A method and system for recommending bank wealth management products

Abstract

The present invention provides a method and system for recommending bank financial products. First, a multi-modal dialogue data stream containing voice input signals and text input content generated by the interaction between the user terminal and the intelligent customer service system is collected. Then, semantic recognition of the voice input signals is performed to generate a structured requirement description set, entity extraction and intention parsing are performed on the text input content to generate a multi-dimensional semantic tag set, and the two are aligned and fused in space and time to form a user portrait enhanced feature matrix. Then, based on a deep matching model, it is hierarchically associated and analyzed with the financial product knowledge graph to generate an initial recommendation result queue. Finally, context awareness optimization of the queue is performed according to real-time environment perception data, a dynamically adapted recommendation list is generated and visually rendered and output through a multi-channel interaction interface, thereby realizing accurate and dynamic financial product recommendations by using multi-modal data, in-depth analysis and combining with the real-time environment.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular, to a method and system for recommending bank financial products. Background Art

[0002] With the continuous development of the financial market and the increasing awareness of users' financial management, the types of bank financial products are becoming increasingly diverse, and users are facing greater and greater difficulties in choosing financial products suitable for themselves. In order to help users quickly and accurately find financial products that meet their own needs, a bank financial product recommendation system has emerged.

[0003] In related technologies, most recommendation methods mainly rely on a single type of data for analysis and recommendation. Some recommendation systems only collect basic information of users, such as static data like age and income, and conduct simple product matching based on this. Since the data sources are limited in this way, it is impossible to comprehensively and deeply understand the real financial management needs and preferences of users, resulting in the recommendation results often being unable to accurately meet the diverse needs of users.

[0004] In addition, although some related technologies consider the historical transaction data of users, they are only limited to the superficial analysis of transaction behaviors and do not deeply explore the user intentions and potential needs behind the transactions. For example, simply counting the types of products purchased by users without being able to understand the deep reasons for users to purchase these products, such as for short-term capital appreciation, long-term asset allocation, or other special needs, makes the recommendation lack pertinence and foresight. Summary of the Invention

[0005] In view of the problems mentioned above, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for recommending bank financial products, and the method includes:

[0006] Collecting a multi-modal dialogue data stream generated during the interaction between a user terminal and an intelligent customer service system, where the multi-modal dialogue data stream includes voice input signals and text input content;

[0007] Performing semantic recognition on the voice input signals to generate a structured requirement description set, and at the same time performing entity extraction and intention parsing on the text input content to generate a multi-dimensional semantic label set;

[0008] Performing spatio-temporal alignment and fusion on the structured requirement description set and the multi-dimensional semantic label set to generate a user portrait enhanced feature matrix;

[0009] Based on a deep matching model, performing multi-level correlation analysis on the user portrait enhanced feature matrix and a financial product knowledge graph to generate an initial recommendation result queue;

[0010] The initial recommendation result queue is context-aware optimized according to the real-time environment perception data, a dynamically adapted recommendation list is generated, and a visual rendering output is performed through a multi-channel interactive interface.

[0011] On the other hand, an embodiment of the present invention further provides a bank wealth management product recommendation system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0012] Based on the above aspects, the embodiment of the present invention further improves the accuracy and effectiveness of data processing by collecting multimodal dialogue data streams containing voice input signals and text input content during the interaction between the user terminal and the intelligent customer service system, and then processing and fusing the multimodal data separately. The structured demand description set generated by the semantic recognition of the voice input signal and the multidimensional semantic label set generated by the entity extraction and intent analysis of the text input content are fused through time and space alignment to form a user portrait enhancement feature matrix, which not only integrates the advantages of different modal data, but also mines the potential time and space correlation between data. Then, based on the deep matching model, the user portrait enhancement feature matrix and the financial product knowledge graph are subjected to multi-level correlation analysis, which can make full use of the rich product information and relationship network in the financial product knowledge graph, not only considering the simple direct matching of user needs and products, but also through multi-level deep analysis, it can mine potential financial product combinations that meet the complex needs of users, greatly broadening the breadth and depth of recommendations. Finally, the initial recommendation result queue is context-aware optimized based on real-time environmental perception data, and a dynamically adapted recommendation list is generated and output through a multi-channel interactive interface visualization. The recommendation results can be flexibly optimized based on factors such as the real-time market environment and the user's scenario. The multi-channel interactive interface visualization output improves the user experience. Users can easily obtain recommendation information in the way they are accustomed to, which enhances the practicality and user-friendliness of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the execution flow of the bank financial product recommendation method provided in an embodiment of the present invention.

[0014] Figure 2 It is a schematic diagram of exemplary hardware and software components of the bank financial product recommendation system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1It is a schematic flowchart of a bank financial product recommendation method provided by an embodiment of the present invention. The bank financial product recommendation method will be introduced in detail below.

[0016] Step S110, collect the multimodal dialogue data stream generated during the interaction between the user terminal and the intelligent customer service system, where the multimodal dialogue data stream includes voice input signals and text input content.

[0017] In this embodiment, consider such a scenario. An investor is using a tablet to interact with a financial intelligent customer service system. He is sitting in his home study, and there is some slight ambient noise around him, such as the sound of cars driving from the distant street and the faint sound of the air conditioner running. At this time, he starts to interact with the intelligent customer service system. For example, he first inputs by voice: "I have made various investments in the past few years, including stocks, funds, and some time deposits. Now I want to adjust my investment portfolio. I hope to find a financial product that can diversify risks to a certain extent. My stock investment is mainly concentrated in the technology sector, and the current returns fluctuate greatly. So I want to find a product that has a less strong correlation with stocks. I also have to consider my family's financial situation. I have a child who is about to go to college and may need to withdraw a large sum of money in the next four or five years. So the liquidity of this product cannot be too poor. And although my risk tolerance is not low, I don't want to face huge fluctuations like stock investment anymore." At the same time, he also inputs some supplementary information in the input box of the interaction interface as text input content: "My current total assets are approximately 5 million, of which stock investment accounts for 3 million, fund investment accounts for 1 million, and time deposits are 1 million. I hope the investment amount of the new financial product can be between 500,000 and 1 million. In addition, I am also more concerned about the management fees of financial products and hope to minimize unnecessary management fee expenditures."

[0018] The intelligent customer service system collects the above multimodal dialogue data stream including voice input signals and text input content through the high-precision microphone and text input interface of the tablet. In this process, it is necessary to accurately capture every word, phrase, and sentence in the voice input signal, as well as the exact numbers, expressed conditions, and restrictions in the text input content.

[0019] Step S120, perform semantic recognition on the voice input signal to generate a structured requirement description set, and at the same time perform entity extraction and intention parsing on the text input content to generate a multi-dimensional semantic label set.

[0020] In detail, for the above-collected voice input signal, firstly, frame processing and background noise elimination are performed. For example, the voice input signal can be framed at a certain time interval by a signal processing algorithm, and the interference of noises such as car driving sound and air conditioner running sound in the background can be identified and reduced according to the frequency characteristics of the sound, and the Mel frequency cepstrum coefficients are extracted as acoustic feature vectors. Then, using a pre-trained speech recognition model, which is trained based on a large amount of voice data and general voice data in the financial field, the voice input signal is converted into an intermediate text expression, such as: "I have made a variety of investments in the past, including stock funds and regular deposits. Now I adjust the portfolio and look for low-correlated stocks, decent liquidity, products that can be withdrawn within four or five years, and products with lower risks than stocks." At the same time, the speaker's emotional intensity value can be marked according to the characteristics of the voice tone, speech speed, etc. Assuming that the emotional intensity value is medium in the above example, it means that the investor states the demand more rationally.

[0021] Next, a multi-head attention mechanism is used to complete the contextual semantics of the intermediate text expression. For example, some missing information can be supplemented based on common investment knowledge and language logic in the financial field to generate a corrected complete semantic paragraph: "I have made various investments in stocks, funds, and time deposits in the past few years. Now I want to adjust my investment portfolio. I hope to find a financial product that is not so closely related to stocks, has good liquidity, can be withdrawn in the next four or five years, and has lower risks than stock investments, because my child is about to go to college and needs to withdraw a large amount of funds during this period."

[0022] After that, the domain-specific named entity recognition model can be used to extract the constraint entities of amount, term, and risk category from the complete semantic paragraph. In the above example, the risk constraint can be extracted as "the risk is lower than stock investment" and the term constraint is "within four or five years". There is no explicit mention of direct constraints related to the amount (but the condition that the investable amount is within a certain range is implied under the overall demand). The acoustic feature vector, sentiment intensity value and constraint entity are tensor-joined to generate a structured demand description set with sentiment weights.

[0023] For the text input content, we perform word segmentation and part-of-speech tagging to build a dependency syntactic analysis tree. For example, after "My current total assets are about 5 million, of which 3 million are in stock investment, 1 million are in fund investment, and 1 million are in time deposits. I hope that the investment amount in new financial products can be between 500,000 and 1 million. In addition, I am also concerned about the impact of management fees on financial products, and hope to minimize unnecessary management fee expenditures." After word segmentation and part-of-speech tagging, we build a dependency syntactic analysis tree to clarify the grammatical relationship between each word.

[0024] Then, execute a pattern matching algorithm on the dependency syntactic analysis tree to identify the modified relationship structures containing comparatives and superlatives. In this text, "minimize as much as possible" in "hoping to minimize unnecessary management fee expenditures" is a comparative expression. Extract the user preference intensity indicator and the conditional constraint expression from the modified relationship structure. Here, "minimize as much as possible" is the preference intensity indicator, and "management fee expenditures" is the conditional constraint expression.

[0025] Identify the beneficiary, target object, and negative conditions in the text input content through a semantic role labeling model to obtain the semantic role labeling result. In the above example, the beneficiary is the investor himself / herself, the target object is the financial product, and the negative condition is "unnecessary management fee expenditures" (i.e., not hoping for such a situation). Finally, perform feature encoding on the user preference intensity indicator, the conditional constraint expression, and the semantic role labeling result to generate a multi-dimensional semantic label set.

[0026] Step S130, perform spatio-temporal alignment and fusion on the structured requirement description set and the multi-dimensional semantic label set to generate a user portrait enhanced feature matrix.

[0027] Specifically, first perform timestamp alignment verification on the structured requirement description set and the multi-dimensional semantic label set. Assume that voice input and text input are almost simultaneous, but due to differences in system processing speed, there may be a very small time difference. Thus, it is possible to identify the time overlap interval and the time sequence interval fault between the voice semantic paragraphs in the structured requirement description set and the text semantic paragraphs in the multi-dimensional semantic label set. For example, in the above example, since the two are almost simultaneously input, the time overlap interval almost covers the entire input process, and the time sequence interval fault can be almost ignored.

[0028] Then, input the voice semantic paragraphs and text semantic paragraphs within the time overlap interval into a gated recurrent unit network. Calculate the semantic conflict probability between the voice semantic paragraphs and the text semantic paragraphs at the time sequence interval fault (although very small but assume there is a little) through the forget gate of the gated recurrent unit network. For example, assume that the voice mentions "hoping that the product risk is lower than stock investment", while the text does not explicitly mention the risk comparison with stock investment, and calculate that there may be a certain semantic conflict probability here. Dynamically allocate weights to the voice semantic paragraphs and the text semantic paragraphs based on this semantic conflict probability.

[0029] Extract the historical voice interaction fragments and historical text interaction fragments associated with the current time overlap interval from the user's historical interaction records. Assume that the user has had historical interaction records of discussing investment risks and management fees with the customer service before, and generate a semantic consistency verification vector based on the recurrence frequency of these historical interaction fragments. If the issues of risk control and management fees have been mentioned many times before, then the relevant weights in this semantic consistency verification vector will be relatively high.

[0030] Perform a point - by - point multiplication operation on the semantic consistency verification vector and the output of the gated recurrent unit network to correct the confidence scores of the speech semantic paragraph and the text semantic paragraph after dynamic weight allocation. Perform cross - modal splicing on the speech semantic paragraph and the text semantic paragraph according to the corrected confidence scores to generate a fused cross - modal semantic unit sequence.

[0031] Extract combinations of semantic keywords that continuously appear in the cross - modal semantic unit sequence, such as "risk lower than stocks", "withdrawal within four or five years", "reduce management fee expenditure", etc. Match these combinations of semantic keywords with the product feature tags in the user's historical purchase records to identify implicit demand markers in the cross - modal semantic unit sequence. Assume that the product feature tags in the user's historical purchase records include tags such as the risk level of previously purchased stock products being high - risk, the income type of fund products being medium - income, and the term structure of time deposits being long - term, etc. Through pattern matching, it is found that the keyword combination "risk lower than stocks" has a contrast relationship with the high - risk stock product feature tag, which may imply an implicit demand for medium - risk or low - risk products.

[0032] Insert the implicit demand markers into the corresponding positions in the cross - modal semantic unit sequence to generate a user portrait enhanced feature matrix with temporal context association. This user portrait enhanced feature matrix not only contains the needs explicitly expressed by the user but also incorporates the implicit needs mined from historical records and cross - modal analysis, comprehensively depicting the user's demand portrait in this financial investment scenario.

[0033] Step S140, based on the deep matching model, perform multi - level association analysis on the user portrait enhanced feature matrix and the financial product knowledge graph to generate an initial recommendation result queue.

[0034] Specifically, product attribute nodes and their associated regulatory policy nodes and market condition nodes can be extracted from the financial product knowledge graph. For example, for the product attribute node of wealth management products in the financial product knowledge graph, the associated regulatory policy nodes may include risk management regulations and sales restriction regulations for wealth management products, and the market condition nodes may include the current market interest rate situation and the market supply - demand situation of similar wealth management products, etc.

[0035] Then, based on the bi - directional graph attention network in the deep matching model, calculate the first matching degree between the user portrait enhanced feature matrix and each product attribute node. Assume that the user portrait enhanced feature matrix has clear requirements in terms of risk, term, income, and management fee, etc. The bi - directional graph attention network will comprehensively consider these factors and calculate the matching degree with different wealth management product attribute nodes. For example, for a wealth management product with low risk, a term of three to five years, relatively stable income, and more preferential management fees, a relatively high first matching degree may be calculated.

[0036] Next, traverse the regulatory policy nodes and market condition nodes through the meta-path random walk algorithm to generate the product compliance evaluation coefficient. Determine the set of meta-path rules between the regulatory policy nodes and market condition nodes in the financial product knowledge graph. For example, the meta-path rules from the financial product attribute node to the regulatory policy node may include the conduction path according to product type - regulatory category - specific policy, and the meta-path rules from the financial product attribute node to the market condition node may be the conduction path according to product type - market segment - market indicator.

[0037] On this basis, according to the path type priority list of each meta-path in the set of meta-path rules, perform multi-hop traversal operations on the financial product knowledge graph to generate a sequence of regulatory policy nodes and a sequence of market condition nodes associated with the product attribute node. For example, for financial products, the sequence of regulatory policy nodes obtained by traversal may include the policy node on the recent release of the sales specification of financial products, and the sequence of market condition nodes may include the average yield node of financial products in the current market, etc.

[0038] Then, extract the effective time interval and the constraint effect intensity parameter of each regulatory policy node from the sequence of regulatory policy nodes, and extract the fluctuation amplitude index and the trend direction identifier of each market condition node from the sequence of market condition nodes. Suppose the effective time interval of a certain regulatory policy node is from last year to the present, the constraint effect intensity parameter is medium, the fluctuation amplitude index of the market condition node is small, and the trend direction identifier is stable.

[0039] Then, perform an overlap degree matching between the effective time interval and the current system time to generate the regulatory policy timeliness weight, and perform a multiplication operation between the constraint effect intensity parameter and the regulatory policy timeliness weight to generate a set of dynamically adjusted regulatory policy impact factors. For example, if the current time is within the effective interval of the regulatory policy and close to the middle position, the generated regulatory policy timeliness weight is 0.5, and multiplying it by the constraint effect intensity parameter 0.5 gives 0.25 as the regulatory policy impact factor.

[0040] Next, compare the fluctuation amplitude index with the preset market sensitivity threshold, screen out the market condition nodes that exceed the market sensitivity threshold as significantly fluctuating market condition nodes, and perform a consistency matching between the trend direction identifier and the risk preference label in the user portrait enhancement feature matrix to generate a set of market condition adaptation degrees. Suppose in the above example, there are no market condition nodes that exceed the market sensitivity threshold, and the trend direction identifier matches well with the user risk preference label, generating a good set of market condition adaptation degrees.

[0041] Then, according to the path hop counts of each node in the dynamically adjusted regulatory policy impact factor set, a decreasing propagation weight coefficient is assigned, and the propagation weight coefficient is weighted and summed with the corresponding regulatory policy impact factor to generate a basic product compliance score. Then, according to the occurrence frequency of the significant volatility market nodes and the matching results in the market condition adaptation degree set, the product market stability offset is calculated, and the product market stability offset is superimposed on the basic product compliance score to generate a product compliance evaluation coefficient.

[0042] Finally, the first matching degree and the compliance evaluation coefficient are non-linearly combined to generate a comprehensive product recommendation score. According to the comprehensive recommendation score, the financial products are sorted and screened to generate an initial recommendation result queue. For example, different financial products such as wealth management products and fund products are arranged in descending order of the comprehensive recommendation score to form an initial recommendation result queue, and the products ranked in the front are the products that better meet the user's needs and compliance requirements.

[0043] Step S150, perform context awareness optimization on the initial recommendation result queue according to the real-time environment perception data, generate a dynamically adapted recommendation list, and perform visual rendering and output through a multi-channel interaction interface.

[0044] Specifically, the position sensor data and ambient light intensity parameters of the user terminal (tablet computer) can be obtained. Suppose the position sensor data shows that the user is in the study at home and the ambient light intensity is moderate. Calculate the user's current attention level based on this information. Since the user is in a quiet study environment, the attention level is relatively high. At the same time, detect the device network latency index and interface rendering frame rate to evaluate the terminal interaction response ability. Suppose the network latency is low and the interface rendering frame rate is normal, indicating that the terminal interaction response ability is good.

[0045] Adjust the display granularity level of the initial recommendation result queue according to the attention level to generate a dynamic information density parameter. Since the user's attention level is high and more detailed information can be provided, increase the display granularity level and improve the dynamic information density parameter. Optimize the loading priority strategy of the initial recommendation result queue according to the terminal interaction response ability. Because the terminal interaction response ability is good, more recommended product information can be loaded preferentially.

[0046] Apply the dynamic information density parameter and the loading priority strategy to the initial recommendation result queue to generate a dynamically adapted recommendation list.

[0047] Create a 3D visualization space in the interactive interface, and map the recommended products in the dynamically adapted recommendation list to three-dimensional icons of different shapes. For example, map products with lower risks to cube icons, and products with higher risks to triangular pyramid icons. Dynamically adjust the color saturation and rotation speed of the three-dimensional icons according to the product risk level. The icons of low-risk products have lower color saturation and slower rotation speed; the icons of high-risk products have higher color saturation and faster rotation speed.

[0048] Build an interactive timeline control to respond to the user's sliding operation and update the spatio-temporal distribution of the recommendation results in real time. When the user gazes at any three-dimensional icon for a duration exceeding the threshold duration, trigger a multi-level information expansion animation. For example, when the user gazes at the icon of a financial product for more than 3 seconds, an animation display of multi-level information such as the detailed income curve and risk assessment report of the product will pop up.

[0049] Generate a focus heat map based on eye tracking in the edge area of the interface, and adjust the layout density of the recommendation elements in real time. If the user's sight is more concentrated in the left area of the interface, then appropriately increase the layout density of the recommendation elements in the left area to better display the product information that the user may be more interested in.

[0050] Based on the above steps, in the embodiment of the present invention, by collecting the multi-modal dialogue data stream including voice input signals and text input content during the interaction between the user terminal and the intelligent customer service system, and then processing and fusing the multi-modal data respectively, the accuracy and effectiveness of data processing are further improved. The structured requirement description set generated by semantic recognition of the voice input signal, and the multi-dimensional semantic label set generated by entity extraction and intention parsing of the text input content are fused through spatio-temporal alignment to form a user portrait enhanced feature matrix, which not only integrates the advantages of different modal data, but also mines the potential spatio-temporal correlation between data. Then, based on the deep matching model, a multi-level correlation analysis is performed between the user portrait enhanced feature matrix and the financial product knowledge graph, which can make full use of the rich product information and relationship network in the financial product knowledge graph, not only considering the simple matching between the user's needs and the product directly, but also through multi-level in-depth analysis, mining potential financial product combinations that meet the user's complex needs, greatly broadening the breadth and depth of recommendations. Finally, the initial recommendation result queue is optimized with context awareness according to the real-time environment perception data, and a dynamically adapted recommendation list is generated and visualized and rendered through a multi-channel interactive interface, which can flexibly optimize the recommendation results according to factors such as the real-time market environment and the user's current scenario. The visualization rendering output of the multi-channel interactive interface improves the user experience, and the user can conveniently obtain the recommendation information in their own accustomed way, enhancing the practicality and user-friendliness of the recommendation system.

[0051] In a possible implementation manner, step S120 includes:

[0052] Step S121, performing frame processing and background noise elimination on the speech input signal, and extracting Mel-frequency cepstral coefficients as acoustic feature vectors.

[0053] In this embodiment, in the scenario described above, since there are some background noises in the investor's study, such as the sound of cars driving in the distant street and the faint sound of air conditioning, the voice input signal can be divided into frames according to the set time interval, which is determined according to the characteristics and processing requirements of the voice input signal, for example, one frame every 10 milliseconds. Then, for each frame signal, the interference of background noise is eliminated by a signal processing algorithm, and the signal processing algorithm is analyzed and processed based on the characteristics of sound frequency, energy and other aspects. After the background noise is eliminated, the system extracts the Mel frequency cepstral coefficient as an acoustic feature vector. The Mel frequency cepstral coefficient is an acoustic feature representation based on the auditory characteristics of the human ear, which reflects the energy distribution of the voice input signal in different frequency ranges. Through calculation, the voice input signal can be converted from the time domain to the Mel frequency domain, and then its cepstral coefficient is calculated. This process involves windowing, fast Fourier transform, Mel filter bank filtering and logarithmic operation of the voice input signal, and finally obtains an acoustic feature vector, which contains important acoustic feature information of the voice input signal.

[0054] Step S122: convert the speech input signal into an intermediate text representation through a pre-trained speech recognition model, and annotate the speaker's emotion intensity value.

[0055] In this embodiment, the pre-trained speech recognition model is trained based on a large amount of speech data in the financial field and general speech data. After receiving the processed speech input signal, the speech recognition model analyzes the acoustic feature vector and gradually converts it into an intermediate text representation using the multi-layer structure of the neural network. In this process, different layers of the speech recognition model will learn different levels of features in the speech input signal, from the underlying acoustic features to the high-level semantic features, and finally generate an intermediate text representation, such as "I have made various investments in the past, including stock funds and regular deposits. Now I adjust the portfolio and look for low-correlated stocks, products with good liquidity, available for withdrawal within four or five years, and lower risks than stocks." At the same time, the speaker's emotional intensity value can be annotated according to the characteristics of the voice tone, speech speed, etc. In this scenario, the investor states the demand more rationally, and the emotional intensity value is annotated as medium. The annotation of the emotional intensity value is obtained by analyzing the acoustic features of the voice, such as the fundamental frequency, intensity, and duration. For example, by analyzing the stability of the tone, the normality of the speech speed, etc., if the tone is relatively stable, the speech speed is normal, and there is no obvious excitement or frustration, it can be determined as a medium emotional intensity value.

[0056] Step S123, using a multi-head attention mechanism to perform contextual semantic completion on the intermediate text representation to generate a corrected complete semantic paragraph.

[0057] In this embodiment, the multi-head attention mechanism is a mechanism that can simultaneously pay attention to multiple different representation subspaces in the text. In this scenario, for the intermediate text representation, the multi-head attention mechanism can be considered from multiple aspects such as common investment knowledge and language logic in the financial field. For example, it can be based on the knowledge of the correlation between different products in financial investment, as well as the common language structures used by people when expressing investment needs. For expressions such as "find low-correlation stocks, decent liquidity, withdrawable within four or five years, and products with lower risks than stocks", some missing information can be supplemented, such as supplementing "because I have a child who is about to go to college, I need to withdraw a large amount of funds during this period" and other information, thereby generating a corrected complete semantic paragraph "I have made a variety of investments such as stocks, funds and time deposits in the past few years. Now I want to adjust my investment portfolio. I hope to find a financial product that is not so strongly correlated with stocks, has good liquidity, can be withdrawn in the next four or five years, and has lower risks than stock investments, because I have a child who is about to go to college and needs to withdraw a large amount of funds during this period." In this process, the multi-head attention mechanism obtains information from different representation subspaces by performing a multi-dimensional analysis of the relationship between each word and other words in the intermediate text representation, thereby achieving contextual semantic completion.

[0058] Step S124, using a domain-specific named entity recognition model to extract constraint entities of amount, term, and risk category from the complete semantic paragraph.

[0059] In this embodiment, the domain-specific named entity recognition model is a model trained specifically for the financial field. In the complete semantic paragraph "I have made a variety of investments such as stocks, funds, and time deposits in the past few years. Now I want to adjust my investment portfolio. I hope to find a financial product that is not so strongly correlated with stocks, has good liquidity, can be withdrawn in the next four or five years, and has lower risk than stock investment, because I have a child who is about to go to college and needs to withdraw a large amount of money during this period." The model will recognize that the risk constraint is "lower risk than stock investment" and the term constraint is "within four or five years", and there is no explicit mention of direct constraints related to the amount in this statement (but the condition that the investable amount is within a certain range is implied under the overall demand). This recognition process is based on the model's feature representation and pattern learning of entities such as amount, term, and risk that are common in the financial field. These constraint condition entities are accurately extracted by performing feature analysis and pattern matching on each word in the complete semantic paragraph.

[0060] Step S125: Perform tensor concatenation on the acoustic feature vector, the emotional intensity value, and the constraint condition entity to generate the structured requirement description set with emotional weights.

[0061] In this embodiment, a tensor is a high-dimensional data representation form. First, the acoustic feature vector contains the acoustic feature information of the speech input signal, the emotional intensity value represents the emotional state of the speaker, and the constraint condition entity defines the key constraint information in the investment requirements. When performing tensor concatenation, these three parts can be combined into a new tensor structure according to certain rules. For example, assume that the acoustic feature vector is an n-dimensional vector, the emotional intensity value can be represented as a single numerical value, and the constraint condition entity can be converted into an m-dimensional vector (where n and m are determined according to the actual number of features). Perform a multiplication operation on each element in the vector of the emotional intensity value and the constraint condition entity to obtain a constraint condition entity vector with emotional weights, and then combine this constraint condition entity vector with emotional weights and the acoustic feature vector according to the rules of tensor concatenation to form a new tensor, which is the structured requirement description set with emotional weights. This structured requirement description set comprehensively includes the acoustic features, emotional information, and constraint condition entity information in the semantic meaning of the speech input signal.

[0062] In a possible implementation manner, step S120 further includes:

[0063] Step S126: Perform word segmentation and part-of-speech tagging on the text input content, construct a dependency syntactic analysis tree, and execute a pattern matching algorithm on the dependency syntactic analysis tree to identify the modified relationship structures containing comparatives and superlatives.

[0064] For example, in the previously set scenario, the text input content entered by the investor is "My current total assets are approximately 5 million, of which 3 million is invested in stocks, 1 million is invested in funds, and 1 million is in fixed deposits. I hope the investment amount in the new financial product can be between 500,000 and 1 million. In addition, I am also relatively concerned about the management fees of financial products and hope to minimize unnecessary management fee expenditures." Then, the text can be segmented according to predefined word segmentation rules. For example, words such as "I", "current", "total assets", "approximately", "is" are segmented separately. Part-of-speech tagging is to determine the part of speech of each word. For example, "I" is a pronoun, "total assets" is a noun, "approximately" is an adverb, etc. On this basis, a dependency syntactic analysis tree is constructed. The dependency syntactic analysis tree describes the syntactic relationship between words. For example, there is a subject-predicate relationship between "total assets" and "is", and an object relationship between "5 million" and "is". Each word in the entire sentence is connected through this dependency relationship to form an analysis tree, which clearly shows the grammatical structure of the text.

[0065] In addition, in the above text, the phrase "hoping to minimize unnecessary management fee expenditures" contains a modified relationship structure with a comparative degree. The pattern matching algorithm searches for structures that conform to the comparative and superlative patterns in the dependency syntactic analysis tree. This process is carried out by matching with predefined syntactic patterns for the comparative and superlative degrees. For example, for the comparative degree, possible patterns include structures with words such as "more", "comparatively", "as much as possible", etc. plus a verb or an adjective. When the structure "hoping to minimize unnecessary management fee expenditures" is recognized, it is determined as a modified relationship structure containing a comparative degree.

[0066] Step S127: Extract the user preference intensity indicator and the conditional constraint expression from the modified relationship structure.

[0067] For example, in the structure "hoping to minimize unnecessary management fee expenditures", "minimize as much as possible" is the user preference intensity indicator, which indicates the degree of the investor's expectation for reducing management fee expenditures. "Management fee expenditures" is the conditional constraint expression, which is an important condition that the investor is concerned about, that is, the impact of the financial product on management fees. The extraction process is based on the semantic analysis of the words in the modified relationship structure. According to the predefined semantic rules, the word indicating the degree is determined as the preference intensity indicator, and the modified noun or noun phrase is determined as the conditional constraint expression.

[0068] Step S128: Identify the beneficiary, target object, and negative conditions in the text input content through a semantic role annotation model to obtain the semantic role annotation result.

[0069] For example, in the entire text input content, the beneficiary is the investor himself / herself because all investment decisions and expectations are centered around the investor's own asset status and needs. The target object is the financial product, which is the object of the investor's investment operation. The negative condition is "unnecessary management fee expenditures", which means that the investor does not want this situation to occur. The semantic role annotation model determines these roles by analyzing the lexical semantics and syntactic structure in the text. For example, for the determination of the beneficiary, the model analyzes the initiator of the action and the recipient of the benefit in the sentence. In investment-related expressions, the investor is the subject of various investment behaviors and also the bearer of investment returns or impacts, so it is determined as the beneficiary; for the target object, the financial product is determined as the target object according to the relationship between investment-related verbs such as "invest" and related nouns; for the negative condition, it is determined by identifying words expressing negative meanings such as "unnecessary" and the object it modifies.

[0070] Step S129: Perform feature encoding on the user preference intensity indicator, conditional constraint expression, and semantic role annotation result to generate the multi-dimensional semantic label set.

[0071] In this scenario, first, the user preference intensity indicator "minimize" is quantized and encoded. For example, according to a pre-set preference intensity rating scale, "minimize" may be encoded as a relatively high intensity value, indicating a strong expectation of the investor to reduce management fee expenses. The conditional constraint expression "management fee expenses" can be encoded as a specific code, which corresponds to the relevant concept of management fees in the financial field. The beneficiary "the investor himself / herself", the target object "financial product", and the negative condition "unnecessary management fee expenses" are also encoded as corresponding codes respectively. These encodings are combined together according to certain rules to form a multi-dimensional semantic tag set. This multi-dimensional semantic tag set can represent the semantic information in the text input content from multiple dimensions, including information such as the user's preference intensity, the conditional constraints of concern, and the relevant roles. For example, this multi-dimensional semantic tag set may represent the preference intensity value in one dimension, the type of target object in another dimension, and the negative conditions in other dimensions, comprehensively depicting the semantic characteristics of the text input content through the above multi-dimensional representation method.

[0072] In a possible implementation manner, step S130 includes:

[0073] Step S131, perform timestamp alignment verification on the structured requirement description set and the multi-dimensional semantic tag set, and identify the time overlap interval and the time sequence interval fault between the speech semantic paragraphs in the structured requirement description set and the text semantic paragraphs in the multi-dimensional semantic tag set.

[0074] In the previously set scenario, since the investor's speech input and text input are almost simultaneous, timestamps can be marked for the speech input and text input respectively. The determination of the time overlap interval is achieved by comparing the start time and end time of the speech semantic paragraph and the text semantic paragraph. In this scenario, assume that the speech input starts at time t1 and ends at time t2, the text input starts at time t1 + Δt1 (Δt1 is very small, indicating that the text input may start slightly later than the speech input but is almost simultaneous), and ends at time t2 + Δt2 (Δt2 is also very small). Then the time overlap interval is from t1 + Δt1 to t2. And the time sequence interval fault is very small in this almost simultaneous input situation. For example, due to minor differences in internal system processing, there may be a very small inconsistency in the processing progress of the speech semantic paragraph and the text semantic paragraph at a certain extremely short moment, but overall this fault can be almost ignored.

[0075] Step S132: Input the speech semantic paragraph and the text semantic paragraph within the time overlap interval into the gated recurrent unit network. Calculate the semantic conflict probability between the speech semantic paragraph and the text semantic paragraph at the time series interval fault through the forget gate of the gated recurrent unit network, and perform dynamic weight allocation on the speech semantic paragraph and the text semantic paragraph based on the semantic conflict probability.

[0076] In this embodiment, the gated recurrent unit network is a special recurrent neural network structure, and the included forget gate is used to control the degree of information forgetting. In this scenario, for the speech semantic paragraph and the text semantic paragraph within the time overlap interval, the forget gate analyzes the semantic relationship between the two at the time series interval fault. For example, if the speech semantic paragraph mentions that "the risk is lower than stock investment" and the text semantic paragraph does not explicitly mention the risk comparison with stock investment, there may be a semantic conflict. The forget gate can analyze the degree of this semantic difference according to the predefined semantic conflict calculation rules, so as to calculate the semantic conflict probability. Suppose that according to the degree of semantic difference and the preset calculation rules, the calculated semantic conflict probability is p (this p value is obtained through comprehensive analysis of various factors such as lexical semantics, syntactic structure, and semantic association in the financial field in the speech semantic paragraph and the text semantic paragraph, and is a value between 0 and 1). Perform dynamic weight allocation on the speech semantic paragraph and the text semantic paragraph based on this semantic conflict probability. Suppose the initial weight of the speech semantic paragraph is w1 and the initial weight of the text semantic paragraph is w2. Then, according to the semantic conflict probability p, the adjusted weight of the speech semantic paragraph w1' = w1 * (1 - p), and the adjusted weight of the text semantic paragraph w2' = w2 * p. This dynamic weight allocation adjusts the importance of the speech semantic paragraph and the text semantic paragraph in subsequent processing according to the possibility of semantic conflict.

[0077] Step S133: Extract the historical speech interaction segment and the historical text interaction segment associated with the current time overlap interval from the user's historical interaction record, and generate a semantic consistency verification vector according to the recurrence frequency of the historical speech interaction segment and the historical text interaction segment.

[0078] In this scenario, the investor has a historical interaction record of discussing investment risks and management fees with the customer service before. The parts associated with the overlapping time interval with the current time can be filtered out from these historical interaction records. For example, the investor mentioned risk control and management fees in a previous discussion about portfolio adjustment. These relevant historical voice interaction segments and historical text interaction segments will be extracted. For each historical interaction segment, the frequency of its repeated occurrence can be counted. Suppose the historical voice interaction segment about risk control has occurred m times in the past n relevant interactions, then its repeated occurrence frequency is m / n. According to the above repeated occurrence frequency, a value is assigned to each relevant historical interaction segment, and these values are combined in a certain order to form a semantic consistency verification vector. For example, if there are three relevant historical interaction segments with repeated occurrence frequencies of f1, f2, and f3 respectively, then the semantic consistency verification vector may be represented as [f1, f2, f3]. This semantic consistency verification vector reflects the relevant information about the semantic consistency degree between the historical interaction and the current interaction.

[0079] Step S134, perform a pointwise multiplication operation on the semantic consistency verification vector and the output of the gated recurrent unit network to correct the confidence scores of the speech semantic passage and the text semantic passage after the dynamic weight allocation.

[0080] In this embodiment, the output of the gated recurrent unit network contains the information of the processed speech semantic passage and text semantic passage. Suppose the speech semantic passage information output by the gated recurrent unit network is v1, and the text semantic passage information is v2, and their corresponding confidence scores are s1 and s2 respectively (the confidence score is a measure of the reliability of the speech semantic passage and text semantic passage information, and the initial value may be based on the system's default settings or some preliminary processing results in the early stage). The semantic consistency verification vector is [c1, c2, c3] (here, for the sake of simplicity, it is assumed that the semantic consistency verification vector has three elements). After performing the pointwise multiplication operation, the corrected confidence score of the speech semantic passage s1' = s1 * c1, and the corrected confidence score of the text semantic passage s2' = s2 * c2 (here, only the first two elements of the vector are taken for example calculation, and the actual situation will be comprehensively calculated according to the complete definition and processing rules of the vector). This correction operation uses historical interaction information to adjust the confidence scores of the current speech semantic passage and text semantic passage, making the result more accurate and reliable.

[0081] Step S135, perform cross-modal splicing on the speech semantic passage and the text semantic passage according to the corrected confidence scores to generate a fused cross-modal semantic unit sequence.

[0082] In this embodiment, after obtaining the corrected confidence scores s1' and s2', the speech semantic paragraph and the text semantic paragraph are combined according to a certain splicing rule. For example, if a higher confidence score indicates that the semantic information of this part is more reliable, then the proportion of the speech semantic paragraph and the text semantic paragraph in the splicing result can be determined according to the proportion of the confidence scores. Suppose the speech semantic paragraph is represented as x1 after a certain transformation, and the text semantic paragraph is represented as x2 after transformation. The spliced cross-modal semantic unit sequence can be represented as the result of mixing according to the confidence score ratio, such as y = (s1' * x1 + s2' * x2) / (s1' + s2') (this is just a simple example calculation method, and the actual splicing rule can consider more factors such as the structure and semantics of the speech semantic paragraph and the text semantic paragraph). This cross-modal semantic unit sequence integrates the semantic information of both speech and text modalities and is reasonably combined according to the confidence scores.

[0083] Step S136, extract the combination of semantic keywords that continuously appear in the cross-modal semantic unit sequence, perform pattern matching between the combination of semantic keywords and the product feature tags in the user's historical purchase records, and identify the implicit demand markers in the cross-modal semantic unit sequence.

[0084] In the foregoing scenario, there may be combinations of semantic keywords such as "risk lower than stocks", "withdrawal within four or five years", "reduce management fee expenditure" in the cross-modal semantic unit sequence. From the user's historical purchase records, the product feature tags of the products previously purchased by the investor include tags such as the risk level of the previously purchased stock product being high risk, the income type of the fund product being medium income, and the term structure of the time deposit being long term. For each combination of semantic keywords, it can be pattern-matched with the product feature tags in the historical purchase records. For example, for the keyword combination "risk lower than stocks", there is a comparison relationship with the high-risk product feature tag of stocks, which may imply an implicit demand for medium-risk or low-risk products. The semantic relationship can be analyzed and judged according to the preset pattern matching rules. Suppose according to a series of matching rules and calculations (these rules and calculations involve various factors such as semantic similarity and logical judgment of semantic relationships), it is determined that the probability of the keyword combination "risk lower than stocks" having an implicit demand for medium-risk products is q1 (q1 is a value between 0 and 1, representing the probability of this implicit demand). By performing such pattern matching on all combinations of semantic keywords, the corresponding implicit demand markers and their probabilities for each keyword combination are determined.

[0085] Step S137, insert the implicit demand markers into the corresponding positions of the cross-modal semantic unit sequence to generate the enhanced feature matrix of the user portrait with temporal context association.

[0086] In this embodiment, after determining each implicit demand tag and its probability, according to the probability level or other preset insertion rules, these implicit demand tags are inserted into the corresponding positions in the cross-modal semantic unit sequence. For example, if the probability q1 of the implicit demand tag for medium-risk products corresponding to "risk lower than stocks" is relatively high, then this implicit demand tag is inserted near the semantic part related to risk in the cross-modal semantic unit sequence. After inserting all the determined implicit demand tags according to the rules, a user profile enhancement feature matrix with temporal context association is generated. This user profile enhancement feature matrix not only contains the explicit information in the speech semantic paragraph and the text semantic paragraph, but also incorporates the implicit demand information mined from historical records and semantic analysis, comprehensively depicting the demand profile of investors in this financial investment scenario.

[0087] In one possible implementation manner, step S136 includes:

[0088] Step S1361, extract the set of product feature tags of all the purchased financial products from the user's historical purchase records, where the set of product feature tags includes product risk level tags, income type tags, and term structure tags.

[0089] In this embodiment, in the previously set scenario, the investor has purchased various financial products before. For example, he has purchased a stock product before, and its risk level tag is high risk; he has purchased a fund product, and the income type tag is medium income; he has also purchased a time deposit, and the term structure tag is long term. These tags completely describe the investment characteristics of the investor in different financial products in the past, providing an important reference basis for subsequent analysis.

[0090] Step S1362, perform part-of-speech tagging and stop word filtering on the semantic keyword combinations in the cross-modal semantic unit sequence, retain the core semantic units containing noun phrases and adjective phrases, and generate a dimension-reduced target keyword sequence.

[0091] For example, in a cross-modal semantic unit sequence, there may be semantic keyword combinations such as "risk lower than stocks", "withdrawal within four or five years", "reduce management fee expenditure", etc. These semantic keyword combinations can be part-of-speech tagged. For example, "risk" is a noun, "lower than" is a preposition, and "stocks" is a noun, so as to determine the part of speech of each word. Then stop word filtering is performed. Prepositions like "than" belong to stop words in this scenario and are filtered out. After processing, the core semantic units containing noun phrases (such as "risk", "stocks") and adjective phrases (if there are relevant adjective phrases) are retained to generate a target keyword sequence after dimensionality reduction. For example, after processing, a target keyword sequence composed of core semantic units such as "risk stocks", "withdrawal in four or five years" may be obtained. This dimensionality reduction operation helps to focus on key semantic information and improve the accuracy and efficiency of subsequent matching.

[0092] Step S1363: Perform item-by-item pattern matching between each noun phrase in the target keyword sequence and the product risk level label, identify the risk preference synonym cluster implicit in the noun phrase, and generate a risk level mapping relationship table.

[0093] For example, in the noun phrase "risk stocks" in the target keyword sequence, "risk" is associated with the risk level label of the previously mentioned high-risk stock product. Thus, through pre-defined pattern matching rules for risk-related vocabulary, a synonym cluster related to "risk" can be identified, such as words that may be semantically similar to "risk" like "danger", "uncertainty", etc. Suppose in this scenario, after comprehensive pattern matching, the synonym cluster related to "risk" is identified as including "danger" and "uncertainty", and the association between "risk" and the high-risk stock product indicates that the investor has an understanding of high risks and expresses an expectation of a risk lower than that of stocks in the current demand. Based on this information, a risk level mapping relationship table is generated, which records the mapping relationship between the risk-related vocabulary in the noun phrase and the risk level, such as "risk - high risk (stocks), expectation lower than this risk", and at the same time records the occurrence frequency of each synonym cluster. For example, "risk" appears 1 time, "danger" appears 0 times in other relevant expressions, "uncertainty" appears 0 times, etc.

[0094] Step S1364: Based on the semantic similarity calculation between the adjective phrase and the income type label, screen out the income descriptors that meet the preset similarity threshold, and associate the corresponding income type labels to generate an income demand mapping relationship table.

[0095] Assume that there are some profit-related adjective phrases (if any) in the target keyword sequence. The semantic similarity between these adjective phrases and profit type labels (such as medium profit) can be calculated. The calculation of semantic similarity is based on the lexical semantic knowledge in the financial field and predefined semantic calculation rules. For example, if there is an adjective phrase such as "stable", the similarity can be calculated according to the semantic association between "stable profit" and medium profit in the financial field. Assume that the preset similarity threshold is 0.6. If the calculated semantic similarity between "stable" and medium profit is 0.7, meeting the preset threshold, then "stable" is selected as a profit descriptor and associated with the corresponding medium profit type label to generate a profit demand mapping relationship table. The table records the corresponding relationship between the profit descriptor and the profit type label and the relevant calculation results, such as "stable - medium profit, similarity is 0.7".

[0096] Step S1365, traverse the time unit identifiers of the term structure label, extract consecutive word pairs containing numerical values and time units from the target keyword sequence, and perform an interval inclusion check on the consecutive word pairs with the allowable value range of the term structure label to generate a term matching relationship table.

[0097] For example, in the previously mentioned term structure label of a time deposit is long term, and its allowable value range may be 3 years and above. From the target keyword sequence, such as the consecutive word pair "withdrawal in four or five years", where "four or five years" is a combination containing a numerical value and a time unit. Then, this consecutive word pair can be subjected to an interval inclusion check with the allowable value range of the term structure label of the time deposit. Since "four or five years" is within the range of 3 years and above, it meets the interval inclusion requirement. By checking all such consecutive word pairs, a term matching relationship table is generated. The table records the matching relationship between the consecutive word pairs and the term structure label and the check result, such as "four or five years - time deposit (long term), meets interval inclusion".

[0098] Step S1366, generate a multi-dimensional demand intensity vector based on the occurrence frequency of the synonym clusters in the risk level mapping relationship table, the emotional polarity intensity of the profit descriptors in the profit demand mapping relationship table, and the interval matching degree of the consecutive word pairs in the term matching relationship table.

[0099] For example, in the risk level mapping relationship table, the frequency of occurrence of "risk" is 1. In the income demand mapping relationship table, if the income descriptor "stable" is regarded as a positive description, its emotional polarity intensity can be set to +1 (assuming that the emotional polarity intensity of positive descriptions is +1 and that of negative descriptions is -1). In the term matching relationship table, the interval matching degree of "four or five years - time deposit (long term)" is 1 (indicating a complete match). Based on this information, a multi-dimensional demand intensity vector is generated, such as [1 (risk frequency), +1 (income emotional polarity intensity), 1 (term matching degree)]. This multi-dimensional demand intensity vector comprehensively describes the demand intensity of investors in terms of risk, income, and term from multiple dimensions.

[0100] Step S1367: Input the multi-dimensional demand intensity vector into the time decay function, and calculate the decayed demand weight value according to the purchase timestamps of each product feature label in the user's historical purchase records.

[0101] Suppose the time when the investor purchased high-risk stocks is t1, and the time interval from the current time is Δt1; the time when the investor purchased medium-income funds is t2, and the time interval from the current time is Δt2; the time when the investor purchased long-term time deposits is t3, and the time interval from the current time is Δt3. The time decay function is preset according to the changing rules of the financial market and the changing characteristics of investors' demands over time. For example, the time decay function may be set such that as the time interval increases, the demand weight decreases at a certain ratio. When calculating the decayed demand weight corresponding to high-risk stocks, first determine the corresponding decay coefficient on the time decay function curve according to Δt1. Suppose in the time decay function, when the decay coefficient corresponding to Δt1 is 0.8 (this decay coefficient is calculated based on the definition of the time decay function and the value of Δt1, and the calculation process involves evaluating the time decay function, such as calculating according to the function expression, the value of Δt1, and the relevant parameters in the function). Then the decayed demand weight value of high-risk stocks in the risk dimension is 1 (risk frequency) * 0.8 = 0.8. Similarly, calculate the decayed demand weight values of medium-income funds and long-term time deposits in the income and term dimensions. Suppose the decay coefficient corresponding to medium-income funds is 0.9, and the calculated decayed demand weight value in the income dimension is +1 (income emotional polarity intensity) * 0.9 = 0.9; the decay coefficient corresponding to long-term time deposits is 0.7, and the calculated decayed demand weight value in the term dimension is 1 (term matching degree) * 0.7 = 0.7.

[0102] Step S1368: Perform weighted fusion on the risk level mapping relationship table, the income demand mapping relationship table, and the term matching relationship table according to the demand weight values to generate a latent demand probability distribution matrix.

[0103] Taking the risk level mapping relationship table as an example, for the records related to high-risk stocks, the previous record "Risk - High Risk (Stock), Expectation Lower than this Risk" is weighted according to the decayed demand weight value of 0.8. Assume that the weights of each item in the original risk level mapping relationship table are defaulted to 1, and the new weight relationship is obtained after weighting. Similarly, each item in the return demand mapping relationship table and the maturity matching relationship table is weighted according to their respective decayed demand weight values. Then these three weighted tables are fused to generate a latent demand probability distribution matrix. Each element in this latent demand probability distribution matrix synthesizes the information of three dimensions: risk, return, and maturity, and takes into account the impact of time decay. For example, a certain row in the latent demand probability distribution matrix may be expressed as [0.8 (risk weight value), 0.9 (return weight value), 0.7 (maturity weight value), latent demand probability value (calculated according to the weight value and other rules)].

[0104] Step S1369, extract the row vectors whose probability values exceed the dynamic threshold from the latent demand probability distribution matrix, perform semantic compatibility verification on the corresponding product feature labels and the context positions of the cross-modal semantic unit sequence, generate an insertable latent demand tag queue, and insert the tags that meet the semantic compatibility conditions in the latent demand tag queue into the gaps of the syntax tree nodes of the cross-modal semantic unit sequence in descending order of weight, generating an enhanced semantic unit sequence containing explicit demand and latent demand.

[0105] Assume that the dynamic threshold is set to 0.6, and the row vectors with probability values exceeding 0.6 are filtered out from the latent demand probability distribution matrix. For example, the product feature labels corresponding to the filtered row vectors are medium-risk products (associated according to risk weight value and other information), medium-return products (associated according to return weight value and other information), etc. Then these product feature labels are subjected to semantic compatibility verification with the context positions of the cross-modal semantic unit sequence. For example, at the context position where "risk is lower than that of stocks" is mentioned in the cross-modal semantic unit sequence, the product feature label of medium-risk products is semantically compatible because it meets the investor's expectation of risk lower than that of stocks. By performing such semantic compatibility verification on all the filtered product feature labels, an insertable latent demand tag queue is generated, and the product feature labels that meet the semantic compatibility conditions are arranged in a certain order (such as from high to low according to the probability value) in the queue.

[0106] On this basis, in the implicit demand tag queue, if the weight of the medium-risk product is the highest, then first insert the implicit demand tag of the medium-risk product into the risk-related position in the gap of the syntax tree nodes of the cross-modal semantic unit sequence. For example, if there is a node related to risk description in the syntax tree of the cross-modal semantic unit sequence, insert the tag of the medium-risk product into the gap near this node. Insert other eligible implicit demand tags in descending order of weight. Finally, generate an enhanced semantic unit sequence that includes explicit demands (such as the clear expressions in the original cross-modal semantic unit sequence) and implicit demands (the demands represented by the newly inserted tags). This enhanced semantic unit sequence more comprehensively reflects the needs of investors, including both the needs clearly expressed by investors and the implicit needs obtained from historical purchase records and semantic analysis, providing richer and more accurate input information for subsequent analysis and recommendation operations.

[0107] In a possible implementation manner, step S1367 includes:

[0108] Step S1367-1, extract a set of product feature tags including product risk level tags, income type tags, and term structure tags from the user's historical purchase records, and synchronously obtain the purchase timestamp corresponding to each product feature tag.

[0109] In the previously set scenario, the investor's historical purchase records include previously purchased stock products with a high-risk product risk level tag and a purchase timestamp of t1; the income type tag of the fund product is medium income, and the purchase timestamp is t2; the term structure tag of the time deposit is long term, and the purchase timestamp is t3. These product feature tags and their corresponding purchase timestamps completely record the investor's past investment situation.

[0110] Step S1367-2, perform a time difference operation on the current system time and the purchase timestamp to generate a time decay interval sequence for each product feature tag.

[0111] Assume that the current system time is T. Then for the stock product, the time decay interval is Δt1 = T - t1; for the fund product, the time decay interval is Δt2 = T - t2; for the time deposit, the time decay interval is Δt3 = T - t3. These time decay intervals reflect the length of time elapsed from the purchase of the product to the current time. The longer the duration, the greater the possible attenuation effect on the demand. Through such calculations, a time decay interval sequence [Δt1, Δt2, Δt3] for each product feature tag is obtained.

[0112] Step S1367-3, according to the length value of the time decay interval sequence, find the corresponding decay slope parameter on the preset time decay function curve to generate a decay coefficient vector for each product feature tag.

[0113] The preset time decay function curve is preset based on the changing rules of the financial market and the changing characteristics of investors' demands over time. For stock products, the position of the time decay interval Δt1 on the time decay function curve determines a decay slope parameter k1. This search process is determined according to the definition of the time decay function curve and the value of Δt1. For example, if the time decay function curve is a monotonically decreasing curve, when the value of Δt1 is large, the corresponding decay slope parameter k1 will result in a greater degree of decay. Similarly, for Δt2 of fund products and Δt3 of time deposits, the corresponding decay slope parameters k2 and k3 are found on the time decay function curve respectively. Thus, a decay coefficient vector [k1, k2, k3] of each product feature label is generated.

[0114] Step S1367-4: Perform a dimension-by-dimension dot product operation on the demand intensity components of the risk level mapping relationship table, the return demand mapping relationship table, and the term matching relationship table in the multi-dimensional demand intensity vector with the corresponding decay coefficient vector to generate a preliminarily decayed demand weight vector.

[0115] In the previous step, the multi-dimensional demand intensity vector may be [1 (risk frequency), +1 (return sentiment polarity intensity), 1 (term matching degree)]. For the demand intensity component 1 (risk frequency) of the risk level mapping relationship table, a dot product operation is performed with the decay coefficient k1 corresponding to stock products, and the demand weight of the risk dimension after preliminary decay is obtained as 1*k1. Similarly, for the demand intensity component +1 (return sentiment polarity intensity) of the return demand mapping relationship table, a dot product operation is performed with the decay coefficient k2 corresponding to fund products, and the demand weight of the return dimension after preliminary decay is obtained as +1*k2; for the demand intensity component 1 (term matching degree) of the term matching relationship table, a dot product operation is performed with the decay coefficient k3 corresponding to time deposits, and the demand weight of the term dimension after preliminary decay is obtained as 1*k3. In this way, a preliminarily decayed demand weight vector [1*k1, +1*k2, 1*k3] is generated.

[0116] Step S1367-5: Normalize the preliminarily decayed demand weight vector to keep the total demand weights of the risk level mapping relationship table, the return demand mapping relationship table, and the term matching relationship table constant, and generate a normalized demand weight vector.

[0117] Suppose the demand weight vector after preliminary attenuation is [w1, w2, w3], and the total demand weight is S = w1 + w2 + w3. The purpose of normalization is to ensure that after considering time attenuation, the demand weights in each dimension are consistent in terms of the total sum, meeting the requirements of the overall demand weight structure. Each component in the normalized demand weight vector is calculated as follows: For the risk dimension, the normalized weight is w1' = w1 / S; for the return dimension, the normalized weight is w2' = w2 / S; for the time-to-maturity dimension, the normalized weight is w3' = w3 / S. In this way, the normalized demand weight vector [w1', w2', w3'] is generated.

[0118] Step S1367-6, perform a probability space superposition of the normalized demand weight vector and the explicit demand markers in the user profile enhancement feature matrix to generate the final demand weight values that integrate the time decay effect.

[0119] Specifically, in the user profile enhancement feature matrix, the explicit demand markers contain the demand information clearly expressed by the investors. For example, in the explicit demand markers related to risk, a certain attitude or requirement towards risk may be clearly expressed. Superpose the normalized demand weight vector [w1', w2', w3'] and these explicit demand markers in the probability space. This superposition process is based on the specific rules and logics of demand analysis in the financial field, comprehensively considering the mutual relationship between the demand weights after considering the time decay effect and the explicit demand markers. For example, if an explicit demand marker has a specific weight representation in the risk dimension, then it is superposed with the normalized risk dimension demand weight w1' for calculation to obtain the final demand weight value that integrates the time decay effect. The same operation is also performed for the return and time-to-maturity dimensions. Finally, the final demand weight values in multiple dimensions such as risk, return, and time-to-maturity that comprehensively integrate the time decay effect are obtained. These final demand weight values can more accurately reflect the current demand situation of the investors considering the influence of time factors, providing a more accurate basis for subsequent analysis and decision-making.

[0120] In a possible implementation manner, step S140 includes:

[0121] Step S141, extract the product attribute nodes and their associated regulatory policy nodes and market condition nodes from the financial product knowledge graph.

[0122] In the previously set scenario, in the financial product knowledge graph, taking wealth management products as an example, the product attribute nodes include attribute information such as product risk level, income type, investment term, etc. For this wealth management product, the associated regulatory policy nodes may include nodes of risk management regulations issued by financial regulatory agencies for wealth management products, such as nodes specifying the risk assessment criteria and risk control requirements for wealth management products; nodes of sales restriction regulations, such as nodes restricting sales channels and sales targets. The market condition nodes may include nodes of the current market interest rate situation, which reflects the impact of the overall market interest rate level on the income of wealth management products; nodes of the market supply and demand situation of similar wealth management products, indicating the balance state of the demand and supply of wealth management products in the market. These nodes describe the characteristics of wealth management products, the regulatory constraints they are subject to, and the market environment they are in from different aspects.

[0123] Step S142: Based on the bi-directional graph attention network in the deep matching model, calculate the first matching degree between the enhanced user profile feature matrix and each product attribute node.

[0124] The enhanced user profile feature matrix contains the demand information of investors in multiple aspects such as risk, income, and term, as well as the implicit demands mined from historical purchase records and semantic analysis. The bi-directional graph attention network comprehensively considers the relationship between this information and the product attribute nodes of wealth management products. For example, for the risk level attribute of a wealth management product, if the enhanced user profile feature matrix shows that the investor expects a medium-risk product, and the risk level of the wealth management product is medium risk, the bi-directional graph attention network will quantitatively evaluate this matching situation according to the pre-set risk level matching rules. For the income type, if the investor expects stable income, and the income type provided by the wealth management product conforms to the characteristics of stable income, corresponding quantitative evaluation will also be carried out. Similar evaluations are also carried out for other attributes such as investment term. Through the comprehensive evaluation of all product attribute nodes, calculate the first matching degree between the enhanced user profile feature matrix and each product attribute node. This first matching degree is a quantitative value, reflecting the matching degree between user needs and product attributes. For example, it may be a value between 0 and 1, where 0 indicates complete mismatch and 1 indicates complete match.

[0125] Step S143: Traverse the regulatory policy nodes and market condition nodes through the meta-path random walk algorithm to generate a product compliance evaluation coefficient.

[0126] In a possible implementation manner, step S143 includes:

[0127] Step S1431: Determine the set of meta-path rules between the regulatory policy nodes and market condition nodes in the financial product knowledge graph. The set of meta-path rules includes a regulatory policy conduction path starting from the product attribute node and a market condition conduction path starting from the product attribute node.

[0128] Taking wealth management products as an example, for the regulatory policy conduction path, there may be a conduction path from the wealth management product attribute node to the financial regulatory agency node and then to the specific regulatory policy clause node; for the market condition conduction path, it may be a conduction path from the wealth management product attribute node to the market sector node and then to the market condition indicator node. The above meta-path rules clarify the path rules for reaching the regulatory policy nodes and market condition nodes from the product attribute node in the financial product knowledge graph.

[0129] Step S1432: According to the path type priority list of each meta-path in the set of meta-path rules, perform a multi-hop traversal operation on the financial product knowledge graph to generate a regulatory policy node sequence and a market condition node sequence associated with the product attribute node.

[0130] Suppose that in the regulatory policy conduction path, the path type from the wealth management product attribute node to the financial regulatory agency node directly regulating the product has a higher priority. When performing the multi-hop traversal operation, first find the corresponding financial regulatory agency node along this high-priority path, and then generate a regulatory policy node sequence based on other regulatory policy clause nodes associated with this agency node. For example, first find the relevant regulatory department node for regulating wealth management products, and then find the regulatory policy clause nodes such as risk management and sales norms related to this wealth management product issued by the relevant regulatory department to form a regulatory policy node sequence. For the market condition node sequence, according to the priority order of the market condition conduction path, starting from the wealth management product attribute node, find the corresponding market sector node, such as the financial product market sector node, and then find the market condition nodes such as market interest rates and supply and demand conditions under this sector to form a market condition node sequence.

[0131] Step S1433: Extract the effective time interval and constraint effectiveness intensity parameter of each regulatory policy node from the regulatory policy node sequence, and extract the fluctuation amplitude index and trend direction identifier of each market condition node from the market condition node sequence.

[0132] For the risk management regulation node in the regulatory policy node sequence, its effective time interval may start from a specific date until now, and the constraint effectiveness intensity parameter is set according to the strictness of the policy. For example, if the regulation is very strict, the constraint effectiveness intensity parameter may be set to a relatively high value. For the market interest rate situation node in the market condition node sequence, the fluctuation range indicator can be determined by analyzing the fluctuation range of the market interest rate over a period of time, and the trend direction identifier can be set according to the rising or falling trend of the interest rate. For example, a positive identifier is set for the rising trend and a negative identifier is set for the falling trend.

[0133] Step S1434: Match the overlap degree between the effective time interval and the current system time to generate the regulatory policy timeliness weight, and perform a multiplication operation on the constraint effectiveness intensity parameter and the regulatory policy timeliness weight to generate a dynamically adjusted regulatory policy impact factor set.

[0134] Suppose the effective time interval of the risk management regulation node starts from t1 until now, and the current system time is T. If the time span from t1 to T accounts for a relatively large proportion of the entire effective time interval, it indicates that the timeliness of this policy is relatively strong. According to the pre-set calculation rule, a relatively high regulatory policy timeliness weight is generated, such as 0.8. Multiply this timeliness weight by the constraint effectiveness intensity parameter (assumed to be 0.6), and the dynamically adjusted regulatory policy impact factor is 0.8 * 0.6 = 0.48. Such calculations are performed for each node in the regulatory policy node sequence to generate a dynamically adjusted regulatory policy impact factor set.

[0135] Step S1435: Compare the fluctuation range indicator with the preset market sensitivity threshold, screen out the market condition nodes that exceed the market sensitivity threshold as significant fluctuation market condition nodes, and perform a consistency match between the trend direction identifier and the risk preference label in the user profile enhancement feature matrix to generate a market condition adaptability set.

[0136] Suppose the preset market sensitivity threshold is a specific fluctuation range value. For the market interest rate situation node, if its fluctuation range indicator exceeds this market sensitivity threshold, then this node is determined as a significant fluctuation market condition node. For the trend direction identifier, if the risk preference label in the user profile enhancement feature matrix shows that investors prefer a stable investment environment, and the trend direction identifier of the market interest rate is downward (indicating relative market stability), then a relatively high adaptability value, such as 0.7, is recorded in the market condition adaptability set, indicating a high consistency between the market condition and the investor's risk preference.

[0137] Step S1436: Based on the hop counts of each node in the dynamically adjusted regulatory policy impact factor set, assign a decreasing propagation weight coefficient, and perform a weighted sum of the propagation weight coefficient and the corresponding regulatory policy impact factor to generate a basic product compliance score.

[0138] In the dynamically adjusted regulatory policy impact factor set, assume that the hop count of a certain regulatory policy node is 1 (close to the product attribute node). According to the pre-set decreasing propagation weight coefficient assignment rule, assign a relatively high propagation weight coefficient, such as 0.5. Perform a weighted sum of this propagation weight coefficient and the regulatory policy impact factor of this node (assumed to be 0.48) to obtain a partial sum. Perform such an operation for each node in the set, and add up all the partial sums to generate a basic product compliance score.

[0139] Step S1437: Calculate the product market stability offset based on the occurrence frequency of the significant volatility market condition node and the matching results in the market condition adaptability set, and superimpose the product market stability offset on the basic product compliance score to generate the product compliance evaluation coefficient.

[0140] Assume that the occurrence frequency of the significant volatility market condition node is 3 times. According to the pre-set calculation rule, this occurrence frequency may cause a certain market stability offset. If the average adaptability in the market condition adaptability set is relatively high, it may mitigate the impact of this offset. Calculate the product market stability offset according to the specific calculation rule (involving comprehensive consideration of factors such as occurrence frequency and adaptability), for example, it is -0.1 (indicating a certain negative impact on the basic compliance score). Superimpose this offset on the basic product compliance score. For example, if the basic product compliance score is 0.8, the product compliance evaluation coefficient after superimposition is 0.8 - 0.1 = 0.7.

[0141] Step S144: Perform a non-linear combination of the first matching degree and the compliance evaluation coefficient to generate a comprehensive product recommendation score.

[0142] The non-linear combination is based on the pre-set combination rule, considering the different importance and mutual relationship of the first matching degree and the compliance evaluation coefficient in product recommendation. For example, a weighted power combination method may be adopted, adding a certain power (such as square) of the first matching degree and a certain power (such as cube) of the compliance evaluation coefficient according to a certain weight. Assume that the first matching degree is 0.6 and the compliance evaluation coefficient is 0.7. Calculate the comprehensive product recommendation score according to the set combination rule. This comprehensive product recommendation score comprehensively reflects the matching degree between user needs and product attributes and the compliance situation of the product.

[0143] Step S145, sort and filter financial products according to the comprehensive recommendation score to generate the initial recommendation result queue.

[0144] For example, the wealth management product with the highest score is ranked at the front of the queue, and the products with lower scores are ranked behind. This initial recommendation result queue is the result obtained from the multi-level correlation analysis of the user portrait enhanced feature matrix and the financial product knowledge graph based on the deep matching model, so as to provide financial product recommendations that better meet the needs of investors and market conditions.

[0145] In a possible implementation manner, step S150 includes:

[0146] Step S151, obtain the position sensor data and ambient light intensity parameters of the user terminal, calculate the user's current attention level, and detect the device network latency index and interface rendering frame rate to evaluate the terminal interaction response ability.

[0147] In this embodiment, in the previously set scenario, the investor uses a tablet computer, and the position sensor data shows that the investor is located in the study at home. Assume that through further analysis of the position sensor data, such as according to the historical usage pattern of the study (if the system has relevant records) and current time and other factors, it is determined that this is a relatively quiet and less interfering environment, which may imply that the investor has a high attention level. The ambient light intensity parameter shows that the light is moderate, neither too bright nor too dark, which also helps the investor maintain a good attention level. Based on the above information, according to the pre-set attention level calculation rule, it is calculated that the user's current attention level is high.

[0148] The device network latency index is obtained by measuring information such as the transmission time of network data packets. Assume that the detected network latency is low, which means that the data transmission speed is fast and information such as recommendation results can be obtained in a timely manner. The interface rendering frame rate reflects the speed of interface update. Detecting that the interface rendering frame rate is normal indicates that the interface can display various elements smoothly. According to the network latency index and interface rendering frame rate, evaluate the terminal interaction response ability according to the pre-defined evaluation rule. Since the network latency is low and the interface rendering frame rate is normal, it is determined that the terminal interaction response ability is good.

[0149] Step S152, adjust the display granularity level of the initial recommendation result queue according to the attention level to generate a dynamic information density parameter, and optimize the loading priority policy of the initial recommendation result queue according to the terminal interaction response ability.

[0150] In this embodiment, since the attention level of investors is relatively high, more detailed and in-depth information can be provided. In the initial recommended result queue, for the recommended information of each financial product, which might originally be just a simple overview, more details can now be added, such as detailed earnings calculation methods, risk analysis reports, historical performance data, etc. According to the degree of such adjustment, a dynamic information density parameter is generated according to preset rules. For example, if the detail level of the information is increased from the basic level to the advanced level, the dynamic information density parameter might increase from 1.0 (the basic value, representing the normal information density) to 1.5, indicating a 50% increase in information density.

[0151] Due to the good terminal interaction response ability, more recommended product information can be loaded more actively, and it can be loaded in an order that is more conducive to users obtaining information. For example, for those products with a higher degree of match with the investors' needs, more detailed information can be loaded first, while for products with a slightly lower degree of match, it can be loaded gradually in the background. The optimization of this loading priority strategy is based on the evaluation result of the terminal interaction response ability, ensuring that information can be provided as efficiently as possible without affecting the user experience.

[0152] Step S153: Apply the dynamic information density parameter and the loading priority strategy to the initial recommended result queue to generate a dynamically adapted recommended list.

[0153] In the initial recommended result queue, the recommended information of each financial product is adjusted according to the dynamic information density parameter, increasing or decreasing the amount of information displayed. For example, for a wealth management product that is ranked relatively high in the queue and has a high degree of match, according to the dynamic information density parameter of 1.5, originally only basic information such as earnings type and risk level was displayed, but now detailed earnings curves for the past three years, risk volatility analysis under different market environments, etc. are also added. At the same time, according to the loading priority strategy, the detailed information of this wealth management product will be loaded first and displayed in a more prominent position. In this way, each product in the initial recommended result queue is processed to generate a dynamically adapted recommended list, which is more in line with the current attention level of investors and the terminal interaction response ability, and can provide more personalized and effective recommended information.

[0154] Step S154: Create a three-dimensional visualization space in the interaction interface, map the recommended products in the dynamically adapted recommended list to three-dimensional icons of different shapes, and dynamically adjust the color saturation and rotation speed of the three-dimensional icons according to the product risk level.

[0155] For example, in the three-dimensional visualization space of the interactive interface, for each financial product in the dynamically adapted recommendation list, a unique three-dimensional icon is created for it. For example, for a time deposit product with relatively low risk, it is mapped to a cube icon. The shape of the cube gives people a feeling of stability and reliability, which is in line with the characteristics of low risk and stable returns of time deposits. For a stock fund product with relatively high risk, it is mapped to a triangular pyramid icon. The shape of the triangular pyramid is relatively sharp, suggesting its relatively high-risk characteristics.

[0156] For example, for a time deposit product with low risk, set the color saturation of its three-dimensional icon to be relatively low. For example, use light blue, and the rotation speed is relatively slow, such as 10 degrees per minute. This relatively low color saturation and slow rotation speed give people a visual feeling of smoothness and safety, intuitively reflecting the low-risk characteristics of the product. For a stock fund product with high risk, set the color saturation of its three-dimensional icon to be relatively high, such as using red, and the rotation speed is relatively fast, such as 30 degrees per minute. The red color and relatively fast rotation speed can attract the user's attention and at the same time convey the characteristics of high risk and high volatility of the product.

[0157] Step S155, construct an interactive timeline control to update the spatio-temporal distribution of the recommendation results in real time in response to the user's sliding operation. When the user's gaze duration on any three-dimensional icon exceeds the threshold duration, trigger a multi-level information expansion animation.

[0158] An interactive timeline control can be constructed in the interactive interface. This timeline control is associated with the time-related information of the recommended products, such as the investment term of the product, the time distribution of historical performance, etc. When the user slides on the timeline, the system updates the spatio-temporal distribution of the recommendation results in real time according to the position where the user slides. For example, if the user slides the timeline slider to the position of the next year, the expected return, risk, etc. of the recommended products in this future time period can be readjusted according to the financial market prediction model (if any), and the updated recommendation results are displayed in real time on the interface.

[0159] When the user's gaze duration on any three-dimensional icon exceeds the threshold duration, trigger a multi-level information expansion animation. Suppose the threshold duration is set to 3 seconds. When the user gazes at the three-dimensional icon of a financial product for more than 3 seconds, a multi-level information expansion animation can be triggered. This multi-level information expansion animation may first show the more detailed income structure of this financial product, such as the expected income ratio for different terms; then show the risk decomposition information, such as the proportion of market risk, credit risk, etc. in the total risk; finally, it may show some market trend analysis or expert comments related to this product. This multi-level information expansion animation enables users to quickly obtain more in-depth information when they focus on a certain product, without having to actively search for this information.

[0160] Step S156, generating a focus heat map based on eye tracking in the edge area of ​​the interface, and adjusting the layout density of the recommended elements in real time.

[0161] For example, through eye tracking technology, it is possible to monitor the movement trajectory and stop position of the user's gaze on the interface, and generate a focus heat map in the edge area of ​​the interface. For example, if the user's gaze is more concentrated on the left area of ​​the interface, a higher heat value will be displayed on the focus heat map of the left area. According to the heat distribution of the focus heat map, the layout density of the recommended elements is adjusted in real time. If the heat in the left area is higher, it means that the user may be more interested in the recommended elements in this area, then the layout density of the recommended elements in the left area can be appropriately increased, such as displaying more recommended product stereo icons or related information in the left area, while reducing the layout density of the right area, so as to better display the product information that the user may be more interested in, and improve the efficiency and experience of users in obtaining information.

[0162] Figure 2 The schematic diagram shows exemplary hardware and software components of the bank financial product recommendation system 100 provided by some embodiments of the present invention that can implement the concept of the present invention. For example, the processor 120 can be used in the bank financial product recommendation system 100 and used to perform the functions of the present invention.

[0163] The bank financial product recommendation system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the bank financial product recommendation method of the present invention. Although the present invention only shows one server, for convenience, the functions described in the present invention can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0164] For example, the bank financial product recommendation system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the bank financial product recommendation system 100 may also include program instructions stored in ROM, RAM, or other types of non-temporary storage media, or any combination thereof. The method of the present invention can be implemented according to these program instructions. The bank financial product recommendation system 100 also includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0165] For ease of explanation, only one processor is described in the bank financial product recommendation system 100. However, it should be noted that the bank financial product recommendation system 100 in the present invention may also include multiple processors. Therefore, the steps performed by one processor described in the present invention may also be jointly performed or separately performed by multiple processors. For example, if the processor of the bank financial product recommendation system 100 performs step A and step B, it should be understood that step A and step B may also be jointly performed by two different processors or separately performed in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0166] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned bank financial product recommendation method is implemented.

[0167] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A method for recommending bank wealth management products, characterized in that, The method includes: Collecting a multi-modal dialogue data stream generated during the interaction between a user terminal and an intelligent customer service system, where the multi-modal dialogue data stream includes a voice input signal and text input content; Performing semantic recognition on the voice input signal to generate a structured requirement description set, and at the same time performing entity extraction and intention parsing on the text input content to generate a multi-dimensional semantic label set; Performing spatio-temporal alignment and fusion on the structured requirement description set and the multi-dimensional semantic label set to generate a user portrait enhanced feature matrix; Based on a deep matching model, performing multi-level correlation analysis on the user portrait enhanced feature matrix and a financial product knowledge graph to generate an initial recommendation result queue; Performing context-aware optimization on the initial recommendation result queue according to real-time environment perception data, generating a dynamically adapted recommendation list and visualizing and rendering the output through a multi-channel interaction interface; The performing spatio-temporal alignment and fusion on the structured requirement description set and the multi-dimensional semantic label set to generate a user portrait enhanced feature matrix includes: Performing timestamp alignment verification on the structured requirement description set and the multi-dimensional semantic label set, and identifying the time overlap interval and the time sequence interval fault between the voice semantic paragraphs in the structured requirement description set and the text semantic paragraphs in the multi-dimensional semantic label set; Inputting the voice semantic paragraphs and text semantic paragraphs within the time overlap interval into a gated recurrent unit network, calculating the semantic conflict probability between the voice semantic paragraphs and the text semantic paragraphs at the time sequence interval fault through the forget gate of the gated recurrent unit network, and dynamically allocating weights to the voice semantic paragraphs and the text semantic paragraphs based on the semantic conflict probability; Extracting historical voice interaction fragments and historical text interaction fragments associated with the current time overlap interval from the user's historical interaction records, and generating a semantic consistency verification vector according to the recurrence frequency of the historical voice interaction fragments and the historical text interaction fragments; Performing a point-wise multiplication operation on the semantic consistency verification vector and the output of the gated recurrent unit network to correct the confidence scores of the voice semantic paragraphs and the text semantic paragraphs after the dynamic weight allocation; Performing cross-modal splicing on the voice semantic paragraphs and the text semantic paragraphs according to the corrected confidence scores to generate a fused cross-modal semantic unit sequence; Extracting continuously occurring semantic keyword combinations in the cross-modal semantic unit sequence, performing pattern matching between the semantic keyword combinations and product feature labels in the user's historical purchase records, and identifying implicit demand markers in the cross-modal semantic unit sequence; Inserting the implicit demand markers into the corresponding positions of the cross-modal semantic unit sequence to generate the user portrait enhanced feature matrix with temporal context association; The performing multi-level correlation analysis on the user portrait enhanced feature matrix and a financial product knowledge graph based on a deep matching model to generate an initial recommendation result queue includes: Extracting product attribute nodes and their associated regulatory policy nodes and market condition nodes from the financial product knowledge graph; Based on the bi-directional graph attention network in the deep matching model, calculating the first matching degree between the user portrait enhanced feature matrix and each product attribute node; Traverse the regulatory policy nodes and market condition nodes through the meta-path random walk algorithm to generate product compliance evaluation coefficients; Non-linearly combine the first matching degree and the compliance evaluation coefficient to generate a comprehensive product recommendation score; Sort and filter financial products according to the comprehensive recommendation score to generate the initial recommendation result queue.

2. The bank financial product recommendation method according to claim 1, wherein Performing semantic recognition on the voice input signal to generate a structured requirement description set, including: Perform frame processing and background noise cancellation on the voice input signal, and extract Mel frequency cepstral coefficients as acoustic feature vectors; Convert the voice input signal into an intermediate text expression through a pre-trained speech recognition model, and label the emotional intensity value of the speaker; Adopt a multi-head attention mechanism to complete the context semantics of the intermediate text expression to generate a corrected complete semantic paragraph; Use a domain-specific named entity recognition model to extract constraint condition entities of amount, term, and risk category from the complete semantic paragraph; Perform tensor splicing on the acoustic feature vector, emotional intensity value, and constraint condition entity to generate the structured requirement description set with emotional weights.

3. The bank financial product recommendation method according to claim 2, wherein, Performing entity extraction and intention parsing on the text input content to generate a multi-dimensional semantic label set, including: Perform word segmentation and part-of-speech tagging on the text input content to construct a dependency syntactic analysis tree; Execute a pattern matching algorithm on the dependency syntactic analysis tree to identify a modification relationship structure containing comparative and superlative degrees; Extract user preference intensity indicators and conditional constraint expressions from the modification relationship structure; Identify the beneficiary, target object, and negative conditions in the text input content through a semantic role annotation model to obtain a semantic role annotation result; Perform feature encoding on the user preference intensity indicator, conditional constraint expression, and semantic role annotation result to generate the multi-dimensional semantic label set.

4. The bank financial product recommendation method according to claim 1, wherein Performing pattern matching on the semantic keyword combination and the product feature label in the user's historical purchase record to identify the implicit demand markers in the cross-modal semantic unit sequence, including: Extract the product feature label set of all purchased financial products from the user's historical purchase record, and the product feature label set includes product risk level labels, income type labels, and term structure labels; Perform part-of-speech tagging and stop word filtering on the semantic keyword combination in the cross-modal semantic unit sequence, and retain the core semantic units containing noun phrases and adjective phrases to generate a dimension-reduced target keyword sequence; Perform item-by-item pattern matching on each noun phrase in the target keyword sequence and the product risk level label to identify the risk preference synonym clusters implicit in the noun phrase, and generate a risk level mapping relationship table; Based on the semantic similarity calculation between the adjective phrase and the income type label, filter out the income descriptors that meet the preset similarity threshold, and associate the corresponding income type labels to generate an income demand mapping relationship table; Traverse the time unit identifiers of the term structure tags, extract consecutive word pairs containing numerical values and time units from the target keyword sequence, and perform interval inclusion verification between the consecutive word pairs and the allowable value intervals of the term structure tags to generate a term matching relationship table; Generate a multi-dimensional demand intensity vector based on the occurrence frequency of synonym clusters in the risk level mapping relationship table, the emotional polarity intensity of income descriptors in the income demand mapping relationship table, and the interval matching degree of consecutive word pairs in the term matching relationship table; Input the multi-dimensional demand intensity vector into a time decay function, and calculate the decayed demand weight values according to the purchase timestamps of each product feature tag in the user's historical purchase records; Perform weighted fusion on the risk level mapping relationship table, the income demand mapping relationship table, and the term matching relationship table according to the demand weight values to generate a latent demand probability distribution matrix; Extract the row vectors with probability values exceeding the dynamic threshold from the latent demand probability distribution matrix, and perform semantic compatibility verification on the corresponding product feature tags and the context positions of the cross-modal semantic unit sequence to generate an insertable latent demand marker queue; Insert the markers that meet the semantic compatibility conditions in the latent demand marker queue into the gaps of the syntax tree nodes of the cross-modal semantic unit sequence in descending order of weight to generate an enhanced semantic unit sequence containing explicit and latent demands.

5. The bank financial product recommendation method according to claim 4, wherein The step of inputting the multi-dimensional demand intensity vector into a time decay function and calculating the decayed demand weight values according to the purchase timestamps of each product feature tag in the user's historical purchase records includes: Extract a set of product feature tags containing product risk level tags, income type tags, and term structure tags from the user's historical purchase records, and synchronously obtain the purchase timestamp corresponding to each product feature tag; Perform a time difference operation between the current system time and the purchase timestamp to generate a time decay interval sequence for each product feature tag; According to the length values of the time decay interval sequence, find the corresponding decay slope parameters on the preset time decay function curve to generate a decay coefficient vector for each product feature tag; Perform a dimension-by-dimension dot product operation on the demand intensity components of the risk level mapping relationship table, the income demand mapping relationship table, and the term matching relationship table in the multi-dimensional demand intensity vector and the corresponding decay coefficient vectors respectively to generate a preliminarily decayed demand weight vector; Perform normalization processing on the preliminarily decayed demand weight vector to keep the total demand weights of the risk level mapping relationship table, the income demand mapping relationship table, and the term matching relationship table constant, and generate a normalized demand weight vector; Perform probability space superposition on the normalized demand weight vector and the explicit demand markers in the user portrait enhancement feature matrix to generate the final demand weight values that integrate the time decay effect.

6. The bank financial product recommendation method according to claim 1, wherein The step of traversing the regulatory policy nodes and market condition nodes through the meta-path random walk algorithm to generate a product compliance evaluation coefficient includes: Determine the set of meta-path rules between regulatory policy nodes and market condition nodes in the financial product knowledge graph. The set of meta-path rules includes a regulatory policy conduction path starting from product attribute nodes and a market condition conduction path starting from product attribute nodes; According to the path type priority list of each meta-path in the set of meta-path rules, perform a multi-hop traversal operation on the financial product knowledge graph to generate a regulatory policy node sequence and a market condition node sequence associated with the product attribute nodes; Extract the effective time interval and constraint effect intensity parameters of each regulatory policy node from the regulatory policy node sequence, and extract the fluctuation amplitude index and trend direction identifier of each market condition node from the market condition node sequence; Match the overlap degree between the effective time interval and the current system time to generate a regulatory policy timeliness weight, and perform a multiplication operation on the constraint effect intensity parameter and the regulatory policy timeliness weight to generate a dynamically adjusted set of regulatory policy impact factors; Compare the fluctuation amplitude index with a preset market sensitivity threshold, filter out market condition nodes that exceed the market sensitivity threshold as significantly fluctuating market condition nodes, and perform a consistency match between the trend direction identifier and the risk preference labels in the user portrait enhancement feature matrix to generate a set of market condition adaptation degrees; According to the path hop count of each node in the dynamically adjusted set of regulatory policy impact factors, assign a decreasing propagation weight coefficient, and perform a weighted sum of the propagation weight coefficient and the corresponding regulatory policy impact factor to generate a basic product compliance score; Calculate the product market stability offset according to the occurrence frequency of the significantly fluctuating market condition nodes and the matching results in the set of market condition adaptation degrees, and superimpose the product market stability offset on the basic product compliance score to generate the product compliance evaluation coefficient.

7. The bank financial product recommendation method according to claim 1, characterized in that The context-aware optimization of the initial recommendation result queue according to real-time environment perception data to generate a dynamically adapted recommendation list and visualize and render it through a multi-channel interaction interface includes: Obtain the location sensor data and ambient light intensity parameters of the user terminal, calculate the user's current attention level, and detect the device network latency index and interface rendering frame rate to evaluate the terminal interaction response ability; Adjust the display granularity level of the initial recommendation result queue according to the attention level to generate a dynamic information density parameter, and optimize the loading priority strategy of the initial recommendation result queue according to the terminal interaction response ability; Apply the dynamic information density parameter and the loading priority strategy to the initial recommendation result queue to generate a dynamically adapted recommendation list; And, create a three-dimensional visualization space in the interaction interface, map the recommended products in the dynamically adapted recommendation list to three-dimensional icons of different shapes, and dynamically adjust the color saturation and rotation speed of the three-dimensional icons according to the product risk level; Build an interactive timeline control to respond to user swiping operations and update the spatio-temporal distribution of recommended results in real time. When the user gazes at any three-dimensional icon for a duration exceeding the threshold duration, trigger a multi-level information expansion animation; Generate a focus heat map based on eye tracking in the interface edge area and adjust the layout density of recommended elements in real time.

8. A bank financial product recommendation system, characterized in that, The bank financial product recommendation system includes a processor and a memory. The memory is connected to the processor. The memory is used to store programs, instructions, or codes, and the processor is used to execute the programs, instructions, or codes in the memory to implement the bank financial product recommendation method according to any one of claims 1-7 above.

Images