A user intention mining and active recommendation method for a financial knowledge middle platform

Abstract

The application discloses a user intention mining and active recommendation method for a financial knowledge middle platform, and belongs to the technical field of artificial intelligence. The method acquires current interaction behavior of a user on the financial knowledge middle platform, determines a business scene and historical context, selects associated survey questions from an intention survey question library, acquires knowledge reply content, pre-retrieves candidate recommendation content, embeds the survey questions into the knowledge reply content in a natural language structure, embeds the candidate recommendation content, generates a composite answer text, and outputs the composite answer text to the user. Feedback is received and intention is analyzed to obtain intention and preference. A user portrait is updated, a recommendation strategy is adjusted, a knowledge unit is actively recommended, and the recommendation content is used as a starting point for a new round of intention inquiry. The application realizes deep integration and continuous optimization of intention acquisition and knowledge recommendation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for user intent mining and proactive recommendation for a financial knowledge platform. Background Technology

[0002] The financial knowledge platform deeply integrates multi-dimensional user behavioral data, such as browsing paths, search keywords, holding information, and operation timestamps, to accurately identify users' potential complex needs and intentions, such as inquiring about financial products, learning professional knowledge, or making investment decisions. This requires going beyond simple behavioral records and conducting in-depth intent reasoning. Simultaneously, the platform constructs a structured financial knowledge graph, systematically organizing professional product information, rules and regulations, terminology, and their interrelationships to form a core knowledge foundation. Based on insights into user intent and the support of the knowledge graph, the system can accurately match users' current needs with relevant content resources in the graph in real time. For example, after identifying users with low-risk preferences, it proactively pushes introductions to money market funds or risk assessment guides. The entire process goes beyond simply matching keywords; it emphasizes understanding the deeper meaning and relevance of the content. Furthermore, the system continuously adjusts its intent mining model and recommendation strategy based on user feedback on recommended content, such as clicks, reading time, and satisfaction, achieving self-learning and optimization, thus making the recommendation service increasingly aligned with users' actual needs.

[0003] Existing technologies for obtaining user intent are generally based on existing dialogues and user habits, lacking mechanisms to proactively guide users to express their preferences in real-time interactions. This makes it difficult to seamlessly integrate intent exploration with knowledge services, resulting in insufficient immediacy, accuracy, and naturalness of user intent acquisition in professional scenarios such as finance. Summary of the Invention

[0004] To address the aforementioned problems, this invention aims to solve the issue of insufficient proactiveness in intent mining and recommendation described above. One objective of this invention is to provide a user intent mining and proactive recommendation method for a financial knowledge platform that addresses the problems described above.

[0005] The present invention provides a method for user intent mining and proactive recommendation for a financial knowledge platform, comprising the following steps: Step S1: Obtain the user's current interaction behavior on the financial knowledge platform, and determine the user's current business scenario tag and historical session context based on the current interaction behavior; Step S2: Select survey questions related to the current interaction behavior from a preset intent survey question library; wherein each survey question in the intent survey question library is pre-labeled with a knowledge dimension tag; Step S3: Obtain the knowledge response content corresponding to the user's current interaction behavior, and based on the current business scenario tags, historical conversation context and user profile, pre-retrieve the N financial knowledge units with the highest relevance from the financial knowledge graph as candidate recommended content; integrate the selected survey questions into the knowledge response content in an embedded natural language structure, and embed the candidate recommended content in the form of summary information or interactive controls to generate composite response text; Step S4: Output the composite response text to the user; Step S5: Receive user feedback on the composite response text, and perform intent parsing on the feedback to obtain the user's intent dimension and preference information; the feedback includes responses to embedded survey questions, explicit operations or implicit behaviors on candidate recommended content; Step S6: Update the user profile based on the intent dimension and preference information, and adjust the recommendation strategy in real time based on the user's feedback on the candidate recommended content; based on the updated user profile and real-time feedback data, proactively recommend financial knowledge units to the user, and embed the recommended content as the starting point for a new round of intent exploration into subsequent interactions.

[0006] Furthermore, step S2 involves selecting a survey question related to the current interactive behavior, specifically including: obtaining a financial knowledge graph, which includes knowledge point nodes, business scenario nodes, and their associated edges; and determining the first node in the financial knowledge graph corresponding to the topic involved in the current interactive behavior. For each survey question in the intent survey question base, obtain the second node corresponding to its knowledge dimension label in the financial knowledge graph, and calculate the graph path distance between the first node and the second node. Generate a structural relevance score based on the graph path distance. Calculate the semantic cosine similarity between the text vector of each survey question and the text vector of the topic involved in the current interaction behavior to generate a semantic relevance score. Weight the structural relevance score and the semantic relevance score to obtain a comprehensive relevance score. Select at least one survey question based on the comprehensive relevance score. Simultaneously, based on the co-occurrence frequency of knowledge point nodes in the financial knowledge graph, the business scenario association weight, and the user's historical preference decay factor, the real-time relevance score of each knowledge unit to the current topic is calculated, and the top N candidate recommended content are pre-selected.

[0007] The preferred technical solution is that, in step S3, the embedded natural language structure includes: embedding of causal adverbial clauses, embedding of conditional adverbial clauses, or embedding of alternative questions; the embedding method of the candidate recommended content includes: adding a related reading section at the end of the knowledge answer content, or inserting a clickable summary of the recommended content into the alternative question.

[0008] A preferred technical solution is that, after generating the composite response text in step S3, the method further includes: embedding an interactive selection control corresponding to the survey question into the composite response text, and adding an interactive summary control of the candidate recommended content; dynamically adjusting the option values ​​of the interactive selection control and the sorting of the candidate recommended content according to the user's historical preference weights.

[0009] The preferred technical solution is that the intent parsing of the feedback in step S5 specifically includes: if the feedback is a click on an interactive selection control, the clicked option value is directly mapped to an intent dimension; if the feedback is free text input, referential resolution, term matching, and sentence structure analysis are performed; if the feedback is implicit behavioral feedback, the preference confidence is updated based on the adoption operation of the knowledge response content; if the user clicks on candidate recommended content, the knowledge dimension tag of that content is directly used as a high-confidence intent dimension; if the user ignores a specific recommended content for more than a preset time, the preference weight of the corresponding knowledge dimension is reduced; if the user saves the recommended content, a further investigation question for that content is generated and stored in the dialogue state stack.

[0010] Furthermore, if the feedback is free text input, then referential resolution is performed, and the resolved text entities are matched with terms in the thesaurus. After matching, the process further includes: calculating a confidence score for a successful match; if the confidence score is higher than a first preset threshold, then the intent dimension corresponding to the successfully matched term is used as the parsing result; if the confidence score is between the first preset threshold and a second preset threshold, then a secondary confirmation text is generated and output to the user, the user's confirmation feedback is received, and the confidence score is updated based on the confirmation feedback; if the confidence score is lower than the second preset threshold, then the survey question is marked as pending a follow-up and is re-output to the user in subsequent interactions.

[0011] The preferred technical solution is that, before integrating the selected survey questions into the knowledge response content using an embedded natural language structure in step S3, the method further includes: obtaining the missing preference dimensions in the current user profile, decomposing the missing preference dimensions into multiple sub-questions, and storing the multiple sub-questions into the dialogue state stack in sequence. In subsequent user interactions, sub-questions are sequentially retrieved from the dialogue state stack and embedded into the knowledge response content of the corresponding interaction round. In each round, only one sub-question is embedded, and the timing of embedding the sub-question is determined according to the interaction depth of the current interaction round. When the interaction depth exceeds a preset threshold, the embedding operation is performed. When the user clicks on candidate recommended content, 3-5 in-depth sub-questions are generated based on the knowledge dimension of the content and added to the dialogue state stack.

[0012] A preferred technical solution is that, after receiving the user's feedback on the composite response text in step S5, the method further includes: maintaining the dialogue state stack of the current session and storing the currently output survey question and the corresponding follow-up marker into the dialogue state stack. The system detects the relevance of user feedback to the current survey question. If the user feedback is determined to be irrelevant to the current survey question, a topic transition is initiated. In response to the topic jump, new query content from user feedback is obtained, and corresponding knowledge answer content is generated and output to the user based on the new query content; After outputting the knowledge response content, the system checks whether there is a pending follow-up marker in the dialogue state stack. If there is a pending follow-up marker, the system outputs the incomplete survey questions to the user in a lightweight prompt manner. If the topic jump is triggered by clicking on the candidate recommended content, the system adds the in-depth survey questions for that content to the pending follow-up queue.

[0013] A preferred technical solution is that, when the sparsity of a user's historical profile exceeds a preset threshold, before step S2, the solution further includes: obtaining the user's job role tags and department tags; obtaining an initial set of survey questions from a preset job preference mapping table based on the job role tags; and calculating the expected information gain and business importance weight of each survey question in the initial set of survey questions. The survey questions in the initial survey question set are sorted according to the expected information gain and business importance weight, and the survey questions with higher ranking are selected as the survey questions in step S2; high-frequency knowledge units associated with job role tags are preloaded as initial candidate recommendation content.

[0014] A preferred technical solution is that receiving user feedback on the composite response text in step S5 includes: simultaneously providing the user with a selection option click channel and a voice input channel; when the user's voice input is received, performing voice-to-text processing on the voice input to obtain transcribed text; and parsing the transcribed text using the same intent parsing process as free text input to obtain a first intent parsing result. If the confidence score of the first intent parsing result is lower than a preset threshold, the audio features of the voice input are matched with the preset audio features of each option value of the interactive selection control, and a second intent parsing result is generated based on the audio feature matching result. The first intent parsing result and the second intent parsing result are fused to obtain the final intent dimension; if the voice input is for candidate recommended content, the confidence score is adjusted by additionally detecting the sentiment tendency value in the audio.

[0015] Compared with existing technologies, the user intent mining and proactive recommendation method for financial knowledge platforms of this invention has the following technical effects: 1. This application addresses a user intent mining and proactive recommendation method for a financial knowledge platform. It deeply integrates proactive surveys and proactive recommendations, constructing a collaborative mechanism where surveys are recommendations and recommendations are surveys, achieving seamless integration at the interaction level and mutual enhancement at the functional level. When generating composite response text, this invention not only embeds survey questions into the knowledge response content using an embedded natural language structure but also simultaneously embeds candidate recommendation content pre-retrieved based on user profiles. This allows survey questions and recommended content to be presented side-by-side or nested within the same interactive interface. User selection of survey questions directly triggers the expansion of corresponding recommended content, while clicks, favorites, or ignores of recommended content provide real-time feedback to intent analysis, forming a two-way driven closed loop. This integration mechanism brings significant synergistic effects. Survey results provide precise intent guidance for recommendations, making recommended content more aligned with the user's immediate needs in the current scenario and avoiding blind pushes. Recommended content, as bait for the survey, stimulates user interest by presenting valuable knowledge fragments, significantly increasing user willingness to participate in the survey and the depth of interaction. Compared to the linear model where intent collection and recommendation services are independent in existing technologies, this invention organically couples the two, achieving both understanding user preferences and meeting user needs in a single round of interaction. This significantly reduces the number of interaction rounds and improves the knowledge service efficiency and user experience of the financial knowledge platform.

[0016] 2. By embedding survey questions into the knowledge response content using an embedded natural language structure, a seamless integration of user intent acquisition and knowledge service is achieved. Compared to the abrupt interruptions of existing independent pop-up questionnaires, the survey wording and knowledge response of this invention are logically related semantically. Users can naturally respond to the survey while acquiring knowledge without switching their attention, significantly reducing the interaction burden. Simultaneously, this invention supports the synchronous embedding of candidate recommendation content as summary information or interactive controls into the composite response text, allowing intent exploration and knowledge recommendation to be presented in parallel on the same interface and in the same round of interaction. This forms a collaborative mechanism of survey as a service and recommendation as a survey, effectively improving user acceptance and interaction fluency in the financial knowledge platform scenario.

[0017] 3. This invention constructs a multi-dimensional intent parsing mechanism capable of processing user intent signals in various forms, including free text, clicks on options, and implicit behaviors. Through the resolution of pronouns, mapping of synonyms in the financial field, identification and response to rhetorical questions, and parsing of conditional clauses, this invention significantly improves the ability to understand natural language feedback. Simultaneously, this invention introduces a confidence-based hierarchical processing strategy, calibrating low-confidence parsing results through secondary confirmation or a follow-up mechanism, effectively preventing low-quality intent signals from polluting user profiles. Furthermore, this invention uses user actions such as clicking, saving, and ignoring recommended content as enhanced signals for intent feedback, enabling continuous iterative optimization of intent understanding as interaction deepens. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the user intent mining and proactive recommendation method for a financial knowledge platform provided in a specific embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0020] The following detailed description, with reference to the accompanying drawings and embodiments, illustrates the user intent mining and proactive recommendation method for a financial knowledge platform. This method proactively acquires user intent in a natural and integrated manner during normal knowledge service delivery, and deeply integrates intent acquisition with knowledge recommendation to form a continuously optimized interactive loop, such as... Figure 1 As shown.

[0021] First, step S1 is executed to obtain the user's current interactive behavior on the financial knowledge platform and determine the user's current business scenario label and historical session context based on the current interactive behavior. Specifically, when the user operates on the financial knowledge platform interface, the system collects various interactive data in real time, including but not limited to the menu items clicked by the user, the search keywords entered, the document titles viewed, the page dwell time, the mouse movement trajectory, and the scrolling speed. This data is reported to the backend service in real time through front-end tracking technology. The backend service performs comprehensive analysis on this data based on a preset business scenario classification model to identify the user's current business scenario. For example, when the user opens the "Corporate Loan Approval" module and starts browsing a company's financial statements, the system uses "Corporate Loan Approval" as the current business scenario label; when the user stays on the "Anti-Money Laundering Suspicious Transaction Report" interface and clicks the "Customer Identification" button, the system uses "Anti-Money Laundering Compliance Review" as the business scenario label. At the same time, the system extracts the user's historical interaction records in the current session from the session cache, including previously asked questions, the system's answers, and the user's feedback on the answers, such as clicks, copies, and favorites, forming a structured historical session context. This context information is stored in the session state stack as key-value pairs for later use.

[0022] Next, step S2 is executed, selecting survey questions related to the current interaction behavior from a pre-built intent survey question library. The intent survey question library is pre-built, where each survey question is labeled with a knowledge dimension tag, such as "credit risk-weighted asset calculation," "operational risk capital measurement," and "anti-money laundering customer identification requirements." To accurately select survey questions related to the current topic, this invention constructs a financial knowledge graph, which includes knowledge point nodes, business scenario nodes, and their associated edges. Knowledge point nodes correspond to specific financial knowledge units, such as "LPR," "internal rating method," and "asset management new regulations"; business scenario nodes correspond to specific business processes, such as "pre-loan investigation," "loan approval," "post-loan management," and "anti-money laundering verification"; and associated edges define the relationship types between nodes, such as "belonging relationship," "sequential order," "causal relationship," and "similar relationship."

[0023] When selecting survey questions, the system first determines the first node in the financial knowledge graph corresponding to the topic involved in the current interaction. For example, if a user is currently querying "LPR," the system locates the "Loan Prime Rate" node in the knowledge graph. Then, for each survey question in the intent survey question library, the system obtains the second node corresponding to its knowledge dimension label in the knowledge graph and calculates the graph path distance between the first and second nodes. Path distance refers to the number of edges traversed by the shortest path between two nodes; the shorter the path distance, the closer the two knowledge points are in terms of knowledge structure. Based on the graph path distance, the system generates a structural relevance score; the shorter the distance, the higher the structural relevance score.

[0024] Simultaneously, the system calculates the semantic cosine similarity between the text vectors of each survey question and the text vectors of the topics involved in the current interaction, generating a semantic relevance score. The text vectors can be generated using pre-trained language models, such as BERT or RoBERTa, to convert the text into a high-dimensional vector representation. For example, the text vector of the survey question "Regarding LPR, are you more concerned about its impact on loan interest rate pricing or the regulations on loan repricing cycles?" has a high cosine similarity to the text vector of the current topic "LPR".

[0025] The system weights and fuses structural relevance scores with semantic relevance scores to obtain a comprehensive relevance score. The weighting coefficients can be dynamically adjusted according to the application scenario; for example, in scenarios with strong business continuity, the weight of structural relevance can be appropriately increased. Finally, the system selects the survey question with the highest comprehensive relevance score as the target survey question. For example, when a user queries "LPR," the system might select the survey question, "Regarding LPR, are you more concerned about its impact on loan interest rate pricing or its regulations on loan repricing cycles?" rather than the general question, "How can I help you?" Building upon this foundation, the system also calculates the real-time relevance score of each knowledge unit to the current topic based on the co-occurrence frequency of knowledge point nodes in the financial knowledge graph, the business scenario association weight, and the user's historical preference decay factor, thus pre-selecting the top N candidate recommended content. The co-occurrence frequency reflects the probability of two knowledge points appearing simultaneously in historical interactions; the business scenario association weight reflects the relevance of a knowledge point to the current business scenario; and the user's historical preference decay factor gives higher weight to knowledge points recently adopted by users. Through multi-dimensional comprehensive calculations, the system can pre-select knowledge units highly relevant to the current topic as candidate recommended content.

[0026] Next, step S3 is executed to obtain the knowledge response content corresponding to the user's current interaction behavior, and the selected survey questions are integrated into the knowledge response content in an embedded natural language structure. At the same time, candidate recommendation content is embedded in the form of summary information or interactive controls to generate composite response text.

[0027] The knowledge response content is a direct answer retrieved from the knowledge base based on the user's current interaction behavior. For example, if a user queries "LPR", the system returns the definition of LPR and the latest quote: "LPR is the Loan Prime Rate. The latest quote in June 2024 is 3.45% for 1-year terms and 4.20% for terms of 5 years or more." Embedded natural language structures include three main forms. The first is the embedding of causal adverbial clauses, where the survey question is integrated into the knowledge response content as a causal adverbial clause. For example, the system generates the following compound response text: "LPR, or Loan Prime Rate, was quoted in June 2024 as 3.45% for 1-year loans and 4.20% for loans over 5 years. Since you have recently focused on corporate lending, if you need to understand the impact of LPR decline on interest income of existing loans, I can calculate the interest changes under different floating ratios for you." In this example, the survey question is naturally packaged into a conditional statement, allowing users to decide whether to further express their preferences while acquiring knowledge.

[0028] The second method is embedding conditional adverbial clauses, which involves incorporating the survey question as a conditional adverbial clause into the knowledge response content. For example: "LPR, or Loan Prime Rate, was quoted in June 2024 as 3.45% for 1-year loans and 4.20% for loans of 5 years or more. If you need to assess the impact of LPR decline on interest income of existing loans, I can calculate the interest changes under different floating ratios for you." The third method is embedding alternative questions, which involves transforming the survey questions into alternative questions and integrating them into the knowledge response content. For example: "LPR, or Loan Prime Rate, was quoted in June 2024 as 3.45% for 1-year loans and 4.20% for loans with a term of 5 years or more. Regarding LPR, are you more concerned about its pricing mechanism or the rules for adjusting interest rates on existing loans?" While embedding survey questions, the system also embeds candidate recommendations into the composite response text in various formats. One embedding method is to add a "Related Recommendations" section at the end of the knowledge response content, displaying the titles and summaries of the candidate recommendations, for example: Related recommendations: 1. Analysis of the impact of LPR reform on commercial banks' net interest margin; 2. Guide to setting the repricing cycle for floating rate loans; 3. Recent LPR trends and interpretation of macroeconomic policies.

[0029] Another embedding method is to insert clickable icons of recommended content next to the options in the multiple-choice questions. For example, when the system presents survey questions as multiple-choice questions, a small icon can be added next to each option, which users can click to expand the corresponding recommended content. This method deeply binds the survey with the recommendations; the process of a user selecting an option is simultaneously the process of obtaining relevant recommendations.

[0030] Furthermore, after generating the composite response text, the system embeds interactive selection controls corresponding to the survey questions and presents the options in a sorted order based on the user's historical preference weights. For example, if a user's historical preferences show a high level of interest in content related to "pricing mechanisms," the system will prioritize the "pricing mechanisms" option. Simultaneously, the system sorts the embedded candidate recommendations in descending order of their real-time relevance scores and adds actionable feedback buttons to each recommendation, including functions such as click to view, save, and ignore, allowing users to easily express their preferences for the recommended content.

[0031] Before integrating the survey questions into the knowledge response content using an embedded natural language structure, the system also executes a multi-round progressive survey strategy. Specifically, the system obtains the missing preference dimensions in the current user profile, breaks these missing dimensions down into multiple sub-questions, and stores these sub-questions sequentially into the dialogue state stack. For example, if the user profile lacks information on the "risk management preference" dimension, the system breaks it down into three sub-questions: "credit risk management attention," "market risk management attention," and "operational risk management attention," and stores them sequentially into the dialogue state stack.

[0032] In subsequent user interactions, the system sequentially retrieves sub-questions from the dialogue state stack and embeds them into the knowledge response content of the corresponding interaction round. Only one sub-question is embedded per round to avoid overwhelming the user with too many questions in a single interaction. The timing of sub-question embedding is determined by the interaction depth of the current round; embedding is performed when the interaction depth exceeds a preset threshold. For example, if a user has already interacted more than three times in the conversation, it indicates that the user has entered a "high engagement" state and is more receptive to additional information, at which point the system can embed sub-questions. Furthermore, when a user clicks on candidate recommended content, the system generates 3-5 in-depth sub-questions based on the knowledge dimension of that content and adds them to the dialogue state stack to further explore the user's deep preferences in that area.

[0033] Step S4 outputs the generated composite response text to the user. Before outputting, the system obtains the user's current interaction state on the interface, including mouse movement trajectory, page scroll position, dwell time, and input state. The system determines the user's focus area based on the mouse movement trajectory and page scroll position; for example, if the user is reading a section of content at the top of the page, the focus area is at that position. The dwell time determines the user's level of engagement with the current content; for example, if the user dwells on a section of content for more than 5 seconds, the level of engagement is high. The input state determines whether the user is editing; for example, if the user is typing in an input box, they are in editing mode.

[0034] When the user's attention is focused on the preset recommended display area, the level of engagement exceeds a preset threshold, and the user is not editing, the system will present the composite response text as a non-intrusive card in the sidebar or bottom floating area of ​​the interface. When the user is editing, the system delays output, recommending the text only after the user has completed the editing process, to avoid interrupting the user's workflow. For example, when a user is filling out a loan approval form, the system will wait for the user to complete the current input box before presenting the composite response text.

[0035] Step S5 receives user feedback on the composite response text and performs intent parsing on the feedback to obtain the user's intent dimensions and preference information. The types of feedback are diverse, including responses to embedded survey questions, explicit actions on candidate recommendations, or implicit behaviors. The system performs intent parsing separately for each type of feedback.

[0036] When a user clicks an interactive selection control, the system directly maps the clicked option value to an intent dimension. For example, if a user clicks "Pricing Mechanism," the system records "Loan Pricing Mechanism" as the intent dimension.

[0037] When a user inputs free text, the system performs a series of natural language processing operations. First, it checks for pronouns such as "first," "later," "this," and "that." If pronouns are present, the system retrieves the order of selections recorded in the current session's context stack and performs dereference resolution based on this order, resulting in the resolved text entity. For example, if the system previously provided two options, "pricing mechanism" and "repricing cycle," and the user inputs "first," the system resolves it to "pricing mechanism"; if the user inputs "later," the system resolves it to "repricing cycle."

[0038] The system then retrieves a thesaurus of financial terms and matches the resolved text entities with the terms in the thesaurus. The financial thesaurus contains common synonyms, abbreviations, and idioms used in the financial industry, such as "risk-weighted assets" and "RWA," "capital adequacy ratio" and "CAR," and "anti-money laundering" and "AML." When a user enters "I'm interested in risk-weighted assets," the system matches the standard term "credit risk-weighted assets" from the thesaurus and maps it to the corresponding intent dimension.

[0039] The system also detects the presence of rhetorical questions in the free text, such as "What's the difference?", "Which is better?", and "How to understand?". If a rhetorical question is detected, the system generates a comparative explanation of the current selection and outputs it to the user, then receives user feedback again. For example, if a user asks, "What's the difference between the pricing mechanism and the repricing cycle?", the system replies: "The pricing mechanism determines how interest rates are generated, while the repricing cycle determines when interest rates are adjusted. The former affects loan pricing levels, while the latter affects the frequency of interest rate changes. Based on these differences, which one do you care about more?" After the user provides feedback again, the system continues its analysis.

[0040] The system also detects the presence of conditional clauses in the free text, such as "If I were a retail customer, I would choose A" or "For corporate business, I would focus on B." If a conditional clause exists, the system extracts the conditional variable and the corresponding intent value, recording the conditional variable as a contextual attribute in the user profile. For example, if a user inputs "If I were a retail customer, I would focus on credit risk," the system extracts the conditional variable "Customer Type = Retail" and the intent value "Credit Risk," storing this conditional preference in the user profile. Subsequently, when the user handles retail customer-related business, the system will prioritize pushing credit risk management-related content.

[0041] When users exhibit implicit behaviors without explicit actions, the system also captures and analyzes these implicit behavioral feedbacks. Specifically, if a user adopts a knowledge-based response in a compound response text, such as copying, saving, or forwarding the content, the system obtains the knowledge dimension tag of that response and updates the user's confidence level in the corresponding preference information based on that knowledge dimension tag with a reverse weighting. For example, if a user copies a passage about the "LPR pricing mechanism," the system increases the confidence level of the preference corresponding to "LPR pricing mechanism."

[0042] If a user clicks on or saves recommended content, the system directly uses the knowledge dimension tag of that content as a high-confidence intent dimension. For example, if a user clicks on the recommended content "Analysis of the Impact of LPR Reform on Commercial Banks' Net Interest Margin," the system will record "Impact on Net Interest Margin" as a high-confidence intent dimension.

[0043] If a user ignores a specific candidate recommendation for more than a preset time, such as not interacting with the recommendation within 10 seconds, the system reduces the user preference weight of the knowledge dimension corresponding to that content to avoid repeatedly pushing content that the user is not interested in in subsequent recommendations.

[0044] After matching the resolved text entities with terms in the thesaurus, the system calculates a confidence score for successful matches. The confidence score can be calculated based on multiple factors, including semantic similarity, the determinism of the matching rules, and historical matching accuracy. If the confidence score is higher than a first preset threshold (e.g., 0.8), the system directly uses the intent dimension corresponding to the successfully matched term as the parsing result. If the confidence score is between the first and second preset thresholds (e.g., 0.5 to 0.8), the system generates a secondary confirmation text output to the user, such as "Do you mean credit risk-weighted assets?", receives confirmation feedback from the user, and updates the confidence score based on the confirmation feedback. If the confidence score is lower than the second preset threshold (e.g., 0.5), the system marks the survey question as pending a follow-up and re-outputs it to the user in subsequent interactions to avoid low-quality parsing results contaminating the user profile.

[0045] During the process of receiving user feedback, the system simultaneously provides users with both clickable selection channels and voice input channels, supporting multimodal interaction. When a user's voice input is received, the system first performs speech-to-text processing to obtain transcribed text. Then, it uses the same intent parsing process as for free text input to parse the transcribed text, obtaining a first intent parsing result. If the confidence score of the first intent parsing result is lower than a preset threshold, the system matches the audio features of the voice input with the preset audio features of each option value in the interactive selection control. For example, when a user says "first," the system extracts the audio features of that voice and matches them with the audio templates of each option value, generating a second intent parsing result based on the audio feature matching result. The system then merges the first and second intent parsing results to obtain the final intent dimension. If the voice input is for candidate recommendation content, the system additionally detects the emotional tendency values ​​in the audio, such as excitement, doubt, or indifference, to adjust the confidence score. For example, if a user says "this is good" in an excited tone, the system will assign a higher positive feedback weight.

[0046] After receiving user feedback, the system also performs topic transition state management. Specifically, the system maintains the dialogue state stack of the current session, storing the currently output survey question and its corresponding pending follow-up marker into the dialogue state stack. When user feedback is received, the system detects the relevance between the content of the user feedback and the current survey question, for example, by calculating the semantic similarity between the feedback text and the survey question. If the similarity is lower than a preset threshold, a topic transition is determined to occur.

[0047] In response to topic transitions, the system prioritizes responding to new user queries, retrieving the new queries from user feedback, generating corresponding knowledge-based answers, and outputting them to the user. After outputting the knowledge-based answers, the system checks if there are any pending follow-up markers in the dialogue state stack. If so, it outputs the incomplete survey questions to the user in a lightweight prompt. For example, the system could output, "Regarding the impact of the LPR on loan interest rates mentioned earlier, would you like to learn more?" If the topic transition was triggered by the user clicking on recommended content, the system will also add in-depth survey questions related to that content to the pending follow-up queue, ensuring that no new points of interest for the user are overlooked.

[0048] Step S6 updates the user profile based on intent dimensions and preference information, and adjusts the recommendation strategy in real time based on user feedback to candidate recommended content. The user profile is stored in vector form, with each preference dimension corresponding to a weight value. The weight values ​​are dynamically updated based on the user's historical interactions and feedback. When a user shows positive feedback to a certain type of content, the weight of the corresponding dimension increases; when a user shows negative feedback or ignores the content, the weight of the corresponding dimension decreases.

[0049] Based on updated user profiles and real-time feedback data, the system retrieves new financial knowledge units from the financial knowledge base and proactively recommends them to users. During the recommendation process, the system comprehensively considers factors such as preference weights in the user profile, business scenario tags, job role tags, and content timeliness to rank candidate knowledge units and selects the top-ranked units for recommendation.

[0050] More importantly, the system uses this recommended content as the starting point for a new round of intent exploration, embedding it into subsequent interactions. For example, when a user clicks on the recommended content "Analysis of the Impact of LPR Reform on Commercial Banks' Net Interest Margin," the system not only records the user's interest in "impact on net interest margin," but also generates 3-5 in-depth sub-questions based on the knowledge dimension of this content and adds them to the dialogue state stack, such as "Are you more concerned about the impact of LPR reform on the yield of interest-earning assets or on the cost rate of interest-bearing liabilities?" In subsequent interactions, the system will opportunely embed these in-depth sub-questions into the knowledge-based responses, further exploring the user's deep preferences in this area. In this way, a continuous interactive closed loop is formed: "intent acquisition → proactive recommendation → feedback collection → profile update → re-recommendation → re-exploration."

[0051] For new users or users with sparse historical profiles, this invention also provides a cold start optimization strategy. When the sparsity of a user's historical profile exceeds a preset threshold, before step S2, the system first obtains the user's job role tag and department tag. The job role tag can be obtained from the user's job information when logging in, such as "Corporate Client Manager," "Credit Approval Officer," "Compliance Manager," "Financial Advisor," etc.; the department tag includes "Corporate Finance Department," "Risk Management Department," "Compliance Department," "Private Banking Department," etc.

[0052] Based on job role tags, the system retrieves an initial set of survey questions from a pre-defined job preference mapping table. This table, built upon expert experience and historical data, reflects the knowledge dimensions most frequently addressed by different roles in their respective business operations. For example, the initial survey question set for a corporate client manager might include questions such as, "Are you more focused on corporate credit risk assessment or credit limit calculation?" and "Which industry do you most frequently deal with: manufacturing, trading, or services?" Similarly, the initial survey question set for a compliance manager might include questions such as, "Are you more focused on anti-money laundering regulatory requirements or internal control compliance construction?" and "Which type of regulatory document do you most frequently handle: notices from the State Financial Regulatory Commission or guidelines from the People's Bank of China?"

[0053] The system calculates the expected information gain and business importance weight of each survey question in the initial survey question set. The expected information gain reflects the incremental information in the user profile that the question can bring, i.e., how much the entropy of the user profile can be reduced through the question; the business importance weight is pre-set by business experts and reflects the importance of the question to business operations. The system sorts the survey questions in the initial survey question set according to the expected information gain and business importance weight, prioritizing the top-ranked survey questions as the survey questions selected in step S2.

[0054] Simultaneously, the system preloads high-frequency knowledge units associated with job role tags as initial candidate recommendations. For example, initial candidate recommendations for corporate account managers include "Management Measures for Working Capital Loans," "Key Points of Corporate Financial Statement Analysis," and "Guidelines for Credit Approval Processes"; initial candidate recommendations for compliance managers include "Guidelines for Anti-Money Laundering Customer Identification," "Management Measures for Reporting Large and Suspicious Transactions by Financial Institutions," and "Compliance Risk Self-Assessment Checklist." Through this job-driven cold start strategy, the system can provide relatively accurate surveys and recommendations for new users in the early stages, effectively alleviating the cold start problem.

[0055] In summary, this invention achieves a complete closed loop by deeply integrating proactive surveys and proactive recommendations. This involves acquiring user intent through natural interaction, optimizing recommendations in real time based on that intent, and using recommendation feedback as the starting point for a new round of intent exploration. In the context of a financial knowledge platform, this method can significantly improve the accuracy of knowledge services and the naturalness of interaction, effectively reducing users' learning costs and operational burden.

[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the article or apparatus that includes that element.

[0057] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. The present invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for user intent mining and proactive recommendation for a financial knowledge platform, characterized in that, Includes the following steps: Step S1: Obtain the user's current interaction behavior on the financial knowledge platform, and determine the user's current business scenario tag and historical session context based on the current interaction behavior; Step S2: Select survey questions related to the current interaction behavior from a preset intent survey question library; wherein each survey question in the intent survey question library is pre-labeled with a knowledge dimension tag; Step S3: Obtain the knowledge response content corresponding to the user's current interaction behavior, and based on the current business scenario tags, historical conversation context and user profile, pre-retrieve the N financial knowledge units with the highest relevance from the financial knowledge graph as candidate recommended content; integrate the selected survey questions into the knowledge response content in an embedded natural language structure, and embed the candidate recommended content in the form of summary information or interactive controls to generate composite response text; Step S4: Output the composite response text to the user; Step S5: Receive user feedback on the composite response text, and perform intent parsing on the feedback to obtain the user's intent dimension and preference information; the feedback includes responses to embedded survey questions, explicit operations or implicit behaviors on candidate recommended content; Step S6: Update the user profile based on the intent dimension and preference information, and adjust the recommendation strategy in real time based on the user's feedback on the candidate recommended content; based on the updated user profile and real-time feedback data, proactively recommend financial knowledge units to the user, and embed the recommended content as the starting point for a new round of intent exploration into subsequent interactions.

2. The user intent mining and proactive recommendation method for a financial knowledge platform as described in claim 1, characterized in that: Step S2 involves selecting survey questions associated with the current interaction behavior, specifically including: Obtain a financial knowledge graph, which includes knowledge point nodes, business scenario nodes, and their associated edges; determine the first node in the financial knowledge graph corresponding to the topic involved in the current interaction behavior; For each survey question in the intent survey question base, obtain the second node corresponding to its knowledge dimension label in the financial knowledge graph, and calculate the graph path distance between the first node and the second node. Generate a structural relevance score based on the graph path distance. Calculate the semantic cosine similarity between the text vector of each survey question and the text vector of the topic involved in the current interaction behavior to generate a semantic relevance score. Weight the structural relevance score and the semantic relevance score to obtain a comprehensive relevance score. Select at least one survey question based on the comprehensive relevance score. Simultaneously, based on the co-occurrence frequency of knowledge point nodes in the financial knowledge graph, the business scenario association weight, and the user's historical preference decay factor, the real-time relevance score of each knowledge unit to the current topic is calculated, and the top N candidate recommended content are pre-selected.

3. The user intent mining and proactive recommendation method for a financial knowledge platform as described in claim 1, characterized in that: In step S3, the embedded natural language structure includes: causal adverbial clause embedding, conditional adverbial clause embedding, or alternative question embedding; the embedding method of the candidate recommendation content includes: adding a related reading section at the end of the knowledge answer content, or inserting a clickable summary of the recommendation content in the alternative question.

4. The user intent mining and proactive recommendation method for a financial knowledge platform as described in claim 1, characterized in that: After generating the composite response text in step S3, the method further includes: An interactive selection control corresponding to the survey question is embedded in the composite response text, and an interactive summary control of the candidate recommended content is added; the option values ​​of the interactive selection control and the sorting of the candidate recommended content are dynamically adjusted according to the user's historical preference weights.

5. The user intent mining and proactive recommendation method for a financial knowledge platform as described in claim 4, characterized in that: The intent parsing of the feedback in step S5 specifically includes: If the feedback is a click on an interactive selection control, the clicked option value is directly mapped to the intent dimension; if the feedback is free text input, referential resolution, terminology matching, and sentence structure analysis are performed; if the feedback is implicit behavioral feedback, the preference confidence is updated based on the adoption operation of the knowledge response content; if the user clicks on candidate recommended content, the knowledge dimension tag of that content is directly used as a high-confidence intent dimension; if the user ignores a specific recommended content for more than a preset time, the preference weight of the corresponding knowledge dimension is reduced; if the user saves the recommended content, a further investigation question for that content is generated and stored in the dialogue state stack.

6. The user intent mining and proactive recommendation method for a financial knowledge platform as described in claim 5, characterized in that: If the feedback is free text input, then referential resolution is performed, and the resolved text entities are matched with terms in the thesaurus; after matching, the process also includes: Calculate the confidence score for a successful match; If the confidence score is higher than the first preset threshold, the intent dimension corresponding to the successfully matched term will be used as the parsing result. If the confidence score is between a first preset threshold and a second preset threshold, a secondary confirmation text is generated and output to the user, the user's confirmation feedback is received, and the confidence score is updated according to the confirmation feedback. If the confidence score is lower than the second preset threshold, the survey question will be marked as pending a follow-up and will be re-output to the user in subsequent interactions.

7. The user intent mining and proactive recommendation method for a financial knowledge platform as described in claim 5, characterized in that: Before integrating the selected survey questions into the knowledge response content using an embedded natural language structure in step S3, the method further includes: obtaining the missing preference dimensions in the current user profile, breaking down the missing preference dimensions into multiple sub-questions, and storing the multiple sub-questions into the dialogue state stack in sequence. In subsequent user interactions, sub-questions are sequentially retrieved from the dialogue state stack and embedded into the knowledge response content of the corresponding interaction round. In each round, only one sub-question is embedded, and the timing of embedding the sub-question is determined according to the interaction depth of the current interaction round. When the interaction depth exceeds a preset threshold, the embedding operation is performed. When the user clicks on candidate recommended content, 3-5 in-depth sub-questions are generated based on the knowledge dimension of the content and added to the dialogue state stack.

8. The user intent mining and proactive recommendation method for a financial knowledge platform as described in claim 7, characterized in that: After receiving the user's feedback on the composite response text in step S5, the method further includes: maintaining the dialogue state stack of the current session and storing the currently output survey question and the corresponding follow-up marker into the dialogue state stack. The system detects the relevance of user feedback to the current survey question. If the user feedback is determined to be irrelevant to the current survey question, a topic transition is initiated. In response to the topic jump, new query content from user feedback is obtained, and corresponding knowledge answer content is generated and output to the user based on the new query content; After outputting the knowledge response content, the system checks whether there is a pending follow-up marker in the dialogue state stack. If there is a pending follow-up marker, the system outputs the incomplete survey questions to the user in a lightweight prompt manner. If the topic jump is triggered by clicking on the candidate recommended content, the system adds the in-depth survey questions for that content to the pending follow-up queue.

9. The user intent mining and proactive recommendation method for a financial knowledge platform as described in claim 1, characterized in that: When the sparsity of a user's historical profile exceeds a preset threshold, before step S2, the method further includes: obtaining the user's job role tag and department tag; obtaining an initial survey question set from a preset job preference mapping table based on the job role tag; and calculating the expected information gain and business importance weight of each survey question in the initial survey question set. The survey questions in the initial survey question set are sorted according to the expected information gain and business importance weight, and the survey questions with higher ranking are selected as the survey questions in step S2; high-frequency knowledge units associated with job role tags are preloaded as initial candidate recommendation content.

10. The user intent mining and proactive recommendation method for a financial knowledge platform as described in claim 4, characterized in that: The step S5 of receiving user feedback on the composite response text includes: simultaneously providing the user with a selection option click channel and a voice input channel; when the user's voice input is received, performing voice-to-text processing on the voice input to obtain transcribed text; and parsing the transcribed text using the same intent parsing process as free text input to obtain a first intent parsing result. If the confidence score of the first intent parsing result is lower than a preset threshold, the audio features of the voice input are matched with the preset audio features of each option value of the interactive selection control, and a second intent parsing result is generated based on the audio feature matching result. The first intent parsing result and the second intent parsing result are fused to obtain the final intent dimension; if the voice input is for candidate recommended content, the confidence score is adjusted by additionally detecting the sentiment tendency value in the audio.

Images