A method for analyzing financial user sentiment and psychological portraits based on a large language model

Abstract

The present invention discloses a method for analyzing the emotion and psychological portrait of financial users based on a large language model. The method includes collecting user text data; preprocessing the collected user text data; using a large language model to perform emotion analysis and psychological feature modeling on the user text data to obtain the user's emotion feature vector and psychological feature vector; combining the user's emotion feature vector and psychological feature vector with a risk scoring model to generate a comprehensive risk score; and visually displaying the user portrait and risk score through visualization. The present invention combines unstructured data with structured data, and uses a large language model to perform sentence-by-sentence emotion classification and psychological feature modeling on user text data, thereby achieving unified feature processing and user portrait generation, significantly improving the integrity and multi-dimensional coverage of user portraits, and providing a more reliable basis for risk assessment.

Description

Technical Field

[0001] The present invention relates to the field of financial risk control technology, and in particular to a method for analyzing emotions and psychological portraits of financial users based on a large language model. Background Art

[0002] In the field of financial risk control, traditional user portrait generation methods mainly rely on users' historical transaction records and credit scores, which can reflect users' financial health and repayment ability to a certain extent. However, existing technologies have many shortcomings. First, the utilization rate of unstructured data is low, such as customer complaint records, customer service chat content, etc. These data often contain important emotional and psychological characteristics of users, but they have not been fully explored in traditional risk control models. Secondly, the lack of emotional and psychological modeling capabilities makes it difficult for risk control models to capture users' emotional fluctuations and psychological tendencies, thereby affecting the accuracy of risk assessment. In addition, traditional models are mostly based on static feature analysis, lack dynamic update capabilities, and are difficult to cope with real-time changes in user behavior and emotions. Summary of the invention

[0003] The purpose of the present invention is to provide a method for analyzing the emotion and psychological portrait of financial users based on a large language model in view of the deficiencies in the prior art.

[0004] To achieve the above objectives, the present invention provides a method for analyzing the emotion and psychological portrait of financial users based on a large language model, comprising:

[0005] Step 1: Collect user text data;

[0006] Step 2: pre-process the collected user text data;

[0007] Step 3: Use the large language model to perform sentiment analysis and psychological feature modeling on the user's text data to obtain the user's sentiment feature vector and psychological feature vector;

[0008] Step 4: Combine the user's emotional feature vector and psychological feature vector with the risk scoring model to generate a comprehensive risk score;

[0009] Step 5: Visualize the user profile and risk score.

[0010] Furthermore, the user text data includes structured data and unstructured data, the structured data includes the user's transaction records and historical credit scores; the unstructured data includes the user's conversation records with customer service, transaction notes, and complaint texts.

[0011] Furthermore, it is characterized in that the preprocessing method is as follows:

[0012] A denoising algorithm is used to remove meaningless characters in user text data, and spelling errors in user text data are corrected through a spelling correction model based on the edit distance algorithm. At the same time, the user text data is segmented into sentences with the help of natural language processing tools, and word segmentation is performed in combination with rules and statistical methods. The two-way maximum matching algorithm is then used to optimize the Chinese word segmentation effect. The key semantic components in the user text data are identified through a pre-trained qualified part-of-speech tagging model, and entity recognition technology is used to extract entity information in the user text data.

[0013] Furthermore, it is characterized in that the step 3 specifically comprises:

[0014] The preprocessed user text data is input into the large language model sentence by sentence for analysis to obtain the sentiment classification result corresponding to each sentence, and the corresponding sentiment feature vector is converted into a quantitative value in the range of 0-1 according to the probability distribution; at the same time, the contextual attention mechanism is used to deeply understand the sentiment level in the user text data, so as to more comprehensively capture the sentiment characteristics of the user text data, and by analyzing the language style of the user text data, extract the syntactic complexity, rhetoric and vocabulary diversity characteristics to infer the user's psychological state, and then generate the user's psychological feature vector.

[0015] Furthermore, the emotion classification results include anger, anxiety and satisfaction.

[0016] Furthermore, the method of generating the comprehensive risk score is as follows:

[0017]

[0018] in, is the generated comprehensive risk score, n is the number of sentiment feature vectors, is the i-th sentiment feature vector, is the risk weight corresponding to the i-th emotional feature vector, m is the number of psychological feature vectors, is the jth psychological feature vector, is the risk weight corresponding to the jth psychological characteristic vector, is the number of traditional financial feature vectors, is the kth traditional financial feature vector, is the risk weight corresponding to the kth financial characteristic vector.

[0019] Furthermore, it also includes:

[0020] Risk weights are determined based on subsequent user behavior data, time, and industry experience. Make dynamic adjustments, specifically, adjust risk weights based on time The specific methods are as follows:

[0021]

[0022] in, is the risk weight adjusted according to time, α is the adjustment parameter, and t is the time interval of sentiment classification results;

[0023] Adjust risk weights based on subsequent user behavior data The specific methods are as follows:

[0024]

[0025] in, To adjust the risk weight based on subsequent user behavior data, To adjust the step size, is the amount of change in the comprehensive risk score;

[0026] Adjust risk weights based on industry experience The specific methods are as follows:

[0027] For different user groups, adjust weights based on the statistical laws of historical data.

[0028] Furthermore, in step 5, the comprehensive risk score is displayed through a scoring dashboard, and segmented color warnings are used to intuitively reflect the risk level.

[0029] Beneficial effects: 1. The present invention combines unstructured data with structured data to achieve unified feature processing and user portrait generation. Compared with the limitation of traditional risk control models that rely on a single data source, the present invention uses a deep learning model for feature extraction and fusion, which significantly improves the integrity and multi-dimensional coverage of user portraits, providing a more reliable basis for risk assessment;

[0030] 2. This application uses a large language model (LLM) to perform sentence-by-sentence sentiment classification and psychological feature modeling on user text data, quantifying sentiment feature vectors and psychological tendencies into numerical form and using them in combination; through the contextual attention mechanism, it can accurately capture the sentiment levels and semantic relationships in the text, and generate a multi-dimensional psychological portrait of the user. This analysis method that combines sentiment feature vectors with psychological feature vectors significantly improves the sensitivity of risk assessment to user behavior and emotional changes;

[0031] 3. The present invention introduces a dynamic risk assessment mechanism and dynamically adjusts the weight of the user risk score through time decay technology, so that the recent behavior data has a greater impact on the score. At the same time, combined with the time series model, it can capture the changing trend of user emotions and psychological characteristics in real time and analyze the dynamic change process of user risks. This dynamic risk assessment method overcomes the shortcomings of traditional static risk models and provides strong technical support for real-time risk management;

[0032] 4. The present invention designs a multi-layer perceptron (MLP) model architecture in the feature fusion process, integrating emotional features, psychological features and traditional financial features. Through the feature weighting mechanism, the system can dynamically adjust the feature weights according to different business scenarios. This architecture not only enhances the flexibility of the model, but also improves the accuracy of the comprehensive risk score, providing optimized risk assessment capabilities for different business needs;

[0033] 5. The present invention introduces adversarial training strategies and anomaly detection technologies in risk assessment to identify abnormal fluctuations in user behavior or emotional characteristics. When the system detects a sudden increase in the frequency of a user's anxiety, it will automatically trigger a risk control warning. This highly robust anomaly detection mechanism not only improves the performance of the model in extreme data scenarios, but also provides financial institutions with real-time risk warning capabilities in a rapidly changing environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of a method for analyzing financial user emotions and psychological portraits based on a large language model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. The embodiments are implemented based on the technical solution of the present invention. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0036] like Figure 1 As shown, the embodiment of the present invention provides a method for analyzing emotions and psychological portraits of financial users based on a large language model, including:

[0037] Step 1: Collect user text data. The data sources collected include but are not limited to CRM systems, financial trading platforms, and customer interaction platforms. Support multiple data access methods, which can not only collect streaming data in real time to meet real-time analysis needs, but also efficiently integrate historical data through batch processing mode to provide high-quality input data for subsequent emotional and psychological analysis. The above user text data includes structured data and unstructured data, among which structured data includes:

[0038] Transaction records: consumption amount, frequency, merchant type, etc.

[0039] Historical credit score: such as user credit assessment provided by a third-party agency.

[0040] Unstructured data includes:

[0041] Conversation records between customers and customer service: Contains users' inquiries, complaints or complaints.

[0042] Complaint Text: Detailed description of the problem submitted by the user.

[0043] Transaction Notes: Additional notes filled in by the user when making payment.

[0044] Step 2: Preprocess the collected user text data. Preprocessing is the core step of the text preprocessing module, which uses multiple technical means to ensure the quality and consistency of the input data. The above preprocessing method is as follows:

[0045] Use a denoising algorithm to remove meaningless characters (such as HTML tags and special symbols) from user text data, and use a spelling correction model based on the edit distance algorithm to correct spelling errors in user text data. At the same time, use natural language processing (NLP) tools (such as SpaCy or NLTK) to segment user text data. In the word segmentation and tagging stage, combine rules and statistical methods to segment words, and then use a two-way maximum matching algorithm to optimize the Chinese word segmentation effect. Then, use a pre-trained qualified part-of-speech tagging model to identify key semantic components (such as nouns and verbs) in user text data, and use entity recognition technology to extract entity information (such as user names, place names, and financial institution names) in user text data to provide accurate data input for subsequent analysis.

[0046] Step 3: Use the large language model to perform sentiment analysis and psychological feature modeling on the user's text data to obtain the user's sentiment feature vector and psychological feature vector. The details are as follows:

[0047] The preprocessed user text data is input into the large language model sentence by sentence for analysis to obtain the corresponding sentiment classification results for each sentence. The above sentiment classification results include anger, anxiety and satisfaction. And according to the probability distribution, the sentiment feature vector corresponding to the sentiment classification result is converted into a quantitative value in the range of 0-1; at the same time, the contextual attention mechanism is used to deeply understand the sentiment level in the user text data. For example, "Although the service is average, the overall experience is good" will be divided into "negative" and "positive" emotions, and the corresponding weights will be assigned to each, so as to more comprehensively capture the emotional characteristics of the text. And by analyzing the language style of the user text data, the syntactic complexity, rhetoric and vocabulary diversity features are extracted to infer the user's psychological state. The details are as follows:

[0048] Syntactic complexity: Count the proportion of long sentences and the number of nested sentence levels. A high proportion of long sentences may indicate that the user has a higher ability to think carefully or has a tendency to take risks.

[0049] Lexical diversity: Calculate the variety and density of vocabulary, such as the Type-Token Ratio (TTR). A high lexical richness may reflect the user's openness.

[0050] Rhetorical devices: Count the frequency of use of rhetorical devices such as metaphors, similes, and parallelism. Users with rich rhetoric may show a higher tendency to take risks or be more creative.

[0051] Based on this, a user's psychological feature vector is generated, including but not limited to the following dimensions:

[0052] Risk Tendency: If users frequently mention words such as "investment" and "high return" and their emotions are positive, it may indicate a high risk tendency.

[0053] Frugality: If the user uses more words such as "saving" and "budget" and has a conservative language style, it may reflect a conservative tendency.

[0054] Decision Hesitation: Measures the uncertainty in user decisions by detecting repeated modifications and vague expressions (such as "maybe" and "probably") in sentences.

[0055] Example of psychological feature vector form:

[0056] Psychological Vector= [Risk Tendency=0.85, Frugality=0.65, DecisionHesitation=0.40]

[0057] In addition, the sentence-by-sentence sentiment analysis results can be semantically aggregated with the overall text semantics to generate a higher-level psychological portrait, providing in-depth support for user portraits and risk assessment. For example, "I am worried about investment risks, but I want to try" may indicate the coexistence of high anxiety and high risk-taking tendency.

[0058] Step 4: Combine the user's emotional feature vector and psychological feature vector with the traditional risk scoring model (using traditional financial features such as credit score and historical transaction records) to generate a comprehensive risk score, providing a more comprehensive user portrait for risk assessment. Specifically, a deep learning model (such as multi-layer perceptron, MLP) can be used to comprehensively process multiple features to achieve accurate assessment of user risks. The input layer of the deep learning model integrates emotional features (such as anger intensity), psychological features (such as risk-taking tendency) and traditional financial features (such as credit score); the hidden layer of the deep learning model consists of a two-layer 128-node fully connected network, using the ReLU activation function to handle complex nonlinear relationships; the output layer of the deep learning model generates a comprehensive risk score ranging from 0 to 1. In addition, a dynamic feature weighting mechanism can be used to dynamically assign feature weights according to different scenarios, such as giving psychological features a higher weight in high-frequency trading behavior.

[0059] During model training, the idea of ​​Generative Adversarial Network (GAN) can also be introduced to generate simulated abnormal data (such as extreme emotional distribution or psychological feature values) and classify them during model training. This improves the robustness of the model to abnormal data and ensures that risks can still be accurately assessed in extreme cases. Use weighted mean square error (MSE) as the loss function to increase the weight of emotional and psychological features in the model. The details are as follows:

[0060]

[0061] in, is the loss value calculated by the loss function, is the weight of feature j, is the true risk value of feature j, is the risk prediction value of feature j. The Adam optimizer (Adaptive Moment Estimation) can also be used to adjust model parameters to improve training efficiency. The learning rate can be dynamically reduced according to the validation set error to avoid model overfitting.

[0062] The composite risk score is generated as follows:

[0063]

[0064] Among them, among them, is the generated comprehensive risk score, n is the number of sentiment feature vectors, is the i-th sentiment feature vector, is the risk weight corresponding to the i-th emotional feature vector, m is the number of psychological feature vectors, is the jth psychological feature vector, is the risk weight corresponding to the jth psychological characteristic vector, is the number of traditional financial feature vectors, is the kth traditional financial feature vector, is the risk weight corresponding to the kth financial characteristic vector.

[0065] The above comprehensive risk score The value range is 0 to 1, which is used to assess the risk level of the user. The risk level is divided into: low risk: 0-0.3; medium risk: 0.3-0.7; high risk: 0.7-1.0. Risk weight The range is usually -1 to 1, and is set according to the impact of emotion on risk. For example:

[0066] Negative emotions (such as "anger") are weighted as positive and increase the risk score.

[0067] Positive sentiments (such as "satisfied") are weighted negatively, reducing the risk score.

[0068] Step 5: Visualize the user portrait and risk score to provide accurate support for the decision-making of financial institutions. The user portrait part includes a bar chart of emotional characteristics (such as "anger = 0.7, satisfaction = 0.2") and a radar chart of psychological characteristics (such as "risk-taking tendency" is high and "conservative characteristics" is low), and displays the trend chart of user emotions and psychological characteristics over time to help institutions understand the dynamic changes of user characteristics. The comprehensive risk score can also be displayed through the scoring dashboard, and segmented color warnings are used, green (low risk: 0-0.3), yellow (medium risk: 0.3-0.7), and red (high risk: 0.7-1.0), which intuitively reflect the risk level. In addition, the present invention also supports a detailed decomposition of risk scores, including the contribution ratio of traditional credit scores, emotional characteristic scores, and psychological characteristic scores, to help institutions deeply understand the composition of the scoring results and the risk factors behind them.

[0069] The embodiment of the present invention also includes: according to the subsequent user behavior data, time and industry experience, the risk weight Make dynamic adjustments, specifically, adjust risk weights based on time The specific methods are as follows:

[0070]

[0071] in, is the risk weight adjusted according to time, α is the adjustment parameter, and t is the time interval of the sentiment classification results. The purpose of this adjustment is to adjust the weight of the latest sentiment label to a higher level, and the historical sentiment labels gradually decay.

[0072] Adjust risk weights based on subsequent user behavior data The specific methods are as follows:

[0073]

[0074] in, To adjust the risk weight based on subsequent user behavior data, To adjust the step size, is the change in the comprehensive risk score. The purpose of this adjustment is to modify the weight according to the user's subsequent behavior (such as whether or not a contract has been breached). For example, if a user frequently expresses "anxiety" and then breaches a contract, the weight of the "anxiety" label will be increased.

[0075] Adjust risk weights based on industry experience The specific methods are as follows:

[0076] For different user groups (such as high-income users and ordinary users), weights are adjusted based on the statistical laws of historical data.

[0077] In addition, time series analysis models (such as LSTM or GRU) can be used to model the dynamic changes of user psychological characteristics and capture short-term and long-term trends. It can also analyze the fluctuations of users' psychological states in different periods of time. For example, the peak period of complaints may be accompanied by increased emotional fluctuations. The long-term changes of users' psychological characteristics can also be analyzed through sliding window statistical methods, such as whether conservative tendencies are gradually weakening. It can also identify sudden changes in users' psychological states. For example, a sharp increase in risk-taking tendencies in a short period of time may indicate high-risk behavior.

[0078] For example:

[0079] Example 1: A user submits a loan application, and the financial institution needs to assess the user's risk to determine the loan amount and interest rate.

[0080] Specific analysis process:

[0081] 1. Data collection and preprocessing

[0082] The system collects historical data of users, including their transaction records (high frequency of recent consumption) and credit scores (0.72, which is medium to high).

[0083] Extract the user’s recent conversation records with customer service, the main content of which reflects his doubts and urgent needs regarding the purpose of the loan.

[0084] 2. Analysis of emotional feature vectors and psychological feature vectors

[0085] Sentiment analysis: The system uses a large language model to analyze the conversation content and extract the user's emotional feature vector. It is found that the emotion of "anxiety" frequently appears in the conversation, with an intensity of 0.75.

[0086] Psychological modeling: Generate the user's psychological feature vector through language style analysis:

[0087] Risk-taking tendency: 0.85 (high)

[0088] Decision Hesitation: 0.65 (moderately high)

[0089] Conservative characteristic: 0.40 (low)

[0090] 3. Feature fusion and risk assessment

[0091] The system integrates emotional feature vectors (anxiety = 0.75), psychological feature vectors (risk-taking tendency = 0.85) and traditional credit features (credit score = 0.72).

[0092] The composite risk score generated by the multi-layer perceptron (MLP) is 0.78 (high risk).

[0093] Risk factor decomposition: emotional feature vector accounts for 40%, psychological feature vector accounts for 30%, and traditional credit score accounts for 30%.

[0094] 4. Result output

[0095] The overall risk score is 0.78, and the system classifies the user as high risk.

[0096] Based on the risk score, the system recommends reducing the loan amount to 50% of the expected amount and attaching a higher interest rate to cover potential risks.

[0097] 5. Decision making example

[0098] Output result: The user's comprehensive risk score is high, and it is recommended that the loan amount be reduced from 50,000 yuan to 25,000 yuan, and the interest rate be increased by 50BP.

[0099] Visualization report: The emotional feature bar chart shows that "anxiety" is the main emotion, and the emotional feature vector is 0.75. The psychological feature radar chart shows that "risk-taking tendency" is high and "conservative characteristics" is low.

[0100] Example 2: A user has communicated with customer service several times in the past two weeks and submitted complaints about incorrect account deductions. The financial institution wants to determine whether the user is a fraud risk.

[0101] Specific analysis process:

[0102] 1. Data collection and preprocessing

[0103] The system collects the user's customer service conversation records from the past 14 days and extracts unstructured text.

[0104] At the same time, the user’s historical credit score (0.45, which is low) and transaction records (including multiple large refund requests) are retrieved.

[0105] 2. Analysis of emotional feature vectors and psychological feature vectors

[0106] Sentiment analysis: By analyzing user conversations sentence by sentence, the system found that the frequency of "anger" increased significantly (sentiment feature vector = 0.82), while other emotions (such as "satisfaction") were basically missing.

[0107] Psychological modeling: Generate user psychological feature vectors through language style analysis:

[0108] Risk-taking tendency: 0.70 (high)

[0109] Decision Hesitation: 0.20 (low)

[0110] Conservative characteristic: 0.30 (low)

[0111] 3. Feature fusion and risk assessment

[0112] The system integrates emotional feature vectors (anger = 0.82), psychological feature vectors (risk-taking tendency = 0.70), and traditional credit features (credit score = 0.45).

[0113] A composite risk score was generated using a deep learning model: 0.85 (very high risk).

[0114] Risk factor decomposition: emotional feature vectors account for 50%, psychological feature vectors account for 20%, and traditional credit features account for 30%.

[0115] 4. Result output

[0116] The comprehensive risk score is 0.85. The system classifies the user as an extremely high-risk level and triggers a risk control warning.

[0117] Risk control measures include: further verifying the authenticity of the complaint.

[0118] 5. Decision making example

[0119] Output result: The "anger" emotion in the user's emotional characteristics is strong. Combined with the low credit score and high risk-taking tendency, the system determines that there is a risk of fraud.

[0120] Visual report: The risk score dashboard is displayed in red (high risk range). User portraits show that 50% of users have high anger emotions, high risk-taking tendencies, and low conservative characteristics.

[0121] The above is only a preferred embodiment of the present invention. It should be noted that for ordinary technicians in this technical field, other parts not specifically described belong to the prior art or common knowledge. Several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention.

Claims

1. A method for analyzing the emotion and psychological portrait of financial users based on a large language model, characterized in that: include: Step 1: Collect user text data; Step 2: pre-process the collected user text data; Step 3: Use the large language model to perform sentiment analysis and psychological feature modeling on the user's text data to obtain the user's sentiment feature vector and psychological feature vector; Step 4: Combine the user's emotional feature vector and psychological feature vector with the risk scoring model to generate a comprehensive risk score; Step 5: Visualize the user profile and risk score. The user text data includes structured data and unstructured data. The structured data includes the user's transaction records and historical credit scores; the unstructured data includes the user's conversation records with customer service, transaction notes, and complaint texts. The step 3 specifically includes: The preprocessed user text data is input into the large language model sentence by sentence for analysis to obtain the sentiment classification result corresponding to each sentence, and the corresponding sentiment feature vector is converted into a quantitative value in the range of 0-1 according to the probability distribution; at the same time, the contextual attention mechanism is used to deeply understand the sentiment level in the user text data, so as to more comprehensively capture the sentiment characteristics of the user text data, and by analyzing the language style of the user text data, the syntactic complexity, rhetoric and vocabulary diversity characteristics are extracted to infer the user's psychological state, and then the user's psychological feature vector is generated; The composite risk score is generated as follows: ； in, is the generated comprehensive risk score, n is the number of sentiment feature vectors, is the i-th sentiment feature vector, is the risk weight corresponding to the i-th emotional feature vector, m is the number of psychological feature vectors, is the jth psychological feature vector, is the risk weight corresponding to the jth psychological characteristic vector, is the number of traditional financial feature vectors, is the kth traditional financial feature vector, is the risk weight corresponding to the kth financial characteristic vector.

2. According to claim 1, a method for analyzing financial user emotions and psychological portraits based on a large language model is characterized in that: The pre-processing method is as follows: A denoising algorithm is used to remove meaningless characters in user text data, and spelling errors in user text data are corrected through a spelling correction model based on the edit distance algorithm. At the same time, the user text data is segmented into sentences with the help of natural language processing tools, and word segmentation is performed in combination with rules and statistical methods. The two-way maximum matching algorithm is then used to optimize the Chinese word segmentation effect. The key semantic components in the user text data are identified through a pre-trained qualified part-of-speech tagging model, and entity recognition technology is used to extract entity information in the user text data.

3. According to claim 1, a method for analyzing financial user emotions and psychological portraits based on a large language model is characterized in that: The emotion classification results include anger, anxiety, and satisfaction.

4. According to claim 1, a method for analyzing financial user emotions and psychological portraits based on a large language model is characterized in that: Also includes: Risk weights are determined based on subsequent user behavior data, time, and industry experience. Make dynamic adjustments, specifically, adjust risk weights based on time The specific method is as follows: ； in, is the risk weight adjusted according to time, α is the adjustment parameter, and t is the time interval of sentiment classification results; Adjust risk weights based on subsequent user behavior data The specific method is as follows: ； in, To adjust the risk weight based on subsequent user behavior data, To adjust the step size, is the amount of change in the comprehensive risk score; Adjust risk weights based on industry experience The specific method is as follows: For different user groups, adjust weights based on the statistical laws of historical data.

5. According to claim 1, a method for analyzing financial user emotions and psychological portraits based on a large language model is characterized in that: In step 5, the comprehensive risk score is displayed through a scoring dashboard, and segmented color warnings are used to intuitively reflect the risk level.

Images