Asset allocation method and device based on dynamic and static characteristic data, equipment and medium

Abstract

The invention relates to the technical field of intelligent decision making, can be applied to business system platforms of financial science and technology, medical health and the like, and discloses an asset allocation method, device, equipment and medium based on dynamic and static characteristic data. Obtaining acoustic dynamic characteristics; converting the user voice data into a text, and performing semantic dynamic feature extraction on the text to obtain semantic dynamic features; performing gating attention dynamic weighting on the acoustic dynamic features and the semantic dynamic features to obtain weighted features; performing dynamic evaluation on the weighted features, and generating a warehouse adjustment instruction according to a dynamic evaluation result and a preset static feature data combination; and generating and sending an asset allocation suggestion according to a preset user authorization authority and the warehouse transfer instruction. And through comprehensive analysis of the dynamic features and the static features, dynamic operation of manual decision making is avoided.

Description

Technical Field

[0001] The present invention relates to the field of intelligent decision-making technology, and in particular to an asset allocation method, device, equipment and medium based on dynamic and static feature data. Background Art

[0002] Asset adjustment is a series of methods and means to reconfigure and adjust various types of assets in the investment portfolio based on specific goals, market conditions and relevant conditions of investors. Logically speaking, when investors formulate asset adjustment strategies, the strategy formation factors will not only be affected by static characteristic data such as the investor's current asset status and market volatility data, but also by dynamic characteristic data such as investment dynamic loss situation, expected income protection, and physical health status. For example, if the expected income is highly guaranteed, the investment account surplus is high, and the physical health is good, then the investor will tend to prefer a high-risk investment strategy; if the expected income protection is reduced, the investment account surplus is small or negative, and the physical health is poor, then the investor will tend to prefer a conservative investment strategy.

[0003] However, the traditional asset adjustment and allocation control method generally analyzes risk preferences based on investors' historical trading data and calculates the corresponding risk control value. Then, based on the calculated risk control value, adjustments to the user's asset allocation can be recommended. Historical trading data can only reflect the user's static characteristic data. During the risk control value analysis and calculation process, dynamic characteristic data is not combined for comprehensive analysis, resulting in asset adjustment and allocation control often being inconsistent with actual conditions and prone to errors. Summary of the Invention

[0004] The present invention provides an asset allocation method, device, equipment and medium based on dynamic and static feature data, which avoids the dynamic operation of manual decision-making through comprehensive analysis of dynamic features and static features.

[0005] In a first aspect, an asset allocation method based on dynamic and static feature data is provided, comprising:

[0006] Extracting acoustic features based on pre-acquired user voice data, and performing acoustic dynamic feature extraction on the acoustic features to obtain acoustic dynamic features;

[0007] Converting the user voice data into text, and performing semantic dynamic feature extraction on the text to obtain semantic dynamic features;

[0008] Performing gated attention dynamic weighting on the acoustic dynamic features and the semantic dynamic features to obtain weighted features;

[0009] Dynamically evaluate the weighted features and generate a position adjustment instruction based on the dynamic evaluation results and preset static feature data;

[0010] Generate and send asset allocation recommendations based on the position adjustment instructions in accordance with the preset user authorization permissions.

[0011] In a second aspect, an asset allocation device based on dynamic and static feature data is provided, comprising:

[0012] An extraction module, configured to extract acoustic features based on pre-acquired user voice data;

[0013] An acoustic dynamic extraction module, configured to extract acoustic dynamic features from the acoustic features to obtain acoustic dynamic features;

[0014] A conversion and extraction module, configured to convert the user voice data into text, and perform semantic dynamic feature extraction on the text to obtain semantic dynamic features;

[0015] A weighting module, configured to perform gated attention dynamic weighting on the acoustic dynamic features and the semantic dynamic features to obtain weighted features;

[0016] An evaluation generation module, configured to dynamically evaluate the weighted features and generate a position adjustment instruction based on the dynamic evaluation results and a combination of preset static feature data;

[0017] The sending module is used to generate and send asset allocation suggestions based on the position adjustment instructions in accordance with the preset user authorization permissions.

[0018] 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 asset configuration method based on dynamic and static feature data are implemented.

[0019] 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 asset configuration method based on dynamic and static feature data are implemented.

[0020] In the above-mentioned solution implemented by the asset allocation method, device, computer equipment and storage medium based on dynamic and static feature data, acoustic dynamic features are extracted from acoustic features, which can identify dynamic signals that cannot be transmitted through text, such as voice tremors (anxiety), increased pitch (excitement), and accelerated speech speed (tension). By extracting semantic dynamic features, the implicit risks in professional terms (such as "stop loss" and "leverage") can be identified, and a risk word contribution report can be generated (such as "plunge" contributes a negative dynamic weight of 0.82) to meet compliance review requirements. The weights of acoustic and semantic features are dynamically adjusted based on the context to reduce the acoustic feature weight when environmental noise causes speech recognition errors to prevent misjudgment. Through real-time dynamic evaluation, subtle changes in market dynamics can be captured at high frequency. Compared with the lagging traditional financial indicators, it is more suitable for short-term trading or event-driven strategies, avoiding the lag and dynamic operation of manual decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] 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.

[0022] Figure 1 This is a schematic diagram of an application environment of an asset configuration method based on dynamic and static feature data in one embodiment of the present invention;

[0023] Figure 2 This is a flow chart of an asset allocation method based on dynamic and static feature data in one embodiment of the present invention;

[0024] Figure 3 This is a structural diagram of an asset allocation device based on dynamic and static feature data in one embodiment of the present invention;

[0025] Figure 4 is a structural diagram of a computer device according to an embodiment of the present invention;

[0026] Figure 5 It is another structural schematic diagram of a computer device in one 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 embodiment of the present invention provides an asset allocation method based on dynamic and static feature data, which can be applied in Figure 1 In an application environment, the client communicates with the server through a network. The server can extract acoustic features based on pre-acquired user voice data, and perform acoustic dynamic feature extraction on the acoustic features to obtain acoustic dynamic features; convert the user voice data into text, and perform semantic dynamic feature extraction on the text to obtain semantic dynamic features; perform gated attention dynamic weighting on the acoustic dynamic features and the semantic dynamic features to obtain weighted features; dynamically evaluate the weighted features, and generate a position adjustment instruction based on the dynamic evaluation results and a combination of preset static feature data; generate and send asset allocation recommendations based on the position adjustment instructions according to the preset user authorization rights, and feed the asset allocation recommendations back to the client. The present invention provides an asset allocation device based on dynamic and static feature data. For the asset allocation recommendation business, the device avoids the dynamic operation of manual decision-making through comprehensive analysis of dynamic and static features. The client can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, and portable wearable devices. The server can be implemented using an independent server or a server cluster consisting of multiple servers. The present invention is described in detail below through specific embodiments.

[0029] See also Figure 2 As shown, Figure 2 A flowchart of an asset allocation method based on dynamic and static feature data provided by an embodiment of the present invention includes the following steps:

[0030] S1. Extract acoustic features based on pre-acquired user voice data, and perform acoustic dynamic feature extraction on the extracted results to obtain acoustic dynamic features.

[0031] In an embodiment of the present invention, the extraction refers to extracting the underlying features that reflect the physical properties of the speech from the original speech waveform, and the acoustic dynamic feature extraction refers to further screening, combining or converting features directly related to dynamics on the basis of the underlying acoustic features, so as to characterize the emotional state contained in the speech.

[0032] Specifically, the speech is preprocessed (such as noise reduction and framing), and the underlying physical features (such as MFCC, pitch, energy, etc.) are extracted to form an "acoustic feature vector". Based on the acoustic features, through statistical analysis, dynamic modeling or feature selection, "dynamically sensitive" high-level features (such as pitch fluctuations, energy change trends, etc.), namely acoustic dynamic features, are generated. The dynamic features include characteristic data such as expected income security and physical health status, and the static features include characteristic data such as current asset status and market volatility data.

[0033] In specific medical and health scenarios, acoustic dynamic features can reflect the physiological state of the vocal organs. For example, the voice of Parkinson's patients may show tremors, slow speech speed and other characteristics. By analyzing these dynamic changes, it can assist in the early screening and disease monitoring of neurological diseases; the dynamic distribution of the voice spectrum of patients with chronic throat diseases or respiratory diseases may be abnormal, and combined with acoustic features, it can provide an objective basis for disease diagnosis.

[0034] In the FinTech scenario, dynamic features such as the tone and intensity of a customer's voice during communication can reflect their emotional state (such as anxiety and anger). Combined with acoustic feature analysis, financial institutions can more accurately judge the customer's credit risk. For example, in pre-loan review, the tension-related features in the customer's voice may be associated with the willingness to repay, providing a supplementary dimension for credit scoring.

[0035] In the embodiment of the present invention, extracting acoustic features based on pre-acquired user voice data includes:

[0036] Dividing the pre-acquired user voice data into multiple groups of frame voice signals according to a preset time length;

[0037] Performing frequency domain power spectrum analysis on the frame speech signal to obtain a frequency domain power spectrum corresponding to the frame speech signal;

[0038] Calculating the energy coefficient of the frequency domain power spectrum according to a preset Mel frequency band to obtain the energy coefficient of the frequency domain power spectrum;

[0039] Performing dynamic range adjustment on the energy coefficient to obtain an adjustment coefficient;

[0040] Perform differential feature extraction on the adjustment coefficient to obtain acoustic features.

[0041] In an embodiment of the present invention, the frequency domain power spectrum analysis refers to converting a short-time frame speech signal from the time domain to the frequency domain to reveal the distribution of speech energy at different frequencies; the energy coefficient calculation refers to converting a linear frequency domain power spectrum into a Mel-frequency domain energy distribution that conforms to human auditory perception; the dynamic range adjustment refers to compressing the dynamic range of the energy coefficient to make it more consistent with the logarithmic perception characteristics of human hearing; the differential feature extraction refers to extracting the temporal change information of the feature to supplement the deficiencies of the static feature (energy coefficient).

[0042] Specifically, the original speech frame (e.g., 20ms length) is multiplied by a window function (e.g., Hamming Window) to reduce spectrum leakage. The window function formula is as follows:

[0043]

[0044] Wherein, n=0, 1, ..., N-1, N is the frame length, and w(n) is the window function value.

[0045] In detail, the windowed frame signal is fast Fourier transformed to obtain the frequency domain amplitude spectrum corresponding to the frame speech signal, and then the modulus of the frequency domain amplitude spectrum is squared to obtain the frequency domain power spectrum corresponding to the frame speech signal. Subsequently, the energy coefficient of the frequency domain power spectrum is calculated according to a preset Mel frequency band. When the energy coefficient of the frequency domain power spectrum is obtained, the linear frequency is converted into the Mel frequency, wherein the linear frequency is the frequency representation obtained after the windowed frame signal is Fourier transformed. M triangular filters (usually M=20-40) are evenly distributed on the Mel frequency axis, and each filter is mapped back to the linear frequency axis. The power spectrum P(k) is compared with each Mel filter H. m (k) and sum to get the Mel-band energy formula as follows:

[0046]

[0047] Where m = 1, 2, ..., M;

[0048] Furthermore, each frame of speech corresponds to a Mel-band energy vector E = [E(1), E(2), ..., E(M)]. The natural logarithm or common logarithm of each Mel-band energy vector is taken to obtain the adjusted coefficient vector C. The first-order difference of the adjustment coefficients C(m) of adjacent frames is calculated. The specific formula is as follows:

[0049]

[0050] Among them, t is the frame index, N is the smoothing window size (usually N = 2 or 3), and then the first-order difference is further differentiated to obtain acceleration information. The adjusted coefficient vector feature, first-order difference, and second-order difference corresponding to each frame of speech are concatenated into the final feature vector, that is, the acoustic feature.

[0051] In the embodiment of the present invention, extracting acoustic dynamic features from the acoustic features to obtain acoustic dynamic features includes:

[0052] Performing fundamental frequency feature analysis on the acoustic feature to obtain a fundamental frequency feature, and determining a fundamental frequency disturbance degree according to the fundamental frequency feature;

[0053] Performing a spectrum feature analysis on the acoustic feature to obtain a spectrum feature, and determining a spectrum slope according to the spectrum feature;

[0054] The acoustic features, fundamental frequency disturbance degree and spectrum tilt are fused to obtain acoustic dynamic features.

[0055] In an embodiment of the present invention, the fundamental frequency feature analysis refers to extracting pitch-related features from acoustic features, the spectrum feature analysis refers to analyzing the energy distribution characteristics of the speech spectrum and quantifying the relative relationship between high-frequency and low-frequency energy, the determination of the fundamental frequency disturbance degree refers to calculating the relative rate of change of adjacent fundamental frequency periods, the determination of the spectrum inclination refers to calculating the rate at which the spectrum energy decreases with increasing frequency, and the feature fusion refers to complementing the advantages of features in different dimensions to construct a more comprehensive acoustic dynamic feature representation.

[0056] Specifically, the autocorrelation method is used to calculate the autocorrelation function of the acoustic feature. The peak position corresponds to the fundamental frequency period. The fundamental frequency is then separated through cepstrum analysis (the cepstrum peak corresponds to the fundamental frequency period). The deep learning method can be used to directly predict the fundamental frequency from the waveform and perform smoothing to obtain a more accurate fundamental frequency feature. The fundamental frequency disturbance degree is calculated using the following formula:

[0057]

[0058] Where N is the total number of speech cycles in the fundamental frequency detection window (for example, if 5 fundamental frequency cycles are detected in a 100ms speech frame, then N=5), P i Indicates the duration of the i-th fundamental frequency cycle (unit: milliseconds), Represents the average duration of all fundamental frequency cycles in the window, J ppq It represents the local volatility of the fundamental frequency period. A larger value indicates more intense dynamic fluctuations in speech (e.g., unstable fundamental frequency during anxiety).

[0059] Furthermore, the acoustic features are subjected to spectral feature analysis, specifically extracting spectral features such as Mel-frequency cepstral coefficients, spectral entropy, and spectral flatness, and then calculating the spectral slope. Specifically, the spectrum is divided into a low-frequency region (such as 0-1kHz) and a high-frequency region (such as 1-8kHz), and the energy ratio of the two regions is calculated; or a linear regression is performed on the spectrum, and the slope is the spectral slope. The specific formula is as follows:

[0060]

[0061] Among them, E hign 、E low is the energy in the high-frequency and low-frequency regions, and SpectralSlope represents the slope of the spectrum;

[0062] It can also be calculated using the following formula:

[0063]

[0064] Among them, S(k) represents the amplitude of the speech signal at the kth frequency point in the frequency domain, K represents the total number of frequency domain components (such as taking the first 512 frequency points), and Slope represents the inclination of the spectrum energy distribution. A larger value indicates a higher proportion of high-frequency energy (such as increased high-frequency energy in anger).

[0065] Furthermore, each feature is normalized (such as Z-score normalization) to eliminate dimensional differences, and acoustic features (such as MFCC), fundamental frequency perturbation (scalar), and spectrum tilt (scalar) are spliced ​​by dimension. The fusion feature = [MFCC1, MFCC2, ..., MFCC D , Jitter, SpectralSlope], where D is the dimension of MFCC (e.g. 13 dimensions), and the final fusion feature dimension is D+2. It should be noted that weights are assigned to different features according to the importance of the task, and the acoustic dynamic features are finally obtained.

[0066] In the embodiment of the present invention, by adjusting the dynamic range of the energy coefficient, high energy values ​​can be compressed, low energy values ​​can be amplified, and the robustness of the feature can be increased. At the same time, the rhythmic changes of the synthesized speech can be controlled by differential feature extraction to make it more natural.

[0067] In an embodiment of the present invention, during a customer service call, the customer's dynamic emotions of anger or anxiety are detected in real time through acoustic features (such as sudden changes in pitch, faster speaking speed, and increased volume), and the call is automatically transferred to senior customer service, thereby reducing the complaint rate caused by dynamic emotional problems.

[0068] S2. Convert the user voice data into text, and extract semantic dynamic features from the text to obtain semantic dynamic features.

[0069] In the embodiment of the present invention, the semantic dynamic feature extraction refers to analyzing and extracting semantic information that can reflect the dynamic state of the speaker from the text content.

[0070] Specifically, it breaks through the direct analysis of audio signals, converts the speech content into understandable text, and extracts the speaker's subjective emotions, attitudes and intentions from the text, providing semantic-level evidence for dynamic recognition.

[0071] In medical and health scenarios, by analyzing the frequency of occurrence of negative words (such as "despair", "insomnia", and "pain"), emotional intensity (such as "very irritable"), and semantic themes (such as repeated mentions of "suicidal thoughts") in patient voice texts, doctors can be assisted in identifying potential psychological problems and shortening the diagnosis cycle.

[0072] In the specific scenarios of financial technology, by analyzing the semantic features in the customer consultation voice text (such as "stable" and "low risk" corresponding to conservative investors, "high yield" and "long-term returns" corresponding to aggressive types), combined with dynamic strength (such as "very concerned about risks"), financial products are automatically matched to improve conversion rates.

[0073] In the embodiment of the present invention, extracting semantic dynamic features from the text to obtain semantic dynamic features includes:

[0074] Construct a domain risk dictionary based on the corpus of a preset domain, and annotate the TF-IDF risk coefficient of each word in the domain risk dictionary;

[0075] Perform word segmentation on the text to obtain text segmentation, and construct a context embedding vector for each text segmentation using a pre-acquired financial BERT model;

[0076] Evaluate the importance weight of each text segmentation based on the context embedding vector;

[0077] The importance weight and the TF-IDF risk coefficient are weightedly fused to obtain a semantic dynamic feature.

[0078] In an embodiment of the present invention, the construction refers to screening out risk-related professional vocabulary from a preset financial field corpus to form the basis of the dictionary, the annotation refers to quantifying the importance and specificity of each risk word in the field corpus, the construction of the context embedding vector refers to converting the segmented text into a semantic vector containing context information, the evaluation refers to evaluating the semantic contribution of each segmentation in the text, and the weighted fusion refers to combining the dictionary risk coefficient with the context importance to form a more comprehensive semantic dynamic feature.

[0079] Specifically, professional texts in the financial field (such as financial reports, research reports, news, and regulatory documents) are collected as training data, and unsupervised methods (such as TF-IDF and TextRank) are used to identify high-frequency professional terms (such as "default", "liquidity crisis", and "leverage ratio"). Combined with domain knowledge (such as annotation by financial analysts), risk-related vocabulary (such as "Ponzi scheme" and "bankruptcy") is manually supplemented, and the extracted vocabulary is organized into a dictionary, with each word corresponding to a unique identifier.

[0080] Furthermore, the frequency of each word in a single document is counted (number of occurrences / total number of words in the document) to obtain the term frequency (TF), and the proportion of the number of documents containing the word in the total number of documents is counted. The logarithm and then the inverse are taken (rare words have high IDF values) to obtain the inverse document frequency (IDF). TF is multiplied by IDF to obtain the TF-IDF value of each word. The coefficient is adjusted according to the experience of domain experts (such as adding additional weight to core words such as "systemic risk") to finally form the risk coefficient of each word.

[0081] Further, use a word segmentation tool (such as jieba) to split the text into individual words, add special markers (such as [CLS], [SEP]) before and after the word segmentation sequence to form an input format acceptable to the model, input the processed word segmentation sequence into the pre-trained financial BERT model, the model outputs the vector representations of each word segmentation at different hidden layers, and extract the vectors fused by the last layer or multiple layers as context embedding vectors (each word corresponds to a 768-dimensional vector).

[0082] Further, use the attention weights自带 by the BERT model, which reflect the association strength between word segmentations. For each word segmentation, calculate the sum of its attention weights with other word segmentations; map the word segmentation results to a domain risk dictionary (such as "explosion of landmine" corresponding to the risk words in the dictionary to obtain its TF-IDF risk coefficient). For each word segmentation, multiply its context importance weight by the TF-IDF risk coefficient in the dictionary (such as the risk coefficient of "explosion of landmine" 0.8 × context weight 0.9 = 0.72), and combine the fusion weights of all word segmentations into a vector to form the semantic dynamic feature of the text.

[0083] In the embodiments of the present invention, by constructing a domain risk dictionary, a unified "dictionary" of risk vocabulary in the financial field is formed, ensuring that the risk indicators of different texts (such as customer communication records, transaction contracts) can be compared horizontally, and at the same time, the attention weights can quantify the "semantic contribution degree" of vocabulary in the context.

[0084] In the embodiments of the present invention, by semantic analysis, the implicit risk vocabulary associations in the text are discovered (such as the combination of "medical dispute" + "misdiagnosis" + "compensation" reflects legal risks), potential problems are warned in advance. In the quality inspection of medical record writing, semantic features such as "operation error" and "incomplete record" are extracted, and documents that may cause medical accidents are automatically identified, reducing compliance risks.

[0085] S3. Perform gated attention dynamic weighting on the acoustic dynamic feature and the semantic dynamic feature to obtain a weighted feature.

[0086] In the embodiments of the present invention, the gated attention dynamic weighting refers to automatically judging the importance of different features through an attention mechanism and dynamically adjusting their weights through a gating mechanism, and finally outputting a more representative weighted feature.

[0087] Specifically, through intelligent algorithms, when the model processes voice data, it can not only simultaneously utilize the sound features of the voice (such as intonation, speech rate) and the text content (such as vocabulary sentiment), but also automatically judge the importance of the two according to the specific context and dynamically allocate weights. For example, in a noisy environment, it relies more on semantic features, and when the text semantics is ambiguous, it relies more on acoustic features, and finally outputs a high-quality feature after fusion.

[0088] In specific healthcare scenarios, such as emergency telephone consultations, the acoustic features of the patient's voice (such as rapid breathing and crying) and semantic content (such as "chest pain" and "bleeding") are combined and dynamically weighted to quickly determine the degree of urgency and prioritize resource allocation to avoid delays due to missed text keywords or misjudgment of voice signals.

[0089] In financial scenarios, based on the acoustic characteristics of customers during consultation (such as hesitation, pauses, and changes in tone) and semantic keywords (such as "conservative" and "high return"), risk preferences are analyzed after dynamic weighting, and products are accurately recommended and the wording is adjusted to avoid resistance caused by excessive sales promotion.

[0090] In the embodiment of the present invention, performing gated attention dynamic weighting on the acoustic dynamic features and the semantic dynamic features to obtain weighted features includes:

[0091] Aligning the acoustic dynamic features and the semantic dynamic features and concatenating them to obtain a joint feature vector;

[0092] Linearly projecting the joint eigenvector according to a preset weight matrix, and converting the projection result into a gating weight value through a nonlinear activation function and a preset bias parameter;

[0093] The acoustic dynamic features and the semantic dynamic features are weightedly fused according to the gating weight value to obtain weighted features.

[0094] In an embodiment of the present invention, the aligned dimension splicing refers to adjusting the acoustic dynamic features and the semantic dynamic features to the same dimension, and then connecting them in series according to the dimension to form a joint feature vector. The linear projection refers to the process of performing a linear transformation on the joint feature vector using a preset weight matrix to obtain an intermediate variable. The conversion refers to transforming the linear projection result through a nonlinear activation function, and converting it into a gated weight value in combination with the bias parameter.

[0095] Specifically, the acoustic dynamic feature is marked as A, and the semantic dynamic feature is marked as S. and D A ≠D S , high-dimensional features are compressed through fully connected layers (FC) or pooling operations, such as A′=FC A (A)∈R D , S′=FC S (S)∈R D , so as to achieve the purpose of dimensional alignment of acoustic dynamic features and semantic dynamic features. If D A =D S=D, indicating that the dimensions of the acoustic dynamic features and the semantic dynamic features have been aligned, the alignment step is skipped directly, and the aligned features are concatenated by dimension. If the acoustic features (100 dimensions) and the semantic features (100 dimensions) are aligned, they are spliced ​​into a 200-dimensional vector.

[0096] Furthermore, a learnable weight matrix is ​​defined (e.g., a matrix size of K×(D1+D2)), where K is usually equal to the number of modes, i.e., K=2, corresponding to the acoustic and semantic weights respectively), and the joint feature vector is multiplied by the weight matrix to obtain an intermediate vector of length K. For example, a 200-dimensional joint vector is multiplied by a 2×200 weight matrix to obtain a 2-dimensional intermediate vector, which represents the original weight scores of the acoustic and semantic features respectively, where the intermediate vector refers to the projection result.

[0097] Furthermore, the original weight scores of the intermediate vectors are converted into probability values ​​or coefficients that can be used for weighting. A nonlinear activation function (such as Softmax or Sigmoid) is used on the intermediate vectors to compress the values ​​into a specific range. For example, if Softmax is used, the 2-dimensional intermediate vectors will be converted into two numbers between 0 and 1, representing the weight ratios of acoustic and semantic features respectively (such as acoustic weight 0.7 and semantic weight 0.3). The preset bias parameters are added to the activation function operation, and the weight baseline is fine-tuned (such as letting the model pay more attention to a certain modality by default). Finally, the gated weight value is output to form a vector containing the weights of each modality.

[0098] Furthermore, according to the gating weight value, the acoustic and semantic features are dynamically combined to generate the final weighted feature. The gating weight value is respectively corresponded to the original acoustic dynamic feature and the semantic dynamic feature. For example, if the acoustic feature is 100-dimensional, the semantic feature is 100-dimensional, and the gating weight is [0.7, 0.3], then the weight of the acoustic feature is 0.7, and the weight of the semantic feature is 0.3. The feature vector of each modality is multiplied by the corresponding weight value, and then added to obtain the final weighted feature vector, which is expressed by the formula: weighted feature = acoustic feature × acoustic weight + semantic feature × semantic weight.

[0099] In the embodiment of the present invention, the conversion of the projection result into a gated weight value using a nonlinear activation function and a preset bias parameter includes:

[0100] Add the preset bias parameters to the projection result to obtain the added projection result;

[0101] Performing nonlinear mapping on the added projection result through a nonlinear activation function to obtain an initial gating weight value;

[0102] The initial gating weight value is normalized to obtain a gating weight value.

[0103] In the embodiment of the present invention, the adding refers to directly adding a preset bias parameter (which can be understood as a learnable constant term) to the result after linear projection to adjust the numerical distribution, and the nonlinear mapping refers to using a nonlinear activation function (such as Sigmoid, ReLU, etc.) to transform the numerical value after adding the bias to introduce a nonlinear relationship.

[0104] Specifically, taking the result [0.4, -0.3, 0.9] after adding the bias as an example, if the Sigmoid function is used, each value will be compressed to the (0, 1) interval: 0.4 → about 0.6; -0.3 → about 0.42; 0.9 → about 0.71, and a set of values ​​after nonlinear transformation (such as [Sigmoid result: 0.6, 0.42, 0.71]) is obtained, which is called the "initial gated weight value". Then all the initial weight values ​​are added together to get the total (0.6 + 0.42 + 0.71 = 1.73). Each initial weight value is divided by the total to get the normalized weight value. The sum of the normalized weight values ​​is 1, which can be directly used for weighted fusion.

[0105] In an embodiment of the present invention, after unifying the dimensions of acoustic dynamic features and semantic dynamic features, the model does not need to design independent branches for different modalities, which reduces the complexity of the network structure and improves training efficiency. At the same time, the spliced ​​joint vector contains both auditory and language information, avoiding the one-sidedness of a single modality. The acoustic dynamic features and semantic dynamic features are weightedly fused according to the gated weight value, and the modal contribution can be dynamically adjusted to improve robustness, while suppressing invalid modalities and enhancing effective information.

[0106] In an embodiment of the present invention, data-driven intelligent fusion is achieved by training and learning the modal importance patterns in different scenarios, rather than relying on manually preset rules. At the same time, "simple addition of multiple modalities" is upgraded to "dynamic focus on key information", ultimately improving the accuracy, robustness and interpretability of dynamic analysis.

[0107] S4. Dynamically evaluate the weighted features, and generate a position adjustment instruction based on the dynamic evaluation results and the preset static feature data combination.

[0108] In an embodiment of the present invention, the dynamic evaluation refers to the use of weighted features (dynamic information that integrates acoustics and semantics) to determine the dynamic tendency and intensity expressed by the current input content, and the combination generation refers to the generation of a specific asset allocation adjustment plan (position adjustment instruction) based on the dynamic evaluation results and preset static feature data.

[0109] Specifically, through previous steps such as feature alignment and weighted fusion, we obtain weighted features that comprehensively reflect dynamics, dynamically evaluate the weighted features, output quantitative dynamic results, combine the dynamic results with market volatility (such as the severity of current stock market fluctuations), and generate position adjustment instructions through the preset strategy model.

[0110] In the embodiment of the present invention, the generating of the position adjustment instruction based on the dynamic evaluation result and the preset static feature data combination includes:

[0111] Obtaining real-time market data, and calculating the market volatility level according to a preset period based on preset static feature data and the real-time market data;

[0112] Constructing a two-dimensional decision matrix, taking the dynamic assessment results and the market volatility level as input parameters, mapping them into the two-dimensional decision matrix, and determining a joint risk level;

[0113] The preset position adjustment rule library is matched according to the joint risk level to obtain the position adjustment instruction.

[0114] In an embodiment of the present invention, the acquisition refers to the real-time collection of financial market data (such as stock prices, futures contracts, exchange rates, etc.) from channels such as exchanges and data service providers; the calculation refers to quantifying the degree of market volatility and dividing it into different levels (such as low, medium, and high volatility) based on real-time market data and preset algorithms; the construction refers to creating a matrix with "dynamic assessment results" and "market volatility level" as coordinate axes for mapping the joint risk level; the mapping refers to using the dynamic assessment results and market volatility level as coordinates to locate the corresponding cells in the decision matrix to obtain the joint risk level; the matching refers to selecting a corresponding position adjustment strategy from a preset rule set based on the joint risk level.

[0115] Specifically, access the real-time market information system (such as Bloomberg Terminal, Wind Financial Terminal) through the API interface, WebSocket protocol or data subscription service, obtain data fields such as price, trading volume, buy and sell orders at a preset frequency (such as seconds, minutes), eliminate outliers (such as price gaps, sudden changes in trading volume), and process missing values ​​(such as interpolation filling).

[0116] Furthermore, the standard deviation of price returns within a preset period (such as 20 days or 60 days) is calculated, and the market expected volatility is inferred through option pricing models (such as Black-Scholes), and the volatility values ​​are mapped to discrete levels (such as 0-10% for low volatility, 10%-20% for medium volatility, and >20% for high volatility). For example, if the 20-day standard deviation of the S&P 500 index return is 15%, it corresponds to the "medium volatility" level.

[0117] Furthermore, when constructing a two-dimensional decision matrix, each matrix cell predefines a joint risk level (such as "high risk", "medium risk", and "low risk"). The dynamic assessment results and the market volatility level are used as coordinates to locate the corresponding cells in the decision matrix to obtain the joint risk level. If the dynamic assessment is "negative" and the market volatility level is "high", it corresponds to the "negative-high volatility" cell in the matrix ("extremely high risk" in the example).

[0118] In addition, a real-time correlation model between dynamic state and asset volatility can be constructed:

[0119] Δw cash =β·(E t -0.5)·σ t

[0120] Among them, E t represents the current user's dynamic anxiety index, normalized to [0,1], E t >0.8 indicates high anxiety, σ t Indicates market volatility, calculated as annualized standard deviation (e.g. σ t =0.2 represents a volatility of 20%), β represents an adjustment factor, which is set to β=0.15 in this solution to control the amplitude of the position adjustment.

[0121] Real-time rebalancing strategy:

[0122] When the dynamic anxiety index E t >0.8 and market volatility σ t When it is >20%, the proportion of cash assets will be automatically increased;

[0123] Implement a reverse strategy when the dynamics and market volatility are below the threshold:

[0124] When E t ≤0.8 and σ t When ≤20%:

[0125] Reduce the cash ratio: adjust inversely according to the formula, Δw cash May be negative (such as E t =0.3,σ t =15%, then Δw cash =-0.15×0.2×0.15=-0.0045, which means the cash ratio is reduced by 0.45%).

[0126] Increase allocation to risky assets: Use the released cash to increase the weight of high-volatility assets such as stocks and futures to increase return potential. The specific formula is as follows:

[0127]

[0128] in, represents the configuration weight of the i-th asset at time t, η represents the learning rate, which controls the speed of weight adjustment (such as η=0.01), L(E t ,σ t ) represents the joint loss function, which measures the degree of matching between dynamics and market fluctuations (the smaller L is, the more adapted the dynamics are to the market status).

[0129] In one embodiment, after performing dynamic state evaluation, the device uses the newly generated speech dialogue data to fine-tune the local model θ local , calculate the gradient ▽θ local , generate Encrypt(▽θ through Paillier homomorphic encryption local ), aggregate the encrypted gradients of N devices, decrypt and average them, and update the global model:

[0130]

[0131] in, represents the global model parameters after the tth round of federated learning, The model gradient of the local training on the i-th user device. N represents the total number of devices participating in the current round of federated learning. Decrypt represents the decryption operation based on homomorphic encryption to ensure that the gradient is decrypted before aggregation.

[0132] When updating the global model, the gradients of all devices are aggregated at a fixed time point every day to update θ glocal ,The updated model will be used for dynamic identification of all users the next day.

[0133] In an embodiment of the present invention, the two independent dimensions of dynamic assessment results and market volatility level are converted into a quantifiable and comparable joint risk indicator in matrix form, avoiding the one-sidedness of single-dimensional decision-making. At the same time, there is no need to analyze the original data row by row, and conclusions can be obtained directly through the correspondence between matrix cells. This is suitable for high-frequency trading or real-time risk control scenarios. Through matrix mapping, a large amount of original data can be condensed into a few risk levels, reducing the interference of redundant information on decision-making.

[0134] In an embodiment of the present invention, real-time dynamic evaluation can capture subtle changes in market dynamics (such as dynamic mutations before and after policy releases) at a high frequency. Compared with traditional financial indicators with lags (such as quarterly financial reports), it is more suitable for short-term trading or event-driven strategies. It automatically generates instructions through preset rules, avoiding the lag and dynamic operations of manual decision-making, and is suitable for high-frequency trading or large-scale asset portfolio management.

[0135] S5. Generate and send asset allocation suggestions based on the position adjustment instructions in accordance with the preset user authorization permissions.

[0136] In an embodiment of the present invention, the sending refers to the process of delivering the generated asset allocation suggestion (i.e., the specific execution plan of the position adjustment instruction) to the user or the relevant execution system through a specific channel in accordance with pre-set rules and user-authorized permissions.

[0137] Specifically, users grant the system specific operating permissions (such as viewing, suggesting, executing, etc.) through agreements or system settings in advance. The scope of permissions is set by the user independently and must comply with regulatory requirements (such as anti-money laundering and investor suitability management). Ordinary investors may only authorize "receiving position adjustment suggestions" (without direct trading permissions), and professional investors or institutional users may authorize "automatic execution of low-risk position adjustment instructions" and perform corresponding operations based on the user's authorized permissions.

[0138] It can be seen that in the above scheme, for the asset allocation recommendation business, acoustic features are extracted based on the pre-acquired user voice data, and acoustic dynamic features are extracted from the acoustic features to obtain acoustic dynamic features; the user voice data is converted into text, and semantic dynamic features are extracted from the text to obtain semantic dynamic features; the acoustic dynamic features and the semantic dynamic features are dynamically weighted by gated attention to obtain weighted features; the weighted features are dynamically evaluated, and a position adjustment instruction is generated based on the dynamic evaluation results and the preset static feature data combination; according to the preset user authorization authority, asset allocation recommendations are generated and sent according to the position adjustment instructions, and the dynamic operation of manual decision-making is avoided through comprehensive analysis of dynamic features and static features.

[0139] 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.

[0140] In one embodiment, an asset allocation device based on dynamic and static feature data is provided, and the asset allocation device based on dynamic and static feature data corresponds one-to-one with the asset allocation method based on dynamic and static feature data in the above embodiment. Figure 3 As shown, the asset configuration device based on dynamic and static feature data includes an extraction module 101, an acoustic dynamic extraction module 102, a conversion extraction module 103, a weighting module 104, an evaluation generation module 105, and a sending module 106. The functional modules are described in detail as follows:

[0141] Extraction module 101, used to extract acoustic features based on pre-acquired user voice data;

[0142] An acoustic dynamic extraction module 102 is configured to extract acoustic dynamic features from the acoustic features to obtain acoustic dynamic features;

[0143] The conversion and extraction module 103 is used to convert the user voice data into text and extract semantic dynamic features from the text to obtain semantic dynamic features;

[0144] A weighting module 104 is configured to perform gated attention dynamic weighting on the acoustic dynamic features and the semantic dynamic features to obtain weighted features;

[0145] Evaluation generation module 105, configured to dynamically evaluate the weighted features and generate a position adjustment instruction based on the dynamic evaluation results and preset static feature data;

[0146] The sending module 106 is used to generate and send asset allocation suggestions based on the position adjustment instructions according to the preset user authorization permissions.

[0147] In one embodiment, when extracting acoustic features based on pre-acquired user voice data, the extraction module 101 is configured to:

[0148] Dividing the pre-acquired user voice data into multiple groups of frame voice signals according to a preset time length;

[0149] Performing frequency domain power spectrum analysis on the frame speech signal to obtain a frequency domain power spectrum corresponding to the frame speech signal;

[0150] Calculating the energy coefficient of the frequency domain power spectrum according to a preset Mel frequency band to obtain the energy coefficient of the frequency domain power spectrum;

[0151] Performing dynamic range adjustment on the energy coefficient to obtain an adjustment coefficient;

[0152] Perform differential feature extraction on the adjustment coefficient to obtain acoustic features.

[0153] In one embodiment, when the acoustic dynamics extraction module 102 extracts the acoustic dynamics features from the acoustic features to obtain the acoustic dynamics features, it is configured to:

[0154] Performing fundamental frequency feature analysis on the acoustic feature to obtain a fundamental frequency feature, and determining a fundamental frequency disturbance degree according to the fundamental frequency feature;

[0155] Performing a spectrum feature analysis on the acoustic feature to obtain a spectrum feature, and determining a spectrum slope according to the spectrum feature;

[0156] The acoustic features, fundamental frequency disturbance degree and spectrum tilt are weightedly fused to obtain acoustic dynamic features.

[0157] In one embodiment, when the conversion and extraction module 103 extracts semantic dynamic features from the text to obtain semantic dynamic features, it is configured to:

[0158] Construct a domain risk dictionary based on the corpus of a preset domain, and annotate the TF-IDF risk coefficient of each word in the domain risk dictionary;

[0159] Perform word segmentation on the text to obtain text segmentation, and construct a context embedding vector for each text segmentation using a pre-acquired financial BERT model;

[0160] Evaluate the importance weight of each text segmentation based on the context embedding vector;

[0161] The importance weight and the TF-IDF risk coefficient are weightedly fused to obtain a semantic dynamic feature.

[0162] In one embodiment, when performing gated attention dynamic weighting on the acoustic dynamic features and the semantic dynamic features to obtain weighted features, the weighting module 104 is configured to:

[0163] Aligning the acoustic dynamic features and the semantic dynamic features and concatenating them to obtain a joint feature vector;

[0164] Linearly projecting the joint eigenvector according to a preset weight matrix, and converting the projection result into a gating weight value through a nonlinear activation function and a preset bias parameter;

[0165] The acoustic dynamic features and the semantic dynamic features are weightedly fused according to the gating weight value to obtain weighted features.

[0166] In one embodiment, when converting the projection result into a gated weight value through a nonlinear activation function and a preset bias parameter, it is used to:

[0167] Add the preset bias parameters to the projection result to obtain the added projection result;

[0168] Performing nonlinear mapping on the added projection result through a nonlinear activation function to obtain an initial gating weight value;

[0169] The initial gating weight value is normalized to obtain a gating weight value.

[0170] In one embodiment, when generating a position adjustment instruction based on a combination of dynamic evaluation results and preset static feature data, the evaluation generation module 105 is configured to:

[0171] Obtaining real-time market data, and calculating the market volatility level according to a preset period based on preset static feature data and the real-time market data;

[0172] Constructing a two-dimensional decision matrix, taking the dynamic assessment results and the market volatility level as input parameters, mapping them into the two-dimensional decision matrix, and determining a joint risk level;

[0173] The preset position adjustment rule library is matched according to the joint risk level to obtain the position adjustment instruction.

[0174] The present invention provides an asset allocation device based on dynamic and static feature data. For the asset allocation recommendation business, acoustic features are extracted based on pre-acquired user voice data, and acoustic dynamic features are extracted from the acoustic features to obtain acoustic dynamic features; the user voice data is converted into text, and semantic dynamic features are extracted from the text to obtain semantic dynamic features; the acoustic dynamic features and the semantic dynamic features are dynamically weighted with gated attention to obtain weighted features; the weighted features are dynamically evaluated, and a position adjustment instruction is generated based on the dynamic evaluation result and a preset static feature data combination; asset allocation recommendations are sent according to the position adjustment instruction in accordance with the preset user authorization authority, and the dynamic operation of manual decision-making is avoided through comprehensive analysis of dynamic features and static features.

[0175] The specific definition of an asset allocation device based on dynamic and static feature data can be found in the definition of an asset allocation method based on dynamic and static feature data above and will not be further elaborated here. Each module in the aforementioned asset allocation device based on dynamic and static feature data can be implemented in whole or in part through 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.

[0176] 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 4 As shown. The computer device includes a processor, memory, network interface and 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 non-volatile and / or volatile storage media and 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 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 implements the functions or steps on the server side of an asset configuration method based on dynamic and static feature data.

[0177] 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 5As shown. The computer device includes a processor, memory, a network interface, a display screen, and an 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 executed by the processor, the computer program implements the client-side functions or steps of an asset configuration method based on dynamic and static feature data.

[0178] 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:

[0179] Extracting acoustic features based on pre-acquired user voice data, and performing acoustic dynamic feature extraction on the acoustic features to obtain acoustic dynamic features;

[0180] Converting the user voice data into text, and performing semantic dynamic feature extraction on the text to obtain semantic dynamic features;

[0181] Performing gated attention dynamic weighting on the acoustic dynamic features and the semantic dynamic features to obtain weighted features;

[0182] Dynamically evaluate the weighted features and generate a position adjustment instruction based on the dynamic evaluation results and preset static feature data;

[0183] Generate and send asset allocation recommendations based on the position adjustment instructions in accordance with the preset user authorization permissions.

[0184] 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:

[0185] Extracting acoustic features based on pre-acquired user voice data, and performing acoustic dynamic feature extraction on the acoustic features to obtain acoustic dynamic features;

[0186] Converting the user voice data into text, and performing semantic dynamic feature extraction on the text to obtain semantic dynamic features;

[0187] Performing gated attention dynamic weighting on the acoustic dynamic features and the semantic dynamic features to obtain weighted features;

[0188] Dynamically evaluate the weighted features and generate a position adjustment instruction based on the dynamic evaluation results and preset static feature data;

[0189] Generate and send asset allocation recommendations based on the position adjustment instructions in accordance with the preset user authorization permissions.

[0190] 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.

[0191] 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).

[0192] 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.

[0193] The above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. If software tools or components other than those of the company appear in the application embodiments, they are merely used for illustration and do not represent actual use. Although the present invention has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above-mentioned embodiments, or make equivalent replacements for some of the technical features therein. However, 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 embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. An asset allocation method based on dynamic and static feature data, characterized in that: include: Extracting acoustic features based on pre-acquired user voice data, and performing acoustic dynamic feature extraction on the acoustic features to obtain acoustic dynamic features; Converting the user voice data into text, and performing semantic dynamic feature extraction on the text to obtain semantic dynamic features; Performing gated attention dynamic weighting on the acoustic dynamic features and the semantic dynamic features to obtain weighted features; Dynamically evaluate the weighted features and generate a position adjustment instruction based on the dynamic evaluation results and preset static feature data; Generate and send asset allocation recommendations based on the position adjustment instructions in accordance with the preset user authorization permissions.

2. The asset allocation method based on dynamic and static feature data according to claim 1, characterized in that: The extracting of acoustic features based on pre-acquired user voice data includes: Dividing the pre-acquired user voice data into multiple groups of frame voice signals according to a preset time length; Performing frequency domain power spectrum analysis on the frame speech signal to obtain a frequency domain power spectrum corresponding to the frame speech signal; Calculating the energy coefficient of the frequency domain power spectrum according to a preset Mel frequency band to obtain the energy coefficient of the frequency domain power spectrum; Performing dynamic range adjustment on the energy coefficient to obtain an adjustment coefficient; Perform differential feature extraction on the adjustment coefficient to obtain acoustic features.

3. The asset allocation method based on dynamic and static feature data according to claim 1, characterized in that: The extracting acoustic dynamic features from the acoustic features to obtain acoustic dynamic features includes: Performing fundamental frequency feature analysis on the acoustic feature to obtain a fundamental frequency feature, and determining a fundamental frequency disturbance degree according to the fundamental frequency feature; Performing a spectrum feature analysis on the acoustic feature to obtain a spectrum feature, and determining a spectrum slope according to the spectrum feature; The acoustic features, fundamental frequency disturbance degree and spectrum tilt are weightedly fused to obtain acoustic dynamic features.

4. The asset allocation method based on dynamic and static feature data according to claim 1, characterized in that: The extracting of semantic dynamic features from the text to obtain semantic dynamic features includes: Construct a domain risk dictionary based on the corpus of a preset domain, and annotate the TF-IDF risk coefficient of each word in the domain risk dictionary; Perform word segmentation on the text to obtain text segmentation, and construct a context embedding vector for each text segmentation using a pre-acquired financial BERT model; Evaluate the importance weight of each text segmentation based on the context embedding vector; The importance weight and the TF-IDF risk coefficient are weighted and fused to obtain a semantic dynamic feature.

5. The asset allocation method based on dynamic and static feature data according to claim 1, characterized in that: The gated attention dynamic weighting of the acoustic dynamic features and the semantic dynamic features to obtain weighted features includes: Aligning the acoustic dynamic features and the semantic dynamic features and concatenating them to obtain a joint feature vector; Linearly projecting the joint eigenvector according to a preset weight matrix, and converting the projection result into a gating weight value through a nonlinear activation function and a preset bias parameter; The acoustic dynamic features and the semantic dynamic features are weightedly fused according to the gating weight value to obtain weighted features.

6. The asset allocation method based on dynamic and static feature data according to claim 5, characterized in that: The projection result is converted into a gated weight value through a nonlinear activation function and a preset bias parameter, including: Add the preset bias parameters to the projection result to obtain the added projection result; Performing nonlinear mapping on the added projection result through a nonlinear activation function to obtain an initial gating weight value; The initial gating weight value is normalized to obtain a gating weight value.

7. The asset allocation method based on dynamic and static feature data according to claim 1, characterized in that: The generating of the position adjustment instruction based on the dynamic evaluation result and the preset static feature data combination includes: Obtaining real-time market data, and calculating the market volatility level according to a preset period based on preset static feature data and the real-time market data; Constructing a two-dimensional decision matrix, taking the dynamic assessment results and the market volatility level as input parameters, mapping them into the two-dimensional decision matrix, and determining a joint risk level; The preset position adjustment rule library is matched according to the joint risk level to obtain the position adjustment instruction.

8. An asset allocation device based on dynamic and static feature data, characterized in that: include: An extraction module, configured to extract acoustic features based on pre-acquired user voice data; An acoustic dynamic extraction module, configured to extract acoustic dynamic features from the acoustic features to obtain acoustic dynamic features; A conversion and extraction module, configured to convert the user voice data into text, and perform semantic dynamic feature extraction on the text to obtain semantic dynamic features; A weighting module, configured to perform gated attention dynamic weighting on the acoustic dynamic features and the semantic dynamic features to obtain weighted features; An evaluation generation module, configured to dynamically evaluate the weighted features and generate a position adjustment instruction based on the dynamic evaluation results and a combination of preset static feature data; The sending module is used to generate and send asset allocation suggestions based on the position adjustment instructions in accordance with the preset user authorization permissions.

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 asset configuration method based on dynamic and static feature data 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 asset configuration method based on dynamic and static feature data as described in any one of claims 1 to 7 are implemented.

Images