Large model information pushing method and system based on industry corpus

Abstract

The invention relates to the technical field of machine learning, and discloses a large model information pushing method and system based on industry corpora, and the method comprises the steps: firstly obtaining positive and negative behavior record sequences of a target user, extracting behavior features, generating a behavior weight, and constructing a multi-dimensional user portrait vector; then obtaining a semantic embedding vector based on a pre-trained industry language model, generating a positive and negative preference embedding vector through similarity threshold adjustment, and obtaining an adjusted push content set according to the positive and negative preference embedding vector; and finally, intelligently sorting the candidate push contents to generate a precise push content sequence. According to the method, the instant demand and the emotion fluctuation of the user are accurately captured, so that accurate and dynamic personalized recommendation is realized, the user experience is effectively improved, and the information fatigue is reduced.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and more specifically, to a method and system for pushing large model information based on industry corpus. Background Art

[0002] With the popularization of the Internet and mobile devices, the application of information push technology in various industries has received widespread attention, especially in industries such as finance and e-commerce. Personalized push technology has become an important means to enhance user experience, optimize user services and increase business conversion rates.

[0003] The patent application with publication number CN116489216A discloses an information push method, device, electronic device, storage medium and program product, which determines whether to push a specific product based on the customer's preference information and push permission information. The method determines whether to push the product to the target user based on the user's first preference information and push permission information. Although this method can reduce the blindness of the push to a certain extent and avoid the push of irrelevant information, it still has limitations: first, the method relies on preset push permission information and cannot dynamically adjust the recommended content according to the user's real-time behavior or emotional fluctuations; second, the determination of product push is limited to matching with customer preference information, and lacks comprehensive analysis of more dimensions such as user emotions and behavioral status. Therefore, how to more accurately capture the user's dynamic needs and provide real-time personalized push through more fine-grained sentiment analysis and semantic processing is still an unresolved problem.

[0004] The patent application with publication number CN110460674A discloses an information push method, device and system, which ensures the security and integrity of information by signing and verifying the push information. This method uses a push gateway to generate signature information and verifies the security of information transmission through a public key and private key mechanism. Although this method is of great significance in the security of information transmission, it mainly focuses on security issues in the push process, and does not involve how to improve the personalization and accuracy of the push content. This method fails to make full use of user behavior data, sentiment analysis and other information to optimize the push content, nor does it consider how to improve the relevance of the push content through intelligent algorithms. Therefore, although the security of information transmission is guaranteed, the quality and personalization of the push content are still the shortcomings of current technology.

[0005] The push methods in the existing technology are often limited to static data analysis or judgment based on simple rules, lacking dynamic response to users' real-time behaviors and emotional fluctuations, resulting in the problem of over-simplification and uniformity of push content. Summary of the invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a large-model information push method and system based on industry corpus, which deeply explores user behavior preferences, integrates industry corpus to build a domain language model, dynamically optimizes push content, and improves the accuracy and personalization of information push.

[0007] To achieve the above object, the present invention provides the following technical solutions: The large model information push method based on industry corpus includes: Obtain a positive behavior record sequence and a negative behavior record sequence of a target user; extract a first behavior feature from the positive behavior record sequence, and extract a second behavior feature from the negative behavior record sequence; generate a first behavior weight of the first behavior feature and a second behavior weight of the second behavior feature; generate a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature, and the second behavior weight; Construct a pre-trained industry language model, and obtain a first semantic embedding vector and a second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjust the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtain an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector; Based on the adjusted push content set, a third candidate push content set is obtained; the contents in the third candidate push content set are intelligently sorted to generate an accurate push content sequence.

[0008] Furthermore, the obtaining of the target user's positive behavior record sequence and negative behavior record sequence includes: Construct an industry corpus, and obtain first historical behavior data of the target user in the industry field from the industry corpus; the first historical behavior data includes n1 behavior types of the target user, and the n1 behavior types include positive behaviors and negative behaviors; n1 is a positive integer; According to the first historical behavior data, a positive behavior record sequence and a negative behavior record sequence are obtained.

[0009] Further, extracting the first behavior feature from the positive behavior record sequence includes: Perform word segmentation on the text content in the positive behavior record sequence, and divide the text content into a word sequence of the smallest semantic unit; Perform part-of-speech tagging on the word sequence after word segmentation; Identify named entities in text content based on part-of-speech tagging; Extract keywords and subject terms from word sequences; Keywords, subject terms and named entity words are used as the first line features.

[0010] Further, the generating of the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature includes: Set the basic weights of different behavior types among n1 behavior types and construct the initial behavior weight matrix; According to the first historical behavior data, statistically analyzing the behavior intensity distribution of different behavior types; Constructing a behavior intensity quantification function, and mapping the behavior intensity distribution of different behavior types into weight adjustment coefficients of corresponding behavior types according to the behavior intensity quantification function; Multiply the basic weights in the initial behavior weight matrix by the corresponding weight adjustment coefficient to generate an adjusted behavior weight matrix; In the adjusted behavior weight matrix, the weight corresponding to the positive behavior is used as the first behavior weight of the first behavior feature, and the weight corresponding to the negative behavior is used as the second behavior weight of the second behavior feature.

[0011] Furthermore, generating a multi-dimensional user portrait vector includes: According to the first behavior characteristics and the first behavior weight, construct the positive preference dimension; According to the second behavior characteristics and the second behavior weight, the negative preference dimension is constructed; The positive preference dimension and the negative preference dimension are combined to generate a multi-dimensional user portrait vector.

[0012] Further, generating the positive preference embedding vector and the negative preference embedding vector includes: Calculate the similarity between the first semantic embedding vector and the second semantic embedding vector to obtain a positive and negative preference similarity score D1; If the positive and negative preference similarity score D1 is greater than θ1, the embedding vector adjustment process is triggered to obtain the positive preference embedding vector and the negative preference embedding vector.

[0013] Furthermore, the embedding vector adjustment process includes: Step S2221, randomly sampling a noise vector, adjusting the first semantic embedding vector according to the noise vector, obtaining an adjusted first semantic embedding vector, and marking it as a positive preference embedding vector; Step S2222, adjusting the second semantic embedding vector according to the noise vector to obtain an adjusted second semantic embedding vector, which is marked as a negative preference embedding vector; Step S2223, calculate the similarity D2 between the positive preference embedding vector and the negative preference embedding vector, and determine whether D2 is less than or equal to θ1. If so, complete the embedding vector adjustment process; otherwise, return to step S2221, re-randomly sample the noise vector, and repeat steps S2221-S2223 until D2≤θ1.

[0014] Furthermore, the adjusted push content set includes: Use the positive preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus to generate the first candidate push content set; Use the negative preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus to generate a second set of candidate push content; Based on the first candidate push content set and the second candidate push content set, an adjusted push content set is obtained.

[0015] Generating the first candidate push content set includes: Convert the text content in the industry corpus into corresponding semantic embedding vectors; Calculate the similarity between the positive preference embedding vector in the multi-dimensional user portrait vector and the semantic embedding vector of the text content in the industry corpus to obtain a first similarity score; The text contents with the first similarity scores in the top K in the industry corpus are screened to generate the first set of candidate push contents.

[0016] Further, obtaining an adjusted push content set based on the first candidate push content set and the second candidate push content set includes: Obtaining semantic embedding vectors of text content in the first candidate push content set and the second candidate push content set; According to the semantic embedding vectors of the text contents in the first candidate push content set and the second candidate push content set, the semantic similarity S2 between each text content in the first candidate push content set and all text contents in the second candidate push content set is calculated to form a similarity matrix; In the similarity matrix, content pairs whose semantic similarity S2 is higher than the second similarity threshold θ2 are selected and marked as highly overlapping content; The marked highly overlapping contents are removed from the first candidate push content set to obtain an adjusted push content set.

[0017] A large model information push system based on industry corpus, which is used to implement the above-mentioned large model information push method based on industry corpus, and the system includes: Portrait construction module: used to obtain the positive behavior record sequence and negative behavior record sequence of the target user; extract the first behavior feature from the positive behavior record sequence and extract the second behavior feature from the negative behavior record sequence; generate the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature; generate a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature and the second behavior weight; Push content generation module: used to build a pre-trained industry language model, obtain a first semantic embedding vector and a second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjust the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtain an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector; Push content sorting module: based on the adjusted push content set, obtain a third candidate push content set; intelligently sort the contents in the third candidate push content set to generate an accurate push content sequence.

[0018] An electronic device includes a memory, a central processing unit, and a computer program stored in the memory and executable on the central processing unit. When the central processing unit executes the computer program, the above-mentioned large model information push method based on industry corpus is implemented.

[0019] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the above-mentioned large model information push method based on industry corpus.

[0020] Compared with the prior art, the present invention has the following beneficial effects: By combining users' real-time behavior data, sentiment analysis, and semantic embedding technology, the present invention can capture changes in users' needs in real time, dynamically adjust recommended content, and ensure that each push is highly consistent with the user's immediate interests and emotional state. This process avoids the static nature of traditional push systems and the problem of over-reliance on historical data, making the recommended content not only consistent with users' long-term interests, but also able to quickly respond to changes in users' current needs, thereby greatly improving the relevance and accuracy of recommendations.

[0021] In addition, by using semantic similarity calculation and deduplication mechanism, the present invention effectively avoids information redundancy and information fatigue, ensuring that the pushed content is both rich and diverse and not repetitive. By optimizing the push strategy through reinforcement learning, the pre-trained industry language model can continuously adjust the order of recommended content based on real-time feedback from users, so that the pushed content not only meets the interests of users, but also provides the most appropriate content at the best time, further improving user participation and conversion rate. The introduction of reinforcement learning enables the push strategy to self-optimize and adapt to market changes and dynamic changes in user behavior, thereby continuously improving recommendation effects and user satisfaction.

[0022] Through multi-dimensional data analysis and intelligent optimization, the present invention not only improves the quality and personalization of pushed content, but also enhances the ability to respond to changes in market and user needs, and can provide more accurate, efficient and personalized services. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 It is a principle flow chart of the large model information push method based on industry corpus in the present invention; Figure 2 A flow chart of a method for extracting a first behavior feature from a positive behavior record sequence in a large model information push method based on industry corpus of the present invention; Figure 3 A flow chart of a method for generating a first behavior weight of a first behavior feature and a second behavior weight of a second behavior feature in a large model information push method based on industry corpus of the present invention; Figure 4 A flowchart of a method for generating a first candidate push content set in a large model information push method based on industry corpus of the present invention; Figure 5 A flowchart of a method for obtaining an adjusted set of pushed content in a large model information push method based on industry corpus of the present invention; Figure 6 This is a functional module diagram of the large model information push system based on industry corpus in the present invention. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] Example 1

[0027] See also Figure 1 As shown, this embodiment provides a large model information push method based on industry corpus, including: Step S1000, obtaining a positive behavior record sequence and a negative behavior record sequence of a target user; extracting a first behavior feature from the positive behavior record sequence, and extracting a second behavior feature from the negative behavior record sequence; generating a first behavior weight of the first behavior feature and a second behavior weight of the second behavior feature; generating a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature, and the second behavior weight; Furthermore, step S1000 includes: Step S1100, obtaining a positive behavior record sequence and a negative behavior record sequence of a target user; Furthermore, step S1100 includes: Step S1110, constructing an industry corpus, and obtaining first historical behavior data of the target user in the industry field from the industry corpus; the first historical behavior data includes n1 behavior types of the target user, and the n1 behavior types include positive behaviors and negative behaviors; n1 is a positive integer; Step S1120: obtaining a positive behavior record sequence and a negative behavior record sequence according to the first historical behavior data.

[0028] Specifically, the purpose of step S1100 is to identify and extract positive and negative behavior records by analyzing the historical behavior of the target user. These behavior records will provide data support for subsequent product push and service recommendations, whether it is pushing real-time industry analysis and exclusive financial products to investors in the financial field, or recommending suitable products to customers in other consumer fields. In this way, more accurate personalized services can be achieved.

[0029] An industry corpus refers to a large set of user behavior data collected in a certain industry (such as finance, consumption, and medical care), which can represent the behavioral characteristics and preferences of most users in the industry. This corpus is the basic data source for analysis, recommendation, and prediction. Building an industry corpus refers to the process of collecting and organizing all relevant user behavior data in a certain industry (such as the financial field, the consumer field, etc.). The industry corpus not only contains the specific behavior records of the target users, but also includes the behavior data of all users in the industry, which is widely representative. By analyzing the industry corpus, the specific behavior patterns of the target users can be mined, and then suitable products or services can be recommended for them.

[0030] In the financial sector, the industry corpus includes the following data: Transaction behavior data: including investors’ stock transactions, fund purchases, bond transactions, etc.

[0031] Financial product purchase history: financial products, insurance products, etc. purchased by the user in the past.

[0032] Investment preferences: investors’ risk tolerance, preferred investment areas, etc.

[0033] Industry information interaction: industry analysis reports, investment strategy articles, etc. that users have read and interacted with.

[0034] In other consumer sectors, the industry corpus includes: Shopping behavior data: types of goods purchased by users, purchase frequency, consumption amount, etc.

[0035] Search and browsing history: users’ search and browsing history for goods or services, as well as their interest in specific products.

[0036] Social interaction records: user interaction records on social platforms or with brand customer service, such as likes, comments, consultations, etc.

[0037] The first historical behavior data refers to the first-hand behavior data of the target user in a certain industry field, which is usually a detailed record of all the user's behaviors in the past period of time. It includes the user's interaction record with the industry, reflecting the user's preferences, needs and potential interests.

[0038] By building an industry corpus, service providers in the financial or consumer fields can fully understand the historical behavior of target users and identify their preferences, needs and pain points. In the financial field, this helps to accurately recommend financial products and investment opportunities based on the user's investment history and preferences; in the consumer field, it can recommend suitable goods or services to customers, avoid over-selling, and improve the accuracy of recommendations and user satisfaction.

[0039] For example, suppose user A purchased stock funds and bond funds through a bank in the past year and made a fixed monthly deposit. These behavior records are marked as "conservative investors" in the industry corpus. Similarly, user B frequently traded high-risk stocks in the past year and frequently adjusted his investment portfolio. These behaviors are marked as "aggressive investors." Based on these marks, low-risk financial products can be recommended for user A, and high-risk stock funds can be recommended for user B.

[0040] Step S1120 divides positive behavior and negative behavior records based on the first historical behavior data. The purpose of this process is to classify the user's behavior, so as to accurately determine which behaviors meet the needs of the target user and which behaviors may bring risks or adverse effects. Positive behaviors are usually those that show the user's good credit, preferences or needs, while negative behaviors are behaviors that are disadvantageous or risky to the user.

[0041] Positive behavior refers to user behavior that meets expectations or is beneficial. For example, in the financial field, positive behavior includes users making continuous deposits, repaying on time, and investing in financial products with moderate risks; in the consumer field, positive behavior includes users frequently buying healthy foods, being highly loyal, and giving good reviews to products.

[0042] Negative behavior refers to user behavior that is contrary to expectations or may cause risks. For example, in the financial field, negative behavior includes users’ overdue repayments, excessive high-risk investments, credit card arrears, etc.; in the consumer field, negative behavior includes users’ frequent returns, poor product ratings, frequent complaints, etc.

[0043] By identifying and recording positive and negative behaviors, it is possible to accurately classify and evaluate users, and provide them with personalized services. For users in the financial field, identifying positive behaviors can help recommend financial products that are suitable for their risk tolerance, while negative behaviors can prompt the system to warn of potential defaults or high-risk investment behaviors. In the consumer field, positive behaviors can help recommend goods or services that meet user needs, while negative behaviors can help the system avoid recommending too many products that may cause user dissatisfaction.

[0044] Step S1200, extracting a first behavior feature from a positive behavior record sequence, and extracting a second behavior feature from a negative behavior record sequence; generating a first behavior weight of the first behavior feature and a second behavior weight of the second behavior feature; Further, step S1200 includes: Step S1210, extracting a first behavior feature from a positive behavior record sequence; The purpose of step S1210 is to extract useful features from the positive behavior record sequence to provide data support for subsequent intelligent marketing decisions. The feature extraction process involves in-depth analysis of the user behavior text content, extracting keywords, subject terms and named entities, and thus establishing the first behavior features that can reflect user needs, interests and preferences. These features will be used to recommend personalized financial products or services to users.

[0045] Furthermore, if Figure 2 As shown, step S1210 includes: Step S1211, performing word segmentation processing on the text content in the positive behavior record sequence, dividing the text content into a word sequence of the smallest semantic unit; Specifically, the text content in the sequence of positive behavior records is first segmented. Word segmentation refers to the division of continuous text into independent minimum semantic units, namely words or phrases. Since words in natural language can have different combinations and meanings, the word segmentation process can extract the independent meaning of each word or phrase in the context. This is the first step in processing user behavior data and analyzing their intentions. Word segmentation is a technology in text processing that divides coherent sentences in the text into vocabulary units. The quality of word segmentation directly affects subsequent semantic analysis and information extraction. In language processing, a semantic unit refers to the smallest unit with independent meaning. In Chinese, it is usually a word or phrase.

[0046] Key steps of word segmentation: Text cleaning: Remove punctuation marks, meaningless characters (such as carriage returns, line breaks), etc.

[0047] Segmentation: Divide the sentences in the text into individual words according to spaces, punctuation, etc.

[0048] Standardization: Normalize some synonyms or variant words into a unified form.

[0049] For example, user A enters "I want to know the difference between stock funds and bond funds" in a financial management APP. During the word segmentation process, this text is divided into the following word sequence: "I / want / to know / stock funds / and / bond funds / of / difference".

[0050] Word segmentation is a basic step in natural language processing. It converts the user's natural language behavior data into a format that the machine can understand and analyze. By dividing the user's text content into individual word sequences, the user's needs and intentions can be analyzed more precisely, laying a foundation for subsequent feature extraction and push systems.

[0051] Step S1212, perform part-of-speech tagging on the segmented word sequence; Specifically, performing part-of-speech tagging on the segmented word sequence aims to assign a part-of-speech tag (such as noun, verb, adjective, etc.) to each word. Part-of-speech tagging helps to understand the grammatical role and meaning of words in a sentence, thereby enabling more accurate information extraction and semantic analysis. Part-of-speech tags are identifiers assigned to each word, indicating the grammatical role of the word in the sentence, such as noun (NN), verb (VB), adjective (JJ), etc.

[0052] For example, in the sentence "I want to know the difference between stock funds and bond funds": "I" is a pronoun (PRP); "want" is a verb (VB); "know" is a verb (VB); "stock funds" is a noun phrase (NP); "and" is a conjunction (CC); "bond funds" is a noun phrase (NP); "of" is a particle (PART); "difference" is a noun (NN); Part-of-speech tagging can help accurately understand the role of each word in a sentence, facilitating the extraction of keywords and topics in subsequent steps. Without part-of-speech tagging, it may be impossible to accurately distinguish whether the "fund" in "bond fund" is a noun or a word with other possible meanings. Through part-of-speech tagging, key nouns (such as types of financial products) and verbs (such as investment, purchase, etc.) can be identified from the text, laying a foundation for precisely recommending financial products.

[0053] For example, user C enters: "I am interested in low-risk funds". After part-of-speech tagging: "I" is a pronoun (PRP); "in" is a preposition (IN); "low-risk" is an adjective (JJ); "funds" is a noun (NN); "am interested" is a verb phrase (VP); Through these tags, the system can further extract keywords related to "funds" and "low-risk".

[0054] Step S1213, based on part-of-speech tagging, identify named entities in the text content; Specifically, based on the part-of-speech tagging results, identify named entities in the text. Named entities usually refer to specific, independently meaningful entities, such as company names, place names, dates, product names, etc. In the financial field, the identification of named entities can help the system identify key information such as the names of investment products and company names.

[0055] The identification of named entities helps accurately extract key information from the text. Especially in the financial field, identifying specific entities such as funds, stocks, and bank names can help the system better understand users' needs and provide more accurate product recommendations.

[0056] Step S1214, extract keywords and topic words from the word sequence; Specifically, extract keywords and topic words related to users' needs from the word sequence. Keywords are the core words that can represent users' needs. For example, in the investment field, keywords may include "stocks", "funds", "bonds", etc. Topic words are the main topics or directions refined from the entire behavior record, which can summarize the topic of users' needs, such as "low-risk financial products", "stock fund investment". By extracting keywords and topic words, the core intention of users can be understood, and based on this, appropriate products can be recommended.

[0057] Extracting keywords and subject terms helps accurately identify the main needs of users, and provides personalized recommendations to users based on this information. If users frequently mention "bond funds" and "low risk", it can be inferred that the user prefers low-risk investment products, and thus recommend related financial products. This not only improves the accuracy of recommendations, but also greatly improves the user experience. For example, user E often enters "low-risk funds", "bond funds", and "long-term investment" in his behavior records on the financial platform. These words will be extracted as keywords and subject terms, and low-risk, long-term bond funds can be recommended to users based on this information.

[0058] Step S1215: taking keywords, subject words and named entity words as first behavior features.

[0059] Specifically, step S1215 combines the keywords, subject words and named entity words extracted in step S1214 as the first behavioral features. These features will be used to describe the user's investment preferences, needs and interests, and provide an important basis for subsequent personalized recommendations and intelligent marketing decisions. By using keywords, subject words and named entity words as the first behavioral features, the user's core needs and preferences can be accurately captured. This characterization process enables subsequent intelligent marketing to more accurately meet user needs, improve the matching degree of recommendations and user satisfaction. For example, assuming that user F enters "low-risk bond fund", the keywords extracted by the system are "low risk" and "bond fund". These words, as the first characteristics of user behavior, can be used to push low-risk bond fund products that meet their preferences.

[0060] Step S1220, extracting a second behavior feature from the negative behavior record sequence; the second behavior feature includes irrelevant words, negative sentiment words, and ignored entity words; Specifically, in the financial field, especially in the banking industry, the negative behavior records of users can reveal the potential risk of user dissatisfaction, complaints and churn. Therefore, step S1220 extracts the second behavior feature from the negative behavior record sequence, aiming to accurately construct a negative profile of the user by analyzing the user's dissatisfaction and negative behavior, helping to predict user churn, optimize service experience and improve user relationship management.

[0061] Negative behavior records usually include user complaints, negative feedback, bad reviews, service cancellations, or unsatisfactory interactions. In this step, the extracted second behavior features include the following three types: Irrelevant words: In negative behavior records, users may mention content that is irrelevant to the business itself. Irrelevant words can be situational descriptions, background, or irrelevant information that is irrelevant to the service. For example, a user may mention "it rained too much yesterday" or "the traffic jam on the road is very serious" when making a complaint. Although these words appear in the records, they are irrelevant to the analysis of negative behavior and therefore need to be filtered out to avoid interfering with sentiment analysis or behavior prediction.

[0062] Negative sentiment words: Negative sentiment words are words that directly express negative emotions in user feedback. For example, "disappointment", "anger", and "dissatisfaction" are typical negative sentiment words. Extracting negative sentiment words helps to accurately capture the user's emotional state and quantify the user's attitude. By identifying negative sentiment words, banks can effectively evaluate the user's emotional fluctuations, so as to take countermeasures in advance and avoid user loss.

[0063] Ignore entity words: In some cases, users may mention specific services or products, but these entities do not dominate the user's dissatisfaction. Ignore entity words usually refer to specific service or product entities mentioned in negative feedback that are irrelevant to user sentiment. For example, a user may mention "the application is bad" but does not clearly explain why the application is bad, or only mention the specific product name when complaining. At this time, these entity information should be ignored because they do not directly affect the user's negative sentiment.

[0064] By accurately extracting irrelevant words, negative sentiment words, and ignoring entity words, it is possible to remove irrelevant information when analyzing negative user feedback, thereby ensuring more accurate sentiment analysis and behavior prediction. This processing helps provide an effective basis for user relationship management, reduce the risk of user churn, and make optimization decisions in a timely manner.

[0065] Step S1230: Generate a first behavior weight of the first behavior feature and a second behavior weight of the second behavior feature.

[0066] Furthermore, if Figure 3 As shown, step S1230 includes: Step S1231, setting basic weights of different behavior types among n1 behavior types to construct an initial behavior weight matrix; Step S1232, based on the first historical behavior data, statistically analyzing the behavior intensity distribution of different behavior types; Step S1233, constructing a behavior intensity quantification function, and mapping the behavior intensity distribution of different behavior types to weight adjustment coefficients of corresponding behavior types according to the behavior intensity quantification function; Step S1234, multiplying the basic weights in the initial behavior weight matrix by the corresponding weight adjustment coefficients to generate an adjusted behavior weight matrix; Step S1235: In the adjusted behavior weight matrix, the weight corresponding to the positive behavior is used as the first behavior weight of the first behavior feature, and the weight corresponding to the negative behavior is used as the second behavior weight of the second behavior feature.

[0067] Specifically, step S1230 generates an adjusted behavior weight matrix and assigns corresponding weights to the first behavior feature (positive behavior) and the second behavior feature (negative behavior) based on the user's behavior data. These weights will affect the personalized recommendation strategy and be used to optimize the recommendation of financial services, such as AI robot intelligent marketing that replaces traditional customer service staff and salesmen in financial fields such as banks.

[0068] Behavior types are classified based on the user's historical behavior data, usually including positive behaviors (such as purchasing financial products, repaying on time, etc.) and negative behaviors (such as overdue repayments, frequent consultations, etc.). The basic weight reflects the initial importance of each behavior type. In the initial stage, positive behaviors are usually given higher weights, while negative behaviors are given lower weights. For example: positive behaviors (such as repaying on time, frequently purchasing low-risk products) may obtain higher basic weights (such as 0.8 or 1.0), indicating that these behaviors have a greater impact on subsequent recommendations. Negative behaviors (such as overdue repayments, frequent returns) may obtain lower basic weights (such as 0.2 or 0.3), indicating that these behaviors have less impact on the push system and should be suppressed. Based on these basic weights, an initial behavior weight matrix is ​​constructed, which will assign a preliminary weight value to each behavior type.

[0069] By setting a basic weight for each behavior type, the importance of each behavior type can be reasonably defined, ensuring that positive behaviors are given priority and negative behaviors are less likely to interfere with recommendations. For example, positive behaviors (such as on-time repayments) directly reflect the user's credit status, while negative behaviors (such as overdue repayments) may indicate the user's credit risk and reduce the probability of their participation in certain financial products.

[0070] Behavior intensity refers to the frequency, duration or engagement of a user in a certain behavior type, reflecting the user's level of commitment to a specific behavior. Behaviors with high frequency, high duration and high engagement are considered "higher intensity", while behaviors with low frequency, short duration and low engagement are considered "lower intensity". Behavior intensity distribution describes the distribution of user investment intensity in different types of behaviors, which helps identify user preferences and activity levels.

[0071] For example: High-intensity behavior: Users frequently purchase financial products and continuously invest large amounts every month.

[0072] Low-intensity behavior: Users occasionally browse the bank’s financial product pages but never make an actual purchase.

[0073] By counting the intensity of these behaviors, it is possible to identify which types of behaviors are the user's main behaviors and which behaviors are only minor behaviors that occur occasionally. Statistical behavior intensity distribution can help the system identify the user's activity level and preferences. For financial product push systems, the system should give priority to users' behaviors with higher intensity (such as frequent purchases of financial products or on-time repayments) in order to provide users with more personalized and accurate recommendations. In contrast, behaviors with lower intensity can be considered secondary information and have less impact. For example, the behavior data of user B is as follows: Purchase of financial products: once a month, with a large investment amount; Consulting on loan products: once a quarter, but no actual application; By statistically analyzing the intensity of these behaviors, we can conclude that user B has a higher intensity in purchasing financial products and a lower intensity in loan consultation. The recommendation strategy can be adjusted based on these intensity distributions.

[0074] The behavior intensity quantification function determines how to adjust the basic weight according to the user's behavior intensity. For example, the behavior intensity can be quantified by the following simple linear function: Adjustment factor ;

[0075] Among them, the behavior intensity is the actual performance of the user in a certain behavior type, and the maximum behavior intensity is the maximum value of all user behaviors in this behavior type. This function can convert the behavior intensity into an adjustment coefficient, thereby adjusting the weights in the initial weight matrix. For example, if a user's purchase of financial products has a behavior intensity of 8 (assuming the maximum intensity is 10), then the adjustment coefficient of this behavior type will be 0.8, indicating that this behavior has a greater impact on the recommendation.

[0076] By constructing a behavior intensity quantification function, the system can flexibly adjust the weight according to the actual behavior intensity of the user, so that behaviors with higher intensity can have a greater impact on the push system. For example, users who frequently purchase financial products have higher behavior intensity and a larger adjustment coefficient. The system should assign a higher weight to this behavior, so that it is more inclined to recommend financial products when recommending. For example, the behavior intensity of user C's purchase of financial products is 7, and the maximum intensity is 10, so the adjustment coefficient is 0.7, and the basic weight of purchasing financial products can be adjusted according to this adjustment coefficient.

[0077] Multiply the basic weights in the initial behavior weight matrix by the corresponding weight adjustment coefficient. This process makes the weight of each behavior type more accurately reflect the user's behavior intensity and participation, thereby ensuring a higher degree of personalization in subsequent recommendations. Through this operation, the weights of different behavior types can be dynamically adjusted so that users' high-intensity behaviors receive higher recommendation priority. For example, if user A's behavior intensity of purchasing financial products is high, the adjusted weight will be greater than other behavior types, and related products will be recommended first.

[0078] Step S1235 finally generates weights for the two behavioral features to provide support for personalized recommendations. By assigning weights for positive and negative behaviors to the two features respectively, it is possible to accurately judge the user's preferences and potential risks based on the user's behavioral data, and adjust the recommendation strategy accordingly. For example, financial products related to positive behaviors (such as financial products, credit cards) can be recommended first, while negative behaviors (such as high-risk investments) can be appropriately avoided. For example, if user A exhibits a strong positive behavior of purchasing financial products, the adjusted weight is 0.8, which will become the first behavior weight of the first behavior feature; if user A has a low overdue repayment record, the adjusted negative behavior weight is 0.3, which will become the second behavior weight of the second behavior feature.

[0079] Step S1300, generating a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature and the second behavior weight.

[0080] Furthermore, step S1300 includes: Step S1310, constructing a positive preference dimension according to the first behavior feature and the first behavior weight; Step S1320, constructing a negative preference dimension according to the second behavior feature and the second behavior weight; Step S1330, integrating the positive preference dimension and the negative preference dimension to generate a multi-dimensional user portrait vector.

[0081] Specifically, the goal of step S1300 is to generate a multi-dimensional user portrait vector based on the user's behavior data, which integrates the user's behavior characteristics in financial services and the weights of these behaviors. Financial institutions, such as banks, usually collect user behavior data, such as deposits, loans, investments, payment behaviors, etc., and weight these data to understand the user's preferences, needs, risk tendencies and other characteristics. This process not only helps banks identify user needs, but also helps them provide personalized services or products and make more targeted marketing decisions in the future.

[0082] The first behavior feature usually refers to the user's positive behavior data, such as the user's deposit, loan application, investment purchase, credit card consumption, etc. Positive behavior reflects the user's preference or interest in financial products, and shows their financial needs and activity. For example, users frequently deposit money, invest in financial products, or actively participate in specific promotional activities, which are all positive behaviors. In actual applications, different behaviors may contribute to different degrees to user profiles, so weights need to be assigned to these behaviors. The assignment of behavior weights depends on the frequency, amount, duration, and importance of the behavior. For example, a large deposit may reflect the user's dependence on bank services more than an occasional small transaction, so its weight is higher; while frequent small consumption or short-term loans may represent the user's general needs and have a lower weight. By combining the first behavior feature with the first behavior weight, a positive preference dimension, that is, the user's positive behavior preference, can be constructed. This dimension can reflect the user's core needs and interests. For example, users who frequently invest in funds indicate a strong interest in investment and financial management, and banks can infer the user's demand for financial products based on this. Constructing a positive preference dimension helps to accurately identify users' interests and needs in specific financial products, thereby achieving personalized marketing and improving user satisfaction and loyalty. For example, banks can recommend relevant financial products and loan offers to users based on their deposit and investment behaviors, thereby improving user experience and product sales results.

[0083] The second behavior feature usually refers to the user's negative behavior data, such as user complaints, service cancellations, loan application rejections, negative reviews, etc. Negative behavior usually reflects user dissatisfaction with financial services or the potential risk of churn. The weight of negative behavior also depends on the frequency, severity and potential impact of the behavior on user churn. For example, when a user complains multiple times in a short period of time or threatens to close an account, the weight should be larger; while an occasional minor complaint should have a smaller weight. By combining the second behavior feature and the second behavior weight, a negative preference dimension can be constructed, that is, the user's behavioral tendency in negative situations. For example, if a user frequently complains about the bank's services or has posted negative reviews on the bank's platform, it indicates that the user has strong negative emotions and may distrust certain bank services. Constructing a negative preference dimension helps to identify user dissatisfaction and churn risks in a timely manner, thereby avoiding further expansion of negative emotions. Timely intervention can be made at the early stage of user dissatisfaction to mitigate negative impacts and prevent user churn by improving services and providing compensation. In addition, the negative preference dimension can also help analyze and optimize service processes, improve user experience, and reduce user complaints.

[0084] In step S1330, the positive preference dimension and the negative preference dimension are fused to form a comprehensive multi-dimensional user portrait vector. This fusion process combines the positive and negative behaviors of the user according to certain weights by weighted sum to obtain a comprehensive user portrait. For example, the positive preference dimension and the negative preference dimension can occupy different weight ratios respectively, or be dynamically adjusted according to their impact on the risk of user churn. The fused user portrait vector contains the user's positive preference, negative emotion, and the comprehensive performance of the two. This multi-dimensional user portrait vector not only shows the user's needs and interests, but also reveals their emotional attitudes and behavior patterns. This comprehensive vector provides strong data support for subsequent product recommendations, personalized services, and marketing decisions. Through the fusion of positive and negative preference dimensions, the user's behavior and emotional attitude can be fully understood, and a more accurate user portrait can be constructed. For example, the positive preference dimension reflects the user's interest in financial products, while the negative preference dimension reveals the user's potential dissatisfaction with the service. The fused portrait can help banks formulate more effective user management strategies, avoid blindly promoting products that users are not interested in, and adjust service quality in a timely manner to avoid user churn.

[0085] Step S2000, constructing a pre-trained industry language model, obtaining a first semantic embedding vector and a second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjusting the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtaining an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector; Furthermore, step S2000 includes: Step S2100, constructing a pre-trained industry language model, and obtaining a first semantic embedding vector and a second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; Furthermore, step S2100 includes: Step S2110, constructing a pre-trained industry language model; Specifically, in the financial field, especially the banking industry, the pre-trained industry language model can help automate tasks such as user service, promotion, and user sentiment analysis. By learning the professional terminology, knowledge background, and user behavior data of the financial industry, the model can better understand and generate natural language content related to financial business.

[0086] The architecture of pre-trained industry language models is usually based on the Transformer structure, such as BERT, GPT, and T5. The Transformer architecture can effectively handle long-distance dependencies through the self-attention mechanism, so it performs particularly well when dealing with complex language patterns in the financial field. Different Transformer variants are suitable for different task scenarios: Encoder-Decoder architectures (such as T5) are good at generation tasks, such as generating bank recommendations or automatic replies; Encoder-only architectures (such as BERT) are more suitable for understanding tasks, such as sentiment analysis, user question classification, etc.; and Decoder-only architectures (such as GPT) are widely used in generation tasks and dialogue generation.

[0087] Pre-training industry language models needs to process text inputs that are closely related to the industry, such as user consultation records, complaint content, bank product descriptions, and user transaction data. During the input data preprocessing and embedding process, the model can gradually learn the professional language features and expressions of the financial industry. The pre-training of the model relies on a large amount of financial field data, including banking service conversations, user feedback, industry reports, and investment analysis articles. These data need to be cleaned and labeled to remove irrelevant content in order to retain vocabulary and sentiment expressions related to the financial industry. During the training process, the model's goal is usually a language modeling task (such as Masked Language Modeling, MLM) or an autoregressive generation task (such as Autoregressive Modeling), which gradually learns context information by predicting masked vocabulary or generating the next part of the text.

[0088] In model training, the input data is usually preprocessed text, which may be presented in the form of a single sentence or multiple sentences. For example, the format of "user consultation: question, bank reply: reply content" can be used as a typical input. Depending on the task, the true label may be a "positive" or "negative" label in a sentiment analysis task, or it may be the correct reply corresponding to the input text in a text generation task. After the training is completed, the results of the model output mainly include two forms: one is the semantic embedding vector, which can capture the deep information of the text content and help the model understand the emotions and intentions in the financial field; the other is the generated content. For generation tasks, the model can directly generate automated replies, such as generating accurate answers to user inquiries in user service scenarios.

[0089] Building a pre-trained industry language model can enable banks and other financial institutions to better understand users' needs, emotional states, and intentions, thereby providing more accurate services or recommendations. For automated customer service and salesperson replacement in the financial industry, the model can automatically generate personalized responses, reduce the pressure on manual customer service, and improve response speed and accuracy.

[0090] Step S2120, inputting the positive preference dimension in the multi-dimensional user portrait vector into the pre-trained industry language model to obtain a first semantic embedding vector; Specifically, the positive preference dimension in the user portrait is input into the pre-trained industry language model to obtain the first semantic embedding vector. The positive preference dimension usually includes the needs and interests shown by the user in past behavior. For example, if the user frequently purchases a certain financial product (such as funds, loans, etc.), it indicates that they have a high interest in these products. This information is converted into text descriptions, such as "the user has a strong interest in fund investment" or "the user often applies for personal loans." These texts are input into the pre-trained industry language model, and the model generates the first semantic embedding vector by processing these texts and applying its built-in semantic knowledge. The semantic embedding vector is a high-dimensional vector that represents the deep semantic information about the user's preferences. It not only contains the user's behavioral characteristics, but also includes information about industry terms and sentiment tendencies. By inputting the positive preference dimension into the pre-trained industry language model, it is possible to have a deeper understanding of the user's needs and interests and provide personalized financial service recommendations for them. The first semantic embedding vector can provide accurate data support for subsequent intelligent customer service, marketing push, etc., thereby improving user experience and bank service efficiency.

[0091] Step S2130, input the negative preference dimension of the multi-dimensional user portrait vector into the pre-trained industry language model to obtain a second semantic embedding vector.

[0092] Specifically, the negative preference dimension usually includes the user's negative behavior data, such as complaint records, unsatisfactory feedback, rejected services, etc. This information is converted into text descriptions, such as "users are dissatisfied with the bank's services" or "users have complained about high service fees". After these texts are input into the pre-trained industry language model, the model will generate a second semantic embedding vector based on its inherent semantic understanding ability. This vector contains semantic information of negative emotions, such as users' aversion to certain bank services or distrust of certain products. By inputting the negative preference dimension into the pre-trained industry language model, users' dissatisfaction and negative emotions can be better identified. The second semantic embedding vector generated by the model provides banks with a basis for predicting user churn, and helps banks to optimize services and recover users in a timely manner based on negative emotions.

[0093] Step S2200, adjusting the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; Further, step S2200 includes: Step S2210, calculating the similarity between the first semantic embedding vector and the second semantic embedding vector to obtain a positive and negative preference similarity score D1; Calculating the similarity between the first semantic embedding vector and the second semantic embedding vector includes: ; in: Part 1 (Gaussian weight term): : The first semantic embedding vector and the second semantic embedding vector The squared Euclidean distance is used to measure the overall geometric distance between two vectors in the semantic space. The larger the geometric distance of the embedded vectors, the more significant the semantic difference between the two.

[0094] : Gaussian weight term, using the exponential decay of the negative square distance, emphasizes similar situations with smaller distances, while rapidly decaying those with larger distances. It can give higher similarity weights to semantic embedding points with closer distances, reducing the influence of distant points.

[0095] : An exponential function with the base e of the natural logarithm as its base.

[0096] Part 2 (Weight Modification Item): : Dynamic weight modifier, combined with the dynamic weight of each dimension A weighted sum of the local distances of the embedding vectors is performed.

[0097] : The dimension size of the embedding vector (usually 300~768).

[0098] : No. The dynamic weight of each dimension indicates the importance of the dimension in the calculation of semantic similarity. Through industry corpus analysis, the semantic distribution variance corresponding to each dimension is statistically analyzed: It is It is used to reduce the impact of high-variance dimensions on similarity calculation and highlight the stable contribution of low-variance dimensions.

[0099] : No. The square of the local distance between two vectors in dimensions is used to quantify the local difference.

[0100] This item can combine the dynamic weights of each dimension to provide a more refined semantic characterization capability for the overall similarity.

[0101] Part 3 (angle similarity term): : The angle between two vectors (expressed in radians), obtained by the inverse cosine function of cosine similarity. is the dot product of two vectors, and is their Euclidean norm. The smaller the angle, the more similar the two vectors are; the larger the angle, the less similar the two vectors are.

[0102] : Normalize the angle to , so that it can directly participate in weighting.

[0103] : The inverse form of angular similarity, the closer to 1, the more similar, and the closer to 0, the less similar.

[0104] The semantic consistency of the two vectors in direction is captured through the angle correction of cosine similarity.

[0105] Part 4 (Semantic Deviation Correction): : The semantic shift correction term combines the dynamic ratio of the first-order norm and the second-order norm to correct the overall semantic shift of the vector.

[0106] : The first-order norm (Manhattan distance) of the first semantic embedding vector and the second semantic embedding vector, which is used to measure the semantic difference between the two vectors.

[0107] : The sum of the Euclidean norms of two vectors, used to measure the overall semantic centrality of the two vectors.

[0108] : semantic offset correction coefficient, a hyperparameter that controls the contribution of the semantic offset term to the final similarity. Value range: . By adjusting the parameters on the validation set, we select the optimal value that can optimize the recommendation performance.

[0109] : Normalization adjustment coefficient, a hyperparameter used to balance the normalization effect of the offset term. Value range: . Obtained by adjusting the parameters of the validation set.

[0110] The semantic shift correction term can capture the overall semantic shift trend of the embedding vector and optimize the similarity distortion problem caused by directionality and local weight.

[0111] when As it increases, the Gaussian weight term Rapid decay, making Lower. Manhattan distance When it increases, the semantic shift correction term increases, further reducing When the angle between the two vectors increases, the cosine similarity decreases, resulting in an angle correction term Decrease, thereby reducing . Weights for high variance dimensions Lower, the impact on similarity calculation is weakened; the weight of low variance dimension Higher, making the local distance The impact of this has increased.

[0112] This formula combines geometric distance, dynamic weight, directionality, and semantic offset correction to more comprehensively measure the similarity of semantic embedding vectors. It not only considers the overall distance of the vector, but also combines the importance of local dimensions and directional consistency, and has the ability to describe similarity from multiple angles. Dynamic weight distribution and the normalized adjustment factor The introduction of makes the formula adaptable to different scenarios and the semantic characteristics of the embedding vector. and Regulation can adjust the sensitivity of similarity in different tasks (such as recommendation or filtering). The Gaussian weight term can quickly attenuate abnormal vectors (too far away) and reduce the impact of outliers on similarity calculation. The semantic offset correction term can correct the distortion of similarity caused by directional and local deviations.

[0113] Step S2220, setting a similarity threshold θ1, if the positive and negative preference similarity scores D1 are greater than θ1, triggering the embedding vector adjustment process to obtain a positive preference embedding vector and a negative preference embedding vector.

[0114] The embedding vector adjustment process includes: Step S2221, randomly sampling a noise vector, adjusting the first semantic embedding vector according to the noise vector, obtaining an adjusted first semantic embedding vector, and marking it as a positive preference embedding vector; Step S2222, adjusting the second semantic embedding vector according to the noise vector to obtain an adjusted second semantic embedding vector, which is marked as a negative preference embedding vector; Step S2223, calculate the similarity D2 between the positive preference embedding vector and the negative preference embedding vector, and determine whether D2 is less than or equal to θ1. If so, complete the embedding vector adjustment process; otherwise, return to step S2221, re-randomly sample the noise vector, and repeat steps S2221-S2223 until D2≤θ1.

[0115] Specifically, a user profile is usually composed of multiple dimensions, including the user's positive preferences (such as preference for certain financial products) and negative preferences (such as dissatisfaction with certain services). The purpose of step S2220 is to ensure that when making personalized recommendations for users, there is a clear distinction between the semantic embedding vectors of positive and negative preferences, so as to accurately reflect the user's needs and emotional state.

[0116] The similarity threshold θ1 is a preset value used to determine whether the similarity of the positive and negative preference embedding vectors exceeds an acceptable range. This value is usually adjusted based on experience or through cross-validation to ensure the balance of the model. For example, if θ1 is set to 0.8, when the positive and negative preference similarity is greater than 0.8, it means that the two vectors are too close and need to be adjusted. By setting the threshold θ1, it is possible to avoid excessive overlap of semantic information in the positive and negative preference dimensions, ensuring that the model can recognize the multi-dimensional behavioral characteristics of users rather than merging positive and negative behaviors into the same preference.

[0117] In step S2221, the first semantic embedding vector is adjusted by sampling a noise vector, with the aim of increasing the difference between the positive and negative preference embedding vectors to make them more distinguishable. The noise vector is a random vector, usually sampled from a standard normal distribution, i.e., a normal distribution with a mean of 0 and a variance of 1. The introduction of noise can help adjust the spatial distribution of the vector, thereby avoiding the embedding vectors being too similar. The noise vector is weighted and synthesized with the first semantic embedding vector to obtain a new second semantic embedding vector. This process is similar to perturbing the original vector, thereby adjusting its position in the vector space so that the adjusted first semantic embedding vector is more different from the second semantic embedding vector. By introducing a noise vector and adjusting the first semantic embedding vector, overly similar embedding vectors can be effectively avoided, ensuring that there is sufficient semantic difference between positive and negative preferences. This is crucial for accurately modeling users' positive and negative emotions, especially in personalized push systems, which can help the model more accurately identify user needs and emotions, thereby optimizing services and recommendations.

[0118] Similar to step S2221, the goal of step S2222 is to adjust the second semantic embedding vector so that the distance between it and the first semantic embedding vector is further increased, thereby ensuring that they are significantly distinguished in the semantic space. The noise vector used here is the same as that in step S2221. By adjusting the second semantic embedding vector, the distinction between positive and negative preferences is further enhanced. The adjustment of the second semantic embedding vector helps to more clearly distinguish the user's negative emotional information, such as dissatisfaction with certain financial products or services, thereby improving the model's ability to understand user needs.

[0119] Step S2223 continuously adjusts the first semantic embedding vector and the second semantic embedding vector until the similarity between the positive and negative preferences meets the preset threshold, thereby ensuring that the final user portrait is more accurate. In this way, it can help provide more accurate personalized services and product recommendations, avoid confusion between positive and negative preferences, and improve user satisfaction and loyalty.

[0120] Step S2300: Obtain an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector.

[0121] Furthermore, step S2300 includes: Step S2310, loading the industry corpus, using the positive preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus, and generating a first candidate push content set; Furthermore, if Figure 4 As shown, step S2310 includes: Step S2311, converting the text content in the industry corpus into corresponding semantic embedding vectors; Step S2312, calculating the similarity between the positive preference embedding vector in the multi-dimensional user portrait vector and the semantic embedding vector of the text content in the industry corpus to obtain a first similarity score; Step S2313: Filter the text contents with the first similarity scores in the top K in the industry corpus to generate a first set of candidate push contents.

[0122] Specifically, the goal of step S2310 is to filter out the text content that best matches the user's needs by calculating the semantic similarity between the user's positive preference embedding vector and the relevant text in the industry corpus, thereby generating a candidate push content set. This process mainly relies on semantic embedding technology to convert text into vector representation and match it by calculating similarity. In this way, personalized recommended content can be provided based on the user's preferences, thereby improving user satisfaction.

[0123] The industry corpus contains industry-related text data, such as product introductions, terms of service, FAQs, market reports, investment advice, etc. These texts are the basis for communication with users and represent the diversity and professionalism of products. The process of loading an industry corpus usually includes reading text files, database queries, etc., to ensure that the text content in the corpus is complete and can be efficiently semantically retrieved. By loading the industry corpus, you can obtain rich industry knowledge and user interaction content, providing basic support for subsequent personalized push. The quality of the corpus directly affects the accuracy of the push content and user experience.

[0124] Step S2311 converts the text content in the industry corpus into corresponding semantic embedding vectors. Semantic embedding vectors are a text representation method that can map text to a high-dimensional vector space so that semantically similar texts are closer in the vector space. Common semantic embedding models include Word2Vec, GloVe, BERT, etc. Through semantic embedding, texts of different lengths and different expressions can be converted into a unified vector representation, which is convenient for subsequent similarity calculations.

[0125] Step S2312 calculates the similarity between the positive preference embedding vector in the multidimensional user portrait vector and the semantic embedding vector of the text content in the industry corpus to obtain a first similarity score. Similarity calculation is a measure of the semantic proximity of two vectors. Commonly used similarity metrics include cosine similarity, Euclidean distance, etc. Cosine similarity calculates the cosine value of the angle between two vectors. The value range is between [-1,1]. The larger the value, the higher the similarity. Euclidean distance calculates the straight-line distance between two vectors in space. The smaller the distance, the higher the similarity. By calculating the similarity score, the text content that best matches the user's needs can be effectively screened out based on the user's positive preferences. The core advantage of this process is that it can capture the deep semantic connection between user needs and text content, thereby achieving personalized recommendations and improving user experience and engagement.

[0126] Step S2313 sorts the text content in the corpus according to the similarity calculation results obtained in step S2312, and selects the top K texts with the highest similarity as candidate push content. The selection of the K value requires a balance between the quality and diversity of the pushed content: a K value that is too small may cause the pushed content to be too limited, while a K value that is too large may introduce irrelevant content. The optimal K value can usually be determined based on user feedback and business needs through AB testing and other methods. By screening out the content that is most relevant to user needs, it can ensure that the pushed text information is more in line with the user's interests and needs, thereby improving the accuracy of the recommendation and user satisfaction. In addition, this process can also optimize push efficiency and reduce interference from irrelevant content.

[0127] For example, assume that user A has shown a preference for low-risk investment products in past financial activities and is more concerned about the user service quality of the bank. In step S2310, the positive preference embedding vector of user A has been generated through his behavior data. Next, the industry corpus contains text descriptions of multiple investment products and services.

[0128] Step S2311: The text in the industry corpus (for example: "Low-risk fund product Q", "User service optimization", etc.) is converted into a semantic embedding vector.

[0129] Step S2312: Calculate the similarity between the positive preference embedding vector of user A and the semantic embedding vector of each text. Assume that the text "low-risk fund product Q" has a high similarity with the positive preference of user A, while "user service optimization" has a relatively low similarity.

[0130] Step S2313: Filter out the top K texts in similarity score ranking. Assuming that the first two texts are “Low-risk fund product Q” and “User service optimization”, these two texts will constitute the first candidate push content set.

[0131] Through this method, banks can accurately push personalized content that meets user needs, improve user engagement, and enhance user satisfaction and loyalty to banking services.

[0132] Step S2320, using the negative preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus to generate a second set of candidate push content; Further, step S2320 includes: Step S2321, performing similarity calculation on the negative preference embedding vector in the multi-dimensional user portrait vector and the semantic embedding vector of the text content in the industry corpus to obtain a second similarity score; Step S2322: Filter the text contents with the second similarity scores ranked in the top M in the industry corpus to generate a second set of candidate pushed contents.

[0133] Specifically, the purpose of step S2320 is to filter out text content related to user dissatisfaction, negative emotions or rejection from the industry corpus based on the user's negative preference embedding vector. By calculating the similarity between the negative preference embedding vector and each text in the corpus, a text set matching the user's emotions and dissatisfaction is generated. The goal of this process is to avoid pushing content that may cause negative emotions in users, and to help banks better identify users' pain points and needs and improve service quality.

[0134] In step S2321, the similarity between the user's negative preference embedding vector and the semantic embedding vector of each text in the industry corpus is first calculated. Similar to step S2312, the similarity between the negative preference embedding vector and the text content is calculated here, with the aim of finding content related to user dissatisfaction and negative emotions. The similarity calculation is completed by comparing the negative preference embedding vector with the semantic embedding vector of each text in the industry corpus. Commonly used similarity measurement methods include cosine similarity, Euclidean distance, etc. By calculating the similarity between the negative preference embedding vector and the text, the text content that matches the user's negative emotions can be accurately identified. This process is crucial for the push system because it can effectively prevent the push of content that may cause user dissatisfaction or distrust. In this way, banks can optimize push content and better manage user emotions.

[0135] In step S2322, according to the second similarity score calculated in step S2321, the top M text contents that best match the user's negative preference are screened out, and these texts are used as the second candidate push content set. These text contents are usually related to the user's negative emotions or needs, and are intended to avoid pushing content that may exacerbate negative emotions or dissatisfaction. By screening out the content that best matches the user's negative preference, it is possible to avoid pushing content that arouses user disgust, thereby improving user satisfaction and loyalty. This step helps to optimize the user experience, especially in terms of service quality and product push, and helps banks avoid the risk of user churn.

[0136] Step S2330: obtaining an adjusted push content set based on the first candidate push content set and the second candidate push content set.

[0137] Furthermore, if Figure 5 As shown, step S2330 includes: Step S2331, obtaining semantic embedding vectors of text contents in the first candidate push content set and the second candidate push content set; Step S2332, calculating the semantic similarity S2 between each text content in the first candidate push content set and all text content in the second candidate push content set based on the semantic embedding vectors of the text content in the first candidate push content set and the second candidate push content set, to form a similarity matrix; Step S2333, in the similarity matrix, filter out content pairs whose semantic similarity S2 is higher than a second similarity threshold θ2, and mark them as highly overlapping content; Step S2334: remove the marked highly overlapping contents from the first candidate push content set to obtain an adjusted push content set.

[0138] Specifically, the goal of step S2330 is to optimize the final push content by combining the first candidate push content set and the second candidate push content set to remove duplicate or highly overlapping content. This step mainly calculates the semantic similarity between the content, screens out highly duplicate content, and performs deduplication processing to ensure that the content finally pushed to the user is diversified and personalized, and avoids excessively repeated pushes.

[0139] The semantic embedding vector mentioned in step S2331 is a technology for mapping text to a high-dimensional space. Through semantic embedding, the text content can be represented as a real number vector of a fixed dimension, in which semantically similar texts are closer in the vector space. Common semantic embedding models include Word2Vec, GloVe, and BERT. These models learn the semantic representation of words or sentences through unsupervised or semi-supervised training on large-scale corpora. By obtaining the semantic embedding vectors of the candidate push content, the similarity between them can be calculated in the vector space, providing a basis for subsequent overlapping content detection.

[0140] Step S2332 calculates the semantic similarity between each text content in the first candidate push content set and all text content in the second candidate push content set to form a similarity matrix. Semantic similarity can be obtained by calculating metrics such as cosine similarity or Euclidean distance between two semantic embedding vectors. Cosine similarity calculates the cosine value of the angle between vectors, and the value range is [-1,1]. The larger the value, the closer the vector direction is, that is, the more similar the text semantics are. Euclidean distance calculates the straight-line distance between vectors, and the smaller the distance, the more similar the text semantics are. By constructing a similarity matrix, the degree of semantic overlap between the two groups of content can be more clearly understood. This step helps to accurately identify the relationship between the first candidate push content and the second candidate push content, so as to make more reasonable screening decisions and avoid excessive duplication or high correlation of content.

[0141] Step S2333 selects content pairs whose semantic similarity exceeds the threshold in the similarity matrix according to the preset second similarity threshold θ2, and marks them as highly overlapping content. The value of the second similarity threshold θ2 needs to be adjusted according to the specific application scenario and the characteristics of the pushed content. If the threshold is set too high, it may not be possible to effectively identify overlapping content; if the threshold is set too low, semantically related but non-duplicate content may be filtered out. Therefore, it is necessary to adjust parameters and test on actual data to select a suitable threshold to balance the accuracy and recall of filtering. By removing highly overlapping content, it is possible to avoid pushing duplicate information, thereby improving the diversity and personalization of push and avoiding excessive push of similar content. This not only improves the user experience, but also increases the accuracy and effectiveness of the push system.

[0142] Step S2334 removes the highly overlapping content marked in step S2333 from the first candidate push content set. The removal operation ensures that the final push content set does not contain repeated or highly similar texts, thereby ensuring that the content pushed to the user is both diverse and targeted. Through this step, it is possible to avoid excessive pushes to users, reduce information fatigue, and improve user engagement and loyalty.

[0143] For example, assume that user C shows great interest in low-risk financial products, but is dissatisfied with excessively high bank fees.

[0144] Step S2331: the text in the first candidate push set includes "recommendation of low-risk financial products" and "fund investment product Y", and the text in the second candidate push set includes "new policy on handling fees" and "service quality improvement plan".

[0145] Step S2332: Calculate the similarity between each first candidate push text and all second candidate push texts to obtain a similarity matrix. Assume that the similarity between "recommendation of low-risk financial products" and "new policy on handling fees" is 0.85, which exceeds the second similarity threshold θ2.

[0146] Step S2333: Filter out content pairs with a similarity higher than a second similarity threshold θ2, such as “recommendation of low-risk financial products” and “new policy on handling fees” are marked as “highly overlapping content”.

[0147] Step S2334: Remove "recommendation of low-risk financial products" from the first candidate push set, generate an adjusted push content set, and ensure that the remaining content does not contain repeated or highly similar text.

[0148] By removing content that is too similar, the bank ensures that each push contains more unique options, avoiding duplication and information overload. In this way, users will receive more different content that meets their needs and interests, thereby improving their experience. By eliminating duplicate push content, users will not be bothered by the same information. By filtering the push content for semantic similarity, it can ensure that each piece of information recommended to users is highly relevant and targeted. Whether it is a recommendation for financial products or a notification of service quality improvement, it can better match the user's preferences and needs, enhancing the personalization and effectiveness of the push content. In a financial environment with a large amount of information, it is crucial to avoid repeatedly pushing the same recommended content. Information fatigue can cause users to lose interest in recommendations or even develop negative emotions. Through this deduplication process, banks can increase users' attention and satisfaction with push content.

[0149] Step S3000: Based on the adjusted push content set, a third candidate push content set is obtained; and the contents in the third candidate push content set are intelligently sorted to generate an accurate push content sequence.

[0150] Furthermore, step S3000 includes: Step S3100, constructing and training a semantic adversarial generative network based on the adjusted push content set; Further, step S3100 includes: Step S3110, marking the adjusted push content set as a positive sample, marking the second candidate push content set as a negative sample, and constructing a training data set for a semantic adversarial generative network; Step S3120: training the semantic generative adversarial network according to the training data set of the semantic generative adversarial network.

[0151] Specifically, the goal of step S3100 is to construct and train a semantic adversarial generative network (SAGN) through the adjusted push content set to generate personalized recommendation content that better meets user needs and emotions. The design of the semantic adversarial generative network is inspired by the generative adversarial network (GANs), which uses adversarial training to generate more natural and highly personalized push content through the competition of two networks (generator and discriminator). The challenge of personalized recommendation lies in how to dynamically adjust the recommended content according to the user's preferences, needs and emotional state. The semantic adversarial generative network can help banks automatically generate recommended content that meets their needs based on the user's positive and negative emotional preferences, thereby improving the accuracy of recommendations and user satisfaction.

[0152] In step S3110, a training data set for the semantic adversarial generative network is prepared. This data set consists of positive samples and negative samples: Positive samples: come from the adjusted set of pushed content. These contents are the pushed texts that have been deduplicated and optimized, representing the recommended content that meets the needs and preferences of users. When generating training data, positive samples should include all the content that is ultimately pushed to users, because they represent the content that the system considers to be the most relevant and most in line with user needs.

[0153] Negative samples: These are from the second candidate push content set. These are usually push information that does not match user needs or content that may cause negative emotions. The second candidate push content set usually includes content related to user negative preferences, such as excessively high fees, poor user service quality, and other information that does not meet user preferences.

[0154] Construct a training dataset: Input: the adjusted push content set (positive samples) and the second candidate push content set (negative samples).

[0155] Output: A dataset labeled with positive and negative samples for training semantic adversarial generative networks.

[0156] These labeled samples are provided to the training network, so that the generator can learn to generate push content that meets user needs, and the discriminator can effectively distinguish between positive and negative samples, thereby guiding the generator to optimize its generation strategy.

[0157] By constructing a training data set of positive and negative samples, sufficient training materials can be provided for the semantic adversarial generative network, ensuring that the generated recommendation content not only meets the user's preferences, but also avoids pushing irrelevant or offensive content. Such training can make the push system more accurate and personalized in practical applications.

[0158] In step S3120, the semantic adversarial generative network is trained using the training data set constructed in step S3110. The training process includes adversarial training of the generator and the discriminator: Generator: The task of the generator is to generate personalized recommendation content based on the user's profile (including positive and negative preferences). It learns from the training data set how to generate content that matches user needs, and continuously optimizes the generation strategy to make the generated content as close to the positive sample as possible at the semantic level.

[0159] Discriminator: The task of the discriminator is to judge whether the content generated by the generator is "real" or "fake". Real content usually comes from positive samples (i.e., the adjusted push content set), while fake content comes from the content generated by the generator. Through continuous learning, the discriminator can better distinguish between positive samples and fake samples, thereby guiding the generator to optimize the generation strategy.

[0160] Training process: Adversarial training: The generator and the discriminator compete with each other during the training process. The generator improves its generation strategy to try to generate content similar to the positive sample, while the discriminator improves its recognition ability by distinguishing real content from fake content. Through continuous iteration, the generator and the discriminator make progress together, and the generator can eventually generate high-quality recommendation content.

[0161] Loss function: The loss function in the training process usually includes generator loss and discriminator loss. The generator loss is used to measure the quality of the generated content, while the discriminator loss is used to measure its ability to distinguish between real content and fake content.

[0162] By training the semantic adversarial generative network, we can achieve higher personalization in the pushed content and generate content that better meets user needs and emotional states. After training, the generator can easily generate personalized recommendations for different user preferences, while the discriminator helps ensure the high quality and relevance of the recommended content. Through adversarial training, the generated content can reach a high level in diversity and accuracy, greatly improving user satisfaction.

[0163] Step S3200, obtaining the real-time behavior data of the target user, and obtaining a third candidate push content set based on the real-time behavior data and a semantic adversarial generation network; Specifically, in step S3200, based on the real-time behavior data of the target user, combined with the trained semantic adversarial generative network, new push content is generated. The core of this step is to dynamically adjust the recommended content by using the user's latest behavior and preferences to ensure that the push content is as consistent as possible with the user's current interests and needs.

[0164] Real-time behavior data refers to the user's instant interaction information, usually including the user's transaction records, click behavior, browsing history, consultation records, etc. These data reflect the user's current interests, needs and emotional state, and can help predict the products or services that the user will need in the future. For example, if the user has recently browsed the financial product page frequently, then he or she may have a strong interest in investment products. Real-time behavior data is usually obtained through various data collection tools or technologies (such as Web log analysis, App behavior tracking, user interaction data analysis, etc.).

[0165] The semantic adversarial generative network (SAGN) updates the user's profile based on real-time behavioral data and generates content that best matches the user's interests. For example, if a user has recently invested in a low-risk fund, SAGN may generate push content that recommends products related to the fund. In this way, SAGN ensures that the push content not only matches the user's historical preferences, but also adapts to their current behavior patterns. By combining real-time behavioral data and semantic adversarial generative networks, it is possible to provide instant, dynamic, personalized recommendations. This behavior-based data-driven push can increase user engagement because it ensures that the recommended content is timely and relevant, thereby improving the user experience and reducing the risk of user churn.

[0166] Step S3300: input the third candidate push content set into a pre-built push strategy reinforcement learning model to obtain an optimal push strategy; intelligently sort the contents in the third candidate push content set according to the optimal push strategy to generate an accurate push content sequence.

[0167] Specifically, step S3300 inputs the generated third candidate push content set into the pre-built push strategy reinforcement learning model to generate the optimal push strategy. The reinforcement learning model simulates different push strategies and optimizes according to preset goals to determine which strategy can maximize user satisfaction and push effect.

[0168] The push strategy reinforcement learning model is a machine learning model that uses a reinforcement learning (RL) algorithm to automatically adjust the push strategy. Reinforcement learning is a learning method based on reward feedback. The model tries different push strategies to obtain the best result. The core of the reinforcement learning model is the reward and punishment mechanism: the execution of each push strategy will bring a "reward" or "penalty" based on the user's response (such as clicks, purchases, browsing time, etc.). In the end, the model will learn which push strategies are most effective and use them in practical applications. The input of the push strategy reinforcement learning model is the third candidate push content set, and the output is the optimal push strategy.

[0169] Reinforcement learning process: Reward and Penalty Mechanism: The reinforcement learning model is trained based on the user's response to the pushed content. For example, if the user clicks on a certain recommended content, the push strategy for that recommended content will be rewarded; if the user does not click, they will be punished. Through continuous feedback and adjustment, the model eventually learns how to optimize the push strategy to maximize user engagement and satisfaction.

[0170] Optimization goal: The goal of a push strategy is to maximize the long-term value of the user, not just a single interaction. For example, if the recommended financial product helps the user improve the return on investment or enhances the user's trust in the bank, then the recommendation will be considered a successful push strategy.

[0171] By optimizing the push strategy reinforcement learning model, we can achieve the best decision in the selection and sorting of push content, ensuring that the recommended content can maximize the user's interests and participation. This adaptive and intelligent push strategy can significantly improve user satisfaction, increase user activity, and effectively reduce the push of invalid or overly repetitive content.

[0172] According to the optimal push strategy, the third candidate push content set is intelligently sorted to generate the final accurate push content sequence. This sorting process is not just a simple sorting by relevance, but a comprehensive consideration of multiple factors, such as the timeliness of the content, user interests, push history, etc., to ensure that each push content is pushed to the user at the best time and in the most appropriate context.

[0173] Key factors for smart sorting: Timeliness: User needs may change at any time, so recommended financial products and services need to be updated in a timely manner. For example, if a user has recently viewed a financial product, relevant updated information should be displayed first.

[0174] Long-term interests of users: In addition to real-time behavior, long-term interests of users also need to be considered. For example, if a user invests in low-risk financial products for a long time, then relevant low-risk investment products will be pushed first.

[0175] User historical interaction data: Based on the historical data of user interaction with the bank (such as purchase records, service feedback, etc.), the model can infer which content may better meet the user's long-term needs.

[0176] Through intelligent sorting, users can be provided with an optimized push content sequence, so that each push message is delivered to the user at the best time and in the best context. This optimization can significantly improve user engagement, purchase conversion rate and satisfaction, while reducing excessive push and information fatigue.

[0177] Example 2

[0178] This embodiment provides a large model information push method based on industry corpus based on the first embodiment, including: Step S1233, constructing a behavior intensity quantification function, and mapping the behavior intensity distribution of different behavior types to weight adjustment coefficients of corresponding behavior types according to the behavior intensity quantification function; The behavior intensity quantification function includes:

[0179] in: Indicates the behavior type The weight adjustment factor.

[0180] Indicates the behavior type The basic weight is a preset constant that reflects the general importance of the behavior type. It can be set based on business experience and historical data analysis.

[0181] Indicates the behavior type The frequency of a behavior is the number of times a user performs a behavior within a certain time frame. This can be obtained by counting user behavior logs.

[0182] Indicates the behavior type The time span is the time range in which the user performs the behavior. It can be obtained by calculating the time difference between the first and last time the user performs the behavior.

[0183] It is the first scaling factor, which is used to adjust the magnitude of the ratio of behavior frequency to time span. It can be set according to the distribution characteristics of the actual data.

[0184] Indicates the behavior type The average duration of a behavior reflects the user's involvement in performing the behavior. For example, for browsing behavior, the average browsing duration can be used to measure it; for commenting behavior, the average number of words in the comment can be used to measure it.

[0185] is the second scaling factor, which is used to adjust the magnitude of the average duration. It can be set according to the distribution characteristics of the actual data.

[0186] Indicates the behavior type The content relevance reflects the degree of match between the content involved in the behavior and the user's interests. It can be measured by calculating the similarity between content features and user preference features.

[0187] is the third scaling factor, which is used to adjust the magnitude of content relevance. It can be set according to the distribution characteristics of actual data.

[0188] This formula takes into account multiple factors that affect the intensity of behavior and generates a weight adjustment coefficient that can reflect the intensity of behavior through nonlinear combination. Specifically: Logarithmic function The ratio of behavior frequency to time span is mapped to a bounded interval, and increases with the increase of the ratio, reflecting the positive impact of the density of behavior on intensity.

[0189] Hyperbolic tangent function Map the average duration to , and increases monotonically with the increase of duration, reflecting the positive impact of user participation on intensity.

[0190] Exponential function Map content relevance to , and increases monotonically with the increase of correlation, reflecting the positive impact of content matching on intensity.

[0191] Base Weight It plays a role in regulating the overall level of weight adjustment coefficients of different behavior types, reflecting the differences in a priori importance of different behavior types.

[0192] In general, this formula can calculate a comprehensive weight adjustment coefficient based on the multi-dimensional characteristics of user behavior. When a user shows high frequency, long duration, high participation and high relevance to a certain behavior type, the weight adjustment coefficient of this behavior type will increase accordingly, reflecting the user's strong interest in this type of content. Conversely, if the user is not active or interested in a certain behavior type, the weight adjustment coefficient will decrease accordingly. In this way, the user's preference for different types of content can be more comprehensively and accurately portrayed, providing better weight information for subsequent personalized recommendations.

[0193] It should be noted that this formula contains multiple scaling factors , its role is to normalize inputs of different orders of magnitude while maintaining the monotonicity of the function, so that the weight adjustment coefficients of different behavior types are in a similar numerical range, which is convenient for subsequent weight comparison and fusion. In practical applications, it is necessary to adjust these scaling factors appropriately according to the distribution characteristics of the data to achieve the best quantification effect. In addition, this formula is a possible form of behavior intensity quantification function. In practice, the formula can be further improved and optimized according to business needs and data characteristics.

[0194] Example 3 This embodiment provides a large model information push system based on industry corpus on the basis of embodiment 1, such as Figure 6 As shown, including: Portrait construction module: used to obtain the positive behavior record sequence and negative behavior record sequence of the target user; extract the first behavior feature from the positive behavior record sequence and extract the second behavior feature from the negative behavior record sequence; generate the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature; generate a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature and the second behavior weight; Push content generation module: used to build a pre-trained industry language model, obtain a first semantic embedding vector and a second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjust the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtain an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector; Push content sorting module: based on the adjusted push content set, obtain a third candidate push content set; intelligently sort the contents in the third candidate push content set to generate an accurate push content sequence.

[0195] In the portrait construction module, the step of obtaining the positive behavior record sequence and the negative behavior record sequence of the target user includes: Step S1110, constructing an industry corpus, and obtaining first historical behavior data of the target user in the industry field from the industry corpus; the first historical behavior data includes n1 behavior types of the target user, and the n1 behavior types include positive behaviors and negative behaviors; n1 is a positive integer; Step S1120: obtaining a positive behavior record sequence and a negative behavior record sequence according to the first historical behavior data.

[0196] In the portrait construction module, extracting the first behavior feature from the positive behavior record sequence includes: Step S1211, performing word segmentation processing on the text content in the positive behavior record sequence, dividing the text content into a word sequence of the smallest semantic unit; Step S1212, performing part-of-speech tagging on the word sequence after word segmentation; Step S1213, identifying named entities in the text content based on part-of-speech tagging; Step S1214, extracting keywords and subject words from the word sequence; Step S1215: taking keywords, subject words and named entity words as first behavior features.

[0197] In the portrait construction module, the generating of the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature includes: Step S1231, setting basic weights of different behavior types among n1 behavior types to construct an initial behavior weight matrix; Step S1232, based on the first historical behavior data, statistically analyzing the behavior intensity distribution of different behavior types; Step S1233, constructing a behavior intensity quantification function, and mapping the behavior intensity distribution of different behavior types to weight adjustment coefficients of corresponding behavior types according to the behavior intensity quantification function; Step S1234, multiplying the basic weights in the initial behavior weight matrix by the corresponding weight adjustment coefficients to generate an adjusted behavior weight matrix; Step S1235: In the adjusted behavior weight matrix, the weight corresponding to the positive behavior is used as the first behavior weight of the first behavior feature, and the weight corresponding to the negative behavior is used as the second behavior weight of the second behavior feature.

[0198] In the portrait construction module, generating a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature, and the second behavior weight includes: Step S1310, constructing a positive preference dimension according to the first behavior feature and the first behavior weight; Step S1320, constructing a negative preference dimension according to the second behavior feature and the second behavior weight; Step S1330, integrating the positive preference dimension and the negative preference dimension to generate a multi-dimensional user portrait vector.

[0199] In the push content generation module, obtaining the first semantic embedding vector and the second semantic embedding vector includes: Step S2110, constructing a pre-trained industry language model; Step S2120, inputting the positive preference dimension in the multi-dimensional user portrait vector into the pre-trained industry language model to obtain a first semantic embedding vector; Step S2130, input the negative preference dimension of the multi-dimensional user portrait vector into the pre-trained industry language model to obtain a second semantic embedding vector.

[0200] In the push content generation module, generating the positive preference embedding vector and the negative preference embedding vector includes: Step S2210, calculating the similarity between the first semantic embedding vector and the second semantic embedding vector to obtain a positive and negative preference similarity score D1; Step S2220, setting a similarity threshold θ1, if the positive and negative preference similarity scores D1 are greater than θ1, triggering the embedding vector adjustment process to obtain a positive preference embedding vector and a negative preference embedding vector.

[0201] The embedding vector adjustment process includes: Step S2221, randomly sampling a noise vector, adjusting the first semantic embedding vector according to the noise vector, obtaining an adjusted first semantic embedding vector, and marking it as a positive preference embedding vector; Step S2222, adjusting the second semantic embedding vector according to the noise vector to obtain an adjusted second semantic embedding vector, which is marked as a negative preference embedding vector; Step S2223, calculate the similarity D2 between the positive preference embedding vector and the negative preference embedding vector, and determine whether D2 is less than or equal to θ1. If so, complete the embedding vector adjustment process; otherwise, return to step S2221, re-randomly sample the noise vector, and repeat steps S2221-S2223 until D2≤θ1.

[0202] In the push content generation module, the adjusted push content set includes: Step S2310, loading the industry corpus, using the positive preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus, and generating a first candidate push content set; Step S2320, using the negative preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus to generate a second set of candidate push content; Step S2330: obtaining an adjusted push content set based on the first candidate push content set and the second candidate push content set.

[0203] The step S2310 includes: Step S2311, converting the text content in the industry corpus into corresponding semantic embedding vectors; Step S2312, calculating the similarity between the positive preference embedding vector in the multi-dimensional user portrait vector and the semantic embedding vector of the text content in the industry corpus to obtain a first similarity score; Step S2313: Filter the text contents with the first similarity scores in the top K in the industry corpus to generate a first set of candidate push contents.

[0204] The step S2320 includes: Step S2321, performing similarity calculation on the negative preference embedding vector in the multi-dimensional user portrait vector and the semantic embedding vector of the text content in the industry corpus to obtain a second similarity score; Step S2322: Filter the text contents with the second similarity scores ranked in the top M in the industry corpus to generate a second set of candidate pushed contents.

[0205] The step S2330 includes: Step S2331, obtaining semantic embedding vectors of text contents in the first candidate push content set and the second candidate push content set; Step S2332, calculating the semantic similarity S2 between each text content in the first candidate push content set and all text content in the second candidate push content set based on the semantic embedding vectors of the text content in the first candidate push content set and the second candidate push content set, to form a similarity matrix; Step S2333, in the similarity matrix, filter out content pairs whose semantic similarity S2 is higher than a second similarity threshold θ2, and mark them as highly overlapping content; Step S2334: remove the marked highly overlapping contents from the first candidate push content set to obtain an adjusted push content set.

[0206] In the push content sorting module, obtaining the third candidate push content set includes: Step S3100, constructing and training a semantic adversarial generative network based on the adjusted push content set; Step S3200: acquiring the real-time behavior data of the target user, and obtaining a third candidate push content set based on the real-time behavior data and a semantic adversarial generation network.

[0207] The step S3100 includes: Step S3110, marking the adjusted push content set as a positive sample, marking the second candidate push content set as a negative sample, and constructing a training data set for a semantic adversarial generative network; Step S3120: training the semantic generative adversarial network according to the training data set of the semantic generative adversarial network.

[0208] In the push content sorting module, generating an accurate push content sequence includes: Step S3300: input the third candidate push content set into a pre-built push strategy reinforcement learning model to obtain an optimal push strategy; intelligently sort the contents in the third candidate push content set according to the optimal push strategy to generate an accurate push content sequence.

[0209] Example 4 This embodiment discloses an electronic device, which may include one or more processors and one or more memories. The memories store computer-readable codes, which, when executed by the one or more processors, may execute the large model information push method based on industry corpus as described above.

[0210] The method or system according to the embodiment of the present application can also be implemented with the help of the architecture of an electronic device. The electronic device may include a bus, one or more CPUs, a read-only memory (ROM), a random access memory (RAM), a communication port connected to a network, an input / output component, a hard disk, etc. A storage device in the electronic device, such as a ROM or a hard disk, can store the large model information push method based on industry corpus provided in the present application. The large-model information push method based on industry corpus may, for example, include: obtaining a positive behavior record sequence and a negative behavior record sequence of a target user; extracting a first behavior feature from the positive behavior record sequence, and extracting a second behavior feature from the negative behavior record sequence; generating a first behavior weight for the first behavior feature and a second behavior weight for the second behavior feature; generating a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature, and the second behavior weight; constructing a pre-trained industry language model, and obtaining a first semantic embedding vector and a second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjusting the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtaining an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector; obtaining a third candidate push content set based on the adjusted push content set; intelligently sorting the contents in the third candidate push content set to generate an accurate push content sequence.

[0211] Furthermore, the electronic device may also include a user interface. Of course, the architecture disclosed in the present invention is only exemplary, and when implementing different devices, one or more components in the electronic device disclosed in the present invention may be omitted according to actual needs.

[0212] Example 5 This embodiment discloses a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor, the large model information push method based on industry corpus of the embodiment of the present application can be executed. The storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory (cache). Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0213] In addition, according to an embodiment of the present application, the process described above with reference to the flowchart may be implemented as a computer software program. For example, the present application provides a non-temporary machine-readable storage medium, which stores machine-readable instructions, and the machine-readable instructions can be executed by a processor to execute instructions corresponding to the method steps provided in the present application, for example: obtaining a positive behavior record sequence and a negative behavior record sequence of a target user; extracting a first behavior feature from the positive behavior record sequence, and extracting a second behavior feature from the negative behavior record sequence; generating a first behavior weight for the first behavior feature and a second behavior weight for the second behavior feature; generating a multidimensional user portrait vector based on the first behavior feature, the first behavior weight, the second behavior feature, and the second behavior weight; constructing a pre-trained industry language model, and obtaining a first semantic embedding vector and a second semantic embedding vector based on the multidimensional user portrait vector and the pre-trained industry language model; adjusting the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtaining an adjusted push content set based on the positive preference embedding vector and the negative preference embedding vector; obtaining a third candidate push content set based on the adjusted push content set; intelligently sorting the content in the third candidate push content set to generate an accurate push content sequence. When the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are performed.

[0214] The methods, systems, and devices of the present application may be implemented in many ways. For example, the methods, systems, and devices of the present application may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present application may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers recording media storing programs for executing the method according to the present application.

[0215] In addition, the parts of the above-mentioned technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0216] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A large model information push method based on industry corpus, characterized in that: The method comprises: Obtain a positive behavior record sequence and a negative behavior record sequence of a target user; extract a first behavior feature from the positive behavior record sequence, and extract a second behavior feature from the negative behavior record sequence; generate a first behavior weight of the first behavior feature and a second behavior weight of the second behavior feature; generate a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature, and the second behavior weight; Construct a pre-trained industry language model, and obtain a first semantic embedding vector and a second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjust the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtain an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector; Based on the adjusted push content set, a third candidate push content set is obtained; the contents in the third candidate push content set are intelligently sorted to generate an accurate push content sequence.

2. The large model information push method based on industry corpus according to claim 1 is characterized in that: The step of obtaining the target user's positive behavior record sequence and negative behavior record sequence includes: Construct an industry corpus, and obtain first historical behavior data of the target user in the industry field from the industry corpus; the first historical behavior data includes n1 behavior types of the target user, and the n1 behavior types include positive behaviors and negative behaviors; n1 is a positive integer; According to the first historical behavior data, a positive behavior record sequence and a negative behavior record sequence are obtained.

3. The large model information push method based on industry corpus according to claim 1 is characterized in that: Extracting the first behavior feature from the positive behavior record sequence includes: Perform word segmentation on the text content in the positive behavior record sequence, and divide the text content into a word sequence of the smallest semantic unit; Perform part-of-speech tagging on the word sequence after word segmentation; Identify named entities in text content based on part-of-speech tagging; Extract keywords and subject terms from word sequences; Keywords, subject terms and named entity words are used as the first line features.

4. The large model information push method based on industry corpus according to claim 2 is characterized in that: The generating of the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature comprises: Set the basic weights of different behavior types among n1 behavior types and construct the initial behavior weight matrix; According to the first historical behavior data, statistically analyzing the behavior intensity distribution of different behavior types; Constructing a behavior intensity quantification function, and mapping the behavior intensity distribution of different behavior types into weight adjustment coefficients of corresponding behavior types according to the behavior intensity quantification function; Multiply the basic weights in the initial behavior weight matrix by the corresponding weight adjustment coefficient to generate an adjusted behavior weight matrix; In the adjusted behavior weight matrix, the weight corresponding to the positive behavior is used as the first behavior weight of the first behavior feature, and the weight corresponding to the negative behavior is used as the second behavior weight of the second behavior feature.

5. The large model information push method based on industry corpus according to claim 1 is characterized in that: Generating a multi-dimensional user portrait vector includes: According to the first behavior characteristics and the first behavior weight, construct the positive preference dimension; According to the second behavior characteristics and the second behavior weight, the negative preference dimension is constructed; The positive preference dimension and the negative preference dimension are combined to generate a multi-dimensional user portrait vector.

6. The large model information push method based on industry corpus according to claim 1 is characterized in that: Generating a positive preference embedding vector and a negative preference embedding vector comprises: Calculate the similarity between the first semantic embedding vector and the second semantic embedding vector to obtain a positive and negative preference similarity score D1; If the positive and negative preference similarity score D1 is greater than θ1, the embedding vector adjustment process is triggered to obtain the positive preference embedding vector and the negative preference embedding vector.

7. The large model information push method based on industry corpus according to claim 6 is characterized in that: The embedding vector adjustment process includes: Step S2221, randomly sampling a noise vector, adjusting the first semantic embedding vector according to the noise vector, obtaining an adjusted first semantic embedding vector, and marking it as a positive preference embedding vector; Step S2222, adjusting the second semantic embedding vector according to the noise vector to obtain an adjusted second semantic embedding vector, which is marked as a negative preference embedding vector; Step S2223, calculate the similarity D2 between the positive preference embedding vector and the negative preference embedding vector, and determine whether D2 is less than or equal to θ1. If so, complete the embedding vector adjustment process; otherwise, return to step S2221, re-randomly sample the noise vector, and repeat steps S2221-S2223 until D2≤θ1.

8. The large model information push method based on industry corpus according to claim 2 is characterized in that: The adjusted push content set includes: Use the positive preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus to generate the first candidate push content set; Use the negative preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus to generate a second set of candidate push content; Based on the first candidate push content set and the second candidate push content set, an adjusted push content set is obtained; Generating the first candidate push content set includes: Convert the text content in the industry corpus into corresponding semantic embedding vectors; Calculate the similarity between the positive preference embedding vector in the multi-dimensional user portrait vector and the semantic embedding vector of the text content in the industry corpus to obtain a first similarity score; The text contents with the first similarity scores in the top K in the industry corpus are screened to generate the first set of candidate push contents.

9. The large model information push method based on industry corpus according to claim 8 is characterized in that: The adjusting the push content set based on the first candidate push content set and the second candidate push content set comprises: Obtaining semantic embedding vectors of text content in the first candidate push content set and the second candidate push content set; According to the semantic embedding vectors of the text contents in the first candidate push content set and the second candidate push content set, the semantic similarity S2 between each text content in the first candidate push content set and all text contents in the second candidate push content set is calculated to form a similarity matrix; In the similarity matrix, content pairs whose semantic similarity S2 is higher than the second similarity threshold θ2 are selected and marked as highly overlapping content; The marked highly overlapping contents are removed from the first candidate push content set to obtain an adjusted push content set.

10. A large model information push system based on industry corpus, which is used to implement the large model information push method based on industry corpus according to any one of claims 1 to 9, characterized in that: The system comprises: Portrait construction module: used to obtain the positive behavior record sequence and negative behavior record sequence of the target user; extract the first behavior feature from the positive behavior record sequence and extract the second behavior feature from the negative behavior record sequence; generate the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature; generate a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature and the second behavior weight; Push content generation module: used to build a pre-trained industry language model, obtain a first semantic embedding vector and a second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjust the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtain an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector; Push content sorting module: based on the adjusted push content set, obtain a third candidate push content set; intelligently sort the contents in the third candidate push content set to generate an accurate push content sequence.

Images