is also categorized into three levels: 1, 2, and 3, representing highly relevant, moderately relevant, and irrelevant, respectively.

[0122] In some examples, the timeliness of news ranking model training data can be determined as follows: Obtain the search volume of the query string corresponding to the news ranking model training data across multiple time periods; iterate through any two adjacent time periods and perform the following processing on these two adjacent time periods: fit the search volume of the adjacent time periods to obtain a test statistic; map the test statistic to a confidence level based on a chi-square distribution. Here, the confidence level characterizes the correlation between the end time and the increase in search volume between the two adjacent time periods; the two adjacent time periods corresponding to the highest confidence level are taken as the two adjacent time periods corresponding to the query string. The time interval between the end time of the next time period corresponding to the two adjacent time periods corresponding to the query string and the current time is taken as the timeliness of the news ranking model training data.

[0123] As an example, the number of searches (search volume) of the query string corresponding to the news ranking model training data in multiple time periods is obtained from the web search logs. The news ranking model training data is extracted from the web ranking model training data, which is a portion of the training data in the web search model training data. The search volume at each moment recorded in the web search logs is counted, and the search volume for multiple time periods is obtained based on the statistical time interval. For example, if the statistical time interval is selected as 5 seconds, the search volume for 0-5 seconds from 11:20:00 on April 7, 2021, the search volume for 5-10 seconds from 11:20:00 on April 7, 2021, the search volume for 10-15 seconds from 11:20:00 on April 7, 2021, and so on, is obtained to obtain the search volume of the query string corresponding to the news ranking model training data in multiple time periods.

[0124] For example, a chi-square test can be used for fitting to examine whether the increase in search volume between adjacent time periods is significantly different. Let's assume two adjacent time periods are T. i-1 ~T i T i ~T i+1 Get the time period t0 (T i-1 Time ~T i Time period t1 (T) i Time ~T i+1 The search volume and non-search volume (each occurrence of the query string in the log is counted as one search) are used to construct a four-cell table based on two adjacent time periods. See also Figure 6 , Figure 6 This is a schematic diagram of a four-cell table provided in an embodiment of this application. The chi-square value can be obtained based on the four-cell table. The chi-square value is used as the test statistic, which can detect whether a sudden, explosive fluctuation has occurred in the current time period. See also Figure 7 , Figure 7 This is a schematic diagram of the chi-square distribution provided in an embodiment of this application. Assuming a chi-square value of 6.109, consulting the chi-square distribution, we know that chi-square value 5.024 < chi-square value 6.109 < chi-square value 6.635. The probability corresponding to chi-square value 5.024 is 0.025, and the probability corresponding to chi-square value 6.635 is 0.010. Therefore, we can determine that the probability range of the confidence level between two adjacent time periods (time period t0 and time period t1) and the increase in search volume corresponding to the query string is between 1-0.01 and 1-0.025. We statistically analyze the confidence levels of any two adjacent time periods within multiple time periods. Based on the comparison of the probability ranges of multiple confidence levels, the two adjacent time periods corresponding to the highest confidence level are taken as the two adjacent time periods corresponding to the query string (i.e., the two adjacent time periods related to the significant change in the increase of the query string). Thus, T... i+1The time interval between a given moment and the current moment serves as the timeliness of the training data for the news ranking model.

[0125] In this embodiment, the relevance labels of the webpage ranking model training data are migrated to timeliness labels that include timeliness and relevance. This allows for the reuse of the original relevance labels and enhances timeliness. In other words, it can be well compatible with the relevance labels of the webpage ranking model training data and adds some features (timeliness) of the news ranking model for news search scenarios to the relevance labels of the webpage ranking model training data.

[0126] In step 804: The news search server trains the news ranking model based on the news ranking model training data and the timeliness tags.

[0127] In some examples, the number of training data points at different tiers in the news ranking model's training dataset is adjusted according to the model's requirement for timeliness. The model is then repeatedly trained until the desired tone is achieved. This adjustment can be done by copying different numbers of samples from a specific tier within the training set of the news ranking model. For instance, if the goal is to further enhance the news search engine's (which operates through a news ranking model) sensitivity to timeliness, multiple copies of the training data (4-5 tiers) can be made to influence the final news ranking model's bias through sample distribution.

[0128] In other embodiments, the news ranking model can be applied to news recommendation scenarios, where a client (e.g., a news client) receives a news feed pushed by a news recommendation server, and the news is ranked according to the news processing method provided in the embodiments of this application.

[0129] In this embodiment, training data for a news ranking model is automatically extracted from the training data of the webpage ranking model. By enhancing the timeliness features required by the news ranking model, the training data of the webpage ranking model is reused in the training data of the news ranking model to obtain the news ranking model. This achieves the goal of automatically constructing training data for the news ranking model based on the training data of the webpage ranking model, and can maintain the consistency of the user experience under the webpage ranking model and the news ranking model as much as possible.

[0130] The following description continues to illustrate the exemplary structure of the news processing device 233 provided in the embodiments of this application as a software module. In some embodiments, such as Figure 2 As shown, the software modules stored in the news processing device 233 of the memory 230 may include:

[0131] Extraction module 2331 is used to extract a portion of training data that meets the timeliness requirements from the webpage ranking model training data, and use it as training data for the news ranking model; determination module 2332 is used to determine the timeliness label of the news ranking model training data based on the timeliness and relevance label of the news ranking model training data; training model 2333 is used to train the news ranking model based on the news ranking model training data and the corresponding timeliness label; ranking module 2334 is used to rank multiple news information based on the trained news ranking model.

[0132] In some embodiments, the determining module 2332 is further configured to obtain the search volume of the query string corresponding to the news ranking model training data in multiple time periods; traverse any two adjacent time periods in the multiple time periods, perform verification processing on the search volume of any two adjacent time periods to obtain the two adjacent time periods corresponding to the query string; and take the time interval between the end time of the next time period corresponding to the two adjacent time periods and the current time as the timeliness of the news ranking model training data.

[0133] In some embodiments, the determining module 2332 is further configured to perform the following processing on any two adjacent time periods: perform fitting processing on the search volume of the two adjacent time periods to obtain a test statistic; map the test statistic to a confidence level based on a chi-square distribution; wherein the confidence level characterizes the correlation between the end time and the increment of the search volume of the two adjacent time periods; and take the two adjacent time periods corresponding to the maximum confidence level as the two adjacent time periods corresponding to the query string.

[0134] In some embodiments, the determining module 2332 is further configured to obtain the weight value of the timeliness and the weight value of the relevance label; wherein the weight value of the timeliness is greater than the weight value of the relevance label; and perform a weighted summation on the timeliness and the relevance label according to the weight value of the timeliness and the weight value of the relevance label, and use the weighted summation result as the timeliness label of the news ranking model training data.

[0135] In some embodiments, the training module 2333 is further configured to determine the proportion of training data corresponding to the timeliness sensitivity of the news ranking model based on the correspondence between different timeliness sensitivities and different proportions of training data in the news ranking model; wherein the timeliness sensitivity is positively correlated with the amount of target training data, the target training data is the news ranking model training data corresponding to the target level of the timeliness label, and the target level includes at least one of the first half of the timeliness label levels in descending order; and adjust the proportion of different types of training data in the news ranking model training data to conform to the proportion of training data corresponding to the timeliness sensitivity of the news ranking model.

[0136] In some embodiments, the extraction module 2331 is further configured to obtain the timestamp of the webpage ranking model training data; use the time interval between the timestamp and the current time point as the timeliness of the webpage ranking model training data; and use the portion of the webpage ranking model training data whose timeliness is greater than the timeliness threshold as the portion of training data that meets the timeliness requirement.

[0137] In some embodiments, the extraction module 2331 is further configured to obtain the titles of various historical news topics; match the webpage ranking model training data with the titles of the historical news topics to obtain the title matching degree; and use the portion of the webpage ranking model training data with the title matching degree greater than the matching degree threshold as the portion of training data that meets the timeliness requirements.

[0138] In some embodiments, the training module 2333 is further configured to acquire the click-through rate of the webpage ranking model training data, determine the target webpage ranking model training data, wherein the target webpage ranking model training data is the training data in the webpage ranking model training data whose click-through rate is higher than the click-through rate threshold; determine the timeliness label of the target webpage ranking model training data according to the correspondence between different click-through rates and different timeliness labels; and combine the news ranking model training data and the target webpage ranking model training data to obtain the news ranking model training data used to train the news ranking model.

[0139] In some embodiments, the determining module 2332 is further configured to obtain the relevance label of the news ranking model training data by: performing content matching processing on the query string corresponding to the news ranking model training data and the news ranking model training data to obtain the content matching degree; and determining the relevance label corresponding to the content matching degree of the news ranking model training data according to the correspondence between different content matching degrees and different relevance labels.

[0140] This application provides 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 news processing method described above in this application.

[0141] This application provides a computer-readable storage medium storing executable instructions. When these executable instructions are executed by a processor, they cause the processor to perform the news processing method provided in this application. For example, ... Figure 3A , 3B The news processing method shown in 3C.

[0142] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0143] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0144] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0145] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0146] In summary, this application's embodiments extract a portion of training data that meets the timeliness requirements from the webpage ranking model training data and use it as training data for the news ranking model. This automatically extracts training data for the news ranking model from the webpage ranking model training data. Then, based on the timeliness and relevance labels of the news ranking model training data, a timeliness label is determined. This allows the training data from the webpage ranking model to be reused in the news ranking model, ensuring both relevance and timeliness. By reusing the webpage ranking model training data, the timeliness features required by the news ranking model are strengthened, resulting in a news ranking model that maintains consistency in user experience across both the webpage ranking model and the news ranking model. The consensus mechanism among nodes in the blockchain network ensures the reliability of the webpage ranking model training data and relevance labels, as well as the trustworthiness of the calculation process. Because webpage ranking models and news ranking models have different requirements, this study considers the two aspects of timeliness requirements (timeliness and historical trending news) for users during the training process of news ranking models. A timeliness screening process is performed on the training data of webpage ranking models to reduce ineffective training on massive amounts of historical webpage ranking model training data, improving training efficiency and ensuring the timeliness of the training data after the initial screening. Based on the news ranking model, timeliness tags are predicted for news information. The results of ranking multiple news items corresponding to the top-ranked items with descending timeliness tags are then obtained and recommended to the user's terminal. The recommended news information not only maintains relevance to the search query string but also meets the user's timeliness requirements. Considering that some webpage ranking model training data contains timeliness features, relevance and timeliness are integrated to determine relevance tags, adapting to all scenarios of webpage ranking model training data. This achieves the goal of automatically constructing news ranking model training data based on webpage ranking model training data.

[0147] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A news processing method characterized by comprising: The method comprises: a time interval between a timestamp of webpage ranking model training data and a current time point as a timeliness of the webpage ranking model training data; part of the webpage ranking model training data with the timeliness greater than a timeliness threshold as news ranking model training data; determining a timeliness label of the news ranking model training data based on the timeliness of the news ranking model training data and a relevance label of the news ranking model training data; training the news ranking model based on the news ranking model training data and the corresponding timeliness label; sorting a plurality of news information based on the trained news ranking model.

2. The method of claim 1, wherein, Before determining the timeliness label of the news ranking model training data, the method further comprises: obtaining search volumes of a query string corresponding to the news ranking model training data in a plurality of time periods; iterating through any two adjacent time periods in the plurality of time periods, and performing inspection processing on the search volumes of the any two adjacent time periods to obtain two adjacent time periods corresponding to the query string; a time interval between an end time of a later time period of the two adjacent time periods corresponding to the query string and a current time as the timeliness of the news ranking model training data.

3. The method of claim 2, wherein, The inspection processing on the search volumes of the any two adjacent time periods to obtain the two adjacent time periods corresponding to the query string comprises: performing the following processing on any two adjacent time periods: fitting the search volumes of the two adjacent time periods to obtain a test statistic; mapping the test statistic to a confidence degree based on a chi-square distribution; wherein the confidence degree represents a correlation degree between the end time and an increment of the search volumes of the two adjacent time periods; the two adjacent time periods corresponding to the maximum confidence degree as the two adjacent time periods corresponding to the query string.

4. The method of claim 1, wherein, The determination of the timeliness label of the news ranking model training data based on the timeliness of the news ranking model training data and the relevance label of the news ranking model training data comprises: obtaining a weight value of the timeliness and a weight value of the relevance label; wherein the weight value of the timeliness is greater than the weight value of the relevance label; performing weighted summation processing on the timeliness and the relevance label according to the weight value of the timeliness and the weight value of the relevance label, and taking a weighted summation result as the timeliness label of the news ranking model training data.

5. The method of claim 1, wherein, Before training the news ranking model based on the news ranking model training data and the corresponding timeliness label, the method further comprises: determining a training data quantity proportion corresponding to a timeliness sensitivity of the news ranking model according to a corresponding relationship between different timeliness sensitivities and different training data quantity proportions of the news ranking model; wherein the timeliness sensitivity is positively correlated with a quantity of target training data, the target training data being the news ranking model training data corresponding to a target level of the timeliness label, and the target level including at least one of a first half of a level descending order of the timeliness label. Adjust the quantity proportion between different types of training data in the news ranking model training data to conform to the training data quantity proportion corresponding to the timeliness sensitivity of the news ranking model.

6. The method of claim 1, wherein, The method further comprises: Obtaining the titles of various historical news topics; Matching the web page ranking model training data with the titles of the historical news topics to obtain a title matching degree; Part of the web page ranking model training data with a title matching degree greater than a matching degree threshold is used as part of the training data that meets the timeliness requirement, and the part of the training data is used as the news ranking model training data.

7. The method of claim 1, wherein, Before training the news ranking model based on the news ranking model training data and the corresponding timeliness label, the method further comprises: Obtaining the click rate of the web page ranking model training data to determine target web page ranking model training data, which is the training data in the web page ranking model training data with a click rate higher than a click rate threshold; According to the correspondence between different click rates and different timeliness labels, the timeliness label of the target web page ranking model training data is determined; The news ranking model training data and the target web page ranking model training data are combined to obtain the news ranking model training data used to train the news ranking model.

8. The method of claim 1, wherein, The method further comprises: Obtaining the relevance label of the news ranking model training data by: Performing content matching processing on the query string corresponding to the news ranking model training data and the news ranking model training data to obtain a content matching degree; According to the correspondence between different content matching degrees and different relevance labels, the relevance label corresponding to the content matching degree of the news ranking model training data is determined.

9. A news processing apparatus characterized by comprising: The device comprises: An extraction module configured to use the time interval between the timestamp of the web page ranking model training data and the current time point as the timeliness of the web page ranking model training data, and use part of the web page ranking model training data with a timeliness greater than a timeliness threshold as the news ranking model training data; A determination module configured to determine the timeliness label of the news ranking model training data based on the timeliness of the news ranking model training data and the relevance label of the news ranking model training data; A training module configured to train the news ranking model based on the news ranking model training data and the corresponding timeliness label; A ranking module configured to perform ranking processing on multiple news information based on the trained news ranking model.

10. The apparatus of claim 9, wherein, The determination module is further configured to, before determining the timeliness label of the news ranking model training data, obtain the search volume of the query string corresponding to the news ranking model training data in multiple time periods; The search volume of any two adjacent time periods in the multiple time periods is verified to obtain two adjacent time periods corresponding to the query string; An interval between an end time of a later time period of two adjacent time periods corresponding to the query string and a current time is taken as time-newness of the news ranking model training data.

11. The apparatus of claim 10, wherein, The determination module is further configured to: perform the following processing on any two adjacent time periods: fitting processing is performed on search volumes of the two adjacent time periods to obtain a test statistic; map the test statistic to a reliability based on chi-square distribution; wherein the reliability represents a degree of correlation between the end time and an increment of the search volumes of the two adjacent time periods; two adjacent time periods corresponding to the maximum reliability are taken as the two adjacent time periods corresponding to the query string.

12. The apparatus of claim 9, wherein, The determination module is further configured to: obtain a weight value of the time-newness and a weight value of the relevance label; wherein the weight value of the time-newness is greater than the weight value of the relevance label; perform weighted sum processing on the time-newness and the relevance label according to the weight value of the time-newness and the weight value of the relevance label, and take a weighted sum result as a time-newness label of the news ranking model training data.

13. The apparatus of claim 9, wherein, The training module is further configured to, before training the news ranking model based on the news ranking model training data and a corresponding time-newness label, determine a training data quantity proportion corresponding to a time-newness sensitivity of the news ranking model according to a corresponding relationship between different time-newness sensitivities and different training data quantity proportions of the news ranking model; wherein the time-newness sensitivity is positively correlated with a quantity of target training data, the target training data being the news ranking model training data corresponding to a target level of the time-newness label, and the target level including at least one of a first half of a rank order of the levels of the time-newness label in descending order; adjust a quantity proportion between different types of training data in the news ranking model training data to conform to the training data quantity proportion corresponding to the time-newness sensitivity of the news ranking model.

14. The apparatus of claim 9, wherein, The extraction module is further configured to: obtain titles of various historical news topics; perform matching processing on the web page ranking model training data and the titles of the historical news topics to obtain a title matching degree; take part of the web page ranking model training data with a title matching degree greater than a matching degree threshold as part of training data meeting a time-newness requirement, and take the part of training data as news ranking model training data.

15. The apparatus of claim 9, wherein, The training module is further configured to, before training the news ranking model based on the news ranking model training data and a corresponding time-newness label, obtain a click rate of the web page ranking model training data, determine target web page ranking model training data, and the target web page ranking model training data being training data in the web page ranking model training data with a click rate higher than a click rate threshold; determine a time-newness label of the target web page ranking model training data according to a corresponding relationship between different click rates and different time-newness labels; determine a time-newness label of the target web page ranking model training data according to a corresponding relationship between different click rates and different time-newness labels; The news ranking model training data and the target webpage ranking model training data are combined to obtain the news ranking model training data for training the news ranking model.

16. The apparatus of claim 9, wherein, The determination module is further configured to: The relevance label of the news ranking model training data is obtained in the following manner: The content matching degree is obtained by performing content matching processing on the news ranking model training data corresponding to the query string and the news ranking model training data. According to a corresponding relationship between different content matching degrees and different relevance labels, the relevance label corresponding to the content matching degree of the news ranking model training data is determined.

17. An electronic device, comprising: Comprise: A memory for storing executable instructions; A processor for executing the executable instructions stored in the memory, realizing the news processing method of any one of claims 1 to 8.

18. A computer-readable storage medium, characterized in that, Executable instructions are stored for being executed by the processor, realizing the news processing method of any one of claims 1 to 8.

Images