Emotion recognition-based investment advisor strategy adaptation method and device, equipment and medium

Abstract

The invention relates to the technical field of computers, and discloses an investment advisor strategy adaptation method, device and equipment based on emotion recognition and a medium, and the method comprises the steps: analyzing the voice input data of a user through a multi-mode emotion perception model, capturing the voice emotion characteristics of the user in real time, and precisely determining the current emotion anxiety index. The emotion market coupling model is combined to associate the user emotion with the market volatility, the market volatility is dynamically analyzed, then the asset allocation strategy is determined according to the user emotion anxiety index and the market volatility, timely adjustment of the asset allocation strategy is achieved, the emotion change of an investor can be accurately captured in real time, the investment strategy is timely adjusted, and the investment efficiency is improved. Therefore, the accuracy and timeliness of investment decision making are improved. Especially when the financial market fluctuates severely, the scheme can respond quickly and ensure that an investor can flexibly adjust asset configuration in the optimal warehouse adjustment window period, so that the investment income is effectively improved, and the competitiveness in a complex and changeable financial environment is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method, device, equipment and medium for investment advisory strategy adaptation based on emotion recognition. Background Art

[0002] In today's booming financial landscape, personalized investment recommendation services have become a key area of ​​competition for many financial institutions. As financial markets become increasingly complex and volatile, investors are increasingly demanding accurate, real-time investment advice tailored to their risk preferences.

[0003] Traditional risk assessment methods primarily rely on historical user trading data to infer risk preferences. However, historical trading data only reflects past investment behavior and struggles to capture real-time changes in user sentiment amidst current market fluctuations. Traditional risk assessments often rely on user-completed risk assessment questionnaires, such as the RQA-9. While these questionnaires can provide a preliminary risk categorization, their results are relatively static and fail to capture subtle fluctuations in user sentiment as the market environment evolves. Consequently, it is difficult to adjust investment recommendation strategies to suit users' true risk tolerance and investment intentions.

[0004] To address the shortcomings of relying solely on historical trading data and questionnaire assessments, some technologies attempt to perceive user sentiment by analyzing text chat logs. However, existing sentiment analysis technology, when applied to the financial sector, cannot perceive investor sentiment changes in real time, resulting in a lag in recommended strategies. Financial markets are characterized by rapid market fluctuations, and missing the optimal time to adjust positions can cause investors to miss out on opportunities or be unable to mitigate risks in a timely manner, negatively impacting investment returns.

[0005] Therefore, in the field of financial investment, inefficient investment strategy adjustments due to untimely emotional perception feedback have become a technical problem that needs to be solved urgently. Summary of the Invention

[0006] The present invention provides an investment advisory strategy adaptation method, device, computer equipment and medium based on emotion recognition to solve the technical problem of inefficient investment strategy adjustment caused by untimely emotion perception feedback.

[0007] First, a method for investment advisory strategy adaptation based on emotion recognition is provided, including:

[0008] Based on the multimodal emotion perception model, feature recognition is performed on the user's voice input data to obtain multimodal emotion features;

[0009] Determining the user's current emotional anxiety index based on the multimodal emotional characteristics;

[0010] Using the multimodal emotion feature as input to an emotion-market coupling model, and determining a current market volatility corresponding to the current emotion anxiety index;

[0011] An asset allocation strategy is determined based on the current emotional anxiety index and the current market volatility.

[0012] In a second aspect, an investment advisory strategy adaptation device based on emotion recognition is provided, comprising:

[0013] A feature recognition module is used to perform feature recognition on the user's voice input data based on a multimodal emotion perception model to obtain multimodal emotion features;

[0014] An emotional anxiety index determination module, configured to determine a user's current emotional anxiety index based on the multimodal emotional features;

[0015] a market volatility determination module, configured to use the multimodal emotion feature as an input to an emotion-market coupling model to determine a current market volatility corresponding to the current emotion anxiety index;

[0016] A strategy determination module is used to determine an asset allocation strategy based on the current emotional anxiety index and the current market volatility.

[0017] In a third aspect, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned investment advisory strategy adaptation method based on emotion recognition are implemented.

[0018] In a fourth aspect, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the steps of the above-mentioned investment advisory strategy adaptation method based on emotion recognition are implemented.

[0019] In the solution implemented by the aforementioned emotion recognition-based investment advisory strategy adaptation method, device, computer equipment, and storage medium, a multimodal emotion perception model analyzes user voice input data, capturing the user's voice emotional characteristics in real time and accurately determining the current emotional anxiety index. This is combined with an emotion-market coupling model to link user emotions with market volatility, dynamically analyzing market volatility. Asset allocation strategies are then determined based on the user's emotional anxiety index and market volatility, enabling timely adjustments to these strategies. This allows for more real-time and accurate capture of investor sentiment changes, enabling timely adjustments to investment strategies, and thus improving the accuracy and timeliness of investment decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0021] Figure 1 1 is a schematic diagram of an application environment of an investment advisory strategy adaptation method based on emotion recognition in one embodiment of the present invention;

[0022] Figure 2 A flowchart of a first embodiment of an investment advisory strategy adaptation method based on emotion recognition provided by an embodiment of the present invention;

[0023] Figure 3 A schematic diagram of the data processing flow of the investment advisory strategy adaptation method based on emotion recognition provided by an embodiment of the present invention;

[0024] Figure 4 1 is a structural diagram of an investment advisory strategy adaptation device based on emotion recognition in one embodiment of the present invention;

[0025] Figure 5 is a structural diagram of a computer device in one embodiment of the present invention;

[0026] Figure 6 FIG. 2 is another structural diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0028] The investment advisory strategy adaptation method based on emotion recognition provided by the embodiment of the present invention can be applied in Figure 1 In an application environment, a client communicates with a server via a network. After receiving voice input data collected by the client, the server can perform feature recognition on the user's voice input data based on a multimodal emotion perception model to obtain multimodal emotion features; determine the user's current emotion anxiety index based on the multimodal emotion features; use the multimodal emotion features as input to an emotion-market coupling model to determine the current market volatility corresponding to the current emotion anxiety index; and determine an asset allocation strategy based on the current emotion anxiety index and the current market volatility.

[0029] In this invention, the technical problem of inefficient investment strategy adjustment caused by untimely emotion perception feedback in the financial investment field is addressed. A multimodal emotion perception model is used to analyze user voice input data, capturing the user's voice emotion characteristics in real time and accurately determining the current emotional anxiety index. This is combined with an emotion-market coupling model to link user emotions with market volatility, dynamically analyzing market volatility. Asset allocation strategies are then determined based on the user's emotional anxiety index and market volatility, enabling timely adjustments to asset allocation strategies. This allows for more real-time and accurate capture of investor sentiment changes, allowing for timely adjustments to investment strategies, thereby improving the accuracy and timeliness of investment decisions.

[0030] The client can be, but is not limited to, various personal computers, laptops, smartphones, tablet computers, and portable wearable devices. The server can be implemented as an independent server or a server cluster consisting of multiple servers. The present invention is described in detail below through specific embodiments.

[0031] See also Figure 2 As shown, Figure 2 The flowchart of the first embodiment of the investment advisory strategy adaptation method based on emotion recognition provided by the embodiment of the present invention includes the following steps:

[0032] S101: Based on the multimodal emotion perception model, perform feature recognition on the user's voice input data to obtain multimodal emotion features;

[0033] In one embodiment, user voice input data can be recorded using a voice acquisition device (such as a microphone or mobile phone) at an appropriate sampling rate (such as the common 16kHz, 44.1kHz, etc.) and quantization bit number (such as 16 bits, 24 bits, etc.). Alternatively, voice communication software or voice assistant platforms can be used to obtain user voice input in natural conversation scenarios or specific task scenarios.

[0034] The user voice input data can be real-time collected voice data or pre-recorded voice data.

[0035] In one embodiment, user voice input data primarily includes the user's voice signal itself, which is used to extract acoustic emotion feature information, and the corresponding conversation text, which is used to extract semantic emotion features. Additionally, auxiliary information such as the voice input timestamp and speaking rate may be recorded.

[0036] In one embodiment, a multimodal emotion perception model is a model that integrates multiple modal data (such as speech, text, images, and video) to identify and analyze human emotional states. By leveraging the rich and complementary emotional information carried by these different modal data, a more comprehensive, accurate, and in-depth perception and understanding of emotional states is achieved.

[0037] Specifically, the multimodal emotion perception model receives data inputs of different modalities, such as speech signals, text content, facial images, etc.

[0038] The multimodal emotion perception model uses a corresponding feature extraction method for each modal data type. For speech data, it extracts acoustic emotion features, such as fundamental frequency perturbation and spectral tilt. For text data, it extracts semantic emotion features, such as using the BERT model in the financial field to parse conversation content and obtain the weights of risk-sensitive words. For image or video data, it extracts facial expression features, such as muscle movements in the eyebrows, eyes, and nose.

[0039] Multimodal emotion perception models can design effective fusion strategies, such as gated attention mechanism and weighted averaging, to fuse features of different modalities and form multimodal emotion features that comprehensively reflect the emotional state.

[0040] Multimodal emotion perception models can use machine learning or deep learning algorithms such as classification and regression to analyze, classify or predict emotional states based on multimodal emotion features, and output results such as emotion labels (such as anger, joy, sadness, etc.) or emotion intensity.

[0041] Therefore, the multimodal emotion perception model can be applied to a variety of application fields and scenarios based on emotion recognition, such as intelligent customer service, financial risk assessment, intelligent driving, medical health, education and other fields, and can flexibly select and combine different modal data for emotion perception according to specific application requirements.

[0042] Furthermore, the user's voice input data is obtained; based on the multimodal emotion perception model, the acoustic emotion features and semantic emotion features in the voice input data are identified; based on the attention mechanism algorithm, the acoustic emotion features and the semantic emotion features are fused to obtain the multimodal emotion features.

[0043] In one embodiment, the user's voice input data is pre-processed, such as filtering and denoising, endpoint detection (determining the effective part of the voice), etc., to improve the signal quality.

[0044] Signal processing techniques, such as short-time Fourier transform (STFT) and Mel-Frequency Cepstral Coefficients (MFCC), are used to extract acoustic emotional features from speech signals. These include fundamental frequency (reflecting the pitch of the voice; it increases and fluctuates more when excited), amplitude (related to loudness; it increases with tension), energy (reflecting the intensity of the emotion), spectrum (shape changes under different emotions), duration (for example, sentences become longer when hesitant), fundamental frequency perturbation (emotions such as anger and anxiety increase fundamental frequency perturbation), and spectral tilt.

[0045] For example, by calculating the fundamental frequency value of the speech signal in each short time period, its changing trend and disturbance situation can be analyzed; the distribution of spectrum energy in different frequency bands can be calculated to obtain the spectrum tilt, etc.

[0046] Automatic speech recognition (ASR) technology is used to transcribe speech signals into text. The transcribed text is then fed into a pre-trained semantic analysis model (such as the BERT model in the financial field). Based on the text content and context, the model calculates the TF-IDF weights of risk-sensitive words (such as "plunge" and "stable") as semantic sentiment features, reflecting the importance of sentiment-related semantic information in the text.

[0047] Furthermore, based on a voiceprint feature extraction algorithm, the voiceprint features in the voice input data are extracted to obtain the acoustic emotion features; the voice input data is converted into voice text data; based on a semantic recognition model, the emotional semantic information in the voice text data is identified to obtain the semantic emotion features.

[0048] In one embodiment, voiceprint features may include LPCC (linear predictive cepstral coefficient), MFCC (Mel Frequency Cepstral Coefficient), short-term energy, zero-crossing rate, fundamental frequency, fundamental frequency perturbation, etc. These features can uniquely identify a speaker's voice characteristics, just like a fingerprint.

[0049] For example, LPCC, based on a physical model of the vocal tract, uses linear prediction to analyze speech signals and derive vocal tract parameters. This method describes the resonance characteristics of the vocal tract, which are related to the speaker's speech habits and emotional state. For example, when someone is emotionally agitated, the tension in the vocal tract changes, affecting the LPCC parameters.

[0050] MFCCs mimic the human ear's sensitivity to different frequencies, converting speech signals to the Mel-frequency scale and then calculating cepstral coefficients. MFCCs effectively capture the spectral characteristics of speech and are widely used in fields such as voiceprint recognition and emotion recognition. In emotion recognition, changes in MFCCs can reflect emotional fluctuations; for example, certain MFCC coefficients may increase in anger.

[0051] In one embodiment, acoustic emotion features primarily reflect emotional states from an acoustic perspective, such as changes in fundamental frequency (pitch and fluctuation), energy (sound intensity), spectral tilt (frequency component distribution trend), and zero-crossing rate. For example, when a person is nervous, their voice energy may increase, their zero-crossing rate may also increase, and their fundamental frequency may fluctuate erratically.

[0052] Specifically, the voiceprint feature extraction algorithm is used to extract acoustic emotion features from speech signals, such as fundamental frequency disturbance, spectrum tilt, and other acoustic emotion features, such as Mel-frequency cepstral coefficients (MFCC), linear prediction cepstral coefficients (LPCC), zero crossing rate (ZCR), energy, etc. These features can also reflect the emotional information of speech from different aspects.

[0053] For example, fundamental frequency refers to the basic frequency in a speech signal. It is closely related to the frequency of vocal cord vibration and reflects the pitch of the voice. Fundamental frequency perturbation refers to the degree of small, rapid, and irregular fluctuations in the fundamental frequency over time. This perturbation can, to some extent, reflect the speaker's emotional state. For example, when a person is anxious, nervous, or angry, their fundamental frequency perturbation typically increases.

[0054] First, the speech signal is preprocessed, including filtering and noise removal, voice activity detection, and other methods to improve signal quality and identify the valid parts of the speech. Then, the fundamental frequency (FFB) is calculated using methods based on the autocorrelation function or the cepstral domain. Based on this FFB calculation, its temporal variation is further analyzed, such as the difference between adjacent FFB values ​​or the fluctuation amplitude, to obtain a quantitative indicator of FFB disturbance.

[0055] Among them, the calculation formula of the fundamental frequency disturbance degree can be expressed as:

[0056]

[0057] Among them, F0 i Represents the fundamental frequency value of the i-th frame in the speech signal; i represents the index from the 1st to the N-3th frame (because the calculation needs to consider the current frame and the two frames after it, it can only go to N-3 to ensure that there is enough data for calculation); j represents the average fundamental frequency used to calculate from i-1 to i+2; N represents the total number of frames of the speech signal.

[0058] Spectral tilt reflects the distribution of the speech signal's spectral energy across different frequency bands. Specifically, it indicates the degree to which the spectral amplitude tilts as frequency increases. The spectral tilt of speech varies with different emotions. For example, when someone is angry or excited, the energy in the high-frequency bands may increase, causing the spectral tilt to change.

[0059] Extracting spectral tilt also requires preprocessing the speech signal. Next, using time-frequency analysis methods such as the Short-Time Fourier Transform (STFT) or wavelet transform, the speech signal is converted from the time domain to the frequency domain, obtaining its spectral representation. A specific range of frequency bands is then selected from the spectrum, and the energy or amplitude of each frequency band is calculated. A line or curve is then fitted that reflects the frequency-dependent variation of the spectral amplitude. By calculating parameters such as the slope of this line or curve, the quantified value of the spectral tilt is obtained.

[0060] The calculation formula of the spectrum tilt can be expressed as:

[0061]

[0062] Where S(k) represents the value of the spectral amplitude of the speech signal at the kth frequency component. k represents the frequency component index from 1 to K. K is the total number of frequency components, which usually depends on the spectral resolution and the sampling rate of the speech signal.

[0063] It is understandable that, based on actual application needs, voiceprint feature extraction algorithms can be used to extract other acoustic emotion features, such as fundamental frequency, amplitude, energy, speech rhythm characteristics, and Mel-frequency cepstral coefficients (MFCCs), which together constitute the rich emotional information in the speech signal. By properly extracting and analyzing these acoustic emotion features, accurate acoustic-level emotion input can be provided to the multimodal emotion perception model, thereby achieving precise recognition and analysis of user emotions.

[0064] In one embodiment, while extracting acoustic emotion features, the speech input data can also be preprocessed. The preprocessed speech input data is input into a speech recognition tool (such as Kaldi, DeepSpeech, etc.). The speech recognition tool outputs the corresponding speech-to-text data based on the speech-to-text mapping relationship learned during training. The converted speech-to-text data is preprocessed, including operations such as word segmentation, stop word removal, and part-of-speech tagging. The text is converted into a format suitable for semantic analysis and the key semantic units in the text are extracted.

[0065] Semantic recognition models, such as the BERT model in the financial field, are used to perform semantic analysis on the preprocessed text. These models identify risk-sensitive keywords in the text. These keywords typically include words that may reflect market and investment risks, such as "plunge," "stable," "risk," "return," "loss," and "bubble."

[0066] For the identified risk-sensitive words, their TF-IDF values ​​are calculated as weights to reflect their importance in the text and their influence on sentiment.

[0067] Specifically, we count the number of times each risk-sensitive word appears in a document and divide it by the total number of words in the document to obtain the word's term frequency (TF). Within the context of a financial text corpus, we calculate the IDF value for each risk-sensitive word. IDF is a measure of the general importance of a word within a collection of documents. Specifically, IDF = log(total number of documents / number of documents containing the risk-sensitive word).

[0068] Multiplying the TF value and IDF value of each risk-sensitive word gives its TF-IDF weight, which comprehensively reflects the importance of the risk-sensitive word in the current text and its potential contribution to the text's sentiment.

[0069] Based on the TF-IDF weight of risk-sensitive words, the sentiment tendency of the text (positive, negative, neutral) can be analyzed to obtain a sentiment tendency score or label, which further reflects the emotional tendency contained in the text.

[0070] Among them, the semantic recognition model (such as the BERT model in the financial field) can be a pre-trained language model based on the Transformer architecture. It is pre-trained on large-scale text data in the financial field and learns the rich semantic information and language features of financial text.

[0071] The obtained semantic emotion features are integrated with emotion features of other modalities such as acoustic emotion features to form a more comprehensive and accurate emotion feature representation, providing a basis for subsequent emotion recognition, analysis, and decision-making tasks. For example, in an intelligent financial customer service system, service strategies can be adjusted in a timely manner according to the user's emotional characteristics, or in an investment decision-making support system, asset allocation recommendations can be made based on market sentiment and investor sentiment.

[0072] Furthermore, based on the attention mechanism algorithm, the first weight corresponding to the acoustic emotion feature and the second weight corresponding to the semantic emotion feature are calculated; based on the feature conversion algorithm, the acoustic emotion feature and the semantic emotion feature are converted to the same target dimension; based on the first weight corresponding to the acoustic emotion feature and the second weight corresponding to the semantic emotion feature, the acoustic emotion feature and the semantic emotion feature converted to the same target dimension are weightedly fused to obtain the multimodal emotion feature.

[0073] In one embodiment, if Figure 3 As shown, a cross-modal attention fusion mechanism, such as a gated attention mechanism, can be designed. The weights of acoustic and semantic emotion features in the fused feature vector are calculated using learnable parameters (such as gated control parameters) and activation functions (such as the sigmoid function). During training, the model automatically learns the importance of acoustic and semantic emotion features under different emotional states and assigns appropriate weights to each feature.

[0074] Before weighted fusion of acoustic emotion features and semantic emotion features, they need to be converted into features of the same dimension.

[0075] Specifically, a fully connected layer can be used to transform features, mapping features of different dimensions to the same target dimension d. Based on the calculated weights, the acoustic and semantic emotion features transformed to the same target dimension are weighted and summed to produce the final multimodal emotion features. These fused features combine emotional information from both acoustic and semantic modalities, enabling a more comprehensive and accurate reflection of the user's emotional state.

[0076] Specifically, a gated attention mechanism is designed to dynamically weight acoustic and semantic emotion features:

[0077] α=σ(W g [f voice ;f text ]), f fusion =αf voice +(1-α)f text

[0078] Among them, σ is the Sigmoid function, which maps the input to the (0,1) interval and generates the weight coefficient.

[0079] is a learnable parameter with a size of d×(d v +d t ), d is the dimension of the output feature vector (that is, the fused feature vector f fusion dimension), d v and d t Represent the acoustic emotion feature vector f voice and semantic sentiment feature vector f text Dimensions. voice ;f text ] represents the acoustic emotion feature vector f voice and semantic sentiment feature vector f text α represents the weight coefficient, which is used to dynamically weight the acoustic emotion features and semantic emotion features. voice Represents the acoustic emotion feature vector, which contains the acoustic emotion features of the semantic signal, such as fundamental frequency perturbation and spectrum tilt; f text Represents the semantic emotion feature vector, which contains the semantic emotion features of the speech text, such as the weight of risk-sensitive words. fusion Represents the fused feature vector, i.e., the multimodal emotion feature.

[0080] S102: Determine the user's current emotional anxiety index based on the multimodal emotional characteristics;

[0081] In one embodiment, a sentiment classification model is constructed to map multimodal sentiment features to sentiment categories. This model can be based on machine learning (e.g., support vector machines, random forests) or deep learning (e.g., recurrent neural networks, Transformer architectures). The model inputs are multimodal sentiment features, and the output is sentiment categories (e.g., "anxious," "calm," "angry," etc.) and their corresponding probabilities.

[0082] Based on multimodal emotional features, an emotion classification model can be used to identify the user's current emotion type and identify the type and probability of anxiety-related emotions. For example, if the emotion classification model outputs an 80% probability of "anxiety," this indicates that the user's current emotional state has a high tendency towards anxiety.

[0083] To further quantify the level of anxiety, an anxiety index can be designed, typically ranging from 0 to 1 (or 0% to 100%). This index can be calculated based on the anxiety probability output by the emotion classification model. For example, the anxiety index can directly use the probability value of anxiety, or convert the probability value into an anxiety index through a mapping function.

[0084] Based on the output of the emotion classification model and the quantification method of the anxiety level, the user's current emotional anxiety index is determined. For example, if the model outputs an 80% probability of anxiety, the user's anxiety index can be determined to be 0.8 or 80%.

[0085] Based on a user's anxiety index, intelligent customer service systems can adjust their service strategies to provide more attentive service. In the financial sector, monitoring a user's anxiety index can be used to assess their risk tolerance and the stability of their investment decisions. In mental health monitoring, the anxiety index can help doctors or counselors promptly identify a user's anxiety and provide intervention recommendations.

[0086] S103: Using the multimodal emotion feature as input to an emotion-market coupling model to determine a current market volatility corresponding to the current emotion anxiety index;

[0087] In one embodiment, a sentiment-market coupling model is constructed that can capture the relationship between sentiment features and market volatility. The sentiment-market coupling model can be trained based on historical data, which should include historical multimodal sentiment features and corresponding market volatility data.

[0088] Furthermore, historical sentiment data and historical asset volatility data are collected; based on a feature extraction algorithm, feature extraction is performed on the historical sentiment data and the historical asset volatility data respectively to obtain historical sentiment features and historical market volatility features; based on the time correspondence between the historical sentiment features and the historical market volatility features, the correlation between the sentiment anxiety index and market volatility is determined; based on the correlation between the sentiment anxiety index and market volatility, the sentiment-market coupling model is constructed.

[0089] In one embodiment, historical sentiment data is collected from sources such as past voice interaction records, social media posts, and financial news commentary. This collected historical sentiment data should include user emotional expressions, such as textual descriptions or voice recordings of emotions like anxiety, anger, and joy. Historical asset price volatility data, such as historical stock price volatility and trading volume changes, is collected from historical financial market records. This historical asset volatility data can be obtained from financial databases, market analysis platforms, and other sources.

[0090] In one embodiment, feature extraction is performed on the collected historical data to obtain historical sentiment features and historical market volatility features.

[0091] Specifically, we use the aforementioned acoustic emotion feature extraction algorithms (such as fundamental frequency perturbation and spectral tilt) and semantic emotion feature extraction methods (such as using the BERT model in the financial field to parse conversation content and obtain risk-sensitive word weights) to process historical emotion data and obtain historical emotion features. We also calculate relevant features of historical asset volatility data, such as statistical indicators such as the mean, variance, maximum, and minimum values ​​of volatility, as well as the trend of volatility (such as rising, falling, and stable), to obtain historical market volatility characteristics.

[0092] Align historical sentiment and market volatility features in chronological order, ensuring that each sentiment data point corresponds to a corresponding market volatility data point. Calculate the correlation coefficient between the sentiment anxiety index and market volatility, and analyze the linear or nonlinear relationship between them. For example, you can use the Pearson correlation coefficient to measure linear correlation, or the mutual information coefficient to measure nonlinear correlation.

[0093] A regression model can be constructed with the anxiety index as the independent variable and market volatility as the dependent variable to determine the quantitative relationship between them. For example, a linear regression model can be used to fit the data and obtain the regression coefficient, which can quantify the impact of the anxiety index on market volatility.

[0094] Based on the characteristics of the relationship between the emotional anxiety index and market volatility, select an appropriate model architecture, such as a linear regression model, a support vector machine (SVM) regression model, a long short-term memory network (LSTM) model, or a Transformer architecture model.

[0095] The model is trained using aligned historical sentiment and market volatility data. During training, optimization algorithms (such as gradient descent) adjust the model parameters to accurately capture the relationship between the emotional anxiety index and market volatility. The dataset is divided into training and test sets. After training the model on the training set, it is validated on the test set to evaluate the model's predictive performance. Based on the validation results, the model is optimized, such as adjusting model parameters and adding regularization terms, to improve its accuracy and generalization.

[0096] For example, construct a multimodal emotion feature E t ∈[0,1] and market volatility σ t Real-time correlation model:

[0097]

[0098] in, represents the current weight value of the i-th model parameter at time step t. represents the weight value of the i-th model parameter at the previous moment at time step (t-1). η represents the learning rate, which is a hyperparameter that controls the step size of the model parameter update. Represents the joint loss function of multimodal sentiment features and market volatility, which measures the model's performance on multimodal sentiment features E t and market volatility σ t prediction error. Represents the loss function For the i-th model parameter w i The partial derivative of the loss function w i exp(·) represents the exponential function, which is used to convert the derivative term into a positive number to ensure that the weight update is always in the direction of reducing the loss.

[0099] The current multimodal emotional characteristics (including the emotional anxiety index) are input into the emotional market coupling model to predict the current market volatility in real time.

[0100] Specifically, multimodal sentiment features are fed into a trained sentiment-market coupling model, which then outputs a forecast of current market volatility. This forecast provides decision support for investors. For example, if market volatility is predicted to rise significantly, investors may be advised to adopt a conservative investment strategy, increasing the proportion of cash assets to reduce investment risk.

[0101] In the investment decision-making system, portfolio adjustments are made based on the current emotional anxiety index and corresponding market volatility forecasts to mitigate risk. Financial institutions can utilize the sentiment-market coupling model to monitor market sentiment changes in real time and take proactive risk prevention measures. By analyzing the emotional fluctuations of market participants, market volatility trends can be predicted, providing a reference for market analysis.

[0102] S104: Determine an asset allocation strategy based on the current emotional anxiety index and the current market volatility.

[0103] In one embodiment, multimodal fusion features can be used as input to a sentiment-market coupling model to determine the market volatility corresponding to the current sentiment anxiety index, thereby providing data support for financial decision-making. For example, in portfolio management, asset allocation can be automatically adjusted based on the sentiment anxiety index and market volatility to reduce investment risk and protect investment principal. This helps investors adjust their asset allocation in a timely manner during market instability to avoid significant losses caused by market fluctuations. Automating the monitoring of the sentiment anxiety index and market volatility, along with asset allocation adjustments, improves the timeliness and accuracy of investment decisions.

[0104] For example, financial institutions can use sentiment-market coupling models to monitor market sentiment changes in real time and take proactive risk prevention measures. For example, if they detect a general rise in market sentiment anxiety index, they can adjust their risk exposure in advance to reduce potential losses.

[0105] By analyzing the sentiment of market participants, we can predict market fluctuations and provide a reference for market analysis. For example, by combining the prediction results of the sentiment-market coupling model, we can issue market volatility warning reports to help investors make more informed decisions.

[0106] In the investment decision-making system, the current emotional anxiety index and the corresponding market volatility forecast are used to adjust the investment portfolio and reduce risk. For example, when the emotional anxiety index is high and the market volatility forecast is high, it is recommended to increase the proportion of cash assets.

[0107] In one embodiment, a first threshold corresponding to the emotional anxiety index and a second threshold corresponding to the market volatility are obtained; when the current emotional anxiety index is greater than the first threshold and the current market volatility is greater than the second threshold, a strategy adjustment mechanism of the asset allocation strategy is triggered; when the strategy adjustment mechanism is triggered, the current asset allocation strategy is determined based on the current emotional anxiety index and the current market volatility.

[0108] In one embodiment, based on historical data and market experience, a first threshold corresponding to the emotional anxiety index and a second threshold corresponding to market volatility can be used to trigger different asset allocation strategies. The first and second thresholds can each be multiple, or multiple combinations of the first and second thresholds. Different combinations of the first and second thresholds correspond to different asset allocation strategies.

[0109] For example, a real-time portfolio adjustment strategy can be constructed, that is, the proportion of cash assets can be adjusted in real time according to the current emotional anxiety index and current market volatility:

[0110] Δw cash =β·(E t -0.5)·σ t , β = 0.15 (adjustment factor)

[0111] Where Δw cash Refers to the adjustment amount of the cash-like asset ratio, indicating the proportion of cash-like assets that needs to be increased or decreased at the current time step. β refers to the adjustment factor, which is a preset constant used to control the adjustment amplitude. By setting the value of β, the intensity of the cash-like asset ratio adjustment can be adjusted when the conditions are met. The larger the value, the greater the adjustment amplitude. t Refers to the current emotional anxiety index, which ranges from [0,1] and reflects the current level of emotional anxiety of market participants. t The current market volatility reflects the volatility of the current market, usually expressed as a percentage.

[0112] For example, the emotional anxiety index threshold can be set to 0.8. t >0.8 indicates that the current sentiment is relatively anxious; the market volatility threshold can be set at 20%. t >20% indicates greater market volatility and higher risk.

[0113] In the current mood anxiety index E t >0.8 and market volatility σ t When the cash ratio is greater than 20%, calculate the adjustment amount Δw of the cash assets ratio. cash According to the calculated Δw cash , adjust the proportion of cash assets.

[0114] In one embodiment, based on the adjustment of the proportion of cash-like assets, the proportion of other assets (such as stocks, bonds, etc.) is adjusted accordingly to maintain the total asset allocation at 100%. For example, if the proportion of cash-like assets increases, the proportion of stocks and / or bonds is reduced.

[0115] Execute actual investment operations, such as buying or selling assets, based on the adjusted asset allocation ratio.

[0116] It is understandable that the asset allocation strategy can also be adjusted according to different emotional anxiety index thresholds and market volatility thresholds. For example, when the emotional anxiety index is less than a first threshold and the market volatility is less than a second threshold, the proportion of cash assets can be reduced, and the adjustment amount of the reduction in cash assets can be calculated. Correspondingly, the proportion of stocks and / or bonds can be increased.

[0117] As can be seen, in the above solution, the multimodal emotion perception model, by analyzing user voice input data, can more accurately identify user emotions and promptly reflect their real-time emotional state during market fluctuations. Furthermore, combined with the emotion-market coupling model, it can link user emotions with market volatility, and then determine asset allocation strategies based on the user's emotional anxiety index and market volatility. This strategy determination method based on multimodal emotion perception enables financial institutions to more accurately capture changes in investor sentiment in real time and adjust investment strategies in a timely manner, thereby improving the accuracy and timeliness of investment decisions and avoiding investment losses caused by delayed emotion perception.

[0118] In financial market decision-making, formulating portfolio adjustment strategies based on users' multimodal emotional characteristics and market volatility does require the in-depth participation of massive user emotional data to ensure that the extracted multimodal emotional characteristics are highly accurate, thereby ensuring the universality and application value of the investment market.

[0119] User sentiment data is collected through various channels, including but not limited to financial trading platforms, social media, and market research institutions. This data comes from a wide range of sources, covering different user groups and market scenarios, and improving the representativeness and comprehensiveness of the data.

[0120] It's important to ensure that the data collection process complies with relevant laws and regulations and respects user privacy. When collecting user sentiment data, clearly inform users of the data's purpose and protection measures, and obtain their explicit authorization. Collected data should be anonymized to remove personally identifiable information, ensuring user privacy is protected.

[0121] In the application of multimodal emotion perception and asset allocation, combining privacy-preserving model updates and cold-start optimization steps can ensure user data privacy while improving model performance and user experience.

[0122] In one embodiment, an encrypted gradient of a multimodal emotion perception sub-model running on at least one user device is received; based on a decryption key, the encrypted gradient is decrypted to obtain gradient information of each of the multimodal emotion perception sub-models; the gradient information of each of the multimodal emotion perception sub-models is aggregated, and the multimodal emotion perception model is updated to optimize the multimodal emotion perception model.

[0123] Specifically, each user device runs a multimodal emotion perception sub-model, which is trained locally using multimodal emotion data collected on the user device (such as acoustic emotion features of speech and semantic emotion features of text). This way, the user's original data never leaves the device, protecting user privacy.

[0124] During local training, the multimodal emotion perception sub-model calculates gradient information (i.e., the direction and magnitude of model parameter updates). This gradient information is then encrypted to prevent interception and interpretation during transmission. This encryption process utilizes advanced algorithms to ensure the security of gradient information during transmission.

[0125] The encrypted gradient information is uploaded to the cloud server, where it is decrypted and aggregated. The cloud server uses the decryption key to decrypt the encrypted gradient information uploaded by each user device and then calculates the average of these gradients to update the global emotion recognition model. This process ensures that data from individual user devices is not directly exposed while allowing the multimodal emotion perception model to learn from the gradients of the multimodal emotion perception sub-models across multiple user devices.

[0126] For example, the user device locally trains the multimodal emotion perception sub-model and only uploads the encrypted gradient information. Aggregate to the cloud server:

[0127]

[0128] in, It represents the global model parameters updated in the (t+1)th iteration. After this iterative optimization, the latest global model parameters obtained by all user devices will be used for subsequent model training and prediction tasks.

[0129] Represents the global model parameters at the tth iteration, which are the basis for the model parameters at the beginning of this iteration. All user devices train and update their local models based on this parameter.

[0130] This operation averages the gradients uploaded by N user devices, where N is the total number of user devices participating in this iteration. This operation integrates the contributions of all user devices, ensuring that the updated global model parameters reflect the characteristics of all device data and preventing a single device's data from overly influencing the global model.

[0131] Represents the encrypted gradient calculated for the i-th user device during the t-th iteration The gradient value obtained after the decryption operation.

[0132] It represents the encrypted gradient calculated by the i-th user device in the t-th iteration, which is an encrypted representation of the model parameter update direction and amplitude calculated by the user device based on local data and the current global model parameters.

[0133] This embodiment, through privacy-preserving model updates, aggregates sentiment data from a large number of users, improving the accuracy of multimodal sentiment feature extraction. This makes portfolio adjustment strategies based on sentiment features and market volatility more reliable and effective. Accurate sentiment features facilitate more precise predictions of market fluctuations, thereby optimizing asset allocation decisions. With the continuous collection of user sentiment data and continuous model updates, portfolio adjustment strategies can dynamically adapt to changes in market sentiment and the evolution of user sentiment features, enabling investors to adjust their asset allocations in a timely manner, respond to market fluctuations, and mitigate investment risks.

[0134] For new users, the cloud server initially lacks personalized emotional data. To address this issue, the cloud server can use clustering algorithms to analyze the emotional data of a large number of existing users and generate multiple group emotional profiles. These profiles represent different user groups with similar emotional characteristics.

[0135] When a new user joins, the cloud server assigns them a closest group emotion profile for initial adaptation based on a matching strategy (such as the user's basic characteristics or initial interaction behavior). Over the next seven days, the cloud server gradually collects the new user's emotion data and trains a personalized model based on this data. Over time, the cloud server reduces its reliance on the group emotion profile and increases the weight of the personalized model, until it fully transitions to a personalized emotion recognition model based on the new user's own data.

[0136] Through cold start optimization, new users can quickly obtain relatively accurate emotion recognition services even in the initial stage when personalized data is lacking, thereby allowing more users to participate in the portfolio adjustment strategy based on emotional characteristics, and improving the universality and application scope of the strategy among different user groups.

[0137] This embodiment improves the performance of the emotion recognition model while protecting user privacy through privacy-preserving model updates and cold start optimization, ensuring the accuracy and universality of the portfolio adjustment strategy, and enabling investors to make better-informed investment decisions based on market sentiment and volatility.

[0138] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0139] In one embodiment, a device for adapting investment advisory strategies based on emotion recognition is provided. The device for adapting investment advisory strategies based on emotion recognition corresponds to the method for adapting investment advisory strategies based on emotion recognition in the above embodiment. Figure 4 As shown, the investment advisory strategy adaptation device based on emotion recognition includes: a feature recognition module 201, an emotional anxiety index determination module 202, a market volatility determination module 203 and a strategy determination module 204. The functional modules are described in detail as follows:

[0140] A feature recognition module 201 is used to perform feature recognition on the user's voice input data based on a multimodal emotion perception model to obtain multimodal emotion features;

[0141] Anxiety index determination module 202, configured to determine a user's current anxiety index based on the multimodal emotion features;

[0142] A market volatility determination module 203 is configured to use the multimodal emotion feature as an input to an emotion-market coupling model to determine a current market volatility corresponding to the current emotion anxiety index;

[0143] The strategy determination module 204 is configured to determine an asset allocation strategy based on the current emotional anxiety index and the current market volatility.

[0144] In one embodiment, the feature recognition module 201 includes:

[0145] A data acquisition unit, configured to acquire user's voice input data;

[0146] a feature recognition unit, configured to recognize acoustic emotion features and semantic emotion features in the speech input data based on the multimodal emotion perception model;

[0147] A feature fusion unit is used to fuse the acoustic emotion feature and the semantic emotion feature based on an attention mechanism algorithm to obtain the multimodal emotion feature.

[0148] In one embodiment, the feature recognition unit includes:

[0149] an acoustic emotion feature recognition subunit, configured to extract the voiceprint feature from the speech input data based on a voiceprint feature extraction algorithm to obtain the acoustic emotion feature;

[0150] A speech-to-text conversion subunit, configured to convert the speech input data into speech-to-text data;

[0151] The semantic emotion feature recognition subunit is used to recognize the emotion semantic information in the speech text data based on the semantic recognition model to obtain the semantic emotion feature.

[0152] In one embodiment, the feature fusion unit includes:

[0153] a weight calculation subunit, configured to calculate, based on the attention mechanism algorithm, a first weight corresponding to the acoustic emotion feature and a second weight corresponding to the semantic emotion feature;

[0154] a feature conversion subunit, configured to convert the acoustic emotion feature and the semantic emotion feature into the same target dimension based on a feature conversion algorithm;

[0155] A feature fusion subunit is used to perform weighted fusion on the acoustic emotion feature and the semantic emotion feature converted into the same target dimension based on the first weight corresponding to the acoustic emotion feature and the second weight corresponding to the semantic emotion feature to obtain the multimodal emotion feature.

[0156] In one embodiment, the investment advisory strategy adaptation device based on emotion recognition further includes a coupling model construction module, including:

[0157] Historical data collection unit, used to collect historical sentiment data and historical asset fluctuation data;

[0158] A historical data feature extraction unit, configured to extract features from the historical sentiment data and the historical asset volatility data based on a feature extraction algorithm to obtain historical sentiment features and historical market volatility features;

[0159] a correlation determination unit, configured to determine a correlation between the emotional anxiety index and the market volatility based on a temporal correspondence between the historical emotional characteristics and the historical market volatility characteristics;

[0160] The coupling model construction unit is used to construct the emotion-market coupling model based on the correlation between the emotion anxiety index and the market volatility.

[0161] In one embodiment, the policy determination module 204 includes:

[0162] A threshold value acquisition unit, configured to acquire a first threshold value corresponding to the emotional anxiety index and a second threshold value corresponding to the market volatility;

[0163] a strategy adjustment mechanism triggering unit, configured to trigger a strategy adjustment mechanism of the asset allocation strategy when the current emotional anxiety index is greater than a first threshold and the current market volatility is greater than a second threshold;

[0164] A strategy determination unit is used to determine a current asset allocation strategy based on the current emotional anxiety index and the current market volatility when the strategy adjustment mechanism is triggered.

[0165] In one embodiment, the investment advisory strategy adaptation device based on emotion recognition further includes a model updating module, including:

[0166] An encrypted gradient receiving unit, configured to receive an encrypted gradient of a multimodal emotion perception sub-model running on at least one user device uploaded by the user device;

[0167] A gradient decryption unit, configured to decrypt the encrypted gradient based on a decryption key to obtain gradient information of each of the multimodal emotion perception sub-models;

[0168] A model updating unit is used to aggregate the gradient information of each of the multimodal emotion perception sub-models and update the multimodal emotion perception model to optimize the multimodal emotion perception model.

[0169] This invention provides an investment advisory strategy adaptation device based on emotion recognition. This device analyzes user voice input data using a multimodal emotion perception model, capturing the user's voice emotional characteristics in real time and accurately determining their current emotional anxiety index. This device, combined with an emotion-market coupling model, links user emotions with market volatility, dynamically analyzes market volatility, and then determines an asset allocation strategy based on the user's emotional anxiety index and market volatility. This allows for timely adjustments to asset allocation strategies, enabling more real-time and accurate capture of investor sentiment changes, allowing for timely adjustments to investment strategies, thereby improving the accuracy and timeliness of investment decisions.

[0170] The specific definitions of the emotion recognition-based investment advisory strategy adaptation device can be found in the definitions of the emotion recognition-based investment advisory strategy adaptation method described above and will not be repeated here. Each module in the emotion recognition-based investment advisory strategy adaptation device described above can be implemented in whole or in part via software, hardware, or a combination thereof. Each of these modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0171] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 5As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external client via a network connection. When the computer program is executed by the processor, it realizes the functions or steps on the service side of an investment advisory strategy adaptation method based on emotion recognition.

[0172] In one embodiment, a computer device is provided. The computer device may be a client, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, memory, network interface, display screen and input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it realizes the functions or steps on the client side of a method for adapting investment strategies based on emotion recognition.

[0173] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are performed:

[0174] Based on the multimodal emotion perception model, feature recognition is performed on the user's voice input data to obtain multimodal emotion features;

[0175] Determining the user's current emotional anxiety index based on the multimodal emotional characteristics;

[0176] Using the multimodal emotion feature as input to an emotion-market coupling model, and determining a current market volatility corresponding to the current emotion anxiety index;

[0177] An asset allocation strategy is determined based on the current emotional anxiety index and the current market volatility.

[0178] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0179] Based on the multimodal emotion perception model, feature recognition is performed on the user's voice input data to obtain multimodal emotion features;

[0180] Determining the user's current emotional anxiety index based on the multimodal emotional characteristics;

[0181] Using the multimodal emotion feature as input to an emotion-market coupling model, and determining a current market volatility corresponding to the current emotion anxiety index;

[0182] An asset allocation strategy is determined based on the current emotional anxiety index and the current market volatility.

[0183] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can be found in the relevant descriptions of the server side and the client side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0184] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0185] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0186] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A method for investment advisory strategy adaptation based on emotion recognition, characterized in that: The method comprises: Based on the multimodal emotion perception model, feature recognition is performed on the user's voice input data to obtain multimodal emotion features; Determining the user's current emotional anxiety index based on the multimodal emotional characteristics; Using the multimodal emotion feature as input to an emotion-market coupling model, and determining a current market volatility corresponding to the current emotion anxiety index; An asset allocation strategy is determined based on the current emotional anxiety index and the current market volatility.

2. The investment advisory strategy adaptation method based on emotion recognition according to claim 1, characterized in that: The method of performing feature recognition on the user's voice input data based on the multimodal emotion perception model to obtain multimodal emotion features includes: Get the user's voice input data; Based on the multimodal emotion perception model, identifying acoustic emotion features and semantic emotion features in the speech input data; Based on the attention mechanism algorithm, the acoustic emotion feature and the semantic emotion feature are fused to obtain the multimodal emotion feature.

3. The investment advisory strategy adaptation method based on emotion recognition according to claim 2, characterized in that: The identifying, based on the multimodal emotion perception model, acoustic emotion features and semantic emotion features in the speech input data includes: extracting the voiceprint features from the speech input data based on a voiceprint feature extraction algorithm to obtain the acoustic emotion features; Converting the voice input data into voice text data; Based on a semantic recognition model, the emotional semantic information in the speech text data is identified to obtain the semantic emotional features.

4. The investment advisory strategy adaptation method based on emotion recognition according to claim 2, characterized in that: The method of fusing the acoustic emotion feature and the semantic emotion feature based on the attention mechanism algorithm to obtain the multimodal emotion feature includes: Based on the attention mechanism algorithm, calculating a first weight corresponding to the acoustic emotion feature and a second weight corresponding to the semantic emotion feature; Based on a feature conversion algorithm, the acoustic emotion feature and the semantic emotion feature are converted to the same target dimension; Based on the first weight corresponding to the acoustic emotion feature and the second weight corresponding to the semantic emotion feature, the acoustic emotion feature and the semantic emotion feature converted into the same target dimension are weightedly fused to obtain the multimodal emotion feature.

5. The investment advisory strategy adaptation method based on emotion recognition according to claim 1, characterized in that: Before using the multimodal emotion feature as the input of the emotion-market coupling model to determine the current market volatility corresponding to the current emotion anxiety index, the method further includes: Collect historical sentiment data and historical asset volatility data; Based on a feature extraction algorithm, feature extraction is performed on the historical sentiment data and the historical asset volatility data to obtain historical sentiment features and historical market volatility features; Determining a correlation between the emotional anxiety index and market volatility based on a temporal correspondence between the historical emotional characteristics and the historical market volatility characteristics; Based on the correlation between the emotional anxiety index and market volatility, the emotional market coupling model is constructed.

6. The investment advisory strategy adaptation method based on emotion recognition according to claim 1, characterized in that: The determining of an asset allocation strategy based on the current emotional anxiety index and the current market volatility includes: Obtaining a first threshold corresponding to the emotional anxiety index and a second threshold corresponding to the market volatility; When the current emotional anxiety index is greater than a first threshold and the current market volatility is greater than a second threshold, triggering a strategy adjustment mechanism for the asset allocation strategy; When the strategy adjustment mechanism is triggered, a current asset allocation strategy is determined based on the current emotional anxiety index and the current market volatility.

7. The investment advisory strategy adaptation method based on emotion recognition according to claim 1, characterized in that: After determining the asset allocation strategy based on the current emotional anxiety index and the current market volatility, the method further includes: receiving an encrypted gradient of a multimodal emotion perception sub-model running on at least one user device uploaded by the user device; Decrypting the encrypted gradient based on a decryption key to obtain gradient information of each of the multimodal emotion perception sub-models; Aggregating the gradient information of each of the multimodal emotion perception sub-models, and updating the multimodal emotion perception model to optimize the multimodal emotion perception model.

8. An investment advisory strategy adaptation device based on emotion recognition, characterized in that: The investment advisory strategy adaptation device based on emotion recognition includes: A feature recognition module is used to perform feature recognition on the user's voice input data based on a multimodal emotion perception model to obtain multimodal emotion features; An emotional anxiety index determination module, configured to determine a user's current emotional anxiety index based on the multimodal emotional features; a market volatility determination module, configured to use the multimodal emotion feature as an input to an emotion-market coupling model to determine a current market volatility corresponding to the current emotion anxiety index; A strategy determination module is used to determine an asset allocation strategy based on the current emotional anxiety index and the current market volatility.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the investment advisory strategy adaptation method based on emotion recognition as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the investment advisory strategy adaptation method based on emotion recognition as described in any one of claims 1 to 7 are implemented.

Images