A news text recommendation method and device based on a large model, equipment and a medium

Abstract

The application discloses a news text recommendation method and device based on a large model, equipment and a medium, and relates to the field of artificial intelligence. The method comprises the following steps: performing data cleaning and standardization processing operations on collected news content to be classified and user information to obtain processed news content and processed user information, constructing multi-modal features according to the processed news content and the processed user information; using a target news text classification large model based on the multi-modal features to classify the processed news content to generate classified news content and generate a target semantic vector; using a target user interest recognition large model based on the multi-modal features to perform a preset interest recognition operation on the processed user information to generate a target user interest recognition result; inputting the target semantic vector and the target user interest recognition result into a target news text recommendation large model to obtain a target news text recommendation list, and displaying the news text to a user end. The method can accurately classify news content and provide personalized recommendations.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to a method, apparatus, device, and medium for recommending news text based on a large model. Background Technology

[0002] Currently, traditional solutions mainly rely on shallow statistical features or manually generated rules, resulting in a significant "semantic gap." For example, methods based on word frequency and keyword matching cannot understand the semantic relationship between "artificial intelligence" and "machine learning," and struggle to distinguish the specific reference of "apple" in different contexts, thus limiting classification accuracy. Furthermore, these methods depend on manually maintained rule systems, making them ill-suited for emerging fields and dynamically changing news topics.

[0003] On the other hand, mainstream recommendation algorithms such as collaborative filtering heavily rely on users' historical behavior data, facing a serious "cold start" problem and poor recommendation quality for new users. They also struggle to capture complex contextual semantics, resulting in insufficient relevance of recommended content and low click-through rates. Furthermore, news data is highly time-sensitive; traditional batch processing architectures suffer from high response latency, failing to meet the real-time requirements of minute-level processing and distribution for breaking news. The fusion and unified representation of multi-source heterogeneous data also presents significant engineering challenges.

[0004] In conclusion, how to efficiently and accurately classify news content and provide personalized recommendations is a problem that urgently needs to be solved. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a news text recommendation method, apparatus, device, and medium based on a large model, capable of efficiently and accurately classifying news content and providing personalized recommendations. The specific solution is as follows: Firstly, this application provides a news text recommendation method based on a large model, including: Collect news content to be classified and user information on the news platform. Perform data cleaning and standardization on the collected news content to be classified and user information to obtain processed news content and processed user information. Construct multimodal features based on the processed news content and processed user information. A target news text classification model is constructed. Based on the multimodal features, the target news text classification model is used to classify the processed news content, generate classified news content, and generate a corresponding target semantic vector for each classified news content. Construct a large target user interest recognition model, and use the large target user interest recognition model to perform preset interest recognition operations on the processed user information based on the multimodal features to generate target user interest recognition results; A target news text recommendation model is established. The target semantic vector and the target user interest recognition result are input into the target news text recommendation model to obtain a target news text recommendation list. The news texts in the target news text recommendation list are displayed to the user terminal in a preset order.

[0006] Optionally, the news content to be categorized includes structured news data, unstructured news data, or multimedia data, and the user information includes operation logs and user profile data on the news platform.

[0007] Optionally, the collection of news content to be categorized and user information on the news platform includes: The system connects to the target data source via a pre-defined API to obtain structured news data; the structured news data includes a title, body text, publication time, and category tags. A distributed crawler cluster is used to crawl user-generated content from target social media, and the user-generated content from the target social media is integrated as unstructured news data; Extract text and video subtitles from the images of the news content to be categorized to construct multimedia data; Real-time capture of user activity logs on the news platform; The system collects user registration information on news platforms, social network association data, and device information to create user profile data.

[0008] Optionally, the step of performing data cleaning and standardization processing on the collected news content to be categorized and the user information includes: The news content to be categorized is filtered by using a preset regular expression to filter out advertising links and special characters. The filtered news content is subjected to sentiment polarity analysis using a preset sentiment model. The analysis results are used to determine whether the results contain clickbait keywords and whether the sentiment polarity value of the clickbait keywords is greater than a preset sentiment threshold. If the news content contains clickbait keywords and the sentiment polarity value of the clickbait keywords is greater than a preset sentiment threshold, then the corresponding filtered news content will be marked as the first invalid data. If the news content does not contain clickbait keywords or the sentiment polarity value of the clickbait keywords is less than or equal to a set threshold, then the corresponding filtered news content is marked as the first valid data. The first valid data is subjected to authenticity verification using a pre-set rumor detection model; If the judgment result indicates that the confidence level of the first valid data is less than the preset confidence threshold, then the corresponding first valid data is determined as the second invalid data; If the judgment result indicates that the confidence level of the first valid data is greater than or equal to the preset confidence threshold, then the corresponding first valid data is determined as the second valid data; The text fingerprint of the second valid data is determined by a preset similarity hash algorithm, and the similarity between different text fingerprints is generated. If there is second valid data with a similarity greater than or equal to a preset similarity threshold, then only the second valid data whose publication time meets the preset latest condition will be retained; The retained second valid data is subjected to time format unification and entity alignment operations to obtain the processed news content.

[0009] Optionally, the step of constructing multimodal features based on the processed news content and processed user information includes: Statistical features are constructed based on the processed news content; wherein the statistical features include any one or more of the following: word count, sentence count, punctuation mark ratio, and keyword density; Determine the sentiment score of the processed news content, and determine the corresponding sentiment features based on the sentiment score of the processed news content; Obtain the publication time of the processed news content, and obtain the corresponding time-series features based on the publication time of the processed news content; The semantic features of the processed news content are identified by a pre-trained BERT large model; Identify the operation logs in the processed user information to obtain the user behavior characteristics of the user terminal; Multimodal features are constructed based on the statistical features, the sentiment features, the temporal features, the semantic features, and the user behavior features.

[0010] Optionally, the step of performing a preset interest recognition operation on the processed user information based on the multimodal features using a target user interest recognition model to generate a target user interest recognition result includes: Using a target user interest identification big model based on real-time behavioral data of the user's behavioral characteristics, the semantic feature vector of the user's click on news within a preset time window is extracted, and the semantic feature vector is weighted and fused according to a preset time decay rule to generate a target short-term interest vector. Based on the user behavior characteristics and historical behavior data within a preset historical period, as well as the user profile data, a target long-term interest vector is generated through topic analysis and knowledge graph embedding methods. The target short-term interest vector and the target long-term interest vector are weighted and fused using a preset attention mechanism to generate a fusion result. The fusion result is then corrected based on the target negative feedback behavior list provided by the user terminal to generate the target user interest vector.

[0011] Optionally, the step of inputting the target semantic vector and the target user interest recognition result into the target news text recommendation model to obtain a target news text recommendation list includes: Based on the target user interest identification results and the target semantic vector, a multi-path recall operation is performed to generate an initial news set that includes semantic recall, collaborative filtering recall, and cold start recall. The news in the initial news set is sorted using a preset deep semantic matching model, the predicted score of each news item in the initial news set is output, and an initial sorted list is generated based on the predicted scores of the news items. The initial sorted list is optimized and rearranged to generate a target news text recommendation list; wherein the strategy optimization includes preset diversity control operations, preset real-time enhancement operations, and preset regional adaptation operations.

[0012] Secondly, this application provides a news text recommendation device based on a large model, comprising: The feature construction module is used to collect news content to be classified and user information on the news platform. It performs data cleaning and standardization operations on the collected news content to be classified and user information to obtain processed news content and processed user information. Multimodal features are constructed based on the processed news content and processed user information. The vector generation module is used to construct a target news text classification model, classify the processed news content based on the multimodal features using the target news text classification model, generate classified news content, and generate corresponding target semantic vectors for each classified news content. The result generation module is used to construct a large target user interest recognition model, and based on the multimodal features, to perform preset interest recognition operations on the processed user information using the large target user interest recognition model to generate target user interest recognition results. The text recommendation module is used to build a large-scale model for recommending target news texts. The target semantic vector and the target user interest recognition results are input into the large-scale model to obtain a list of target news text recommendations. The news texts in the list are then displayed to the user in a preset order.

[0013] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the large model-based news text recommendation method described above.

[0014] Fourthly, this application provides a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned news text recommendation method based on a large model.

[0015] In summary, this application first collects news content to be categorized and user information from the news platform. Data cleaning and standardization are then performed on the collected news content and user information to obtain processed news content and processed user information. Multimodal features are constructed based on the processed news content and processed user information. A target news text classification model is then constructed. Based on the multimodal features, the target news text classification model is used to classify the processed news content, generating categorized news content. A corresponding target semantic vector is generated for each categorized news content. A target user interest recognition model is then constructed. Based on the multimodal features, the target user interest recognition model is used to perform preset interest recognition operations on the processed user information to generate target user interest recognition results. Finally, a target news text recommendation model is established. The target semantic vector and the target user interest recognition results are input into the target news text recommendation model to obtain a target news text recommendation list. The news texts in the target news text recommendation list are then displayed to the user in a preset order. As described above, this application first collects the news content to be classified and user information from the news platform. After data cleaning and standardization, corresponding processing results are obtained. Then, multimodal features are constructed based on these results. Subsequently, a large-scale target news text classification model is built. The multimodal features are used to classify the processed news content and generate corresponding target semantic vectors. At the same time, a large-scale target user interest recognition model is built. Based on the multimodal features, a preset interest recognition operation is performed on the processed user information to obtain the target user interest recognition results. Finally, a large-scale target news text recommendation model is established. The target semantic vector and the target user interest recognition results are input into the model to obtain a target news text recommendation list, and the news texts in the list are displayed to the user in a preset order. In this way, by constructing a news text semantic classification model based on BERT (Bidirectional Encoder Representations from Transformers, language representation model), the problem of insufficient semantic understanding in traditional methods is solved. At the same time, a dynamic interest modeling framework that integrates real-time user behavior and long-term preferences is designed to improve recommendation accuracy. In addition, we developed a high-concurrency real-time data processing pipeline to achieve minute-level response from news collection to recommendation, provide an interpretable recommendation mechanism, and enhance users' trust in the recommendation results. Attached Figure Description

[0016] 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.

[0017] Figure 1 This is a flowchart of a news text recommendation method based on a large model disclosed in this application; Figure 2 This is a schematic diagram of the structure of a news text recommendation device based on a large model disclosed in this application; Figure 3 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0018] 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.

[0019] Currently, traditional solutions mainly rely on shallow statistical features or manual rules, resulting in a significant "semantic gap." Furthermore, these methods depend on manually maintained rule systems, making them ill-suited for emerging fields and dynamically changing news topics. On the other hand, mainstream recommendation algorithms such as collaborative filtering heavily rely on users' historical behavior data, facing a severe "cold start" problem and poor recommendation quality for new users. They also struggle to capture complex contextual semantics, leading to insufficient relevance of recommended content and low user click-through rates. In addition, news data is highly time-sensitive; traditional batch processing architectures suffer from high response latency, failing to meet the real-time requirements of minute-level processing and distribution of breaking news. The fusion and unified representation of multi-source heterogeneous data also presents significant engineering challenges. To address these technical problems, this application discloses a news text recommendation method, apparatus, device, and medium based on a large model, capable of efficiently and accurately classifying news content and providing personalized recommendations.

[0020] See Figure 1 As shown in the figure, this invention discloses a news text recommendation method based on a large model, including: Step S11: Collect the news content to be classified and the user information of the user on the news platform. Perform data cleaning and standardization processing on the collected news content to be classified and the user information to obtain the processed news content and processed user information. Construct multimodal features based on the processed news content and processed user information.

[0021] In this embodiment, news content is first collected from multiple data sources. The news content to be categorized includes structured news data, unstructured news data, or multimedia data. The user information includes operation logs and user profile data from the news platform. It is important to note that structured news data is obtained by connecting to the target data source through a preset API (Application Programming Interface). This structured news data includes a title, body text, publication time, and category tags. A distributed crawler cluster is used to crawl user-generated content from the target social media platform, and this user-generated content is integrated as unstructured news data. Text and video subtitles are extracted from the images of the news content to be categorized to construct multimedia data. User operation logs on the news platform are captured in real time. User registration information, social network association data, and device information on the news platform are collected as user profile data. Specifically, the system connects to authoritative data sources via API to obtain formatted news articles including titles, body text, publication time, and category tags; it utilizes a distributed crawler cluster to scrape user-generated content from social media and forums, with a daily collection volume of ≥5 million articles, covering text and images, as well as short video text extracted via OCR (Optical Character Recognition); it extracts text from news images and video subtitles to construct a multimodal text corpus. The collected news content is then used as the subject of classification. Simultaneously, it captures real-time user activity logs on news platforms, such as browsing duration, click locations, collection / sharing / comment records, and search keywords; it also collects user profile data, including registration information such as age, gender, and region, social network association data such as LinkedIn professional information, and device information such as phone model and operating system. These user activity logs and user profile data are then used as user information.

[0022] Furthermore, advertising links and special symbols in the news content to be categorized are filtered using a preset regular expression to obtain filtered news content. A preset sentiment model is used to perform sentiment polarity analysis on the filtered news content to determine whether the analysis result contains clickbait keywords and whether the sentiment polarity value of the clickbait keywords is greater than a preset sentiment threshold. If clickbait keywords are included and their sentiment polarity value is greater than the preset sentiment threshold, the corresponding filtered news content is marked as first invalid data. If clickbait keywords are not included or their sentiment polarity value is less than or equal to a set threshold, the corresponding filtered news content is marked as first valid data. A preset rumor detection model is then used to further analyze the first valid data. The system performs an authenticity check. If the check result indicates that the confidence level of the first valid data is less than a preset confidence threshold, then the corresponding first valid data is determined as the second invalid data. If the check result indicates that the confidence level of the first valid data is greater than or equal to the preset confidence threshold, then the corresponding first valid data is determined as the second valid data. The system uses a preset similarity hash algorithm to determine the text fingerprint of the second valid data and generates similarity between different text fingerprints. If there is second valid data with a similarity greater than or equal to a preset similarity threshold, then only the second valid data whose publication time meets the preset latest condition is retained. The retained second valid data undergoes time format unification and entity alignment operations to obtain the processed news content. Specifically, in the noise filtering step, a preset regular expression pattern `http(s)?: / / [^\s]+` is used to match advertising links. This pattern can capture strings that start with http: / / or https: / / until a whitespace character is encountered. At the same time, an expression is preset to match non-text special symbols. The original news content to be classified is used as input. Regular expressions are applied to scan and replace the text, removing matched advertising links and special characters to generate filtered news content free of this noise. In the sentiment filtering step, a pre-trained RoBERTa sentiment model is used to perform sentiment polarity analysis on the filtered news content. The model outputs the sentiment polarity value of the content. Simultaneously, a table containing typical clickbait headline features such as "Shocking!" and "Must-Read!" is maintained. The filtered news content is checked to see if it contains any of these features, and if so, its sentiment polarity value is greater than a preset threshold of 0.8. If the content simultaneously meets both conditions of "containing a feature word" and "sentiment polarity value > 0.8", it is marked as the first invalid data; otherwise, if the content does not contain a feature word or contains a feature word but its sentiment polarity value is ≤ 0.8, it is marked as the first valid data. In the rumor detection step, a BERT model finely tuned based on the FactCheck dataset is used to determine the authenticity of the first valid data. The model outputs a confidence score representing the authenticity of the content.The confidence score output by the model is compared with a preset threshold. If the confidence score is less than the preset threshold, the content is judged as a rumor and identified as the second invalid data; if the confidence score is greater than or equal to the preset threshold, it is identified as the second valid data. In the duplicate detection step, the SimHash algorithm is used to calculate a fixed-length binary hash value as its text fingerprint for each piece of second valid data. Next, the similarity between all pairwise fingerprints of second valid data is calculated, specifically by calculating the Hamming distance of the fingerprints and converting it into a similarity value. When any two or more data fingerprints are found to have a similarity greater than or equal to a preset similarity threshold, these contents are judged as duplicate news. Then, the publication time of all the news determined to be duplicates is checked, and only the most recently published one is retained as a representative, while the remaining duplicates are removed. In the data standardization step, two processes are performed on the news data retained after duplicate detection. The first is to unify the time format, that is, to parse the publication time strings from different sources and with different formats in the news and convert them into a unified UTC (Coordinated Universal Time) timestamp representation. Then, entity pairing operations are performed. First, the BERT-NER model is used to identify named entities in the news text, such as names of people, organizations, and locations. Next, attempts are made to link these identified entity names to their corresponding unique entity identifiers in the Wikidata knowledge base. During this process, contextual information from the knowledge base is used to eliminate ambiguity regarding entities with the same name; for example, the context of the news content is used to distinguish whether "apple" refers to company entity Q321 or fruit entity Q35328. After completing all the above steps, the final output is the news content after noise filtering, sentiment and realism filtering, deduplication, and standardization.

[0023] Understandably, after standardizing the news content to obtain processed news content, features are constructed from shallow to deep features. First, statistical features are constructed based on the processed news content; these statistical features include any one or more of the following: word count, sentence count, punctuation mark ratio, and keyword density. Next, the sentiment score of the processed news content is determined, and the corresponding sentiment features are determined based on the sentiment score. The publication time of the processed news content is obtained, and the corresponding temporal features are obtained based on the publication time. Semantic features of the processed news content are identified using a pre-trained BERT large model. Operation logs in the processed user information are identified to obtain user behavior features from the user's end. Finally, multimodal features are constructed based on the statistical features, sentiment features, temporal features, semantic features, and user behavior features. Specifically, the process begins by constructing basic features, including statistical features, sentiment features, and temporal features. Statistical features include word count, sentence count, punctuation mark ratio, and keyword density. Sentiment features can be calculated using NLTK SentimentIntensityAnalyzer to obtain positive / negative sentiment scores, with a value range of [-1, 1]. Temporal features include weekdays, hours, seasons, and other periodic features corresponding to the publication time, used to analyze news dissemination patterns. Semantic features are then acquired, including BERT embedding features and topic features. The text is input into the BERT-base model, and the 12th layer CLS (Classification Token Vector) vector is extracted as the global semantic representation. The BERTopic model is used to model the topics of the news collection, such as generating 100 topic vectors and assigning Top 3 topic tags to each news item. Finally, user behavior features are developed. One type is short-term behavior vectors, such as the average BERT CLS vector of news clicks within the past 24 hours, reflecting real-time interests. The other type is long-term interest tags, calculated using an implicit feedback algorithm to determine the user's preference weights for each topic, ranging from [0, 1]. The resulting features are then defined as multimodal features.

[0024] Step S12: Construct a target news text classification model. Based on the multimodal features, use the target news text classification model to classify the processed news content, generate classified news content, and generate corresponding target semantic vectors for each classified news content.

[0025] In this embodiment, the first step is to construct a large-scale target news text classification model. This model can be based on a BERT-base-uncased pre-trained model, namely a 12-layer Transformer, 768-dimensional hidden layers, and 12 attention heads. The input layer uses WordPiece word segmentation, supporting a maximum input of 512 characters, thus determining the backbone network of the target news text classification model. The classification head is determined, and two fully connected layers (Dropout=0.1) are added after the CLS vector. The output is pre-defined to represent news categories, for example, 20 news categories such as technology, finance, sports, and entertainment. The cross-entropy loss function is used. Additionally, multi-task learning is implemented, allowing simultaneous prediction of news topics and sentiment polarity, optimized through a joint loss function. ,in To preset weights, =0.5; The loss function is defined for the news topic. The loss function is the emotional polarity. Let L be the loss function for news text classification; L is the total loss of the model. After completing the model architecture, the model is trained. First, data augmentation is performed using EDA (Easy Data Augmentation) techniques to perform synonym replacement, random insertion, swapping, and deletion operations on the text, improving the model's generalization ability. Then, the learning rate is scheduled using linear warmup, increasing the learning rate from 1e-7 to 5e-5 in the first 10% of steps, with cosine annealing decay. The training period is 8 epochs, and the batch size is 32. After training, the model is evaluated on the THUCNews test set. The classification accuracy reaches 96.3%, and the F1-score is 95.8%, which is 21% better than the traditional SVM model. The complete model training is then completed.

[0026] In addition, in the target news text classification model, a semantic vector retrieval engine is built to store the BERTCLS vectors of all news into a vector database, create an index, and support millisecond-level nearest neighbor queries; the index is updated incrementally on a regular basis, that is, for newly added news, vectors are generated in real time and inserted into the index, and deleted news is physically deleted to avoid ghost data.

[0027] Next, this large model is applied to the processed news content to perform a classification task, assigning one or more category labels to each news item, thereby generating classified news content with category identifiers. In the vectorization generation stage, a semantic vector generation module needs to be designed or employed. This module takes each classified news item as input and converts it into a fixed-dimensional numerical vector, i.e., the target semantic vector, through a specific encoding method or model. The target semantic vector is designed to represent the position and features of the news content in the semantic space, facilitating subsequent similarity calculations, retrieval, or clustering operations. Finally, a unique target semantic vector is generated for each classified news item.

[0028] Step S13: Construct a target user interest recognition model. Based on the multimodal features, use the target user interest recognition model to perform a preset interest recognition operation on the processed user information to generate target user interest recognition results.

[0029] In this embodiment, a large-scale target user interest identification model is constructed. Based on real-time behavioral data of the user's behavioral characteristics, the model extracts semantic feature vectors of news clicks made by the user within a preset time window. These semantic feature vectors are then weighted and fused according to a preset time decay rule to generate a target short-term interest vector. Based on historical behavioral data of the user's behavioral characteristics within a preset historical period and the user profile data, a target long-term interest vector is generated using topic analysis and knowledge graph embedding methods. The target short-term interest vector and the target long-term interest vector are weighted and fused using a preset attention mechanism to generate a fusion result. This fusion result is then corrected based on a target negative feedback behavior list provided by the user to generate the target user interest vector. Specifically, in short-term interest modeling, the semantic feature vectors of news clicks in the processed user information are used, with a window size set to 5 minutes. The following features are statistically analyzed: the distribution of clicked news categories (e.g., technology-related news accounts for 60%); high-frequency entity sets; and the distribution of sentiment tendencies (e.g., positive sentiment news accounts for >70%). For each clicked news item, its BERTCLS vector is extracted and weighted by time decay to generate the target short-term interest vector. , It is the category of the i-th news item; For the preset weight, it can be set to 0.95. In long-term interest modeling, user historical behavior data from the past 30 days is aggregated, and topic distribution is analyzed using a topic model such as BERTopic. The weights of specific topics are then manually adjusted based on user profiles, such as weighting entertainment topics for younger users. Simultaneously, knowledge graph embedding technology maps user profile attributes and news entities to the same vector space, generating target long-term interest vectors that reflect potential semantic relationships. In the interest fusion stage, a multi-head attention mechanism is used to fuse target short-term interest vectors. With the target long-term interest vector : , where Q= K= V= And through dynamic weighting coefficients Balance the contributions of both. The value is automatically adjusted based on the user's recent activity level (e.g., average daily clicks over the past 7 days): highly active users are focused on short-term interests. =0.7), low-activity users focus on long-term interests ( =0.3). Furthermore, based on negative user feedback behaviors, such as clicking "not interested" or having too short a dwell time, the semantic vector of the corresponding news item can be subtracted from the fusion vector with a negative weight, achieving dynamic correction of the interest vector. Finally, the target user's interest vector is generated.

[0030] Step S14: Establish a target news text recommendation model. Input the target semantic vector and the target user interest recognition result into the target news text recommendation model to obtain a target news text recommendation list. Then, display the news texts in the target news text recommendation list to the user terminal in a preset order.

[0031] In this embodiment, a large-scale target news text recommendation model is established. Based on the target user interest identification results and the target semantic vector, a multi-path recall operation is performed to generate an initial news set including semantic recall, collaborative filtering recall, and cold start recall. A preset deep semantic matching model is used to sort the news in the initial news set, outputting the predicted score for each news item in the initial news set, and generating an initial ranking list based on the predicted scores. The initial ranking list is then optimized and rearranged to generate a target news text recommendation list. The strategy optimization includes preset diversity control operations, preset real-time enhancement operations, and preset regional adaptation operations. Specifically, in the recall phase, the large-scale target news text recommendation model performs multi-path recall based on the target user interest vector and the news semantic vector. First, semantic recall is performed, calculating the similarity between the user interest vector and the news semantic vector using a vector retrieval engine, and returning the most relevant Top 200 news items. Next, collaborative filtering recall is performed, using a graph neural network model trained based on the user-news interaction graph to generate implicit representations of users and news, and recalling the Top 100 news items with similar user preferences based on the user representations. Finally, a cold start recall is triggered for new users, returning trending news and news categorized by default based on user profiles to form the initial news set. During the ranking phase, the initial news set is input into a deep semantic matching model. The model simultaneously receives user-side features, such as user interest vectors, activity levels, and historical click distribution; news-side features, such as news semantic vectors, publication time, and click-through rate; and cross-features between the two. The model calculates the matching score between the user and each news article through a multi-layer neural network, which serves as the basis for predicting the click-through rate. The model is optimized to maximize the accuracy of click-through rate prediction, outputting a predicted score for each news article in the initial news set, and generating an initial ranking list based on the descending order of scores. In the re-ranking optimization phase, several strategies are applied to adjust the initial ranking list. First, diversity control is implemented using the MMR (Match Making Rating) algorithm to balance relevance and diversity, ensuring that the proportion of similar content does not exceed a set threshold. Second, real-time performance is enhanced by identifying breaking news through a popularity prediction model; if its popularity exceeds a threshold, its ranking is forcibly increased. Simultaneously, local related news is prioritized based on the user's IP address. Finally, the system determines the final target news text recommendation list after adjusting for diversity, real-time weighting, and regional adaptation, and pushes the recommended news content to the user's terminal based on the target news text recommendation list.

[0032] As described above, this embodiment first collects the news content to be classified and user information on the news platform. After data cleaning and standardization, corresponding processing results are obtained. Then, multimodal features are constructed based on these results. Subsequently, a large-scale target news text classification model is built. The multimodal features are used to classify the processed news content and generate corresponding target semantic vectors. At the same time, a large-scale target user interest recognition model is built. Based on the multimodal features, a preset interest recognition operation is performed on the processed user information to obtain the target user interest recognition results. Finally, a large-scale target news text recommendation model is established. The target semantic vector and the target user interest recognition results are input into the model to obtain a target news text recommendation list, and the news texts in the list are displayed to the user in a preset order. In this way, by constructing a BERT-based news text semantic classification model, the problem of insufficient semantic understanding in traditional methods is solved. At the same time, a dynamic interest modeling framework that integrates real-time user behavior and long-term preferences is designed to improve recommendation accuracy. In addition, a high-concurrency real-time data processing pipeline is developed to achieve minute-level response from news collection to recommendation, providing an interpretable recommendation mechanism and enhancing user trust in the recommendation results.

[0033] As described in the previous embodiment, this application discloses a news text recommendation method based on a large model, which can efficiently and accurately classify news content and provide personalized recommendations. For example, for a 26-year-old user named "Xiao Li," his user profile shows that he is male, works as a programmer, resides in Beijing, and has historical interests leaning towards technology, games, and sports. Next, the news text recommendation method based on a large model for "Xiao Li" will be explained in detail.

[0034] First, this application uses a news crawler to collect news from across the internet in real time via Kafka Topics. A Flink job consumes the Kafka data, performing cleaning, classification, and vector generation operations. Specifically, regular expressions are used to filter advertising links and special characters, and sentiment and rumor detection models are used to remove clickbait and false content. Next, the SimHash algorithm is used for deduplication, retaining the latest reports. Finally, the time format is standardized to UTC timestamps, and BERT-NER is used to identify and link entities, generating standardized news text and corresponding multimodal features. The processed data is stored in Elasticsearch for retrieval and a database for metadata management. Simultaneously, user behavior data such as clicks, browsing, and searches, as well as static profiles such as age, occupation, and region, are collected from Xiao Li. Behavioral logs are cleaned to remove invalid operations such as accidental touches, and integrated with the user profile to form structured multimodal user features.

[0035] Next, the constructed news classification model categorizes each news article based on multimodal features. For example, a news article about "XX Company releases AI chip" is classified as "Technology - Artificial Intelligence" by the model based on its content and entity (XX Company - Q312). Subsequently, the model extracts the BERT semantic vector of the news article as its deep semantic representation.

[0036] Then, a large-scale user interest recognition model is constructed to process Xiao Li's behavioral data. For short-term interests, the model analyzes the news articles he clicked in the last 5 minutes, such as 3 gaming news articles and 1 technology news article, extracts semantic vectors, applies time decay weights, and weights them to generate a short-term interest vector. For long-term interests, the model analyzes his behavior over the past 30 days, discovering that the themes "artificial intelligence" and "e-sports" have higher weights through topic modeling. Combined with his programmer identity, the technology theme is weighted to generate a long-term interest vector. Finally, an attention mechanism is used to fuse the short-term and long-term vectors, and higher weights are given to short-term interests based on their high activity level, generating the final user interest vector. If Xiao Li clicks "not interested" on a "blockchain" news article, the semantic vector of that news article is subtracted from the interest vector to reduce the weight of that topic.

[0037] Finally, based on Xiao Li's user interest vector, the system retrieves the top 200 semantically similar news articles from the vector database; it then uses collaborative filtering to recall the top 100 similar news articles liked by the user; since Xiao Li is not a new user, cold start recall is not triggered. During the ranking phase, the deep matching model integrates user interest vectors, news semantic vectors, timeliness, popularity, and cross-features, such as Xiao Li's historical preference for technology categories and the matching degree between these categories, to calculate the predicted click-through rate for each news article and generate an initial ranking list. In the re-ranking phase, the system uses diversity control to ensure the recommendation list covers multiple themes such as technology, games, and sports; the real-time module inserts local breaking news such as "Sudden thunderstorm in location A" at the top; finally, a personalized recommendation list is generated and stored in an in-memory database, supporting millisecond-level retrieval. This list is then pushed to Xiao Li's news client homepage according to priority, completing the personalized recommendation process.

[0038] See Figure 2 As shown, this embodiment of the invention discloses a news text recommendation device based on a large model, comprising: The feature construction module 11 is used to collect news content to be classified and user information on the news platform. It performs data cleaning and standardization processing on the collected news content to be classified and user information to obtain processed news content and processed user information. Multimodal features are constructed based on the processed news content and processed user information. The vector generation module 12 is used to construct a target news text classification model, classify the processed news content based on the multimodal features using the target news text classification model, generate classified news content, and generate corresponding target semantic vectors for each classified news content. The result generation module 13 is used to construct a target user interest recognition model, and to perform a preset interest recognition operation on the processed user information based on the multimodal features using the target user interest recognition model to generate target user interest recognition results. The text recommendation module 14 is used to establish a target news text recommendation model. The target semantic vector and the target user interest recognition result are input into the target news text recommendation model to obtain a target news text recommendation list, and the news text in the target news text recommendation list is displayed to the user terminal in a preset order.

[0039] As described above, this application first collects news content to be classified and user information from news platforms. After data cleaning and standardization, corresponding processing results are obtained. Then, multimodal features are constructed based on these results. Subsequently, a large-scale target news text classification model is built. The multimodal features are used to classify the processed news content and generate corresponding target semantic vectors. Simultaneously, a large-scale target user interest recognition model is constructed. Based on the multimodal features, a preset interest recognition operation is performed on the processed user information to obtain the target user interest recognition results. Finally, a large-scale target news text recommendation model is established. The target semantic vector and the target user interest recognition results are input into this model to obtain a target news text recommendation list, and the news texts in the list are displayed to the user in a preset order. In this way, by constructing a BERT-based news text semantic classification model, the problem of insufficient semantic understanding in traditional methods is solved. At the same time, a dynamic interest modeling framework integrating real-time user behavior and long-term preferences is designed to improve recommendation accuracy. Furthermore, a high-concurrency real-time data processing pipeline is developed to achieve minute-level response from news collection to recommendation, providing an interpretable recommendation mechanism and enhancing user trust in the recommendation results.

[0040] In some specific implementations, the news content to be categorized includes structured news data, unstructured news data, or multimedia data, and the user information includes operation logs and user profile data on the news platform.

[0041] In some specific implementations, the feature construction module 11 may specifically include: The news data acquisition unit is used to connect to the target data source through a preset API to obtain structured news data; the structured news data includes a title, body text, publication time, and category tags. The news data generation unit is used to use a distributed crawler cluster to crawl user-generated content from target social media and integrate the user-generated content from the target social media as unstructured news data. A multimedia data construction unit is used to extract text and video subtitles from images of the news content to be classified, and construct multimedia data. The operation log capture unit is used to capture user operation logs on the news platform in real time; The user profile data collection unit is used to collect user registration information on news platforms, social network association data, and device information as user profile data.

[0042] In some specific implementations, the feature construction module 11 may specifically include: The news content filtering unit is used to filter advertising links and special symbols in the news content to be classified by using a preset regular expression to obtain filtered news content. The polarity value judgment unit is used to perform sentiment polarity analysis on the filtered news content using a preset sentiment model, and to determine whether the analysis result contains clickbait feature words and whether the sentiment polarity value of the clickbait feature words is greater than a preset sentiment threshold. The first polarity value determination unit is used to mark the corresponding filtered news content as first invalid data if it contains clickbait feature words and the sentiment polarity value of the clickbait feature words is greater than a preset sentiment threshold. The second polarity value determination unit is used to mark the corresponding filtered news content as the first valid data if it does not contain clickbait feature words or the emotional polarity value of the clickbait feature words is less than or equal to a set threshold. The data discrimination unit is used to perform authenticity discrimination on the first valid data through a preset rumor detection model; The first data determination unit is used to determine the corresponding first valid data as second invalid data if the confidence level of the determination result representing the first valid data is less than a preset confidence threshold. The second data determination unit is used to determine the corresponding first valid data as second valid data if the confidence level of the discrimination result representing the first valid data is greater than or equal to a preset confidence threshold. A similarity generation unit is used to determine the text fingerprint of the second valid data through a preset similarity hash algorithm, and to generate the similarity between different text fingerprints; The data retention unit is used to retain only the second valid data whose publication time meets the preset latest condition if there is second valid data with the similarity greater than or equal to the preset similarity threshold; The entity alignment unit is used to perform time format unification and entity alignment operations on the retained second valid data to obtain the processed news content.

[0043] In some specific implementations, the feature construction module 11 may specifically include: The statistical feature construction unit is used to construct statistical features based on the processed news content; wherein the statistical features include any one or more of the following: word count, sentence count, punctuation mark ratio, and keyword density. A sentiment feature determination unit is used to determine the sentiment score of the processed news content and determine the corresponding sentiment features based on the sentiment score of the processed news content. The time-series feature acquisition unit is used to acquire the publication time of the processed news content and obtain the corresponding time-series features based on the publication time of the processed news content. A semantic feature recognition unit is used to recognize the semantic features of the processed news content through a pre-trained BERT large model; A behavior feature acquisition unit is used to identify the operation log in the processed user information in order to obtain the user behavior features of the user terminal. The multimodal feature construction unit is used to construct multimodal features based on the statistical features, the sentiment features, the temporal features, the semantic features, and the user behavior features.

[0044] In some specific implementations, the result generation module 13 may specifically include: The first vector generation unit is used to extract the semantic feature vector of the user's click on news within a preset time window by using the real-time behavioral data of the target user interest identification big model based on the user's behavioral characteristics, and to perform weighted fusion of the semantic feature vector according to the preset time decay rule to generate the target short-term interest vector. The second vector generation unit is used to generate a target long-term interest vector based on the user behavior characteristics within a preset historical period and the user profile data, through topic analysis and knowledge graph embedding methods. The third vector generation unit is used to perform weighted fusion of the target short-term interest vector and the target long-term interest vector through a preset attention mechanism to generate a fusion result, and to correct the fusion result according to the target negative feedback behavior list provided by the user terminal to generate the target user interest vector.

[0045] In some specific implementations, the text recommendation module 14 may specifically include: The news collection generation unit is used to perform multi-path recall operations based on the target user interest identification results and the target semantic vector to generate an initial news collection that includes semantic recall, collaborative filtering recall and cold start recall. The sorting list generation unit is used to sort the news in the initial news set using a preset deep semantic matching model, output the predicted score of each news item in the initial news set, and generate an initial sorting list based on the predicted scores of the news items. The recommendation list generation unit is used to optimize and rearrange the initial sorted list to generate a target news text recommendation list; wherein the strategy optimization includes preset diversity control operations, preset real-time enhancement operations, and preset regional adaptation operations.

[0046] Furthermore, embodiments of this application also disclose an electronic device, Figure 3 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0047] Figure 3 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the large-model-based news text recommendation method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be a computer.

[0048] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0049] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0050] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the large-model-based news text recommendation method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0051] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned news text recommendation method based on a large model. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0052] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its 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 in the method section.

[0053] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0054] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0055] Finally, it should 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 a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. 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.

[0056] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. 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 this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A news text recommendation method based on a large model, characterized in that, include: Collect news content to be classified and user information on the news platform. Perform data cleaning and standardization on the collected news content to be classified and user information to obtain processed news content and processed user information. Construct multimodal features based on the processed news content and processed user information. A target news text classification model is constructed. Based on the multimodal features, the target news text classification model is used to classify the processed news content, generate classified news content, and generate a corresponding target semantic vector for each classified news content. Construct a large target user interest recognition model, and use the large target user interest recognition model to perform preset interest recognition operations on the processed user information based on the multimodal features to generate target user interest recognition results; A target news text recommendation model is established. The target semantic vector and the target user interest recognition result are input into the target news text recommendation model to obtain a target news text recommendation list. The news texts in the target news text recommendation list are displayed to the user terminal in a preset order.

2. The news text recommendation method based on a large model according to claim 1, characterized in that, The news content to be categorized includes structured news data, unstructured news data, or multimedia data, and the user information includes operation logs and user profile data on the news platform.

3. The news text recommendation method based on a large model according to claim 2, characterized in that, The collection of news content to be categorized and user information on the news platform includes: The system connects to the target data source via a pre-defined API to obtain structured news data; the structured news data includes a title, body text, publication time, and category tags. A distributed crawler cluster is used to crawl user-generated content from target social media, and the user-generated content from the target social media is integrated as unstructured news data; Extract text and video subtitles from the images of the news content to be categorized to construct multimedia data; Real-time capture of user activity logs on the news platform; The system collects user registration information on news platforms, social network association data, and device information to create user profile data.

4. The news text recommendation method based on a large model according to claim 3, characterized in that, The data cleaning and standardization processing operations performed on the collected news content to be categorized and the user information include: The news content to be categorized is filtered by using a preset regular expression to filter out advertising links and special characters. The filtered news content is subjected to sentiment polarity analysis using a preset sentiment model. The analysis results are used to determine whether the results contain clickbait keywords and whether the sentiment polarity value of the clickbait keywords is greater than a preset sentiment threshold. If the news content contains clickbait keywords and the sentiment polarity value of the clickbait keywords is greater than a preset sentiment threshold, then the corresponding filtered news content will be marked as the first invalid data. If the news content does not contain clickbait keywords or the sentiment polarity value of the clickbait keywords is less than or equal to a set threshold, then the corresponding filtered news content is marked as the first valid data. The first valid data is subjected to authenticity verification using a pre-set rumor detection model; If the judgment result indicates that the confidence level of the first valid data is less than the preset confidence threshold, then the corresponding first valid data is determined as the second invalid data; If the judgment result indicates that the confidence level of the first valid data is greater than or equal to the preset confidence threshold, then the corresponding first valid data is determined as the second valid data; The text fingerprint of the second valid data is determined by a preset similarity hash algorithm, and the similarity between different text fingerprints is generated. If there is second valid data with a similarity greater than or equal to a preset similarity threshold, then only the second valid data whose publication time meets the preset latest condition will be retained; The retained second valid data is subjected to time format unification and entity alignment operations to obtain the processed news content.

5. The news text recommendation method based on a large model according to claim 4, characterized in that, The construction of multimodal features based on the processed news content and processed user information includes: Statistical features are constructed based on the processed news content; wherein the statistical features include any one or more of the following: word count, sentence count, punctuation mark ratio, and keyword density; Determine the sentiment score of the processed news content, and determine the corresponding sentiment features based on the sentiment score of the processed news content; Obtain the publication time of the processed news content, and obtain the corresponding time-series features based on the publication time of the processed news content; The semantic features of the processed news content are identified by a pre-trained BERT large model; Identify the operation logs in the processed user information to obtain the user behavior characteristics of the user terminal; Multimodal features are constructed based on the statistical features, the sentiment features, the temporal features, the semantic features, and the user behavior features.

6. The news text recommendation method based on a large model according to claim 5, characterized in that, The step of performing a preset interest recognition operation on the processed user information based on the multimodal features using a target user interest recognition model to generate target user interest recognition results includes: Using a target user interest identification big model based on real-time behavioral data of the user's behavioral characteristics, the semantic feature vector of the user's click on news within a preset time window is extracted, and the semantic feature vector is weighted and fused according to a preset time decay rule to generate a target short-term interest vector. Based on the user behavior characteristics and historical behavior data within a preset historical period, as well as the user profile data, a target long-term interest vector is generated through topic analysis and knowledge graph embedding methods. The target short-term interest vector and the target long-term interest vector are weighted and fused using a preset attention mechanism to generate a fusion result. The fusion result is then corrected based on the target negative feedback behavior list provided by the user terminal to generate the target user interest vector.

7. The news text recommendation method based on a large model according to any one of claims 1 to 6, characterized in that, The step of inputting the target semantic vector and the target user interest recognition result into the target news text recommendation model to obtain a target news text recommendation list includes: Based on the target user interest identification results and the target semantic vector, a multi-path recall operation is performed to generate an initial news set that includes semantic recall, collaborative filtering recall, and cold start recall. The news in the initial news set is sorted using a preset deep semantic matching model, the predicted score of each news item in the initial news set is output, and an initial sorted list is generated based on the predicted scores of the news items. The initial sorted list is optimized and rearranged to generate a target news text recommendation list; wherein the strategy optimization includes preset diversity control operations, preset real-time enhancement operations, and preset regional adaptation operations.

8. A news text recommendation device based on a large model, characterized in that, include: The feature construction module is used to collect news content to be classified and user information on the news platform. It performs data cleaning and standardization operations on the collected news content to be classified and user information to obtain processed news content and processed user information. Multimodal features are constructed based on the processed news content and processed user information. The vector generation module is used to construct a target news text classification model, classify the processed news content based on the multimodal features using the target news text classification model, generate classified news content, and generate corresponding target semantic vectors for each classified news content. The result generation module is used to construct a large target user interest recognition model, and based on the multimodal features, to perform preset interest recognition operations on the processed user information using the large target user interest recognition model to generate target user interest recognition results. The text recommendation module is used to build a large-scale model for recommending target news texts. The target semantic vector and the target user interest recognition results are input into the large-scale model to obtain a list of target news text recommendations. The news texts in the list are then displayed to the user in a preset order.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the large-model-based news text recommendation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when the computer programs are executed by a processor, they implement the news text recommendation method based on a large model as described in any one of claims 1 to 7.

Images