A financial content recommendation method, system, storage medium and electronic device

Abstract

This application provides a financial content recommendation method, comprising: identifying articles to be recommended and financial instruments of interest to target users; calculating the explanatory power of anomalies in the articles to be recommended and related information on price fluctuations; calculating the content-positive or content-negative degree of the articles to be recommended for financial assets; determining the investment advisory value of the financial instruments based on the explanatory power of anomalies, the content-positive degree, and the content-negative degree; calculating the content-investment advisory value of the articles to be recommended for target users based on the investment advisory value of each financial instrument contained in the articles to be recommended; and determining the financial recommended content based on the content-investment advisory value corresponding to each article to be recommended. This application introduces investment advisory value as a factor into the financial content recommendation algorithm to achieve accurate delivery of financial content and improve user satisfaction. This application also provides a financial content recommendation system, a computer-readable storage medium, and an electronic device, which have the above-mentioned beneficial effects.

Description

Technical Field

[0001] This application relates to the field of content recommendation, and in particular to a method, system, storage medium, and electronic device for recommending financial content. Background Technology

[0002] For investors, when viewing recommended content, they primarily seek information beneficial to their investment decisions—content with investment advisory value. The investment advisory value of a piece of content mainly consists of two aspects: first, analyzing past and present trends to help investors analyze the reasons for significant price fluctuations in financial instruments; second, analyzing the impact of the latest data, events, and news on the subsequent trends of financial instruments.

[0003] Currently, there is no effective method to quantify the investment advisory value of financial content. Therefore, recommendation algorithms rely on indirect factors such as the popularity, timeliness, and user preferences of the content to be recommended. Consequently, investors do not always see content that is helpful to their investment decisions in content recommendation scenarios. Summary of the Invention

[0004] The purpose of this application is to provide a financial content recommendation method, system, storage medium, and electronic device that incorporates investment advisory value as a factor into the financial content recommendation algorithm to achieve precise delivery of financial content recommendations.

[0005] To address the aforementioned technical problems, this application provides a method for recommending financial content, the specific technical solution of which is as follows:

[0006] Identify the articles to be recommended and the financial targets that the target users are interested in;

[0007] Determine the price fluctuation information of the financial asset, and calculate the explanatory power of the article to be recommended and the price fluctuation information; the explanatory power of the article to be recommended is used to characterize the degree to which the article to be recommended explains the reasons for the price fluctuation of the financial asset.

[0008] Calculate the positive and negative impacts of the recommended articles on financial assets.

[0009] The investment advisory value of the financial target is determined based on the explanatory power of the cause of the fluctuation, the positive impact of the content, and the negative impact of the content.

[0010] The content investment value of the article to be recommended to the target user is calculated based on the investment advisory value of each financial target contained in the article to be recommended.

[0011] The financial recommendation content is determined based on the investment advisory value corresponding to each of the articles to be recommended.

[0012] Optionally, the process of identifying the financial targets that the target users are interested in includes:

[0013] The level of attention that target users pay to financial targets is determined by the average monthly page view frequency, average monthly click volume, and average monthly page view duration of the financial target's details page;

[0014] The financial targets that the target users are interested in are determined based on the level of attention mentioned.

[0015] Optionally, determining the price fluctuation information of the financial target and calculating the explanatory power of the article to be recommended and the price fluctuation information includes:

[0016] Determine the set of relevant content for the financial target;

[0017] The degree of correlation between the price fluctuations and the cause of the latest abnormal event is determined by applying a language model to the prompt words; the degree of correlation includes high correlation, low correlation, and no correlation.

[0018] The original text portions containing all the highly relevant, low-relevance, and unrelated information are deduplicated and summarized to obtain highly relevant extracted text, low-relevance extracted text, and unrelated extracted text.

[0019] The first semantic similarity S1, the second semantic similarity S2, and the third semantic similarity S3 between the article to be recommended and the highly relevant extracted text, the low relevant extracted text, and the unrelated extracted text are calculated using a vector similarity algorithm.

[0020] The first weight coefficient w1 of the first semantic similarity S1, the second weight coefficient w2 of the second semantic similarity S2, and the third weight coefficient w3 of the third semantic similarity S3 are determined respectively.

[0021] The explanatory power of the article to be recommended and the information related to the rise and fall is calculated based on the product of the first semantic similarity S1 and the first weight coefficient w1, the product of the second semantic similarity S2 and the second weight coefficient w2, and the third semantic similarity S3 and the third weight coefficient w3.

[0022] Optionally, the calculation of the content-positive and content-negative impact of the article to be recommended on financial assets includes:

[0023] For each piece of financial content in the aforementioned set of related content, detect the original text content that is favorable or unfavorable to the financial target.

[0024] The second language model is used to determine the content benefit or content negativity of each original text.

[0025] Optionally, the training process of the second language model includes:

[0026] Construct all financial articles for a selected financial target within a set time period;

[0027] Using prompt words, an initial model is applied to extract suspected original text content from the financial articles that is related to the rise and fall of the financial target.

[0028] The suspected original text content was used as the training set.

[0029] The training set is filtered based on the price fluctuations of the selected financial instruments after the set time period to remove suspected original content that contradicts the price fluctuations.

[0030] The initial model is trained based on the training set to obtain the second language model.

[0031] Optionally, after filtering the training set based on the price fluctuations of the selected financial instrument after the set time period to remove similar original text content that contradicts the price fluctuations, the method further includes:

[0032] The initial model is used to determine the target original text content in the training set that can explain the price fluctuations;

[0033] The target original text content is used to construct the final training set.

[0034] Optionally, calculating the content advisory value of the article to be recommended to the target user based on the advisory value of each financial asset contained in the article to be recommended includes:

[0035] The investment advisory value calculation coefficient for each financial target is determined based on the degree of attention the target users pay to each financial target.

[0036] The content advisory value of the article to be recommended to the target user is determined based on the investment advisory value calculation coefficient and the corresponding investment advisory value of each financial target included in the article to be recommended.

[0037] This application also provides a financial content recommendation system, including:

[0038] The object determination module is used to determine the articles to be recommended and the financial targets that the target users are interested in;

[0039] The anomaly explanation degree calculation module is used to determine the rise and fall information related to the financial target, and calculate the anomaly explanation degree between the article to be recommended and the rise and fall information; the anomaly explanation degree is used to characterize the degree to which the article to be recommended explains the reasons for the rise and fall of financial assets;

[0040] The content relevance calculation module is used to calculate the content positivity and stigmatization of the recommended articles for financial assets.

[0041] The investment advisory value calculation module is used to determine the investment advisory value of the financial target based on the explanatory power of the cause of the anomaly, the positive impact of the content, and the negative impact of the content.

[0042] The content recommendation module is used to calculate the content investment value of the article to be recommended to the target user based on the investment value of each financial target contained in the article to be recommended, and to determine the financial recommendation content based on the content investment value corresponding to each article to be recommended.

[0043] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0044] This application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described above when it invokes the computer program in the memory.

[0045] This application provides a financial content recommendation method, comprising: determining articles to be recommended and financial targets of interest to target users; determining the price fluctuation information of the financial targets, and calculating the explanatory power of the anomalies between the articles to be recommended and the price fluctuation information; the explanatory power of the anomalies characterizes the degree to which the articles to be recommended explain the reasons for the price fluctuations of financial assets; calculating the positive and negative content of the articles to be recommended for the financial assets; determining the investment advisory value of the financial targets based on the explanatory power of the anomalies, the positive content, and the negative content; calculating the investment advisory value of the articles to be recommended for the target users based on the investment advisory value of each financial target included in the articles to be recommended; and determining the financial recommended content based on the investment advisory value corresponding to each article to be recommended.

[0046] This application quantifies the investment advisory value of recommended articles for financial targets by calculating the explanatory power of the price fluctuations of financial targets, the explanatory power of the anomalies in the recommended articles and their related price fluctuations, as well as the positive and negative aspects of the content. Furthermore, it determines the investment advisory value of the recommended articles for the target user based on all financial targets included in the articles. By first quantifying the investment advisory value of the recommended articles for different financial targets, and then considering the investment advisory value for the target user, this application incorporates investment advisory value as a factor into the financial content recommendation algorithm, achieving precise delivery of financial content and improving user satisfaction.

[0047] This application also provides a financial content recommendation system, a computer-readable storage medium, and an electronic device, which have the aforementioned beneficial effects, and will not be elaborated here. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0049] Figure 1 A flowchart illustrating a financial content recommendation method provided in this application embodiment;

[0050] Figure 2 This is a schematic diagram of the financial content recommendation system provided in the embodiments of this application;

[0051] Figure 3 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0053] See Figure 1 , Figure 1 A flowchart of a financial content recommendation method provided in this application embodiment, the method including:

[0054] S101: Identify the articles to be recommended and the financial targets that the target users are interested in;

[0055] S102: Determine the price fluctuation information of the financial target, and calculate the explanatory power of the article to be recommended and the price fluctuation information; the explanatory power of the article to be recommended is used to characterize the degree to which the article to be recommended explains the reasons for the price fluctuation of the financial asset.

[0056] S103: Calculate the content-positive and content-negative impact of the recommended articles on financial assets;

[0057] S104: Determine the investment advisory value of the financial target based on the explanatory power of the cause of the fluctuation, the positive impact of the content, and the negative impact of the content;

[0058] S105: Calculate the content investment value of the article to be recommended to the target user based on the investment value of each financial target contained in the article to be recommended;

[0059] S106: Determine the financial recommendation content based on the investment advisory value corresponding to each of the articles to be recommended.

[0060] This document does not specify how to determine the financial targets that the target users are interested in.

[0061] In one feasible approach, the level of attention a target user pays to a financial target can be determined based on the average monthly browsing frequency, average monthly clicks, and average monthly browsing duration of the financial target's details page, and then the financial targets that the target user is interested in can be determined based on the level of attention.

[0062] This involves obtaining a list of financial targets that the target user is interested in. The data format is like this: {Company A: 0.99, Company B: 0.45…}, where the numbers represent the level of interest. A higher level of interest indicates greater user interest in that financial target. The logic for calculating the level of interest is as follows:

[0063] Attention = A1 * Average monthly pageview frequency of the target details page (time-based page) * Average monthly dwell time of the target details page (time-based page) + A2 * Average monthly click-through rate of related content pages of the target * Average monthly dwell time of related content pages of the target.

[0064] The default parameters A1=A2=1, but the specific values ​​of A1 and A2 do not affect the calculation results. If the target user does not have relevant following behavior, currently popular stocks can be used as the user's followed stocks, and an equal level of attention can be set for each stock.

[0065] The financial underlying asset described in this application refers to the asset stipulated in the financial derivatives contract, which may be a physical commodity, financial asset, interest rate, exchange rate, or various composite price indexes, etc.

[0066] Financial underlying assets, as the foundation of financial derivatives, are diverse and their trading methods complex. Understanding and mastering the attributes of these financial underlying assets is crucial for investors in financial markets, as they directly relate to investment returns and risk management. Broadly speaking, financial underlying assets can be physical assets, such as crude oil, wheat, and copper, or financial assets, such as foreign exchange, bank deposits, stocks, and securities. Furthermore, they can also be abstract financial indicators, including interest rates, exchange rates, and related content from various composite price indices. For a news article or commentary, NLP techniques can be used to identify the financial underlying assets mentioned. If a news article mentions a particular financial underlying asset, it can be considered relevant information about that financial underlying asset. Additionally, algorithms (such as TF-IDF) can be used to calculate the correlation between news and the financial underlying asset; only when the correlation exceeds a set threshold is it considered relevant information about that financial underlying asset.

[0067] In step S102, it is necessary to determine the explanatory power of the abnormal fluctuations. This is mainly achieved by summarizing information related to the rise and fall of a certain financial asset over a period of time using a language model. For example, this information can be divided into three categories: highly relevant, low relevant, and unrelevant. Then, the semantic similarity between the article to be recommended and these three categories of content is calculated. Finally, the degree to which the article to be recommended explains the rise and fall of the financial asset is calculated according to the formula, i.e., the explanatory power of the abnormal fluctuations.

[0068] One feasible implementation method is as follows:

[0069] The first step is to determine the set of relevant content for the financial target;

[0070] The second step is to use a language model to determine the degree of correlation between the price fluctuations and the cause of the latest abnormal event by applying the prompt words; the degree of correlation includes high correlation, low correlation, and no correlation.

[0071] The third step is to remove duplicates from all the highly relevant, low-relevance, and unrelated original text portions and summarize them to obtain highly relevant extracted text, low-relevance extracted text, and unrelated extracted text.

[0072] Step 4: Calculate the first semantic similarity S1, the second semantic similarity S2, and the third semantic similarity S3 between the article to be recommended and the highly relevant extracted text, the low relevant extracted text, and the unrelated extracted text using a vector similarity algorithm.

[0073] Step 5: Determine the first weight coefficient w1 of the first semantic similarity S1, the second weight coefficient w2 of the second semantic similarity S2, and the third weight coefficient w3 of the third semantic similarity S3 respectively;

[0074] Step 6: Calculate the explanatory power of the anomaly between the article to be recommended and the information related to the rise and fall based on the product of the first semantic similarity S1 and the first weight coefficient w1, the product of the second semantic similarity S2 and the second weight coefficient w2, the third semantic similarity S3 and the third weight coefficient w3.

[0075] This retrieves a set of asset-related contents within a specific timeframe for the financial asset, including news from authoritative media outlets and opinions expressed by users in online communities. The default timeframe is the past 7 days from today. However, different parameters can be used in different practices. If the target audience is primarily short-term investors, the timeframe can be shorter, as they mainly consider the impact of short-term events and news on the financial asset. Conversely, if the target audience is medium- to long-term investors, the timeframe can be longer, such as the past 3 months or even the past year.

[0076] Using a language model, based on the acquired relevant content set, we summarize three categories of content over a certain period: highly correlated, low correlated, and uncorrelated with the latest abnormal events of the financial target. A language model is a pre-trained language model that learns the complex structure and rules of language by pre-training a Transformer model on a large-scale corpus. Prompting is a way to interact with the language model; by providing the model with a series of input examples (i.e., prompts), it can guide the model to generate the desired output. This method allows the model to utilize its internal knowledge to understand and generate natural language.

[0077] Therefore, it's necessary to determine what recent unusual events have occurred in this financial product. Specifically, this can be achieved by setting significant change rules to define what constitutes a significant change. For example, if focusing on the A-share market, a significant price movement could be defined as follows: if a stock's market capitalization is less than 50 billion yuan and its price rises by more than 7% within a single day, or if its market capitalization exceeds 50 billion yuan and its price rises by more than 5%. Then, based on these rules, it's possible to periodically query which financial assets have triggered these rules, thereby recording the latest unusual events for each asset. Finally, for each piece of content in the asset-related contents set for that financial asset, a language model is used with prompts to determine which parts of the original text are highly correlated with the cause of the latest unusual event, which are lowly correlated, and which are irrelevant. In practice, the language model and prompts used do not have strict formal requirements, as long as they achieve the purpose of content extraction. For example, one prompt could be: "Please determine the extent to which the following content can be considered a reason for the recent continuous rise in Foxconn Industrial Internet's stock price (highly relevant, lowly relevant, irrelevant), and extract the corresponding parts of the original text." Finally, the language model is used to deduplicate and summarize all the highly relevant, lowly relevant, and irrelevant parts of the original text, resulting in extracted text categorized into three types: highly relevant, lowly relevant, and irrelevant. Again, in practice, the language model and prompts used do not have strict formal requirements. To improve the accuracy of the language model in this task, fine-tuning can be used to optimize the performance of the large model. To continuously train the large model and improve its accuracy in this extraction task, search engines are periodically used to query content related to the analysis of the causes of rises, falls, and anomalies in financial stocks.

[0078] After cleaning, the texts were used as the training set for highly relevant texts. Other types of articles were used as the training set for low-relevance and irrelevant texts.

[0079] The main text of the content to be recommended is obtained. Using a vector similarity algorithm, the semantic similarity S1, S2, and S3 are calculated between the content to be recommended and the extracted texts of three categories of content: highly relevant, low relevant, and irrelevant. Vector generation can employ common NLP encoder models such as BERT. Then, cosine distance is used to measure the semantic similarity between the vector of the content to be recommended and the extracted texts of the three categories. Finally, the explanatory power E1 of the content to be recommended for the cause of the financial instrument's fluctuation is calculated.

[0080] E1 = w1 × S1 + w2 × S2 - w3 × S3;

[0081] In the above formula, w1, w2, and w3 are weighting coefficients. No specific limitations are imposed on these coefficients; they can be set by those skilled in the art. By default, w1=w3=2 and w2=1. The specific values ​​of the weights have no impact on the calculation results. The weights can be set as needed in different practices. If you want the algorithm to focus more on highly relevant content and exclude low-relevance or irrelevant content, you can increase the values ​​of w1 and w3 and decrease the value of w2. The logic behind this algorithm is that the more comprehensive and relevant the content related to the anomaly in the main text of a piece of content to be recommended, and the less irrelevant content it contains, the better it can explain the latest anomaly event of that financial asset.

[0082] To calculate the positive and negative content, we can first examine each piece of financial content in the relevant content set to detect whether the original text content is positive or negative for the financial target. Then, we can use a second language model to determine the positive or negative content content of each piece of original text content.

[0083] Obtain a collection of asset-related contents within a specific time frame, including relevant information or community comments about the asset.

[0084] For each piece of content in the relevant content set of the financial target, the large language model is guided by prompt words to identify the original text portions that are strongly positive, moderately positive, strongly negative, moderately negative, or neutral for the target. It should be noted that in the actual application of this application, the large language model and prompt words do not have strict formal requirements, as long as they achieve the purpose of content extraction. For example, one prompt word could be: Please determine which parts of the following original text have a strong positive, moderately positive, neutral, moderately negative, or strongly negative impact on a certain bank, respectively.

[0085] To improve the accuracy of the large language model in this task, fine-tuning can be used to optimize its performance. To this end, an automated training set generation system will be built to obtain training data and continuously improve the large language model's performance in this extraction task. The basic principle is that for a labeled training set containing bullish and bearish information for a specific stock generated on a given day, the system will automatically filter the data based on the actual price movement of that stock over the following days. The specific solution is as follows:

[0086] The first step is to construct all financial articles related to the selected financial target within a set time period;

[0087] The second step is to use prompt words and an initial model to extract suspected original text content from the financial articles that is related to the rise and fall of the financial target.

[0088] The third step is to use the suspected original text content as a training set.

[0089] Step 4: Filter the training set based on the price fluctuations of the selected financial target after the set time period to remove suspected original content that contradicts the price fluctuations.

[0090] Step 5: Train the initial model based on the training set to obtain the second language model.

[0091] When building the training set, we first collect all relevant articles about a specific financial instrument on a particular day. Then, using prompt words, the large language model identifies the parts of these articles that may affect stock price movements (impact is categorized into five types: strongly positive, moderately positive, neutral, moderately negative, and strongly negative). These parts are temporarily used as the training set.

[0092] The training set is then filtered based on actual price movements, specifically the price fluctuations of the selected financial instrument over a set time period, to remove potentially problematic content that contradicts these price movements. For example, let's examine the performance of this financial instrument over the next five days. If the stock price rises significantly (over 7%), the sections previously considered strongly positive are retained as part of the training set. If the stock price rises only slightly (2% to 7%), the generally positive sections are retained. If the stock price changes little (between -2% and 2%), the neutral sections are retained. If the stock price falls somewhat (-2% to -7%), the generally negative sections are retained. If the stock price falls significantly (over -7%), the strongly negative sections are retained.

[0093] Furthermore, the training set can be further validated. Specifically, the initial model can be used to determine the target text content in the training set that can explain the stock price fluctuations, and this target text content can be used to construct the final training set. For example, based on the stock's performance five days later, articles analyzing the reasons for the stock's rise or fall can be searched. If the stock price rises by more than 2%, the reason for the rise is determined; if it falls by more than 2%, the reason for the fall is determined. Next, prompt words are used to allow the second language model to re-examine the target text content previously considered positive or negative to see if it truly matches the found reason analyses. If it matches, this target text content becomes the final training set; if it does not match, it is removed from the training set.

[0094] The original text, categorized as strongly positive, moderately positive, strongly negative, moderately negative, and neutral, is deduplicated and summarized using a second language model, resulting in five categories of extracted text: strongly positive, moderately positive, strongly negative, moderately negative, and neutral. Similar to the previous step, in practice, there are no strict formal requirements for the large language model and prompt words used.

[0095] The main text of the content to be recommended is obtained. Using a vector similarity algorithm, the semantic similarity (S4, S5, S6, S7, S8) is calculated between the content to be recommended and the extracted texts categorized into five types: strongly positive, moderately positive, strongly negative, moderately negative, and neutral. Vector generation can employ common NLP encoder models such as BERT. Then, cosine distance is used to measure the semantic similarity between the vector of the content to be recommended and the extracted texts of each of the five types. Finally, the positive (E2) and negative (E3) content scores for the financial instrument are calculated.

[0096] E2 = w4 × S4 + w5 × S5 - w8 × S8;

[0097] E3 = w6 × S6 + w7 × S7 - w8 × S8;

[0098] In the above formula, w4, w5, w6, w7, and w8 are used as weighting coefficients. No specific limitations are imposed on these weighting coefficients; they can be set by those skilled in the art. The default values ​​are w4=w8=w6=2 and w7=w5=1. The specific values ​​of these weighting coefficients have no impact on the calculation results. In different practices, these values ​​can be set as needed. If you want the algorithm to focus more on strongly positive or negative content and exclude neutral content, you can increase the values ​​of w4, w6, and w8, and decrease the values ​​of w5 and w7. The logic behind this algorithm is that the more comprehensive and comprehensive the positive (or negative) content in the main text of a piece of content to be recommended, and the fewer neutral content it contains, the stronger its effect on analyzing the subsequent trend of the target stock.

[0099] Subsequently, the specific details of the method for calculating the investment advisory value (VFA) of the content to be recommended for the financial asset are as follows: After calculating the explanatory power (E1), the positive impact (E2), and the negative impact (E3) of the content to be recommended for the financial asset's fluctuations using the above method, the formula for calculating the investment advisory value (VFA) is as follows:

[0100] VFA = w9×E1 + w10×E2 - w11×E3;

[0101] In the above formula, w9, w10, and w11 are weighting coefficients. No specific restrictions are placed on these coefficients; they can be set by those skilled in the art. By default, w9 = w10 = w11, and the specific values ​​of w9, w10, and w11 have no impact on the calculation results.

[0102] Considering that target users may be interested in multiple financial instruments, it is necessary to combine the Value at Advisory Value (VFA) of a proposed content with all financial instruments to quantify the investment advisory value of that content for the target user. The specific details of the method for calculating the Value at Advisory Value (VFU) of the proposed content for the target user are as follows:

[0103] The investment advisory value calculation coefficient for each financial target is determined based on the degree of attention the target users pay to each financial target.

[0104] The content advisory value of the article to be recommended to the target user is determined based on the investment advisory value calculation coefficient and the corresponding investment advisory value of each financial target included in the article to be recommended.

[0105] Calculate the Content Investment Value (VFU) of the content to be recommended relative to the target user. The Content Investment Value (VFU) is equal to the weighted sum of the Investment Values ​​(VFA) of all financial instruments followed by the target user. The weights used can be derived from the target user's level of interest in each financial instrument. That is, the more financial instruments a piece of content has high investment value for a given user, the higher its investment value for that user.

[0106] Finally, the financial recommendations can be determined based on the content investment value of each article to be recommended. At this point, a click-through rate (CTR) prediction model (DNN, LR, XgBoost, etc.) can be constructed, incorporating content investment value, timeliness, and user personalization characteristics (such as demographic attributes, investment risk preferences, industry preferences, etc.) as features of the CTR prediction model. The predicted CTR is then used as the final recommendation score to rank the content.

[0107] This application embodiment calculates the investment advisory value of the recommended article for the financial target by calculating the price fluctuation information of the financial target, the explanatory power of the abnormal fluctuation of the recommended article and the price fluctuation information, as well as the content's positive and negative aspects. Then, based on all the financial targets included in the recommended article, the content investment advisory value of the recommended article for the target user is determined. First, the investment advisory value of the recommended article for different financial targets is quantified, and then the content investment advisory value of the recommended article for the target user is considered. The investment advisory value is introduced as a factor into the financial content recommendation algorithm to achieve accurate delivery of financial content and improve user satisfaction.

[0108] See Figure 2 , Figure 2 This is a schematic diagram of the structure of a financial content recommendation system provided in an embodiment of this application. The system includes:

[0109] The object determination module is used to determine the articles to be recommended and the financial targets that the target users are interested in;

[0110] The anomaly explanation degree calculation module is used to determine the rise and fall information related to the financial target, and calculate the anomaly explanation degree between the article to be recommended and the rise and fall information; the anomaly explanation degree is used to characterize the degree to which the article to be recommended explains the reasons for the rise and fall of financial assets;

[0111] The content relevance calculation module is used to calculate the content positivity and stigmatization of the recommended articles for financial assets.

[0112] The investment advisory value calculation module is used to determine the investment advisory value of the financial target based on the explanatory power of the cause of the anomaly, the positive impact of the content, and the negative impact of the content.

[0113] The content recommendation module is used to calculate the content investment value of the article to be recommended to the target user based on the investment value of each financial target contained in the article to be recommended, and to determine the financial recommendation content based on the content investment value corresponding to each article to be recommended.

[0114] Based on the above embodiments, as a preferred embodiment, the object determination module includes:

[0115] The first determining unit is used to determine the target user's level of attention to the financial target based on the average monthly browsing frequency, average monthly click volume, and average monthly browsing duration of the financial target's details page; and to determine the financial target that the target user is paying attention to based on the level of attention.

[0116] Based on the above embodiments, as a preferred embodiment, the anomaly cause explanatory power calculation module is a module for performing the following steps:

[0117] Determine the set of relevant content for the financial target;

[0118] The degree of correlation between the price fluctuations and the cause of the latest abnormal event is determined by applying a language model to the prompt words; the degree of correlation includes high correlation, low correlation, and no correlation.

[0119] The original text portions containing all the highly relevant, low-relevance, and unrelated information are deduplicated and summarized to obtain highly relevant extracted text, low-relevance extracted text, and unrelated extracted text.

[0120] The first semantic similarity S1, the second semantic similarity S2, and the third semantic similarity S3 between the article to be recommended and the highly relevant extracted text, the low relevant extracted text, and the unrelated extracted text are calculated using a vector similarity algorithm.

[0121] The first weight coefficient w1 of the first semantic similarity S1, the second weight coefficient w2 of the second semantic similarity S2, and the third weight coefficient w3 of the third semantic similarity S3 are determined respectively.

[0122] The explanatory power of the article to be recommended and the information related to the rise and fall is calculated based on the product of the first semantic similarity S1 and the first weight coefficient w1, the product of the second semantic similarity S2 and the second weight coefficient w2, and the third semantic similarity S3 and the third weight coefficient w3.

[0123] Based on the above embodiments, as a preferred embodiment, the content relevance calculation module is a module for performing the following steps:

[0124] For each piece of financial content in the aforementioned set of related content, detect the original text content that is favorable or unfavorable to the financial target.

[0125] The second language model is used to determine the content benefit or content negativity of each original text.

[0126] Based on the above embodiments, as a preferred embodiment, the training process of the second language model includes:

[0127] Construct all financial articles for a selected financial target within a set time period;

[0128] Using prompt words, an initial model is applied to extract suspected original text content from the financial articles that is related to the rise and fall of the financial target.

[0129] The suspected original text content was used as the training set.

[0130] The training set is filtered based on the price fluctuations of the selected financial instruments after the set time period to remove suspected original content that contradicts the price fluctuations.

[0131] The initial model is trained based on the training set to obtain the second language model.

[0132] Based on the above embodiments, as a preferred embodiment, it further includes:

[0133] The training set generation module is used to determine the target original text content in the training set that can explain the rise and fall performance using the initial model; and to construct the final training set using the target original text content.

[0134] Based on the above embodiments, as a preferred embodiment, the investment advisory value calculation module includes:

[0135] The coefficient calculation unit is used to determine the investment advisory value calculation coefficient of each financial target based on the target user's level of attention to each financial target.

[0136] The value calculation unit is used to determine the content investment value of the article to be recommended to the target user based on the investment value calculation coefficient and the corresponding investment value of each financial target contained in the article to be recommended.

[0137] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed, can perform the steps provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0138] This application also provides an electronic device, see [link to document]. Figure 3 The present application provides a structural diagram of an electronic device, as shown in the embodiment. Figure 3 As shown, it may include a processor 1410 and a memory 1420.

[0139] The processor 1410 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 1410 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 1410 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 1410 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 1410 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0140] The memory 1420 may include one or more computer-readable storage media, which may be non-transitory. The memory 1420 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In this embodiment, the memory 1420 is used to store at least the following computer program 1421, which, after being loaded and executed by the processor 1410, is capable of implementing the relevant steps in the methods executed by the electronic device side as disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 1420 may also include an operating system 1422 and data 1423, etc., and the storage method may be temporary storage or permanent storage. The operating system 1422 may include Windows, Linux, Android, etc.

[0141] In some embodiments, the electronic device may further include a display screen 1430, an input / output interface 1440, a communication interface 1450, a sensor 1460, a power supply 1470, and a communication bus 1480.

[0142] certainly, Figure 3 The structure of the electronic device shown does not constitute a limitation on the electronic device in the embodiments of this application. In practical applications, the electronic device may include more than [other components]. Figure 3 More or fewer components as shown, or combinations of certain components.

[0143] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. As the system provided in the embodiments corresponds to the method provided in the embodiments, the description is relatively simple; relevant parts can be found in the method section.

[0144] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

[0145] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for recommending financial content, characterized in that, include: Identify the articles to be recommended and the financial targets that the target users are interested in; Determine the price fluctuation information of the financial target, and calculate the explanatory power of the article to be recommended and the price fluctuation information. The explanation of the reasons for the anomaly is used to characterize the degree to which the article to be recommended explains the reasons for the rise and fall of financial assets; Calculate the positive and negative impacts of the recommended articles on financial assets. The investment advisory value of the financial target is determined based on the explanatory power of the cause of the fluctuation, the positive impact of the content, and the negative impact of the content. The content investment value of the article to be recommended to the target user is calculated based on the investment advisory value of each financial target contained in the article to be recommended. The financial recommendation content is determined based on the investment advisory value corresponding to each of the articles to be recommended. The determination of the price fluctuation information of the financial target and the calculation of the explanatory power of the anomalies in the article to be recommended and the price fluctuation information include: Determine the set of relevant content for the financial target; The degree of correlation between the price fluctuations and the cause of the latest abnormal event is determined by applying a language model to the prompt words; the degree of correlation includes high correlation, low correlation, and no correlation. The original text portions containing all the highly relevant, low-relevance, and unrelated information are deduplicated and summarized to obtain highly relevant extracted text, low-relevance extracted text, and unrelated extracted text. The first semantic similarity S1, the second semantic similarity S2, and the third semantic similarity S3 between the article to be recommended and the highly relevant extracted text, the low relevant extracted text, and the unrelated extracted text are calculated using a vector similarity algorithm. The first weight coefficient w1 of the first semantic similarity S1, the second weight coefficient w2 of the second semantic similarity S2, and the third weight coefficient w3 of the third semantic similarity S3 are determined respectively. The explanatory power of the anomaly of the article to be recommended and the information related to the rise and fall is calculated based on the product of the first semantic similarity S1 and the first weight coefficient w1, the product of the second semantic similarity S2 and the second weight coefficient w2, and the third semantic similarity S3 and the third weight coefficient w3. The calculation of the content-positive and content-negative impact of the recommended article on financial assets includes: For each piece of financial content in the aforementioned set of related content, detect the original text content that is favorable or unfavorable to the financial target. The second language model is used to determine the content benefit or content negativity of each original text.

2. The financial content recommendation method according to claim 1, characterized in that, The process of identifying the financial targets that the target users are interested in includes: The level of attention that target users pay to financial targets is determined by the average monthly page view frequency, average monthly click volume, and average monthly page view duration of the financial target's details page; The financial targets that the target users are interested in are determined based on the level of attention mentioned.

3. The financial content recommendation method according to claim 1, characterized in that, The training process of the second language model includes: Construct all financial articles for a selected financial target within a set time period; Using prompt words, an initial model is applied to extract suspected original text content from the financial articles that is related to the rise and fall of the financial target. The suspected original text content was used as the training set. The training set is filtered based on the price fluctuations of the selected financial instruments after the set time period to remove suspected original content that contradicts the price fluctuations. The initial model is trained based on the training set to obtain the second language model.

4. The financial content recommendation method according to claim 3, characterized in that, After filtering the training set based on the price fluctuations of the selected financial instrument after the set time period to remove similar original text content that contradicts the price fluctuations, the process further includes: The initial model is used to determine the target original text content in the training set that can explain the price fluctuations; The target original text content is used to construct the final training set.

5. The financial content recommendation method according to claim 2, characterized in that, The content advisory value of the article to be recommended to the target user is calculated based on the advisory value of each financial instrument contained in the article to be recommended, including: The investment advisory value calculation coefficient for each financial target is determined based on the degree of attention the target users pay to each financial target. The content advisory value of the article to be recommended to the target user is determined based on the investment advisory value calculation coefficient and the corresponding investment advisory value of each financial target included in the article to be recommended.

6. A financial content recommendation system, characterized in that, include: The object determination module is used to determine the articles to be recommended and the financial targets that the target users are interested in; The anomaly explanation degree calculation module is used to determine the rise and fall related information of the financial target and calculate the anomaly explanation degree of the article to be recommended and the rise and fall related information. The explanation of the reasons for the anomaly is used to characterize the degree to which the article to be recommended explains the reasons for the rise and fall of financial assets; The content relevance calculation module is used to calculate the content positivity and stigmatization of the recommended articles for financial assets. The investment advisory value calculation module is used to determine the investment advisory value of the financial target based on the explanatory power of the cause of the anomaly, the positive impact of the content, and the negative impact of the content. The content recommendation module is used to calculate the content investment value of the article to be recommended to the target user based on the investment value of each financial target contained in the article to be recommended, and to determine the financial recommendation content based on the content investment value corresponding to each article to be recommended. The anomaly cause explanation degree calculation module is a module used to perform the following steps: Determine the set of relevant content for the financial target; The degree of correlation between the price fluctuations and the cause of the latest abnormal event is determined by applying a language model to the prompt words; the degree of correlation includes high correlation, low correlation, and no correlation. The original text portions containing all the highly relevant, low-relevance, and unrelated information are deduplicated and summarized to obtain highly relevant extracted text, low-relevance extracted text, and unrelated extracted text. The first semantic similarity S1, the second semantic similarity S2, and the third semantic similarity S3 between the article to be recommended and the highly relevant extracted text, the low relevant extracted text, and the unrelated extracted text are calculated using a vector similarity algorithm. The first weight coefficient w1 of the first semantic similarity S1, the second weight coefficient w2 of the second semantic similarity S2, and the third weight coefficient w3 of the third semantic similarity S3 are determined respectively. The explanatory power of the anomaly of the article to be recommended and the information related to the rise and fall is calculated based on the product of the first semantic similarity S1 and the first weight coefficient w1, the product of the second semantic similarity S2 and the second weight coefficient w2, and the third semantic similarity S3 and the third weight coefficient w3. The content relevance calculation module is used to perform the following steps: For each piece of financial content in the relevant content set, detect the original text content that is favorable or unfavorable to the financial target; use a second language model to determine the content favorability or unfavorability of each piece of original text content.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the financial content recommendation method as described in any one of claims 1-5.

8. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor, when calling the computer program in the memory, implements the steps of the financial content recommendation method as described in any one of claims 1-5.

Images