Reasonable lottery purchase supervision method and device based on emotion fusion analysis and medium

Abstract

The invention discloses a rational lottery buying supervision method and device based on emotion fusion analysis and a medium. The method comprises the following steps: firstly, acquiring multi-modal data generated in a lottery buying process of a user and current lottery buying behavior data of the user in real time, and preprocessing the multi-modal data and the current lottery buying behavior data of the user; then user real-time emotion analysis is carried out according to the type of the multi-modal data, and finally the tendency probability of the current irrational lottery buying behavior of the user is obtained through evaluation according to the real-time emotion intensity value of the user, the emotion type label, the current lottery buying behavior data of the user and an irrational prediction model pre-constructed by the system. And the irrational lottery buying behavior of the user is intervened according to the result. According to the method and the device, flexible intervention of irrational lottery buying behaviors can be realized, and the problem of poor user experience caused by adoption of a fixed rule in irrational lottery buying behavior intervention in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of big data, and in particular to a method, device, and medium for rational lottery purchase supervision based on sentiment fusion analysis. Background Technology

[0002] With the increasing popularity of online lottery sales platforms, more and more users are buying lottery tickets or placing bets online. Preventing irrational lottery-buying behavior has become an important issue for equipment and technology. Currently, intervention is mainly carried out through static rules, such as setting a daily betting limit or forcibly setting a cooling-off period. However, the rules of this method are fixed and cannot be flexibly set according to the user's specific operating status or risk level. Setting the same rules for all users leads to a poor user experience and unsatisfactory intervention effect. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, one of the objectives of this invention is to provide a rational lottery purchase supervision method based on emotion fusion analysis, which can solve the problems of poor user experience caused by the solidification of intervention measures against irrational lottery purchases in the prior art.

[0004] In order to overcome the shortcomings of the prior art, the second objective of this invention is to provide a rational lottery purchase monitoring device based on emotion fusion analysis, which can solve the problems of poor user experience caused by the solidification of intervention measures against irrational lottery purchases in the prior art.

[0005] To overcome the shortcomings of the prior art, the third objective of this invention is to provide a computer-readable storage medium that can solve the problem of poor user experience caused by the solidification of intervention measures against irrational lottery purchases by users in the prior art.

[0006] One of the objectives of this invention is achieved through the following technical solution:

[0007] A rational lottery purchase supervision method based on sentiment fusion analysis, the rational lottery purchase supervision method includes:

[0008] Data acquisition steps: Real-time acquisition of multimodal data generated during the user's lottery purchase process and the user's current lottery purchase behavior data, and preprocessing of the multimodal data and the user's current lottery purchase behavior data;

[0009] Sentiment analysis steps: Based on the type of the multimodal data, corresponding features are extracted from the multimodal data, and the features of each modality are fused based on hierarchical fusion measurement to obtain a fused feature vector. Based on a pre-trained sentiment intensity prediction model, the user's real-time sentiment intensity value is obtained, and the sentiment type label is determined based on the user's real-time sentiment intensity value.

[0010] Intervention steps: Based on the user's real-time emotional intensity value, emotional type label, the user's current lottery purchase behavior data, and the irrational prediction model pre-built by the system, the probability of the user's current irrational lottery purchase behavior is assessed, and intervention is carried out on the user's irrational lottery purchase behavior based on the user's real-time emotional intensity value, emotional type label, and the probability of the user's current irrational betting behavior.

[0011] Furthermore, the intervention step, which intervenes in the user's irrational betting behavior based on the user's real-time emotional intensity value, emotional type label, and the probability of the user's current irrational betting behavior tendency, specifically includes: selecting corresponding intervention strategies from a pre-built intervention strategy adaptation table in the system based on the user's real-time emotional intensity analysis results and the user's current irrational betting behavior tendency probability, and then intervening in the user's irrational betting behavior based on the corresponding intervention strategy; wherein, the intervention strategy adaptation table includes a matching relationship table between user emotional intensity, user irrational betting behavior tendency probability, and intervention strategy.

[0012] Furthermore, the intervention strategies include an emotional interaction guidance strategy, a cooling-off period order locking mechanism strategy, and a dynamic limit adjustment strategy. The emotional interaction guidance strategy includes dynamically generating a guidance script based on emotional tags, generating an AI virtual persona based on the guidance script and the AI ​​digital human, and having the AI ​​virtual persona send corresponding voice content to the user, as well as entering a user dialogue mode to obtain and respond to the user's dialogue content. The cooling-off period order locking mechanism strategy includes hiding the betting button and locking the order, and displaying a public welfare project promotional video. The dynamic limit adjustment strategy includes locking the betting operation and entering an interactive Q&A mode to obtain the user's answers.

[0013] The intervention steps also include: setting the intervention termination condition for the corresponding strategy and ending the current intervention strategy and entering the normal lottery purchase mode when the corresponding intervention termination condition is triggered.

[0014] Furthermore, the intervention steps also include: firstly, constructing a personalized behavioral benchmark based on the target object's historical purchase data, and then constructing a feature set based on a pre-constructed irrational purchase behavior feature set, user sentiment analysis results, real-time purchase behavior data, and the personalized behavioral benchmark to make a preliminary judgment on the user's current lottery purchase behavior; then, based on the feature set, sentiment intensity value, sentiment type label, preliminary judgment results, and a pre-constructed fusion prediction model, deriving the probability of the user's irrational lottery purchase behavior tendency; wherein, the feature set includes sentiment-behavioral deviation features and behavioral temporal anomaly features;

[0015] Furthermore, the fusion prediction model specifically includes: constructing a training dataset based on historical multimodal sample data, sentiment analysis results, feature sets, and manually labeled irrational judgment tags and tendency probability reference values; training the constructed fusion prediction model to obtain the fusion prediction model; the input layer of the fusion prediction model includes sentiment feature branches, behavioral feature branches, and judgment result branches; the feature enhancement layer includes each branch extracting deep features through an independent Transformer encoder to output a multidimensional feature vector, and then using a multi-head self-attention mechanism to calculate the correlation weights between each branch and weighted fusion to obtain a global feature vector; the prediction layer includes outputting the judgment result through a Sigmoid activation function and outputting the tendency probability value through a Softmax activation function.

[0016] Furthermore, the sentiment analysis step specifically includes extracting corresponding features from the multimodal data according to the type of the multimodal data: firstly, obtaining speech data from the multimodal data and recognizing the speech data to obtain text data, and then extracting text data from the multimodal data, extracting keywords from the text data to obtain sentiment words related to the lottery field, generating word vectors through the Word2Vec model, and extracting deep semantic features of the text by combining the BERT model;

[0017] First, speech data is obtained based on the multimodal data, and speech features are extracted from the speech data. The speech features include fundamental frequency, volume, speech rate, spectral centroid, Mel-spectral cepstral coefficients, pitch change rate, and volume fluctuation amplitude.

[0018] First, visual data is obtained based on the multimodal data, and user facial recognition is performed on the visual data to identify user facial features. Then, video emotional features are obtained based on the user facial features. The video emotional features include facial key points, expression action unit intensity, facial texture features, and eye change frequency features. Facial key points include the position and shape of eyebrows, eyes, and mouth.

[0019] First, user behavior emotional characteristics are derived based on user operation behavior data; these characteristics include operation frequency characteristics, operation rhythm characteristics, and preference characteristics; operation frequency characteristics include average click frequency and swipe frequency; operation rhythm characteristics include input interval standard deviation; and preference characteristics include high-frequency access areas.

[0020] Furthermore, the method of fusing features from various modalities based on a hierarchical fusion strategy to obtain a fused feature vector specifically includes: standardizing the features extracted from each modality, constructing an attention mechanism and calculating the sentiment contribution weight of each modal feature, weighting and summing the features of each modality according to the sentiment contribution weight of each modality feature to obtain a preliminary fused feature, and then inputting the preliminary fused feature into a deep learning model to mine the correlation information between the features of each modality to output the fused feature vector;

[0021] The sentiment intensity prediction model is specifically obtained by training the sentiment intensity prediction model based on the support vector regression model using a constructed sentiment intensity labeled dataset; wherein, the sentiment intensity labeled dataset includes historical multimodal data samples and corresponding manually labeled sentiment intensity values;

[0022] Furthermore, the data acquisition steps specifically include: firstly, determining the user's lottery purchase scenario based on the user's current operational behavior, and then acquiring multimodal data based on the user's lottery purchase scenario and data acquisition rules; wherein, when the user's lottery purchase scenario is a number selection scenario, the multimodal data is derived based on the user's terminal operation behavior data, user operation text data, user consultation voice data, and user verification facial video data; when the user's lottery purchase scenario is a bet confirmation scenario, the multimodal data is derived based on the pressure sensor of the user's terminal device's pressed button, the user's heart rate data, the user's voice data, and the user's video data; when the user's lottery purchase scenario is a prize redemption scenario, the multimodal data is derived based on the user's voice data, video data, and user operation data; and when the lottery purchase scenario is a consultation / complaint scenario, the multimodal data is derived based on the user's input text data, voice data, and behavioral feedback data.

[0023] The second objective of this invention is achieved by the following technical solution:

[0024] A rational betting supervision device based on sentiment fusion analysis includes a memory and a processor. The memory stores a rational betting supervision program that runs on the processor. The rational betting supervision program is a computer program. When the processor executes the rational betting supervision program, it implements the steps of the rational betting supervision method based on sentiment fusion analysis as one of the objectives of this invention.

[0025] The third objective of this invention is achieved by the following technical solution:

[0026] A computer-readable storage medium storing a rational betting supervision program thereon, the rational betting supervision program being a computer program, wherein when executed by a processor, the rational betting supervision program implements the steps of a rational betting supervision method based on sentiment fusion analysis as one of the objectives of this invention.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] This invention utilizes a large model to perform sentiment fusion analysis on behavioral data, voice / text data, and visual data generated by users during the lottery purchase process. This allows for the discovery of the probability of users exhibiting irrational lottery purchase tendencies, enabling timely intervention to address such behaviors and achieve responsible supervision, thereby ensuring users maintain a rational consumption perspective. Attached Figure Description

[0029] Figure 1 The flowchart of the rational lottery purchase supervision method based on emotion fusion analysis provided by the present invention. Detailed Implementation

[0030] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0031] Example 1

[0032] This invention collects multimodal data on users' lottery purchases to perform sentiment analysis and predict the probability of irrational purchases based on user behavior. This allows for the adaptation of corresponding intervention strategies to intervene in users' lottery purchasing behavior, prevent impulsive purchases, and achieve responsible supervision within the lottery industry.

[0033] Specifically, the present invention provides a preferred embodiment of a rational lottery purchase supervision method based on emotion fusion analysis, such as... Figure 1 As shown, it includes:

[0034] Step S1: Acquire multimodal data generated during the user's lottery purchase process and the user's current lottery purchase behavior data in real time, and preprocess the multimodal data and the user's current lottery purchase behavior data.

[0035] Specifically, this invention acquires multimodal data from users during the lottery purchasing process, including various types of data such as text data, voice data, and visual data. Text data refers to data input by the user through operating the terminal device or text data obtained through voice recognition. Voice data refers to the user's tone and speaking speed during voice operations. Visual data refers to the user's facial feature data, such as the amplitude or trajectory of movement of facial feature points.

[0036] User behavior data includes betting frequency, betting amount, operation frequency, operation time, and number selection time—all data related to lottery purchasing behavior. Both this multimodal data and user behavior data are obtained through lottery purchasing terminals, which can be online or offline. Online terminals are mobile devices used by users, while offline terminals include self-service kiosks in stores. Different lottery purchasing terminals may differ in how they acquire this multimodal data; the appropriate method should be selected based on actual needs.

[0037] Specifically, when the lottery purchase terminal is an offline terminal, hardware modifications can be made to the offline terminal equipment. This involves adding a miniature camera to the offline terminal to collect user facial video data, thereby obtaining micro-expression and facial feature data. For example, a 1080P wide dynamic range camera can be equipped on the offline lottery purchase terminal to solve problems such as low lighting and complex backgrounds in the sales store, such as uneven lighting and background interference from people. Simultaneously, the camera lens adopts an 85° golden angle of view and is used in conjunction with an infrared supplementary light to acquire user facial feature data, ensuring a facial feature point capture rate of up to 95%. Furthermore, an occlusion detection module is added to the lottery purchase terminal to detect situations where facial features are obscured by masks or other objects worn on the body.

[0038] When the lottery purchase device is an online device, when the user starts or opens the lottery purchase APP, text and voice data are collected through the mobile terminal. At the same time, the user's physiological signals are collected by combining the data from the mobile terminal's built-in sensors and external wristbands.

[0039] Simultaneously, the collected data undergoes preprocessing. Preprocessing includes data cleaning, data alignment, and data completion. Data cleaning removes redundant and invalid data. Data alignment involves synchronizing the collected data according to timestamps. Data completion utilizes an LSTM-based time-series completion model to predict and complete missing time-series data such as heart rate and stress by combining valid data from the preceding and following 10 seconds.

[0040] More preferably, step S1 further includes: firstly determining the user's lottery purchase scenario based on the user's current operational behavior, and then acquiring multimodal data based on the user's lottery purchase scenario and data acquisition rules. The lottery purchase scenario includes number selection scenario, bet confirmation scenario, prize redemption scenario, and lottery purchase scenario.

[0041] Specifically, when a user's lottery purchase scenario involves selecting numbers, multimodal data is derived from user terminal behavior data, user text data, user consultation voice data, and user facial verification video data. When the user's lottery purchase scenario involves bet confirmation, multimodal data is derived from the pressure sensor data of the user's terminal device's pressed buttons, the user's heart rate data, the user's voice data, and the user's video data. When the user's lottery purchase scenario involves prize redemption, multimodal data is derived from the user's voice data, video data, and user operation data. When the lottery purchase scenario involves consultation or complaint, multimodal data is derived from the user's input text data, voice data, and behavioral feedback data. This invention also sets different data collection methods according to different lottery purchase scenarios to make data collection more efficient.

[0042] Step S2: Extract corresponding features from multimodal data according to the type of multimodal data, fuse the features of each modality based on hierarchical fusion measurement to obtain a fused feature vector, obtain the user's real-time emotional intensity value based on a pre-trained emotional intensity prediction model, and determine the emotional type label based on the user's real-time emotional intensity value.

[0043] The dataset types of multimodal data include speech data, text data, and visual data. Therefore, the specific feature extraction of multimodal data according to its type includes:

[0044] First, speech data is derived from multimodal data, and text data is obtained by recognizing the speech data. Then, text data is extracted from multimodal data, and keywords are extracted from the text data to obtain sentiment words related to the lottery field. Word vectors are generated through the Word2Vec model, and deep semantic features of the text are extracted by combining the BERT model.

[0045] First, speech data is obtained from multimodal data, and speech features are extracted from the speech data. Speech features include fundamental frequency, volume, speech rate, spectral centroid, Mel-spectral cepstral coefficients, pitch change rate, and volume fluctuation amplitude.

[0046] First, visual data is obtained from multimodal data, and user facial recognition is performed on the visual data to identify user facial features. Then, video emotional features are obtained based on user facial features. Video emotional features include facial key points, expression action unit intensity, facial texture features, and eye change frequency features. Facial key points include the position and shape of eyebrows, eyes, and mouth.

[0047] Simultaneously, to derive the user's emotional intensity analysis, this invention also extracts features based on the user's current operational behavior data. That is, it derives user behavioral emotional characteristics based on user operational behavior data; these characteristics include operation frequency features, operation rhythm features, and preference features; operation frequency features include average click frequency and swipe frequency; operation rhythm features include the standard deviation of input intervals; and preference features include high-frequency access areas.

[0048] After obtaining the feature data of each modality, this invention further fuses the features of each modality based on a hierarchical fusion strategy to obtain a fused feature vector. Specifically, the features extracted from each modality are standardized, an attention mechanism is constructed, and the sentiment contribution weight of each modality feature is calculated. Based on the sentiment contribution weight of each modality feature, the features of each modality are weighted and summed to obtain preliminary fused features. Then, the preliminary fused features are input into a deep learning model to mine the correlation information between the features of each modality and output the fused feature vector.

[0049] The sentiment intensity prediction model in this invention is specifically obtained by training a sentiment intensity prediction model based on a support vector regression model using a constructed sentiment intensity labeled dataset. The sentiment intensity labeled dataset includes historical multimodal data samples and corresponding manually labeled sentiment intensity values. This pre-trained model is integrated into the system to provide timely analysis of the user's sentiment intensity during the current lottery-buying process.

[0050] Simultaneously, the system determines emotion type labels based on the user's real-time emotion intensity value. These labels include impulsiveness, anger, calmness, sadness, etc., and each label corresponds to a specific range of real-time emotion intensity values. Therefore, the emotion type label can be determined based on the user's real-time emotion intensity value. For example, when the user's real-time emotion intensity value exceeds a first preset value, the emotion type label is "impulsive." Furthermore, the user's emotion level can be calculated based on both the emotion type label and the user's real-time emotion intensity value.

[0051] In addition, the present invention can also generate analysis reports based on the user's real-time emotional intensity value, emotional type, and user emotional level, and support visualization for users to view.

[0052] This invention integrates data from four modalities—voice, text, vision, and user behavior—and fully leverages the complementarity of each modality to overcome the limitations of single-modality data, thereby significantly improving the accuracy and robustness of user emotion intensity analysis.

[0053] Step S3: Based on the user's real-time emotional intensity value, emotional type label, the user's current lottery purchase behavior data, and the irrational prediction model pre-built by the system, assess the probability of the user's current irrational lottery purchase behavior, and intervene in the user's irrational lottery purchase behavior based on the user's real-time emotional intensity value, emotional type label, and the probability of the user's current irrational betting behavior.

[0054] Specifically, user intervention includes: matching the user's real-time emotional intensity analysis results and the probability of the user's current irrational betting behavior tendency with the corresponding intervention strategy from a pre-built intervention strategy matching table in the system, and then intervening in the user's irrational betting behavior according to the corresponding intervention strategy; wherein, the intervention strategy matching table includes a matching relationship table between user emotional intensity, user irrational betting behavior tendency probability and intervention strategy.

[0055] The intervention strategies include emotional interaction and guidance strategies, a cooling-off period order-locking mechanism, and dynamic limit adjustment strategies. Specifically, the emotional interaction and guidance strategy involves dynamically generating guidance scripts based on emotional tags, creating an AI virtual avatar based on the scripts and the AI ​​digital human, and having the AI ​​virtual avatar send corresponding voice content to the user. It also involves entering a user dialogue mode to obtain and respond to the user's dialogue. For example, the AI ​​digital avatar might say to the user: "You are currently quite emotional. Perhaps you should first understand the purpose of the public welfare fund and rationally plan your lottery purchase amount." Simultaneously, it enters a voice dialogue mode, automatically recognizing the user's questions and answers through ASR (Automatic Social Responsiveness) and providing quick responses. When the dialogue duration exceeds a preset time or the user's questions and answers become more calm, the intervention decision-making mode exits and the normal lottery purchase mode resumes.

[0056] The cooling-off period order-locking mechanism includes hiding the betting button and locking the order, as well as displaying promotional videos for public welfare projects. For example, online, the betting button can be hidden during the order-locking period, and videos showcasing the progress of public welfare projects can be displayed. Offline, betting operations can be locked via the terminal, and interactive questions can pop up, such as: "Buying lottery tickets should be done within your means; have you planned your lottery budget for this month?" or other questions related to risk disclosure. Once the user answers correctly or answers multiple questions correctly consecutively, the intervention decision-making mode exits, and the user enters normal lottery purchase mode.

[0057] Dynamic limit adjustment strategies include locking betting operations and entering an interactive Q&A mode to obtain user answers. For example, a baseline limit is set based on the user's historical spending capacity (average monthly betting amount over the past 3 months), and the limit is reduced proportionally during intervention. Simultaneously, the limit adjustment record is synchronized to the blockchain for notarization, and users can view the reason for the adjustment through the app. Furthermore, the above intervention strategies can be combined depending on the specific circumstances; for example, emotional interaction and guidance strategies and cooling-off period order locking mechanisms can be used simultaneously.

[0058] More preferably, the present invention also constructs a relationship table based on emotional intensity and the probability of irrational lottery purchasing tendency with intervention decisions, as shown in Table 1:

[0059] Emotional Level The probability of irrational lottery-buying tendencies is low. The probability of irrational lottery-buying tendencies is medium risk. The probability of irrational lottery-buying tendencies is high. Low Without intervention, push public service science videos or pictures. A pop-up window prompts users to purchase lottery tickets responsibly and remains open for a specified time. Limit adjustment middle Pop-up notification for public service science popularization Cooling-off period corresponding time + preference guidance Cooling-off period duration + AI digital human voice guidance high Limit adjustment Cooling-off period corresponding time + emotional guidance text Order locking time + human customer service intervention

[0060] Table 1

[0061] When the model outputs results, the corresponding intervention strategies are found by combining the emotional intensity analysis and the probability of irrational lottery purchase tendency with Table 1. The intervention strategies are then used to intervene in the user's lottery purchase behavior to achieve responsible supervision of the user.

[0062] Furthermore, step S3 also includes: firstly, constructing a personalized behavior benchmark based on the target object's historical purchase data, and then constructing a feature set based on the pre-constructed irrational purchase behavior feature set, user sentiment analysis results, real-time purchase behavior data, and personalized behavior benchmark to make a preliminary judgment on the user's current lottery purchase behavior; then, based on the feature set, sentiment intensity value, sentiment type label, preliminary judgment results, and the system's pre-constructed fusion prediction model, deriving the probability of the user's irrational lottery purchase behavior tendency.

[0063] The feature set includes emotion-behavior deviation features and behavioral temporal anomaly features. Emotion-behavior deviation features include the coupling coefficient between real-time purchasing behavior frequency and emotion intensity, and the deviation rate between behavioral response delay time and historical benchmark. For example, if the coupling coefficient is greater than or equal to a preset threshold and the deviation rate is greater than or equal to 35%, it is marked as emotion-driven, that is, the behavior is suspected to be irrational.

[0064] Abnormal behavioral characteristics include real-time process compression ratio (real-time process time / historical baseline time), decision duration deviation rate, and the proportion of repeated operation frequency. For example, if any two of these exceed the historical baseline, the behavior is marked as abnormal, meaning the behavior is suspected to be irrational. First, a preliminary judgment is made using the feature set. If the behavior is determined to be suspected to be irrational, it is then input into a fusion prediction model to predict the probability of irrational behavior.

[0065] More preferably, the structure of the fusion prediction model in this invention is as follows: the input layer includes an emotion feature branch, a behavior feature branch, and a judgment result branch; the feature enhancement layer includes each branch extracting deep features through an independent Transformer encoder to output a multi-dimensional feature vector, and then using a multi-head self-attention mechanism to calculate the correlation weights between each branch and weighted fusion to obtain a global feature vector; the prediction layer includes outputting the judgment result through a Sigmoid activation function and outputting the tendency probability value through a Softmax activation function.

[0066] The training process of the fusion prediction model specifically includes: constructing a training dataset by combining historical multimodal sample data, sentiment analysis results, feature sets, and manually labeled irrational judgment tags and tendency probability reference values; and training the constructed fusion prediction model to obtain the fusion prediction model. During model training, the SMOTE-Tomek algorithm can be used to handle class imbalance, ensuring a positive-to-negative sample ratio of 1:2. Simultaneously, the training dataset is divided into training, validation, and test sets in an 8:1:1 ratio, optimized using a joint loss function of cross-entropy loss and mean squared error loss, training until the validation set accuracy is ≥95% and the probability prediction error is ≤±2%. Furthermore, an adversarial training mechanism can be introduced during training, using the FGSM algorithm to enhance model robustness, ensuring that the prediction accuracy under noisy data meets the preset accuracy.

[0067] This invention achieves sentiment analysis of users' lottery purchase process by integrating multimodal data during the purchase process. It combines the results of the sentiment analysis with the user's historical lottery purchase behavior data to predict whether the user's lottery purchase behavior is normal and rational. Based on the analysis results, it intervenes in irrational lottery purchase behavior in a timely manner, thereby achieving flexible supervision of users, improving user experience, and solving the problems of poor user experience caused by the use of fixed intervention strategies in existing technologies.

[0068] Example 2

[0069] A rational betting monitoring device based on sentiment fusion analysis includes a memory and a processor. The memory stores a rational betting monitoring program that runs on the processor. The rational betting monitoring program is a computer program. When the processor executes the rational betting monitoring program, it performs the following steps:

[0070] Data acquisition steps: Real-time acquisition of multimodal data generated during the user's lottery purchase process and the user's current lottery purchase behavior data, and preprocessing of the multimodal data and the user's current lottery purchase behavior data;

[0071] Sentiment analysis steps: Extract corresponding features from multimodal data according to the type of multimodal data, fuse the features of each modality based on hierarchical fusion measurement to obtain a fused feature vector, obtain the user's real-time sentiment intensity value based on a pre-trained sentiment intensity prediction model, and determine the sentiment type label based on the user's real-time sentiment intensity value.

[0072] Intervention steps: Based on the user's real-time emotional intensity value, emotional type label, the user's current lottery purchase behavior data, and the irrational prediction model pre-built by the system, the probability of the user's current irrational lottery purchase behavior is assessed, and intervention is carried out on the user's irrational lottery purchase behavior based on the user's real-time emotional intensity value, emotional type label, and the probability of the user's current irrational betting behavior.

[0073] Furthermore, the intervention steps for intervening in users' irrational betting behavior based on their real-time emotional intensity value, emotional type label, and the probability of their current irrational betting behavior include: selecting corresponding intervention strategies from a pre-built intervention strategy adaptation table in the system based on the user's real-time emotional intensity analysis results and the probability of their current irrational betting behavior; and then intervening in users' irrational betting behavior based on the corresponding intervention strategies; wherein, the intervention strategy adaptation table includes a table of matching relationships between user emotional intensity, user irrational betting behavior probability, and intervention strategies.

[0074] Furthermore, the intervention strategies include an emotional interaction guidance strategy, a cooling-off period order locking mechanism strategy, and a dynamic limit adjustment strategy. The emotional interaction guidance strategy involves dynamically generating guidance scripts based on emotional tags, creating an AI virtual avatar based on the guidance scripts and the AI ​​digital human, and having the AI ​​virtual avatar send corresponding voice content to the user, as well as entering a user dialogue mode to obtain and respond to the user's dialogue content. The cooling-off period order locking mechanism strategy includes hiding the betting button and locking the order, and displaying a public welfare project promotional video. The dynamic limit adjustment strategy includes locking the betting operation and entering an interactive Q&A mode to obtain the user's answers.

[0075] The intervention steps also include: setting the intervention termination condition for the corresponding strategy and ending the current intervention strategy and entering the normal lottery purchase mode when the corresponding intervention termination condition is triggered.

[0076] Furthermore, the intervention steps also include: firstly, constructing a personalized behavioral benchmark based on the target object's historical purchase data, and then constructing a feature set based on a pre-constructed irrational purchase behavior feature set, user sentiment analysis results, real-time purchase behavior data, and personalized behavioral benchmark to make a preliminary judgment on the user's current lottery purchase behavior; then, based on the feature set, sentiment intensity value, sentiment type label, preliminary judgment results, and the system's pre-constructed fusion prediction model, deriving the probability of the user's irrational lottery purchase behavior tendency; wherein, the feature set includes sentiment-behavioral deviation features and behavioral temporal anomaly features;

[0077] Furthermore, the fusion prediction model specifically includes: constructing a training dataset based on historical multimodal sample data, sentiment analysis results, feature sets, and manually labeled irrational judgment tags and tendency probability reference values; training the constructed fusion prediction model to obtain the fusion prediction model; the input layer of the fusion prediction model includes sentiment feature branches, behavioral feature branches, and judgment result branches; the feature enhancement layer includes each branch extracting deep features through an independent Transformer encoder to output multidimensional feature vectors, and then using a multi-head self-attention mechanism to calculate the correlation weights between each branch and weighted fusion to obtain a global feature vector; the prediction layer includes outputting judgment results through a Sigmoid activation function and outputting tendency probability values ​​through a Softmax activation function.

[0078] Furthermore, the sentiment analysis step specifically includes extracting corresponding features from multimodal data based on the type of multimodal data: first, obtaining speech data from multimodal data and recognizing the speech data to obtain text data; then, extracting keywords from the text data to obtain sentiment words related to the lottery field; generating word vectors through the Word2Vec model; and extracting deep semantic features of the text by combining the BERT model.

[0079] First, speech data is obtained from multimodal data and speech features are extracted from the speech data. Speech features include fundamental frequency, volume, speech rate, spectral centroid, Mel-spectral cepstral coefficients, pitch change rate, and volume fluctuation amplitude.

[0080] First, visual data is obtained from multimodal data, and user facial recognition is performed on the visual data to identify user facial features. Then, video emotional features are obtained based on user facial features. Video emotional features include facial key points, expression action unit intensity, facial texture features, and eye change frequency features. Facial key points include the position and shape of eyebrows, eyes, and mouth.

[0081] First, user behavior sentiment characteristics are derived from user operation behavior data. These characteristics include operation frequency, operation rhythm, and preference characteristics. Operation frequency characteristics include average click frequency and swipe frequency. Operation rhythm characteristics include the standard deviation of input interval. Preference characteristics include high-frequency access areas.

[0082] Furthermore, the fusion of features from various modalities based on the hierarchical fusion strategy to obtain a fused feature vector specifically includes: standardizing the features extracted from each modality, constructing an attention mechanism and calculating the sentiment contribution weight of each modality feature, weighting and summing the features of each modality based on the sentiment contribution weight of each modality feature to obtain a preliminary fused feature, and then inputting the preliminary fused feature into a deep learning model to mine the correlation information between the features of each modality to output the fused feature vector;

[0083] The sentiment intensity prediction model is specifically obtained by training the sentiment intensity prediction model based on the support vector regression model using a constructed sentiment intensity labeled dataset; the sentiment intensity labeled dataset includes historical multimodal data samples and corresponding manually labeled sentiment intensity values;

[0084] Furthermore, the data acquisition steps specifically include: first, determining the user's lottery purchase scenario based on the user's current operational behavior, and then acquiring multimodal data based on the user's lottery purchase scenario and data acquisition rules; specifically, when the user's lottery purchase scenario is a number selection scenario, multimodal data is derived based on the user's terminal operation behavior data, user operation text data, user consultation voice data, and user verification facial video data; when the user's lottery purchase scenario is a bet confirmation scenario, multimodal data is derived based on the pressure sensor data of the user's terminal device's pressed buttons, the user's heart rate data, the user's voice data, and the user's video data; when the user's lottery purchase scenario is a prize redemption scenario, multimodal data is derived based on the user's voice data, video data, and user operation data; and when the lottery purchase scenario is a consultation and complaint scenario, multimodal data is derived based on the user's input text data, voice data, and behavioral feedback data.

[0085] Example 3

[0086] A computer-readable storage medium storing a rational betting monitoring program thereon, the rational betting monitoring program being a computer program, which, when executed by a processor, performs the following steps:

[0087] Data acquisition steps: Real-time acquisition of multimodal data generated during the user's lottery purchase process and the user's current lottery purchase behavior data, and preprocessing of the multimodal data and the user's current lottery purchase behavior data;

[0088] Sentiment analysis steps: Extract corresponding features from multimodal data according to the type of multimodal data, fuse the features of each modality based on hierarchical fusion measurement to obtain a fused feature vector, obtain the user's real-time sentiment intensity value based on a pre-trained sentiment intensity prediction model, and determine the sentiment type label based on the user's real-time sentiment intensity value.

[0089] Intervention steps: Based on the user's real-time emotional intensity value, emotional type label, the user's current lottery purchase behavior data, and the irrational prediction model pre-built by the system, the probability of the user's current irrational lottery purchase behavior is assessed, and intervention is carried out on the user's irrational lottery purchase behavior based on the user's real-time emotional intensity value, emotional type label, and the probability of the user's current irrational betting behavior.

[0090] Furthermore, the intervention steps for intervening in users' irrational betting behavior based on their real-time emotional intensity value, emotional type label, and the probability of their current irrational betting behavior include: selecting corresponding intervention strategies from a pre-built intervention strategy adaptation table in the system based on the user's real-time emotional intensity analysis results and the probability of their current irrational betting behavior; and then intervening in users' irrational betting behavior based on the corresponding intervention strategies; wherein, the intervention strategy adaptation table includes a table of matching relationships between user emotional intensity, user irrational betting behavior probability, and intervention strategies.

[0091] Furthermore, the intervention strategies include an emotional interaction guidance strategy, a cooling-off period order locking mechanism strategy, and a dynamic limit adjustment strategy. The emotional interaction guidance strategy involves dynamically generating guidance scripts based on emotional tags, creating an AI virtual avatar based on the guidance scripts and the AI ​​digital human, and having the AI ​​virtual avatar send corresponding voice content to the user, as well as entering a user dialogue mode to obtain and respond to the user's dialogue content. The cooling-off period order locking mechanism strategy includes hiding the betting button and locking the order, and displaying a public welfare project promotional video. The dynamic limit adjustment strategy includes locking the betting operation and entering an interactive Q&A mode to obtain the user's answers.

[0092] The intervention steps also include: setting the intervention termination condition for the corresponding strategy and ending the current intervention strategy and entering the normal lottery purchase mode when the corresponding intervention termination condition is triggered.

[0093] Furthermore, the intervention steps also include: firstly, constructing a personalized behavioral benchmark based on the target object's historical purchase data, and then constructing a feature set based on a pre-constructed irrational purchase behavior feature set, user sentiment analysis results, real-time purchase behavior data, and personalized behavioral benchmark to make a preliminary judgment on the user's current lottery purchase behavior; then, based on the feature set, sentiment intensity value, sentiment type label, preliminary judgment results, and the system's pre-constructed fusion prediction model, deriving the probability of the user's irrational lottery purchase behavior tendency; wherein, the feature set includes sentiment-behavioral deviation features and behavioral temporal anomaly features;

[0094] Furthermore, the fusion prediction model specifically includes: constructing a training dataset based on historical multimodal sample data, sentiment analysis results, feature sets, and manually labeled irrational judgment tags and tendency probability reference values; training the constructed fusion prediction model to obtain the fusion prediction model; the input layer of the fusion prediction model includes sentiment feature branches, behavioral feature branches, and judgment result branches; the feature enhancement layer includes each branch extracting deep features through an independent Transformer encoder to output multidimensional feature vectors, and then using a multi-head self-attention mechanism to calculate the correlation weights between each branch and weighted fusion to obtain a global feature vector; the prediction layer includes outputting judgment results through a Sigmoid activation function and outputting tendency probability values ​​through a Softmax activation function.

[0095] Furthermore, the sentiment analysis step specifically includes extracting corresponding features from multimodal data based on the type of multimodal data: first, obtaining speech data from multimodal data and recognizing the speech data to obtain text data; then, extracting keywords from the text data to obtain sentiment words related to the lottery field; generating word vectors through the Word2Vec model; and extracting deep semantic features of the text by combining the BERT model.

[0096] First, speech data is obtained from multimodal data and speech features are extracted from the speech data. Speech features include fundamental frequency, volume, speech rate, spectral centroid, Mel-spectral cepstral coefficients, pitch change rate, and volume fluctuation amplitude.

[0097] First, visual data is obtained from multimodal data, and user facial recognition is performed on the visual data to identify user facial features. Then, video emotional features are obtained based on user facial features. Video emotional features include facial key points, expression action unit intensity, facial texture features, and eye change frequency features. Facial key points include the position and shape of eyebrows, eyes, and mouth.

[0098] First, user behavior sentiment characteristics are derived from user operation behavior data. These characteristics include operation frequency, operation rhythm, and preference characteristics. Operation frequency characteristics include average click frequency and swipe frequency. Operation rhythm characteristics include the standard deviation of input interval. Preference characteristics include high-frequency access areas.

[0099] Furthermore, the fusion of features from various modalities based on the hierarchical fusion strategy to obtain a fused feature vector specifically includes: standardizing the features extracted from each modality, constructing an attention mechanism and calculating the sentiment contribution weight of each modality feature, weighting and summing the features of each modality based on the sentiment contribution weight of each modality feature to obtain a preliminary fused feature, and then inputting the preliminary fused feature into a deep learning model to mine the correlation information between the features of each modality to output the fused feature vector;

[0100] The sentiment intensity prediction model is specifically obtained by training the sentiment intensity prediction model based on the support vector regression model using a constructed sentiment intensity labeled dataset; the sentiment intensity labeled dataset includes historical multimodal data samples and corresponding manually labeled sentiment intensity values;

[0101] Furthermore, the data acquisition steps specifically include: first, determining the user's lottery purchase scenario based on the user's current operational behavior, and then acquiring multimodal data based on the user's lottery purchase scenario and data acquisition rules; specifically, when the user's lottery purchase scenario is a number selection scenario, multimodal data is derived based on the user's terminal operation behavior data, user operation text data, user consultation voice data, and user verification facial video data; when the user's lottery purchase scenario is a bet confirmation scenario, multimodal data is derived based on the pressure sensor data of the user's terminal device's pressed buttons, the user's heart rate data, the user's voice data, and the user's video data; when the user's lottery purchase scenario is a prize redemption scenario, multimodal data is derived based on the user's voice data, video data, and user operation data; and when the lottery purchase scenario is a consultation and complaint scenario, multimodal data is derived based on the user's input text data, voice data, and behavioral feedback data.

[0102] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A rational lottery purchase supervision method based on emotional fusion analysis, characterized in that, The methods for regulating rational lottery purchases include: Data acquisition steps: Real-time acquisition of multimodal data generated during the user's lottery purchase process and the user's current lottery purchase behavior data, and preprocessing of the multimodal data and the user's current lottery purchase behavior data; Sentiment analysis steps: Based on the type of the multimodal data, corresponding features are extracted from the multimodal data, and the features of each modality are fused based on hierarchical fusion measurement to obtain a fused feature vector. Based on a pre-trained sentiment intensity prediction model, the user's real-time sentiment intensity value is obtained, and the sentiment type label is determined based on the user's real-time sentiment intensity value. Intervention steps: Based on the user's real-time emotional intensity value, emotional type label, the user's current lottery purchase behavior data, and the irrational prediction model pre-built by the system, the probability of the user's current irrational lottery purchase behavior is assessed, and intervention is carried out on the user's irrational lottery purchase behavior based on the user's real-time emotional intensity value, emotional type label, and the probability of the user's current irrational betting behavior.

2. The rational lottery purchase supervision method based on sentiment fusion analysis according to claim 1, characterized in that, The intervention step, which intervenes in the user's irrational betting behavior based on the user's real-time emotional intensity value, emotional type label, and the probability of the user's current irrational betting behavior tendency, specifically includes: selecting corresponding intervention strategies from a pre-built intervention strategy adaptation table in the system based on the user's real-time emotional intensity analysis results and the probability of the user's current irrational betting behavior tendency, and then intervening in the user's irrational betting behavior according to the corresponding intervention strategy; wherein, the intervention strategy adaptation table includes a matching relationship table between user emotional intensity, user irrational betting behavior tendency probability, and intervention strategy.

3. The rational lottery purchase supervision method based on sentiment fusion analysis according to claim 2, characterized in that, The intervention strategies include an emotional interaction guidance strategy, a cooling-off period order locking mechanism strategy, and a dynamic limit adjustment strategy. The emotional interaction guidance strategy involves dynamically generating guidance scripts based on emotional tags, creating an AI virtual human based on the guidance scripts and the AI ​​digital human, and having the AI ​​virtual human send corresponding voice content to the user, as well as entering a user dialogue mode to obtain and respond to the user's dialogue content. The cooling-off period order locking mechanism strategy includes hiding the betting button and locking the order, and displaying a public welfare project promotional video. The dynamic limit adjustment strategy includes locking the betting operation and entering an interactive Q&A mode to obtain the user's answers. The intervention steps also include: setting the intervention termination condition for the corresponding strategy and ending the current intervention strategy and entering the normal lottery purchase mode when the corresponding intervention termination condition is triggered.

4. The rational lottery purchase supervision method based on sentiment fusion analysis according to claim 3, characterized in that, The intervention steps further include: firstly, constructing a personalized behavioral benchmark based on the target object's historical purchase data, and then constructing a feature set based on a pre-constructed irrational purchase behavior feature set, user sentiment analysis results, real-time purchase behavior data, and personalized behavioral benchmark to make a preliminary judgment on the user's current lottery purchase behavior; then, based on the feature set, sentiment intensity value, sentiment type label, preliminary judgment results, and a pre-constructed fusion prediction model, deriving the probability of the user's irrational lottery purchase behavior tendency; wherein, the feature set includes sentiment-behavioral deviation features and behavioral temporal anomaly features.

5. The rational lottery purchase supervision method based on sentiment fusion analysis according to claim 4, characterized in that, The fusion prediction model specifically includes: constructing a training dataset based on historical multimodal sample data, sentiment analysis results, feature sets, and manually labeled irrational judgment tags and tendency probability reference values; training the constructed fusion prediction model to obtain the fusion prediction model; the input layer of the fusion prediction model includes sentiment feature branches, behavioral feature branches, and judgment result branches; the feature enhancement layer includes each branch extracting deep features through an independent Transformer encoder to output a multidimensional feature vector, and then using a multi-head self-attention mechanism to calculate the correlation weights between each branch and weighted fusion to obtain a global feature vector; the prediction layer includes outputting the judgment result through a Sigmoid activation function and outputting the tendency probability value through a Softmax activation function.

6. The rational lottery purchase supervision method based on sentiment fusion analysis according to claim 1, characterized in that, The sentiment analysis step specifically includes extracting corresponding features from the multimodal data according to the type of the multimodal data: first, obtaining speech data from the multimodal data and recognizing the speech data to obtain text data; then, extracting keywords from the text data to obtain sentiment words related to the lottery field; generating word vectors through the Word2Vec model; and extracting deep semantic features of the text by combining the BERT model. First, speech data is obtained based on the multimodal data, and speech features are extracted from the speech data. The speech features include fundamental frequency, volume, speech rate, spectral centroid, Mel-spectral cepstral coefficients, pitch change rate, and volume fluctuation amplitude. First, visual data is obtained based on the multimodal data, and user facial recognition is performed on the visual data to identify user facial features. Then, video emotional features are obtained based on the user facial features. The video emotional features include facial key points, expression action unit intensity, facial texture features, and eye change frequency features. Facial key points include the position and shape of eyebrows, eyes, and mouth. First, user behavior emotional characteristics are derived based on user operation behavior data; these characteristics include operation frequency characteristics, operation rhythm characteristics, and preference characteristics; operation frequency characteristics include average click frequency and swipe frequency; operation rhythm characteristics include input interval standard deviation; and preference characteristics include high-frequency access areas.

7. The rational lottery purchase supervision method based on sentiment fusion analysis according to claim 1, characterized in that, The method of fusing features from various modalities based on a hierarchical fusion strategy to obtain a fused feature vector specifically includes: standardizing the features extracted from each modality, constructing an attention mechanism and calculating the emotional contribution weight of each modal feature, weighting and summing the features of each modality according to the emotional contribution weight of each modality feature to obtain a preliminary fused feature, and then inputting the preliminary fused feature into a deep learning model to mine the correlation information between the features of each modality and output the fused feature vector. The sentiment intensity prediction model is specifically obtained by training the sentiment intensity prediction model based on the support vector regression model using a constructed sentiment intensity labeled dataset; wherein, the sentiment intensity labeled dataset includes historical multimodal data samples and corresponding manually labeled sentiment intensity values.

8. The rational lottery purchase supervision method based on sentiment fusion analysis according to claim 1, characterized in that, The data acquisition steps specifically include: first, determining the user's lottery purchase scenario based on the user's current operational behavior, and then acquiring multimodal data according to the user's lottery purchase scenario and data acquisition rules; wherein, when the user's lottery purchase scenario is a number selection scenario, the multimodal data is derived based on the user's terminal operation behavior data, user operation text data, user consultation voice data, and user verification facial video data; when the user's lottery purchase scenario is a bet confirmation scenario, the multimodal data is derived based on the pressure sensor of the user's terminal device's pressed button, the user's heart rate data, the user's voice data, and the user's video data; when the user's lottery purchase scenario is a prize redemption scenario, the multimodal data is derived based on the user's voice data, video data, and user operation data; when the lottery purchase scenario is a consultation and complaint scenario, the multimodal data is derived based on the user's input text data, voice data, and behavioral feedback data.

9. A rational lottery purchase monitoring device based on sentiment fusion analysis, comprising a memory and a processor, wherein the memory stores a rational lottery purchase monitoring program running on the processor, characterized in that, The rational lottery purchase supervision program is a computer program, and when the processor executes the rational lottery purchase supervision program, it implements the steps of the rational lottery purchase supervision method based on sentiment fusion analysis as described in any one of claims 1-8.

10. A computer-readable storage medium storing thereon a rational lottery purchase monitoring program, characterized in that, The rational lottery purchase supervision program is a computer program, and when the rational lottery purchase supervision program is executed by the processor, it implements the steps of the rational lottery purchase supervision method based on sentiment fusion analysis as described in any one of claims 1-8.

Images