Bank financing product recommendation method and system

Abstract

The invention provides a bank financial management product recommendation method and system, and the method comprises the steps: firstly collecting a multi-modal dialogue data flow which is generated by interaction of a user terminal and an intelligent customer service system and contains voice input signals and text input contents, and then carrying out the semantic recognition of the voice input signals, and generating a structured demand description set; performing entity extraction and intention analysis on text input content to generate a multi-dimensional semantic tag set, performing time-space alignment and fusion on the text input content and the multi-dimensional semantic tag set to form a user portrait enhanced feature matrix, and performing multi-level association 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; and finally, according to the real-time environment perception data, queue context perception optimization is carried out, a dynamic adaptive recommendation list is generated, and visual rendering output is carried out through a multi-channel interaction interface, so that accurate and dynamic financial management product recommendation is realized by utilizing multi-modal data, deep analysis and combination with a 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 financial management among users, the types of bank financial products are becoming more and more diverse, and users are facing increasing difficulties in choosing financial products that suit them. In order to help users quickly and accurately find financial products that meet their needs, the bank financial product recommendation system came into being.

[0003] In related technologies, most recommendation methods rely on a single type of data for analysis and recommendation. Some recommendation systems only collect basic information about users, such as static data such as age and income, and use this as a basis for simple product matching. Due to limited data sources, this method cannot fully and deeply understand users' real financial needs and preferences, resulting in recommendation results that often cannot accurately meet users' diverse needs.

[0004] In addition, some related technologies, although they take into account the historical transaction data of users, are limited to the analysis of the surface of transaction behavior, without digging deeply into the user intentions and potential needs behind the transactions. For example, they simply count the types of products that users have purchased, but fail to understand the deep reasons why users buy these products, such as whether they are for short-term capital appreciation, long-term asset allocation or other special needs, which makes the recommendations lack pertinence and foresight. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for recommending bank financial products, the method comprising: Collecting a multimodal dialogue data stream generated during the interaction between the user terminal and the intelligent customer service system, wherein the multimodal dialogue data stream includes a voice input signal and a text input content; Performing semantic recognition on the voice input signal to generate a structured demand description set, and performing entity extraction and intent analysis on the text input content to generate a multi-dimensional semantic tag set; The structured requirement description set is temporally and spatially aligned and fused with the multidimensional semantic label set to generate a user portrait enhanced feature matrix; Based on the deep matching model, the user portrait enhanced feature matrix and the financial product knowledge graph are subjected to multi-level correlation analysis to generate an initial recommendation result queue; 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.

[0006] 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.

[0007] 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

[0008] 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.

[0009] 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

[0010] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 The figure is a flow chart of a method for recommending bank financial products provided by an embodiment of the present invention. The method for recommending bank financial products is introduced in detail below.

[0011] Step S110, collecting a multimodal dialogue data stream generated during the interaction between the user terminal and the intelligent customer service system, wherein the multimodal dialogue data stream includes a voice input signal and text input content.

[0012] In this embodiment, consider a scenario where an investor is using a tablet computer to interact with the financial intelligent customer service system. He is sitting in his home study room and there are some slight ambient noises around him, such as the sound of cars driving on the street in the distance and the faint sound of air conditioning running. At this point, he began to interact with the intelligent customer service system, for example, first with voice input: "I have made a variety of investments in the past few years, including stocks, funds, and some time deposits. Now I want to adjust my investment portfolio, and I hope to find a financial product that can diversify risks to a certain extent. My stock investments are mainly concentrated in the technology sector, and the current returns are quite volatile, so I want to find a product that is not so strongly correlated 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 amount of funds 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 investments." At the same time, he also entered some supplementary information in the input box of the interactive interface as text input content: "My current total assets are about 5 million, of which 3 million is stock investment, 1 million is fund investment, and 1 million is time deposits. I hope that the investment amount of the new financial product 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." The intelligent customer service system collects the multimodal conversation data stream containing voice input signals and text input content through the tablet computer's high-precision microphone and text input interface. In this process, it is necessary to accurately capture every word, phrase, and sentence in the voice input signal, as well as the precise numbers, expression conditions, and restrictions in the text input content.

[0013] Step S120, performing semantic recognition on the voice input signal to generate a structured demand description set, and performing entity extraction and intent analysis on the text input content to generate a multi-dimensional semantic tag set.

[0014] 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.

[0015] 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." 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.

[0016] 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.

[0017] Then, a pattern matching algorithm is executed on the dependency parse tree to identify the modification relation structure containing comparative and superlative degrees. In this text, "try to reduce as much as possible" in "hope to reduce unnecessary management fee expenditures" is a comparative expression. The user preference strength indicator and conditional constraint expression are extracted from the modification relation structure, where "try to reduce as much as possible" is the preference strength indicator and "management fee expenditures" is the conditional constraint expression.

[0018] The semantic role labeling model is used to identify the beneficiaries, target objects, and negation conditions in the text input content, and the semantic role labeling results are obtained. In the above example, the beneficiary is the investor himself, the target object is the financial product, and the negation condition is "unnecessary management fee expenditure" (i.e., such a situation is not desired). Finally, the user preference strength indicator, conditional constraint expression, and semantic role labeling results are feature encoded to generate a multidimensional semantic label set.

[0019] Step S130, aligning and fusing the structured requirement description set with the multidimensional semantic tag set in time and space to generate a user portrait enhanced feature matrix.

[0020] In detail, firstly, the timestamp alignment check is performed on the structured requirement description set and the multidimensional semantic tag set. Assume that the voice input and text input are almost performed at the same time, but due to the difference in system processing speed, there may be a very small time difference. In this way, the time overlapping intervals and timing interval faults between the voice semantic paragraphs in the structured requirement description set and the text semantic paragraphs in the multidimensional semantic tag set can be identified. For example, in the above example, since the two are input almost at the same time, the time overlapping interval almost covers the entire input process, and the timing interval fault can be almost ignored.

[0021] Then, the speech semantic paragraphs and text semantic paragraphs in the time overlapping interval are input into the gated recurrent unit network. The probability of semantic conflict between the speech semantic paragraphs and the text semantic paragraphs at the temporal interval fault (although it is very small, it is assumed that there is a point) is calculated through the forget gate of the gated recurrent unit network. For example, assuming that the speech mentions "I hope the product risk is lower than stock investment", but the text does not explicitly mention the risk comparison of stock investment, it is calculated that there may be a certain probability of semantic conflict here. Based on the semantic conflict probability, dynamic weights are assigned to the speech semantic paragraphs and text semantic paragraphs.

[0022] Extract historical voice interaction segments and historical text interaction segments associated with the current time overlapping interval from the user's historical interaction records. Assuming that the user has previously discussed investment risks and management fees with customer service, generate a semantic consistency verification vector based on the frequency of recurrence of these historical interaction segments. If risk control and management fees have been mentioned many times before, the relevant weight in the semantic consistency verification vector will be relatively high.

[0023] The semantic consistency verification vector is multiplied point by point with the output of the gated recurrent unit network to correct the confidence scores of the speech semantic paragraph and text semantic paragraph after dynamic weight allocation. The speech semantic paragraph and text semantic paragraph are cross-modally spliced ​​according to the corrected confidence scores to generate a fused cross-modal semantic unit sequence.

[0024] Extract semantic keyword combinations that appear continuously in the cross-modal semantic unit sequence, such as "lower risk than stocks", "withdraw within four to five years", "reduce management fee expenditure", etc. Perform pattern matching on these semantic keyword combinations and 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 the previously purchased stock product is high risk, the fund product return type is medium return, and the term structure of the fixed deposit is long-term. Through pattern matching, it is found that the keyword combination "lower risk than stocks" has a contrasting relationship with the high-risk stock product feature tags, which may imply an implicit demand for medium-risk or low-risk products.

[0025] The implicit demand markers are inserted into the corresponding positions of 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 includes the needs explicitly expressed by users, but also incorporates the implicit needs mined from historical records and cross-modal analysis, comprehensively depicting the user's demand profile in this financial investment scenario.

[0026] Step S140: Based on the deep matching model, the user portrait enhanced feature matrix and the financial product knowledge graph are subjected to multi-level correlation analysis to generate an initial recommendation result queue.

[0027] In detail, product attribute nodes and their associated regulatory policy nodes and market situation nodes can be extracted from the financial product knowledge graph. For example, the product attribute node of wealth management products in the financial product knowledge graph may have associated regulatory policy nodes including risk management regulations and sales restriction regulations for wealth management products, and market situation nodes may include current market interest rates and market supply and demand of similar wealth management products.

[0028] Then, based on the bidirectional graph attention network in the deep matching model, the first matching degree between the user portrait enhancement feature matrix and each product attribute node is calculated. Assuming that there are clear requirements for risk, term, income and management fees in the user portrait enhancement feature matrix, the bidirectional graph attention network will comprehensively consider these factors and calculate the matching degree with the attribute nodes of different financial products. For example, for a financial product with low risk, term of three to five years, relatively stable income and more management fee discounts, a higher first matching degree may be calculated.

[0029] Next, the meta-path walking algorithm is used to traverse the regulatory policy nodes and market situation nodes to generate product compliance assessment coefficients. The meta-path rule set between the regulatory policy nodes and market situation nodes in the financial product knowledge graph is determined. For example, the meta-path rule from the financial product attribute node to the regulatory policy node may include a transmission path according to product type-regulatory category-specific policy, and the meta-path rule from the financial product attribute node to the market situation node may be a transmission path according to product type-market sector-market indicator.

[0030] On this basis, according to the path type priority list of each meta-path in the meta-path rule set, a multi-hop traversal operation is performed on the financial product knowledge graph to generate a regulatory policy node sequence and a market situation node sequence associated with the product attribute node. For example, for financial products, the regulatory policy node sequence obtained by traversal may include the policy node recently released on the sales specifications of financial products, and the market situation node sequence may include the average yield node of financial products currently on the market, etc.

[0031] Then, the effective time interval and constraint strength parameters of each regulatory policy node are extracted from the regulatory policy node sequence, and the volatility index and trend direction identifier of each market node are extracted from the market node sequence. Assume that the effective time interval of a regulatory policy node is from last year to now, the constraint strength parameter is medium, the volatility index of the market node is small, and the trend direction identifier is stable.

[0032] Then, the effective time interval is matched with the current system time to generate the regulatory policy timeliness weight, and the constraint effectiveness strength parameter is multiplied with the regulatory policy timeliness weight to generate a dynamically adjusted regulatory policy impact factor set. For example, if the current time is within the effective time interval of the regulatory policy and close to the middle, the generated regulatory policy timeliness weight is 0.5, which is multiplied by the constraint effectiveness strength parameter 0.5 to obtain 0.25 as the regulatory policy impact factor.

[0033] Next, the volatility index is compared with the preset market sensitivity threshold, and the market nodes that exceed the market sensitivity threshold are screened out as significant volatility nodes, and the trend direction identifier is matched with the risk preference label in the user portrait enhancement feature matrix to generate a market suitability set. Assuming that in the above example, there is no market node that exceeds the market sensitivity threshold, and the trend direction identifier matches the user risk preference label well, a better market suitability set is generated.

[0034] Then, according to the number of path hops 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 the product compliance basic score. Then, according to the frequency of occurrence of significant fluctuation market nodes and the matching results in the market market adaptability set, the product market stability offset is calculated, and the product market stability offset is superimposed on the product compliance basic score to generate the product compliance assessment coefficient.

[0035] Finally, the first matching degree and the compliance assessment coefficient are nonlinearly combined to generate a comprehensive product recommendation score. Financial products are sorted and screened according to the comprehensive recommendation score to generate an initial recommendation result queue. For example, different financial products such as wealth management products and fund products are arranged from high to low according to the comprehensive recommendation score to form an initial recommendation result queue. The products at the front are products that better meet user needs and compliance requirements.

[0036] Step S150 , context-aware optimization is performed on the initial recommendation result queue 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.

[0037] Specifically, the location sensor data and ambient light intensity parameters of the user terminal (tablet) can be obtained. Assume that the location sensor data shows that the user is in the study room at home and the ambient light intensity is moderate. Based on this information, the user's current attention level is calculated. Since the user is in a quiet study environment, the attention level is high. At the same time, the device network delay index and interface rendering frame rate are detected to evaluate the terminal's interactive responsiveness. Assuming that the network delay is low and the interface rendering frame rate is normal, it indicates that the terminal's interactive responsiveness is good.

[0038] Adjust the display granularity level of the initial recommendation result queue according to the attention level and generate dynamic information density parameters. Since users have high attention levels and can provide more detailed information, increase the display granularity level and improve the dynamic information density parameters. Optimize the loading priority strategy of the initial recommendation result queue according to the terminal's interactive response capability. Because the terminal has good interactive response capabilities, more recommended product information can be loaded first.

[0039] The dynamic information density parameters and loading priority strategy are applied to the initial recommendation result queue to generate a dynamically adapted recommendation list.

[0040] Create a three-dimensional 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 pyramid icons. Dynamically adjust the color saturation and rotation speed of the three-dimensional icons according to the risk level of the product. 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.

[0041] Build an interactive timeline control to respond to user sliding operations and update the temporal and spatial distribution of recommendation results in real time. When the user stares at any three-dimensional icon for longer than the threshold, the multi-level information expansion animation is triggered. For example, when the user stares at the icon of a financial product for more than 3 seconds, an animation display of multi-level information such as the detailed yield curve and risk assessment report of the product will pop up.

[0042] Generate a focus heat map based on eye tracking in the edge area of ​​the interface and adjust the layout density of recommended elements in real time. If the user's eyes are more focused on the left area of ​​the interface, then appropriately increase the layout density of recommended elements in the left area to better display product information that the user may be more interested in.

[0043] Based on the above steps, 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 processes and fuses 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 matching between user needs and products directly, 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.

[0044] In a possible implementation, step S120 includes: Step S121, performing frame processing and background noise elimination on the speech input signal, and extracting Mel-frequency cepstral coefficients as acoustic feature vectors.

[0045] 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.

[0046] 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.

[0047] 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.

[0048] 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.

[0049] 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.

[0050] 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.

[0051] 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.

[0052] Step S125, performing tensor concatenation of the acoustic feature vector, the emotion intensity value and the constraint condition entity to generate the structured requirement description set with emotion weight.

[0053] In this embodiment, a tensor is a high-dimensional data representation. First, the acoustic feature vector contains the acoustic feature information of the speech input signal, the emotional intensity value represents the speaker's emotional state, and the constraint entity clarifies the key constraint information in the investment demand. When performing tensor splicing, these three parts can be combined into a new tensor structure according to certain rules. For example, assuming 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 entity can be converted into an m-dimensional vector (where n and m are determined according to the actual number of features). The emotional intensity value is multiplied with each element in the vector of the constraint entity to obtain a constraint entity vector with emotional weights, and then the constraint entity vector with emotional weights is combined with the acoustic feature vector according to the rules of tensor splicing to form a new tensor, which is a structured demand description set with emotional weights. The structured demand description set comprehensively contains the acoustic features, emotional information, and constraint entity information in the semantics of the speech input signal.

[0054] In a possible implementation, step S120 further includes: Step S126, performing word segmentation and part-of-speech tagging on the text input content, constructing a dependency syntactic analysis tree, executing a pattern matching algorithm on the dependency syntactic analysis tree, and identifying modification relationship structures including comparatives and superlatives.

[0055] For example, in the scenario set previously, the text input by the investor is "My current total assets are about 5 million, of which 3 million is stock investment, 1 million is fund investment, and 1 million is time deposit. I hope that the investment amount of 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." Then, this text can be segmented according to predefined segmentation rules, such as separating words such as "I", "currently", "of", "total assets", "approximately", and "is". Part-of-speech tagging determines the part of speech of each word, such as "I" as a pronoun, "total assets" as a noun, and "approximately" as an adverb. On this basis, a dependency syntactic analysis tree is constructed, which describes the syntactic relationship between words. For example, "total assets" and "is" have a subject-predicate relationship, and "5 million" and "is" have an object relationship. The words in the entire sentence are connected to each other through this dependency relationship to form an analysis tree, which clearly shows the grammatical structure of the text.

[0056] In addition, in the above text, "Hope to minimize unnecessary management expenses" contains a comparative modification relationship structure. The pattern matching algorithm searches for structures that conform to comparative and superlative patterns in the dependency syntactic analysis tree. This process is matched through pre-defined comparative and superlative grammatical patterns. For example, for comparatives, possible patterns include words such as "more", "relatively", and "as much as possible" plus verbs or adjectives. When the structure "Hope to minimize unnecessary management expenses" is identified, it is determined to be a modification relationship structure containing a comparative.

[0057] Step S127, extracting the user preference strength indicator and the conditional constraint expression from the modified relationship structure.

[0058] For example, in the structure "hope to minimize unnecessary management fee expenditures", "minimize" is the user preference strength indicator, which shows the degree of investors' expectations for reducing management fee expenditures. "Management fee expenditures" is a conditional constraint expression, which is an important condition that investors pay attention to, namely the impact of financial products on management fees. The extraction process is based on the semantic analysis of the words in the modification relationship structure. According to the pre-set semantic rules, the words indicating the degree are determined as preference strength indicators, and the modified nouns or noun phrases are determined as conditional constraint expressions.

[0059] Step S128, identifying the beneficiaries, target objects and negation conditions in the text input content through a semantic role labeling model to obtain a semantic role labeling result.

[0060] For example, in the entire text input content, the beneficiary is the investor himself, 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 expenditure", which means that the investor does not want this situation to happen. The semantic role labeling 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 the investment-related statements, the investor is the subject of various investment behaviors and the recipient of investment income or impact, so it is determined as the beneficiary; for the target object, the financial product is determined as the target object based on the relationship between investment-related verbs such as "investment" and related nouns; for the negative condition, it is determined by identifying words that express negative meanings such as "unnecessary" and the objects it modifies.

[0061] Step S129 , feature encoding the user preference strength indicator, conditional constraint expression and semantic role labeling result to generate the multi-dimensional semantic tag set.

[0062] In this scenario, the user preference strength indicator "minimize" is first quantified and encoded. For example, according to the pre-set preference strength level scale, "minimize" may be encoded as a higher strength value, indicating that investors have a strong expectation to reduce management fee expenditures. The conditional constraint expression "management fee expenditures" can be encoded as a specific code, which corresponds to the concept of management fee in the financial field. The beneficiary "investor himself", the target object "financial product" and the negative condition "unnecessary management fee expenditures" are also encoded as corresponding codes. These codes are combined according to certain rules to form a multi-dimensional semantic tag set. The multi-dimensional semantic tag set can represent the semantic information in the text input content from multiple dimensions, including the user's preference strength, the conditional constraints of concern, and the relevant roles. For example, the multi-dimensional semantic tag set may represent the preference strength value in one dimension, the type of the target object in another dimension, and the negative condition in other dimensions. The semantic characteristics of the text input content are fully characterized by the above multi-dimensional representation method.

[0063] In a possible implementation, step S130 includes: Step S131, performing a timestamp alignment check on the structured requirement description set and the multidimensional semantic tag set, and identifying the time overlap interval and timing interval fault between the speech semantic paragraphs in the structured requirement description set and the text semantic paragraphs in the multidimensional semantic tag set.

[0064] In the scenario set up previously, since the investor's voice input and text input are almost simultaneous, the voice input and text input can be timestamped separately. The determination of the time overlap interval is achieved by comparing the start time and end time of the voice semantic paragraph and the text semantic paragraph. In this scenario, assuming that the voice input starts at time t1 and ends at time t2, and the text input starts at time t1+Δt1 (Δt1 is very small, indicating that the text input may start slightly later than the voice input, but almost at the same time), and ends at time t2+Δt2 (Δt2 is also very small), then the time overlap interval is from t1+Δt1 to t2. The timing interval fault is very small in the case of almost simultaneous input. For example, due to slight differences in the internal processing of the system, there may be a slight inconsistency in the processing progress of the voice semantic paragraph and the text semantic paragraph at a very short moment, but overall this fault is almost negligible.

[0065] Step S132, input the speech semantic paragraphs and text semantic paragraphs in the time overlapping interval into the gated recurrent unit network, calculate the semantic conflict probability between the speech semantic paragraphs and the text semantic paragraphs at the temporal interval fault through the forget gate of the gated recurrent unit network, and dynamically assign weights to the speech semantic paragraphs and the text semantic paragraphs based on the semantic conflict probability.

[0066] In this embodiment, the gated recurrent unit network is a special recurrent neural network structure, and the forget gate contained in it is used to control the degree of information forgetting. In this scenario, for the speech semantic paragraphs and text semantic paragraphs in the time overlapping interval, the forget gate will analyze the semantic relationship between the two at the time interval fault. For example, the speech semantic paragraph mentions "the risk is lower than stock investment", but the text semantic paragraph does not explicitly mention the risk compared to stock investment, which may cause semantic conflict. The forget gate can analyze the degree of this semantic difference according to the pre-defined semantic conflict calculation rules, thereby calculating the probability of semantic conflict. Assume that according to the degree of semantic difference and the pre-set calculation rules, the probability of this semantic conflict is calculated to be p (this p value is obtained by comprehensive analysis of multiple factors such as lexical semantics, syntactic structure and semantic association in the speech semantic paragraph and text semantic paragraph in the financial field, and is a value between 0 and 1). Dynamic weight allocation is performed on the speech semantic paragraph and the text semantic paragraph based on the semantic conflict probability. Assuming that 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 probability of semantic conflict p, the adjusted speech semantic paragraph weight w1'=w1*(1-p) and text semantic paragraph weight w2'=w2*p. This dynamic weight allocation adjusts the importance of speech semantic paragraphs and text semantic paragraphs in subsequent processing according to the possibility of semantic conflict.

[0067] Step S133, extracting historical voice interaction segments and historical text interaction segments associated with the current time overlapping 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 segments and historical text interaction segments.

[0068] In this scenario, the investor has previously had historical interaction records with customer service discussing investment risks and management fees. The parts associated with the current time overlap interval can be filtered out from these historical interaction records. For example, investors have mentioned risk control and management fees in similar portfolio adjustment discussions before, and these related historical voice interaction fragments and historical text interaction fragments will be extracted. For each historical interaction fragment, its frequency of recurrence can be counted. Assuming that the historical voice interaction fragment about risk control has appeared m times in the past n related interactions, its recurrence frequency is m / n. According to the above recurrence frequency, a numerical value is assigned to each related historical interaction fragment, and these numerical values ​​are combined in a certain order to form a semantic consistency verification vector. For example, if there are three related historical interaction fragments, and their recurrence frequencies are f1, f2, and f3 respectively, then the semantic consistency verification vector may be expressed as [f1, f2, f3]. The semantic consistency verification vector reflects the relevant information about the degree of semantic consistency between the historical interaction and the current interaction.

[0069] Step S134, performing 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 the dynamic weight allocation.

[0070] In this embodiment, the output of the gated recurrent unit network includes the information of the processed speech semantic paragraph and text semantic paragraph. Assume that the speech semantic paragraph information output by the gated recurrent unit network is v1, and the text semantic paragraph information is v2, and the corresponding confidence scores are s1 and s2 respectively (the confidence score is a measure of the reliability of the speech semantic paragraph and the text semantic paragraph information, and the initial value may be based on the default setting of the system or some preliminary processing results in the early stage). The semantic consistency verification vector is [c1, c2, c3] (here, in order to simplify the explanation, it is assumed that the semantic consistency verification vector has three elements). After the point-by-point multiplication operation, the corrected speech semantic paragraph confidence score s1'=s1*c1, and the corrected text semantic paragraph confidence score s2'=s2*c2 (here only the first two elements of the vector are taken for example calculation, and the actual situation will be fully 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 paragraph and text semantic paragraph, so that the results are more accurate and reliable.

[0071] Step S135 , cross-modally concatenating the speech semantic paragraph and the text semantic paragraph according to the corrected confidence score to generate a fused cross-modal semantic unit sequence.

[0072] In this embodiment, after obtaining the corrected confidence scores s1' and s2', the speech semantic paragraphs and the text semantic paragraphs are combined together according to certain splicing rules. For example, if the higher the confidence score, the more reliable the semantic information is, then the proportion of the speech semantic paragraphs and the text semantic paragraphs in the splicing result can be determined according to the ratio of the confidence scores. Assuming that the speech semantic paragraph is represented as x1 after a certain conversion, and the text semantic paragraph is represented as x2 after conversion, 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 rules can take into account more factors such as the structure and semantics of the speech semantic paragraphs and the text semantic paragraphs). The cross-modal semantic unit sequence integrates the semantic information of the two modes of speech and text, and is reasonably combined according to the confidence score.

[0073] Step S136, extracting semantic keyword combinations that appear continuously in the cross-modal semantic unit sequence, performing pattern matching on the semantic keyword combinations and product feature tags in the user's historical purchase records, and identifying implicit demand markers in the cross-modal semantic unit sequence.

[0074] In the above scenario, there may be semantic keyword combinations such as "risk lower than stocks", "withdraw within four or five years", and "reduce management fee expenditure" in the cross-modal semantic unit sequence. From the user's historical purchase records, the product feature tags that investors have previously purchased include tags such as the risk level of the previously purchased stock products is high risk, the fund product return type is medium return, and the term structure of fixed deposits is long-term. For each semantic keyword combination, 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 contrast relationship with the high-risk stock product feature tag, which may imply an implicit demand for medium-risk or low-risk products. This semantic relationship can be analyzed and judged according to the pre-set pattern matching rules. Assume that according to a series of matching rules and calculations (these rules and calculations involve multiple 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, indicating the probability of such implicit demand). By performing such pattern matching on all semantic keyword combinations, the implicit demand tags and their probabilities corresponding to each keyword combination are determined.

[0075] Step S137, inserting the implicit demand marker into the corresponding position of the cross-modal semantic unit sequence to generate the user portrait enhanced feature matrix with temporal context association.

[0076] In this embodiment, after determining each implicit demand marker and its probability, these implicit demand markers are inserted into the corresponding positions of the cross-modal semantic unit sequence according to the probability or other pre-set insertion rules. For example, if the implicit demand marker probability q1 of the medium-risk product corresponding to "risk lower than stocks" is higher, then the implicit demand marker is inserted near the risk-related semantic part in the cross-modal semantic unit sequence. By inserting all the determined implicit demand markers according to the rules, a user portrait enhanced feature matrix with temporal context association is generated. The user portrait enhanced feature matrix not only contains the explicit information in the voice semantic paragraphs and text semantic paragraphs, but also incorporates the implicit demand information mined from historical records and semantic analysis, and comprehensively portrays the demand portrait of investors in this financial investment scenario.

[0077] In a possible implementation, step S136 includes: Step S1361, extracting a product feature tag set of all purchased financial products from the user's historical purchase records, wherein the product feature tag set includes a product risk level tag, a return type tag, and a term structure tag.

[0078] In this embodiment, in the previously set scenario, the investor has previously purchased a variety of financial products. For example, he has previously purchased stock products with a high risk risk level label, fund products with a medium return type label, and time deposits with a term structure label of long-term. These labels fully describe the investor's past investment characteristics in different financial products, providing an important reference for subsequent analysis.

[0079] Step S1362, performing part-of-speech tagging and stop word filtering on the semantic keyword combination in the cross-modal semantic unit sequence, retaining the core semantic units including noun phrases and adjective phrases, and generating a target keyword sequence after dimensionality reduction.

[0080] 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, thus determining the part of speech of each word. Then stop word filtering is performed. A preposition like "than" belongs to stop words in this scenario and is 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.

[0081] 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 synonym cluster of risk preference implied in the noun phrase, and generate a risk level mapping relationship table.

[0082] 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. Assume that in this scenario, after comprehensive pattern matching, the synonym cluster related to "risk" is identified as including "danger", "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 lower risk than 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 is 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.

[0083] Step S1364: Based on the calculation of the semantic similarity 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.

[0084] Assuming that there are some adjective phrases related to income in the target keyword sequence (if any), the semantic similarity between these adjective phrases and income type labels (such as medium income) 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 based on the semantic association between "stable income" and medium income in the financial field. Assuming that the preset similarity threshold is 0.6, if the calculated semantic similarity between "stable" and medium income is 0.7, which meets the preset threshold, then "stable" is screened out as an income descriptor, and the corresponding medium income type label is associated to generate an income demand mapping relationship table. The table records the correspondence between the income descriptor and the income type label and the relevant calculation results, such as "stable-medium income, similarity is 0.7".

[0085] Step S1365, traverse the time unit identifier of the term structure label, extract continuous word pairs containing numerical values ​​and time units from the target keyword sequence, perform interval inclusion check on the continuous word pairs and the allowed value interval of the term structure label, and generate a term matching relationship table.

[0086] For example, the term structure label of the term deposit mentioned earlier is long-term, and its allowed value range may be 3 years and above. From the target keyword sequence, such as the continuous word pair "withdrawal in four or five years", "four or five years" is a combination of numerical values ​​and time units. Then, the continuous word pair can be checked for interval inclusion with the allowed value range of the term structure label of the term deposit. Since "four or five years" is within the range of 3 years and above, the interval inclusion requirement is met. By verifying all such continuous word pairs, a term matching relationship table is generated, which records the matching relationship between the continuous word pairs and the term structure labels and the verification results, such as "four or five years - term deposit (long-term), meets interval inclusion".

[0087] Step S1366, generating a multidimensional demand intensity vector according to the occurrence frequency of synonym clusters in the risk level mapping relationship table, the sentiment polarity strength of the benefit descriptor in the benefit demand mapping relationship table, and the interval matching degree of continuous word pairs in the term matching relationship table.

[0088] For example, in the risk level mapping relationship table, the frequency of occurrence of "risk" is 1. In the yield demand mapping relationship table, if the yield descriptor "stable" is regarded as a positive description, its emotional polarity strength can be set to +1 (here it is assumed that the emotional polarity strength of positive descriptions is +1 and negative descriptions are -1). In the term matching relationship table, the interval matching degree of "four or five years - fixed deposit (long-term)" is 1 (indicating a complete match). Based on this information, a multidimensional demand strength vector is generated, such as [1 (risk frequency), +1 (yield emotional polarity strength), 1 (term matching)]. This multidimensional demand strength vector comprehensively describes the investor's demand strength in terms of risk, return and term from multiple dimensions.

[0089] 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 timestamp of each product feature tag in the user's historical purchase record.

[0090] Assume that the time when the investor buys high-risk stocks is t1, and the time from the current time is Δt1; the time when the investor buys medium-yield funds is t2, and the time from the current time is Δt2; the time when the investor buys long-term time deposits is t3, and the time from the current time is Δt3. The time decay function is pre-set according to the changing laws of the financial market and the changing characteristics of investor demand over time. For example, the time decay function may be set so that the demand weight decreases at a certain ratio as the time interval increases. When calculating the demand weight decay corresponding to high-risk stocks, first determine the corresponding decay coefficient on the time decay function curve according to Δt1. Assume that in the time decay function, when the decay coefficient corresponding to Δt1 is 0.8 (the decay coefficient is calculated based on the definition of the time decay function and the value of Δt1, and the calculation process involves the evaluation of the time decay function, such as calculation based on the expression of the function, the value of Δt1 and the relevant parameters in the function). Then the demand weight value of high-risk stocks after decay in the risk dimension is 1 (risk frequency) * 0.8 = 0.8. Similarly, calculate the demand weight values ​​of medium-yield funds and long-term time deposits after decay in the income and term dimensions. Assuming that the attenuation coefficient corresponding to the medium-yield fund is 0.9, the demand weight value after attenuation in the yield dimension is calculated to be +1 (yield sentiment polarity intensity) * 0.9 = 0.9; the attenuation coefficient corresponding to the long-term time deposit is 0.7, and the demand weight value after attenuation in the term dimension is calculated to be 1 (term matching degree) * 0.7 = 0.7.

[0091] Step S1368, weighted fusion of the risk level mapping relationship table, the income demand mapping relationship table and the term matching relationship table is performed according to the demand weight value to generate an implicit demand probability distribution matrix.

[0092] 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), Expected to be Lower Than This Risk" is weighted according to the attenuated demand weight value of 0.8. Assuming that the default weight of each item in the original risk level mapping relationship table is 1, the new weight relationship is obtained after weighting. Similarly, the items in the return demand mapping relationship table and the term matching relationship table are also weighted according to their respective attenuated demand weight values. Then the three weighted tables are merged to generate the implicit demand probability distribution matrix. Each element in the implicit demand probability distribution matrix integrates the information of the three dimensions of risk, return and term, and takes into account the influence of time attenuation. For example, a row in the implicit demand probability distribution matrix may be expressed as [0.8 (risk weighted value), 0.9 (return weighted value), 0.7 (term weighted value), implicit demand probability value (calculated according to weighted values ​​and other rules)].

[0093] Step S1369, extracting row vectors whose probability values ​​exceed the dynamic threshold from the implicit demand probability distribution matrix, performing semantic compatibility check on the corresponding product feature labels and the context position of the cross-modal semantic unit sequence, generating an insertable implicit demand tag queue, inserting the tags in the implicit demand tag queue that meet the semantic compatibility conditions into the syntax tree node gaps of the cross-modal semantic unit sequence in descending order of weight, and generating an enhanced semantic unit sequence containing explicit demands and implicit demands.

[0094] Assuming that the dynamic threshold is set to 0.6, the row vectors with probability values ​​exceeding 0.6 are screened out from the implicit demand probability distribution matrix. For example, the product feature labels corresponding to the screened row vectors are medium-risk products (obtained by associating with information such as risk weighted values), medium-return products (obtained by associating with information such as return weighted values), etc. Then, these product feature labels are semantically compatible with the context position of the cross-modal semantic unit sequence. For example, in the context position where "risk is lower than stocks" is mentioned in the cross-modal semantic unit sequence, the product feature label of medium-risk products is semantically compatible because it meets investors' expectations that risk is lower than stocks. By performing such a semantic compatibility check on all screened product feature labels, an insertable implicit demand tag queue is generated, in which product feature labels that meet the semantic compatibility conditions are arranged in a certain order (for example, from high to low according to probability values).

[0095] On this basis, in the implicit demand tag queue, if the weight of medium-risk products is the highest, then the implicit demand tag of medium-risk products is first inserted into the risk-related position in the gap between 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, the medium-risk product tag is inserted in the gap near the node. Other implicit demand tags that meet the conditions are inserted in descending order of weight, and finally an enhanced semantic unit sequence containing explicit needs (such as explicit expressions in the original cross-modal semantic unit sequence) and implicit needs (needs represented by the newly inserted tags) is generated. This enhanced semantic unit sequence reflects the needs of investors more comprehensively, including both the needs explicitly 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.

[0096] In a possible implementation, step S1367 includes: Step S1367-1, extracting a set of product feature tags including a product risk level tag, a return type tag and a term structure tag from the user's historical purchase records, and synchronously obtaining a purchase timestamp corresponding to each product feature tag.

[0097] In the scenario set above, the investor's historical purchase records include previously purchased stock products, whose product risk level label is high risk and purchase timestamp is t1; the fund product's return type label is medium return and purchase timestamp is t2; the term structure label of the time deposit is long term and purchase timestamp is t3. These product feature labels and their corresponding purchase timestamps fully record the investor's past investment situation.

[0098] Step S1367-2, performing 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.

[0099] Assuming that the current system time is T, then for stock products, the time decay interval is Δt1=T-t1; for fund products, the time decay interval is Δt2=T-t2; for time deposits, the time decay interval is Δt3=T-t3. These time decay intervals reflect the time from the purchase of the product to the current time. The longer the time, the greater the decay effect on demand. Through such calculations, the time decay interval sequence [Δt1, Δt2, Δt3] of each product feature label is obtained.

[0100] 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 the decay coefficient vector of each product feature label.

[0101] The preset time decay function curve is pre-set based on the changing laws of the financial market and the changing characteristics of investor demand over time. For stock products, the position corresponding to the time decay interval Δt1 on the time decay function curve determines an attenuation slope parameter k1. The search process is determined based on 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 attenuation slope parameter k1 will cause a greater degree of attenuation. Similarly, for Δt2 of fund products and Δt3 of time deposits, the corresponding attenuation slope parameters k2 and k3 are searched on the time decay function curve respectively. Thereby generating the attenuation coefficient vector [k1, k2, k3] of each product feature label.

[0102] Step S1367-4, perform dimensionally point multiplication operations 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 multidimensional demand intensity vector with the corresponding attenuation coefficient vector respectively to generate a demand weight vector after preliminary attenuation.

[0103] In the previous step, the multidimensional demand intensity vector may be [1 (risk frequency), +1 (income sentiment polarity strength), 1 (term matching)]. For the demand intensity component 1 (risk frequency) in the risk level mapping relationship table, dot multiplication operation is performed with the attenuation coefficient k1 corresponding to the stock product, and the initial attenuated risk dimension demand weight is 1*k1. Similarly, for the demand intensity component +1 (income sentiment polarity strength) in the income demand mapping relationship table, dot multiplication operation is performed with the attenuation coefficient k2 corresponding to the fund product, and the initial attenuated income dimension demand weight is +1*k2; for the demand intensity component 1 (term matching degree) in the term matching relationship table, dot multiplication operation is performed with the attenuation coefficient k3 corresponding to the time deposit, and the initial attenuated term dimension demand weight is 1*k3. In this way, the initial attenuated demand weight vector [1*k1, +1*k2, 1*k3] is generated.

[0104] Step S1367-5, normalize the demand weight vector after the initial attenuation, so that the sum of the demand weights of the risk level mapping relationship table, the income demand mapping relationship table and the term matching relationship table remains constant, and generate a normalized demand weight vector.

[0105] Assume that the demand weight vector after initial decay is [w1, w2, w3], and the sum of the demand weights is S=w1+w2+w3. The purpose of normalization is to ensure that after considering time decay, the demand weights of each dimension remain consistent in sum to meet 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 term dimension, the normalized weight is w3'=w3 / S. This generates a normalized demand weight vector [w1', w2', w3'].

[0106] Step S1367-6, probabilistically superimpose the normalized demand weight vector and the explicit demand mark in the user portrait enhancement feature matrix to generate a final demand weight value that incorporates the time attenuation effect.

[0107] In detail, in the user profile enhancement feature matrix, the explicit demand markers contain the demand information explicitly expressed by investors. For example, in the explicit demand markers related to risk, a certain attitude or requirement towards risk may be explicitly expressed. The normalized demand weight vector [w1', w2', w3'] is superimposed with these explicit demand markers in the probability space. This superposition process is based on the specific rules and logic of demand analysis in the financial field, and comprehensively considers the relationship between the demand weight after the time decay effect and the explicit demand marker. For example, if the explicit demand marker has a specific weight representation in the risk dimension, then it is superimposed with the normalized risk dimension demand weight w1' to obtain the final demand weight value after the integration of the time decay effect. The same operation is performed for the return and term dimensions, and finally the final demand weight values ​​in multiple dimensions such as risk, return and term that fully integrate the time decay effect are obtained. These final demand weight values ​​can more accurately reflect the current demand status of investors under the influence of time factors, and provide a more accurate basis for subsequent analysis and decision-making.

[0108] In a possible implementation, step S140 includes: Step S141, extracting product attribute nodes and their associated regulatory policy nodes and market situation nodes from the financial product knowledge graph.

[0109] In the scenario set previously, in the financial product knowledge graph, taking wealth management products as an example, its product attribute nodes contain attribute information such as product risk level, return type, investment period, etc. For this wealth management product, its associated regulatory policy nodes may include risk management regulations nodes issued by financial regulatory agencies for wealth management products, such as nodes that stipulate risk assessment standards and risk control requirements for wealth management products; sales restriction regulations nodes, such as restrictions on sales channels and sales objects. Market situation nodes may include current market interest rate situation nodes, which reflect the impact of the overall market interest rate level on the returns of wealth management products; market supply and demand situation nodes for similar wealth management products, which indicate the balance between demand and supply of wealth management products in the market. These nodes describe the characteristics of wealth management products, regulatory constraints they are subject to, and the market environment in which they are located from different aspects.

[0110] Step S142, based on the bidirectional 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.

[0111] The user profile enhancement feature matrix contains information about investors' needs in terms of risk, return, term, etc., as well as implicit needs mined from historical purchase records and semantic analysis. The bidirectional graph attention network comprehensively considers the relationship between this information and the product attribute nodes of the wealth management product. For example, for the risk level attribute of the wealth management product, if the user profile enhancement feature matrix shows that investors expect medium-risk products, and the risk level of the wealth management product is medium risk, the bidirectional graph attention network will quantitatively evaluate this matching situation according to the pre-set risk level matching rules. For the type of income, if the investor expects stable income, and the income type provided by the wealth management product meets the characteristics of stable income, a corresponding quantitative evaluation will also be performed. Similar evaluations are also performed for other attributes such as the investment term. Through a comprehensive evaluation of all product attribute nodes, the first matching degree between the user profile enhancement feature matrix and each product attribute node is calculated. The first matching degree is a quantitative value that reflects the matching degree between user needs and product attributes. For example, it may be a value between 0 and 1, where 0 indicates a complete mismatch and 1 indicates a complete match.

[0112] Step S143, traverse the regulatory policy nodes and market situation nodes through the meta-path walking algorithm to generate a product compliance assessment coefficient.

[0113] In a possible implementation, step S143 includes: Step S1431, determine the meta-path rule set between the regulatory policy node and the market situation node in the financial product knowledge graph, the meta-path rule set includes a regulatory policy transmission path starting from the product attribute node and a market situation transmission path starting from the product attribute node.

[0114] Taking financial products as an example, for the regulatory policy transmission path, there may be a transmission path from the financial product attribute node to the financial regulatory agency node, and then to the specific regulatory policy clause node; for the market situation transmission path, there may be a transmission path from the financial product attribute node to the market sector node, and then to the market situation indicator node. The above meta-path rules clarify the path rules for how to start from the product attribute node in the financial product knowledge graph and reach the regulatory policy node and market situation node.

[0115] Step S1432, according to the path type priority list of each meta-path in the meta-path rule set, a multi-hop traversal operation is performed on the financial product knowledge graph to generate a regulatory policy node sequence and a market situation node sequence associated with the product attribute node.

[0116] Assume that in the regulatory policy transmission path, the path type priority from the wealth management product attribute node to the financial regulatory agency node that directly regulates the product is higher. When performing a 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 the agency node. For example, first find the node of the relevant regulatory department that regulates the wealth management product, and then find the regulatory policy clause nodes such as risk management and sales specifications related to the wealth management product issued by the relevant regulatory department to form a regulatory policy node sequence. For the market situation node sequence, according to the priority order of the market situation transmission 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 interest rate, supply and demand situation and other market situation nodes under the sector to form a market situation node sequence.

[0117] Step S1433, extracting the effective time interval and constraint effectiveness strength parameters of each regulatory policy node from the regulatory policy node sequence, and extracting the volatility index and trend direction identifier of each market situation node from the market situation node sequence.

[0118] For the risk management regulations node in the regulatory policy node sequence, its effective time interval may be from a specific date to the present, and the constraint strength parameter is set according to the strictness of the policy. For example, if the regulations are very strict, the constraint strength parameter may be set to a higher value. For the market interest rate situation node in the market situation node sequence, the volatility index 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, such as setting the rising trend as a positive identifier and the falling trend as a negative identifier.

[0119] Step S1434, matching the overlap degree of the effective time interval with the current system time, generating a regulatory policy timeliness weight, and multiplying the constraint effectiveness strength parameter with the regulatory policy timeliness weight to generate a dynamically adjusted set of regulatory policy impact factors.

[0120] Assume that the effective time interval of the risk management regulation node is from t1 to the present, and the current system time is T. If the time span between t1 and T accounts for a large proportion of the entire effective time interval, it means that the policy is highly timely. According to the pre-set calculation rules, a higher regulatory policy timeliness weight is generated, such as 0.8. The timeliness weight is multiplied by the constraint effectiveness strength parameter (assuming it is 0.6), and the dynamically adjusted regulatory policy impact factor is 0.8*0.6=0.48. This calculation is performed for each node in the regulatory policy node sequence to generate a dynamically adjusted regulatory policy impact factor set.

[0121] Step S1435, compare the volatility index with a preset market sensitivity threshold, screen out market nodes that exceed the market sensitivity threshold as significant volatility nodes, and perform consistency matching between the trend direction identifier and the risk preference label in the user portrait enhancement feature matrix to generate a market suitability set.

[0122] Assuming that the preset market sensitivity threshold is a specific volatility value, for a market interest rate situation node, if its volatility index exceeds the market sensitivity threshold, then the node is determined to be a significant volatility node. For the trend direction identifier, if the risk preference label in the user portrait enhancement feature matrix shows that investors prefer a stable investment environment, and the trend direction identifier of the market interest rate is a decline (indicating that the market is relatively stable), then a higher fitness value, such as 0.7, is recorded in the market fitness set, indicating that the market situation and the investor's risk preference are highly consistent.

[0123] Step S1436, assigning a decreasing propagation weight coefficient according to the number of path hops of each node in the dynamically adjusted regulatory policy impact factor set, and performing weighted summation of the propagation weight coefficient and the corresponding regulatory policy impact factor to generate a basic product compliance score.

[0124] In the dynamically adjusted regulatory policy impact factor set, assuming that the path hop count of a regulatory policy node is 1 (closer to the product attribute node), a higher propagation weight coefficient, such as 0.5, is assigned according to the pre-set decreasing propagation weight coefficient allocation rule. The propagation weight coefficient is weighted and summed with the regulatory policy impact factor of the node (assuming it is 0.48) to obtain a partial sum. This operation is performed for each node in the set, and all partial sums are added together to generate the basic product compliance score.

[0125] Step S1437, calculate the product market stability offset according to the occurrence frequency of the significant fluctuation market node and the matching results in the market market adaptability set, and add the product market stability offset to the product compliance basic score to generate the product compliance assessment coefficient.

[0126] Assuming that the frequency of occurrence of significant volatility nodes is 3 times, according to the pre-set calculation rules, this frequency may cause a certain market stability deviation. If the average fitness in the market fitness set is high, the impact of this deviation may be mitigated. According to the specific calculation rules (involving comprehensive consideration of multiple factors such as frequency of occurrence and fitness), the product market stability deviation is calculated, for example, -0.1 (indicating a certain negative impact on the basic compliance score). The deviation is superimposed on the product compliance basic score. For example, if the product compliance basic score is 0.8, the product compliance assessment coefficient after superposition is 0.8-0.1=0.7.

[0127] Step S144: nonlinearly combine the first matching degree and the compliance assessment coefficient to generate a comprehensive product recommendation score.

[0128] Nonlinear combination is based on pre-set combination rules, taking into account the different importance and mutual relationship of the first matching degree and the compliance assessment coefficient in product recommendation. For example, a weighted power combination method may be adopted to add a certain power (such as square) of the first matching degree and a certain power (such as cube) of the compliance assessment coefficient according to a certain weight. Assuming that the first matching degree is 0.6 and the compliance assessment coefficient is 0.7, the comprehensive product recommendation score is calculated according to the set combination rules. The comprehensive product recommendation score combines the degree of match between user needs and product attributes and the compliance of the product.

[0129] Step S145, sorting and screening the financial products according to the comprehensive recommendation scores to generate the initial recommendation result queue.

[0130] For example, the financial products with the highest scores are placed at the front of the queue, and the products with lower scores are placed at the back. This initial recommendation result queue is the result of 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 investors with financial product recommendations that better meet their needs and market conditions.

[0131] In a possible implementation, step S150 includes: Step S151, obtain the location sensor data and ambient light intensity parameters of the user terminal, calculate the user's current attention level, detect the device network delay index and interface rendering frame rate, and evaluate the terminal's interactive responsiveness.

[0132] In this embodiment, in the previously set scenario, the investor is using a tablet computer, and the location sensor data shows that the investor is in the study room at home. Assume that through further analysis of the location sensor data, for example, based on the historical usage pattern of the study room (if the system has relevant records) and the current time and other factors, it is determined that this is a relatively quiet environment with less interference, which may imply that the investor has a higher level of attention. The ambient light intensity parameter shows that the light is moderate, neither too bright nor too dark, which also helps the investor maintain good attention. Combining the above information, according to the pre-set attention level calculation rules, it is calculated that the user's current attention level is high.

[0133] The device network delay index is obtained by measuring information such as the transmission time of network data packets. Assuming that the network delay is low, it means that the data transmission speed is fast and the recommendation results and other information can be obtained in time. The interface rendering frame rate reflects the speed of interface update. If the interface rendering frame rate is normal, it means that the interface can display various elements smoothly. According to the network delay index and the interface rendering frame rate, the terminal interactive responsiveness is evaluated according to the pre-defined evaluation rules. Since the network delay is low and the interface rendering frame rate is normal, the terminal interactive responsiveness is judged to be good.

[0134] Step S152, adjusting the display granularity level of the initial recommendation result queue according to the attention level, generating a dynamic information density parameter, and optimizing the loading priority strategy of the initial recommendation result queue according to the terminal interactive response capability.

[0135] In this embodiment, because investors have a higher level of attention, more detailed and in-depth information can be provided. In the initial recommendation result queue, the recommended information for each financial product may have been a simple overview, but now more details can be added, such as detailed return calculation methods, risk analysis reports, historical performance data, etc. According to the degree of this adjustment, a dynamic information density parameter is generated according to preset rules. For example, if the level of detail of the information is increased from the basic level to the high level, the dynamic information density parameter may be increased from 1.0 (basic value, indicating normal information density) to 1.5, indicating that the information density has increased by 50%.

[0136] Due to the good interactive responsiveness of the terminal, more recommended product information can be loaded more actively and in an order that is more conducive to the user's acquisition of information. For example, for those products that are more closely matched with investor needs, more detailed information can be loaded first, while for products with a slightly lower degree of match, they can be loaded gradually in the background. This optimization of the loading priority strategy is based on the evaluation results of the interactive responsiveness of the terminal, ensuring that information is provided as efficiently as possible without affecting the user experience.

[0137] Step S153: Apply the dynamic information density parameter and the loading priority strategy to the initial recommendation result queue to generate a dynamic adaptation recommendation list.

[0138] In the initial recommendation result queue, the recommendation information of each financial product is adjusted according to the dynamic information density parameter to increase or decrease the amount of information displayed. For example, for a financial product that is at the front of the queue and has a high degree of matching, according to the dynamic information density parameter of 1.5, it originally only displayed basic information such as the type of return and risk level, but now also adds detailed information such as the detailed return curve for the past three years and risk volatility analysis under different market environments. At the same time, according to the loading priority strategy, the detailed information of the financial product will be loaded first and displayed in a more prominent position. In this way, each product in the initial recommendation result queue is processed to generate a dynamically adapted recommendation list, which is more in line with the current attention level of investors and the interactive response capabilities of the terminal, and can provide more personalized and effective recommendation information.

[0139] Step S154, creating a three-dimensional visualization space in the interactive interface, mapping the recommended products in the dynamically adapted recommendation list into three-dimensional icons of different shapes, and dynamically adjusting the color saturation and rotation speed of the three-dimensional icons according to the product risk level.

[0140] For example, in the three-dimensional visualization space of the interactive interface, a unique three-dimensional icon is created for each financial product in the dynamically adapted recommendation list. For example, for a fixed deposit product with a lower risk, it is mapped as a cube icon. The shape of the cube gives people a sense of stability and reliability, which is consistent with the characteristics of fixed deposits with low risk and stable returns; for a stock fund product with a higher risk, it is mapped as a triangular pyramid icon. The shape of the triangular pyramid is relatively sharp, suggesting its higher risk characteristics.

[0141] For example, for low-risk time deposit products, the color saturation of the three-dimensional icon is set to be low, such as light blue, and the rotation speed is slow, such as 10 degrees per minute. This low color saturation and slow rotation speed give people a stable and safe visual experience, which intuitively reflects the low-risk characteristics of the product. For high-risk stock fund products, the color saturation of the three-dimensional icon is set to be high, such as red, and the rotation speed is fast, such as 30 degrees per minute. Red and fast rotation speed can attract users' attention, and also convey the high-risk and high-volatility characteristics of the product.

[0142] Step S155, construct an interactive timeline control to update the spatiotemporal distribution of recommendation results in real time in response to the user's sliding operation. When the user gazes at any three-dimensional icon for a time period exceeding a threshold, a multi-level information expansion animation is triggered.

[0143] An interactive timeline control can be built in the interactive interface, which is associated with the time-related information of the recommended product, such as the investment period of the product, the time distribution of historical performance, etc. When the user slides on the timeline, the system updates the time and space distribution of the recommended results in real time according to the position of the user's slide. For example, if the user slides the timeline slider to the position of the next year, the expected return, risk and other information of the recommended product in the future time period can be readjusted according to the forecast model of the financial market (if any), and the updated recommendation results can be displayed in real time on the interface.

[0144] When the user stares at any three-dimensional icon for longer than the threshold time, the multi-level information expansion animation is triggered. Assuming the threshold time is set to 3 seconds, when the user stares at the three-dimensional icon of a financial product for more than 3 seconds, the multi-level information expansion animation can be triggered. The multi-level information expansion animation may first display a more detailed income structure of the financial product, such as the expected income ratio of different terms; then display risk decomposition information, such as the proportion of market risk, credit risk, etc. in the total risk; finally, it may display some market trend analysis or expert comments related to the product. This multi-level information expansion animation allows users to quickly obtain more in-depth information when they are paying attention to a certain product, without having to actively look for this information.

[0145] 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.

[0146] 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.

[0147] 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.

[0148] 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.

[0149] 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.

[0150] 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, so the steps performed by one processor described in the present invention may also be performed jointly or individually 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 performed jointly by two different processors or individually 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 perform steps A and B together.

[0151] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned bank wealth management product recommendation method is implemented.

[0152] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A method for recommending bank financial products, characterized in that: The method comprises: Collecting a multimodal dialogue data stream generated during the interaction between the user terminal and the intelligent customer service system, wherein the multimodal dialogue data stream includes a voice input signal and a text input content; Performing semantic recognition on the voice input signal to generate a structured demand description set, and performing entity extraction and intent analysis on the text input content to generate a multi-dimensional semantic tag set; The structured requirement description set is temporally and spatially aligned and fused with the multidimensional semantic label set to generate a user portrait enhanced feature matrix; Based on the deep matching model, the user portrait enhanced feature matrix and the financial product knowledge graph are subjected to multi-level correlation analysis to generate an initial recommendation result queue; 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.

2. The bank financial product recommendation method according to claim 1, characterized in that: The performing semantic recognition on the voice input signal to generate a structured requirement description set includes: Performing frame processing and background noise elimination on the speech input signal, and extracting Mel frequency cepstral coefficients as acoustic feature vectors; The speech input signal is converted into an intermediate text representation through a pre-trained speech recognition model, and the speaker's emotion intensity value is annotated; A multi-head attention mechanism is used to perform contextual semantic completion on the intermediate text expression to generate a corrected complete semantic paragraph; Extracting constraint entities of amount, term, and risk category from the complete semantic paragraph using a domain-specific named entity recognition model; The acoustic feature vector, the emotion intensity value and the constraint condition entity are tensor-concatenated to generate the structured requirement description set with emotion weight.

3. The bank financial product recommendation method according to claim 2, characterized in that: The entity extraction and intent analysis of the text input content to generate a multi-dimensional semantic tag set includes: Perform word segmentation and part-of-speech tagging on the text input content to construct a dependency syntactic analysis tree; Executing a pattern matching algorithm on the dependency syntax analysis tree to identify modification relationship structures including comparatives and superlatives; Extracting user preference strength indicators and conditional constraint expressions from the modified relation structure; Identify the beneficiaries, target objects and negation conditions in the text input content through a semantic role labeling model to obtain a semantic role labeling result; The user preference strength indicator, conditional constraint expression and semantic role labeling result are feature encoded to generate the multi-dimensional semantic label set.

4. The bank financial product recommendation method according to claim 1, characterized in that: The step of performing spatiotemporal alignment and fusion of the structured requirement description set and the multidimensional semantic tag set to generate a user portrait enhanced feature matrix includes: Performing a timestamp alignment check on the structured demand description set and the multidimensional semantic tag set to identify time overlap intervals and timing interval faults between the speech semantic paragraphs in the structured demand description set and the text semantic paragraphs in the multidimensional semantic tag set; Inputting the speech semantic paragraph and the text semantic paragraph in the time overlapping interval into a gated recurrent unit network, calculating the semantic conflict probability between the speech semantic paragraph and the text semantic paragraph at the temporal interval fault through the forget gate of the gated recurrent unit network, and dynamically assigning weights to the speech semantic paragraph and the text semantic paragraph based on the semantic conflict probability; Extracting historical voice interaction segments and historical text interaction segments associated with the current time overlapping 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 segments and historical text interaction segments; Performing 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 the dynamic weight allocation; Cross-modally concatenating the speech semantic paragraphs and the text semantic paragraphs according to the corrected confidence scores to generate a fused cross-modal semantic unit sequence; Extracting semantic keyword combinations that appear continuously in the cross-modal semantic unit sequence, performing pattern matching on the semantic keyword combinations and product feature tags in the user's historical purchase records, and identifying implicit demand markers in the cross-modal semantic unit sequence; The implicit demand marker is inserted into the corresponding position of the cross-modal semantic unit sequence to generate the user portrait enhanced feature matrix with temporal context association.

5. The bank financial product recommendation method according to claim 4, characterized in that: The pattern matching of the semantic keyword combination with the product feature tags in the user's historical purchase records to identify the implicit demand markers in the cross-modal semantic unit sequence includes: Extracting a product feature label set of all purchased financial products from the user's historical purchase records, wherein the product feature label set includes a product risk level label, a return type label, and a term structure label; Performing part-of-speech tagging and stop word filtering on the semantic keyword combination in the cross-modal semantic unit sequence, retaining the core semantic units including noun phrases and adjective phrases, and generating a target keyword sequence after dimensionality reduction; Performing pattern matching item by item on each noun phrase in the target keyword sequence and the product risk level label, identifying risk preference synonym clusters implied in the noun phrases, and generating a risk level mapping relationship table; Based on the semantic similarity calculation between the adjective phrase and the income type label, the income descriptor that meets the preset similarity threshold is screened out, and the corresponding income type label is associated to generate an income demand mapping relationship table; Traversing the time unit identifiers of the term structure label, extracting continuous word pairs containing numerical values ​​and time units from the target keyword sequence, performing interval inclusion check on the continuous word pairs and the allowed value interval of the term structure label, and generating a term matching relationship table; Generate a multidimensional demand intensity vector according to the occurrence frequency of synonym clusters in the risk level mapping relationship table, the sentiment polarity strength of the benefit descriptor in the benefit demand mapping relationship table, and the interval matching degree of the continuous word pairs in the term matching relationship table; Input the multidimensional demand intensity vector into a time decay function, and calculate the decayed demand weight value according to the purchase timestamp of each product feature tag in the user's historical purchase record; The risk level mapping relationship table, the income demand mapping relationship table and the term matching relationship table are weighted and integrated according to the demand weight value to generate an implicit demand probability distribution matrix; Extracting row vectors whose probability values ​​exceed a dynamic threshold from the implicit demand probability distribution matrix, performing semantic compatibility check between the corresponding product feature labels and the context position of the cross-modal semantic unit sequence, and generating an insertable implicit demand tag queue; The tags in the implicit requirement tag queue that meet the semantic compatibility condition are inserted into the syntax tree node gaps of the cross-modal semantic unit sequence in descending order of weight to generate an enhanced semantic unit sequence containing explicit requirements and implicit requirements.

6. The bank financial product recommendation method according to claim 5, characterized in that: The step of inputting the multi-dimensional demand intensity vector into a time decay function and calculating the decayed demand weight value according to the purchase timestamp of each product feature tag in the user's historical purchase record comprises: Extracting a product feature tag set including a product risk level tag, a return type tag, and a term structure tag from the user's historical purchase records, and synchronously obtaining a purchase timestamp corresponding to each product feature tag; Performing 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 value of the time decay interval sequence, the corresponding decay slope parameter is searched on the preset time decay function curve to generate the decay coefficient vector of each product feature label; Performing a dimension-wise dot multiplication 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 multidimensional demand intensity vector with the corresponding attenuation coefficient vector, respectively, to generate a demand weight vector after preliminary attenuation; Normalizing the demand weight vector after the initial attenuation so that the sum of the demand weights of the risk level mapping relationship table, the income demand mapping relationship table and the term matching relationship table remains constant, and generating a normalized demand weight vector; The normalized demand weight vector is probabilistically superimposed with the explicit demand mark in the user portrait enhanced feature matrix to generate a final demand weight value integrating the time decay effect.

7. The bank financial product recommendation method according to claim 1, characterized in that: The deep matching model is used to perform multi-level correlation analysis on the user portrait enhanced feature matrix and the financial product knowledge graph to generate an initial recommendation result queue, including: Extract product attribute nodes and their associated regulatory policy nodes and market situation nodes from the financial product knowledge graph; Based on the bidirectional graph attention network in the deep matching model, the first matching degree between the user portrait enhanced feature matrix and each product attribute node is calculated; The meta-path walking algorithm is used to traverse regulatory policy nodes and market situation nodes to generate product compliance assessment coefficients; Performing a nonlinear combination of the first matching degree and the compliance assessment coefficient to generate a comprehensive product recommendation score; The financial products are sorted and screened according to the comprehensive recommendation scores to generate the initial recommendation result queue.

8. The bank financial product recommendation method according to claim 7, characterized in that: The meta-path walking algorithm traverses the regulatory policy nodes and market situation nodes to generate product compliance assessment coefficients, including: Determine a meta-path rule set between a regulatory policy node and a market situation node in the financial product knowledge graph, wherein the meta-path rule set includes a regulatory policy transmission path starting from a product attribute node and a market situation transmission path starting from a product attribute node; According to the path type priority list of each meta-path in the meta-path rule set, a multi-hop traversal operation is performed on the financial product knowledge graph to generate a regulatory policy node sequence and a market situation node sequence associated with the product attribute node; Extracting the effective time interval and binding strength parameters of each regulatory policy node from the regulatory policy node sequence, and extracting the volatility index and trend direction identifier of each market situation node from the market situation node sequence; Match the overlap degree of the effective time interval with the current system time to generate the regulatory policy timeliness weight, and multiply the constraint effectiveness strength parameter with the regulatory policy timeliness weight to generate a dynamically adjusted regulatory policy impact factor set; Compare the volatility index with a preset market sensitivity threshold, select market nodes that exceed the market sensitivity threshold as significant volatility nodes, and perform consistency matching between the trend direction identifier and the risk preference label in the user portrait enhancement feature matrix to generate a market suitability set; According to the number of path hops of each node in the dynamically adjusted regulatory policy impact factor set, a decreasing propagation weight coefficient is allocated, and the propagation weight coefficient is weighted and summed with the corresponding regulatory policy impact factor to generate a basic compliance score for the product; According to the occurrence frequency of the significant fluctuation market nodes and the matching results in the market situation adaptability set, the product market stability offset is calculated, and the product market stability offset is superimposed on the product compliance basic score to generate the product compliance assessment coefficient.

9. 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 the real-time environment perception data, generating a dynamically adapted recommendation list and performing visual rendering output through a multi-channel interactive interface, comprises: Obtain the location sensor data and ambient light intensity parameters of the user terminal, calculate the user's current attention level, detect the device network delay index and interface rendering frame rate, and evaluate the terminal's interactive responsiveness; Adjusting the display granularity level of the initial recommendation result queue according to the attention level, generating a dynamic information density parameter, and optimizing the loading priority strategy of the initial recommendation result queue according to the terminal interactive response capability; Applying the dynamic information density parameter and the loading priority strategy to the initial recommendation result queue to generate a dynamic adaptation recommendation list; And, creating a three-dimensional visualization space in the interactive interface, mapping the recommended products in the dynamically adapted recommendation list into three-dimensional icons of different shapes, and dynamically adjusting 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 sliding operations and update the temporal and spatial distribution of recommendation results in real time. When the user stares at any three-dimensional icon for a time period exceeding the threshold, a multi-level information expansion animation is triggered. Generate a focus heat map based on eye tracking in the edge area of ​​the interface and adjust the layout density of recommended elements in real time.

10. A bank financial product recommendation system, characterized in that: The bank wealth management 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 wealth management product recommendation method described in any one of claims 1 to 9 above.

Images