Method and system for realizing multi-round questions and answers under security consultation based on large model

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a method and a system for realizing multi-round questions and answers under security consultation based on a large model, and the method comprises the steps: receiving security consultation data of a target user to generate a dynamic user portrait of the target user, analyzing a first-level intention label and a first-level intention confidence coefficient of the security consultation data, and generating a second-level intention confidence coefficient; determining a second-level intention label of the target user; outputting structured query parameters of the target user; outputting dual-channel data of the security consultation data, and identifying a keyword blacklist, implicit commitments and undisclosed risks of the dual-channel data to filter the dual-channel data to obtain filtered multi-modal knowledge fragments; and generating a multi-dimensional answer of the target user to generate a multi-strategy model and a multi-strategy adjustment parameter and a user portrait update parameter of the dynamic user portrait, and executing multi-round questions and answers of the target user based on the multi-strategy adjustment parameter and the user portrait update parameter. According to the invention, the suggestion accuracy and personalization under security consultation can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a method and system for implementing multi-round question and answer under securities consulting based on large models, belonging to the field of artificial intelligence. BACKGROUND

[0002] Multi-round question and answer under securities consulting refers to the process of multiple interactive communication between investors and securities analysts, investment consultants or other professionals. It allows investors to gradually understand complex financial concepts and market dynamics, while analysts can provide customized advice based on the specific needs of investors. This interactive process helps to build trust and long-term customer relationships.

[0003] Traditional multi-round question and answer methods for securities consulting usually rely on human customer service or analysts to communicate one-on-one with investors. Investors ask questions, and customer service / analysts answer based on their professional knowledge and experience. If the question is complex or requires more information, further questioning will be conducted, and the investor will answer again, and so on. The main disadvantages of this method are low efficiency, high cost, and limited by the breadth and depth of a single consultant's knowledge, making it difficult to provide highly standardized and structured information. At the same time, there is a lack of systematic analysis of user historical behavior and intent, resulting in inaccurate and personalized recommendations. SUMMARY

[0004] The present application provides a method and system for implementing multi-round question and answer under securities consulting based on large models, which aims to improve the accuracy and personalization of recommendations under securities consulting.

[0005] To achieve the above purpose, the present application provides a method for implementing multi-round question and answer under securities consulting based on large models, which includes: Optionally, the generation of the dynamic user portrait of the target user includes: According to the explicit data and implicit data of the target user, analyze the basic risk and sensitivity of the target user; Based on the sensitivity, dynamically adjust the basic risk to obtain the risk level label of the target user; Analyze the financial term understanding, problem complexity and historical interaction depth of the explicit data and implicit data to determine the knowledge level index of the target user; According to the explicit data and implicit data, analyze the semantic confusion degree and intent confidence difference of the target user; And combined with the semantic confusion degree and intent confidence difference, determine the intent ambiguity score of the target user; Combine the risk level label, the knowledge level index and the intent ambiguity score to generate the dynamic user portrait of the target user.

[0006] Optionally, the step of analyzing the semantic perplexity and intent confidence difference of the target user based on the explicit and implicit data includes: Based on the explicit and implicit data, the cosine similarity score and sentence coherence weight of the target user are determined. Based on the cosine similarity score and the sentence coherence weight, the semantic perplexity of the target user is calculated using the following formula: ; in, Indicates the semantic perplexity of the target user. This indicates the total number of statements made by the target user in the current conversation round. The context cosine similarity score represents the target user's context. Indicates the weight of statement coherence. This indicates the current conversation round number of the target user. Sentence text; Based on the explicit and implicit data, analyze the probability of the target user's intent; Based on the intent probability, the intent confidence difference of the target user is determined.

[0007] Optionally, determining the intent confidence difference of the target user based on the intent probability includes: Based on the intent probability, the intent confidence difference of the target user is calculated using the following formula: ; in, The expression of intent has poor confidence. This indicates the number of candidate intents for the target user. Indicates the target user number The first candidate intention in a round of dialogue Indicates the target user number The second candidate intention in a round of dialogue A probability function representing the candidate's intent. This represents the function that takes the maximum value.

[0008] Optionally, the analysis of the primary intent label and primary intent confidence level of the securities consulting data includes: The securities consulting data is preprocessed to obtain processed securities consulting data; Mark the key entities involved in processing securities advisory data; Based on the key entities, the trained intent classification model is used to output the primary intent label of the securities consultation data. analyze a term enhancement confidence and an entity consistency of the primary intent label to determine a primary intent confidence of the primary intent label.

[0009] Optionally, the determining the secondary intent label of the target user according to the primary intent label and the primary intent confidence comprises: when the primary intent confidence meets a preset primary intent confidence threshold, constructing a secondary intent knowledge graph of the primary intent label; analyzing an intent multi-dimensional feature of the target user according to the securities consulting data of the target user; based on the intent multi-dimensional feature, analyzing a secondary intent score of the secondary intent knowledge graph to mark a secondary candidate intent of the target user; analyzing an intent risk coefficient of the secondary candidate intent; determining the secondary intent label of the target user based on the intent risk coefficient.

[0010] Optionally, the outputting the double-channel data of the securities consulting data according to the structured query parameter and the dynamic user portrait comprises: determining a query API of the securities consulting data according to the structured query parameter; determining structured channel data of the securities consulting data through the query API; labeling an outputtable node of a pre-constructed financial graph by using a risk level label and a knowledge level index of the dynamic user portrait; traversing the outputtable node according to the structured query parameter to obtain unstructured channel data; combining the structured channel data and the unstructured channel data to determine the double-channel data of the securities consulting data.

[0011] Optionally, the identifying the keyword blacklist, the implied commitment and the undisclosed risk of the double-channel data comprises: constructing a multi-level blacklist library of the double-channel data; determining potential extension words and potential transformation forms of the multi-level blacklist library; determining a keyword blacklist of the double-channel data according to the multi-level blacklist library, the potential extension words and the potential transformation forms; determining a commitment semantic pattern of the double-channel data to retrieve an implied commitment of the double-channel data; detecting a risk word of the double-channel data; determining a specific risk point of the double-channel data; Determine the undisclosed risk of the double-channel data in combination with the risk word and the specific risk point.

[0012] Optionally, the generating the multi-dimensional answer of the target user according to the filtered multi-modal knowledge fragment and the dynamic user portrait comprises: Generating a generation mode of the filtered multi-modal knowledge fragment according to the knowledge level index of the dynamic user portrait, wherein the generation mode comprises a professional mode and an explanation mode; When the intention ambiguity score of the dynamic user portrait is greater than a preset intention ambiguity threshold, determining a counter-question label of the target user; Generating the multi-dimensional answer of the target user based on the generation mode and the counter-question label.

[0013] To solve the above problems, the application further provides a multi-round question and answer system based on a large model for securities consulting, which comprises: A user portrait construction module is configured to receive securities consulting data of a target user, wherein the securities consulting data comprises explicit data and implicit data, so as to generate a dynamic user portrait of the target user, wherein the dynamic user portrait comprises a risk level label, a knowledge level index and an intention ambiguity score; A user intention analysis module is configured to analyze a first-level intention label and a first-level intention confidence of the securities consulting data, and determine a second-level intention label of the target user according to the first-level intention label and the first-level intention confidence; A query parameter construction module is configured to track multi-round historical dialogue records of the target user, and output a structured query parameter of the target user according to the first-level intention label, the second-level intention label and the multi-round historical dialogue records; A double-channel data query module is configured to output double-channel data of the securities consulting data according to the structured query parameter and the dynamic user portrait, wherein the double-channel data comprises structured channel data and unstructured channel data, identify a keyword blacklist, an implied commitment and an undisclosed risk of the double-channel data, so as to filter the double-channel data and obtain a filtered multi-modal knowledge fragment; An intelligent multi-round question and answer module is configured to generate a multi-dimensional answer of the target user according to the filtered multi-modal knowledge fragment and the dynamic user portrait, and obtain explicit feedback and implicit feedback of the multi-dimensional answer, so as to generate multi-strategy adjustment parameters and user portrait update parameters of the multi-strategy model and the dynamic user portrait, and perform multi-round question and answer of the target user based on the multi-strategy adjustment parameters and the user portrait update parameters.

[0014] Firstly, the introduction of dynamic user portrait greatly improves the personalization and safety of consultation. The precise characterization of risk level label, knowledge level index and intention ambiguity score enables the system to dynamically adjust the consultation strategy according to the user's risk tolerance, professional knowledge background and current intention clarity. For users with high risk level and low knowledge level, the system can automatically filter high-risk product information and use more popular and understandable language for explanation, effectively reducing the potential loss caused by user misoperation or understanding deviation, enhancing the safety of consultation service. For users with ambiguous intention, the system can intervene in clarification in time to avoid providing information deviating from the user's real needs, improving the communication efficiency. Secondly, multi-level intention recognition and structured query parameter generation significantly improve the accuracy and efficiency of information retrieval. By accurately positioning the secondary intention through the first-level intention label and confidence, and generating structured query parameters combined with multi-round historical dialogue records, the system can quickly and accurately filter out the most relevant information from massive data, avoiding the low-efficiency search caused by information overload or inaccurate keyword matching in traditional consultation, so that users can obtain the required answers more quickly, improving the user experience. Thirdly, the dual-channel data fusion and intelligent filtering mechanism ensures the comprehensiveness and reliability of the consultation content. The combination of structured data channel and unstructured knowledge channel ensures that the answer has accurate data support and rich background knowledge. More importantly, the system can actively identify and filter keyword blacklists, implicit promises and undisclosed risks, effectively avoiding information misdirection and potential legal compliance risks, providing a safe and reliable information environment for users. Finally, the synergistic optimization of multi-strategy model and user portrait realizes the continuous improvement and intelligent evolution of consultation service. By obtaining explicit and implicit feedback of users on multi-dimensional answers, the system dynamically adjusts the multi-strategy model parameters and updates the user portrait, so that the system can continuously learn the user's preferences and demand changes, and optimize the answer strategy. Therefore, the present application can improve the accuracy and personalization of securities consultation. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A flowchart of a method for realizing multi-round question and answer under securities consultation based on a large model is provided for an embodiment of the present application. Figure 2 A user portrait update schematic diagram for realizing the method for realizing multi-round question and answer under securities consultation based on a large model is provided for an embodiment of the present application. Figure 3 A module schematic diagram for realizing the method for realizing multi-round question and answer under securities consultation based on a large model is provided for an embodiment of the present application.

[0016] The object realization, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0017] It is to be understood that the specific embodiments described herein are merely illustrative of the present application and do not limit the scope of the application.

[0018] The embodiment of the present application provides a method for realizing multi-round question and answer under security consulting based on a large model. The execution subject of the method for realizing multi-round question and answer under security consulting based on a large model includes but is not limited to at least one of electronic devices such as a server, a terminal and the like which can be configured to execute the method provided by the embodiment of the present application. In other words, the method for realizing multi-round question and answer under security consulting based on a large model can be executed by software or hardware installed in a terminal device or a server device. The server includes but is not limited to a single server, a server cluster, a cloud server or a cloud server cluster and the like.

[0019] Embodiment 1 Referring to Figure 1 FIG. 1 is a flowchart of a method for realizing multi-round question and answer under security consulting based on a large model provided by an embodiment of the present application. In the embodiment, the method for realizing multi-round question and answer under security consulting based on a large model includes: S1, receiving security consulting data of a target user, wherein the security consulting data includes explicit data and implicit data to generate a dynamic user portrait of the target user, wherein the dynamic user portrait includes a risk level label, a knowledge level index and an intention ambiguity score.

[0020] It should be explained that the explicit data includes current consulting question text, historical question and answer records, risk preference actively submitted by the user, and the implicit data includes semantic tendency in the dialogue (such as frequent mentioning of the word "high risk"), interactive behavior (such as response delay time, question modification frequency.

[0021] The dynamic user portrait of the target user generated by the present application can provide a basis for later intention analysis.

[0022] In detail, the generation of the dynamic user portrait of the target user includes: analyzing the basic risk and sensitivity of the target user according to the explicit data and the implicit data of the target user; dynamically adjusting the basic risk based on the sensitivity to obtain the risk level label of the target user; analyzing the financial term understanding degree, the question complexity and the historical interaction depth of the explicit data and the implicit data to determine the knowledge level index of the target user; analyzing the semantic confusion degree and the intention confidence difference of the target user according to the explicit data and the implicit data; and combining the semantic confusion degree and the intention confidence difference, determining the intention ambiguity score of the target user; The dynamic user portrait of the target user is generated in combination with the risk level label, the knowledge level index, and the intention ambiguity score.

[0023] The base risk refers to the inherent and relatively static risk degree of the target user preliminarily evaluated based on the explicit data (such as public investment records, risk tolerance questionnaire answers, asset allocation, etc.) and part of the implicit data (such as the type of consulting questions, the investment varieties concerned, etc.) of the target user, the sensitivity refers to the reaction degree of the target user to specific risk factors or market changes, emotional fluctuations, and the tolerance to changes in personal financial status, the risk level label refers to a comprehensive and dynamic evaluation result of the current overall risk tolerance and risk preference of the user, which is usually presented in the form of a label (such as "low risk", "medium risk", "high risk", "extremely high risk", etc.), the knowledge level index refers to a quantitative indicator of the understanding degree of the target user to the financial knowledge and the stock market, the semantic ambiguity degree refers to the degree of clarity, logical coherence, and whether there is ambiguity or contradiction in the language expression of the target user when expressing the consulting intention, the intention confidence difference refers to the difference between the confidence of the model to the most possible intention of the user and the confidence to other secondary intentions in the intention recognition process, the intention ambiguity score refers to a quantitative indicator of the clarity of the current consulting intention of the user comprehensively evaluated in combination with the semantic ambiguity degree and the intention confidence difference, and the dynamic user portrait refers to a comprehensive set of user characteristics that will be updated over time (especially after each consulting interaction).

[0024] Further, the analysis of the semantic ambiguity degree and the intention confidence difference of the target user according to the explicit data and the implicit data comprises: determining the cosine similarity score and the sentence coherence weight of the target user according to the explicit data and the implicit data; calculating the semantic ambiguity degree of the target user based on the cosine similarity score and the sentence coherence weight by using the following formula: ; wherein, the semantic ambiguity degree of the target user is represented by, the total number of sentences of the target user in the current dialogue round is represented by, the context cosine similarity score of the target user is represented by, the sentence coherence weight is represented by, the text of the i-th sentence of the target user in the current dialogue round is represented by, analyzing the intention probability of the target user according to the explicit data and the implicit data; and determining the intention confidence difference of the target user based on the intention probability. ​

[0025] The cosine similarity score refers to quantifying semantic relevance of the current statement of the target user and the context historical statement, reflecting the coherence of the dialogue, the statement coherence weight refers to quantifying semantic relevance of the current statement of the target user and the context historical statement, reflecting the coherence of the dialogue, and the intent probability refers to quantifying the possibility distribution of each intent in the current dialogue turn of the target user.

[0026] Further, the intent confidence difference of the target user is determined based on the intent probability, comprising: The intent confidence difference of the target user is calculated based on the intent probability by using the following formula: ; Wherein, The intent confidence difference is represented by, The number of candidate intents of the target user is represented by, The first candidate intent in the first dialogue turn of the target user is represented by, The second candidate intent in the first dialogue turn of the target user is represented by, The probability function of the candidate intent is represented by, The maximum value function is represented by. Wherein, the first candidate intent refers to the intent with the highest probability in the current dialogue turn, the second candidate intent refers to the intent with the second highest probability in the current dialogue turn, the probability function refers to the intent probability distribution function output by the machine learning model, and the maximum value function refers to a function for extracting the maximum value in the probability distribution.

[0027] Wherein, the first candidate intent refers to the intent with the highest probability in the current dialogue turn, the second candidate intent refers to the intent with the second highest probability in the current dialogue turn, the probability function refers to the intent probability distribution function output by the machine learning model, and the maximum value function refers to a function for extracting the maximum value in the probability distribution.

[0028] S2, analyze the first intent label and the first intent confidence of the securities consulting data, and determine the second intent label of the target user according to the first intent label and the first intent confidence.

[0029] The analysis of the first intent label and the first intent confidence of the securities consulting data can generate multi-level intent recognition and structured query parameters, which significantly improves the accuracy and efficiency of information retrieval.

[0030] In detail, the analysis of the first intent label and the first intent confidence of the securities consulting data comprises: The securities consulting data is preprocessed to obtain processed securities consulting data; The key entities of the processed securities consulting data are labeled; The first intent label of the securities consulting data is output by using a trained intent classification model according to the key entities; ​analyze a term enhancement confidence and an entity consistency of the primary intent label to determine a primary intent confidence of the primary intent label.

[0031] The processing of the securities consultation data refers to a series of standardization and cleaning operations on the originally received securities consultation data, the key entity refers to specific information units in the securities consultation data that are crucial to understanding user intent and subsequent information retrieval, such as securities-related entities, such as specific stock codes (such as "600036"), stock names (such as "China Merchants Bank"), fund codes (such as "161725"), fund names, index names (such as "Shanghai Composite Index"), industry names (such as "new energy"), time-related entities, such as the time period (such as "the last month", "the second quarter of 2025"), specific dates (such as "yesterday", "July 1st") and the like, the intent classification model refers to a pre-trained artificial intelligence model, which functions to analyze and output the main intent category expressed by the input text (here, the processed securities consultation data, combined with the identified key entity information as auxiliary input), the primary intent label refers to the highest level of intent classification result of the user's securities consultation data output by the intent classification model, the term enhancement confidence refers to the confidence brought by the relevance of specific "terms" in the text to the intent label when calculating the confidence of the primary intent label, the entity consistency refers to the logical matching degree between the "key entity" identified in the processed securities consultation data and the currently predicted "primary intent label", and the primary intent confidence refers to the confidence score of the accuracy of the "primary intent label".

[0032] Optionally, the key entity of the processed securities consultation data is identified according to pre-defined rules and patterns. For example, a rule can be set: "strings starting with '600' followed by 5 digits" can be A-share stock codes; "strings starting with '300' followed by 4 digits" can be GEM stock codes.

[0033] According to the primary intent label and the primary intent confidence, the secondary intent label of the target user can be determined more accurately to determine the secondary intent label of the securities consultation data and improve the overall performance of the system.

[0034] In detail, the determination of the secondary intent label of the target user according to the primary intent label and the primary intent confidence comprises: When the primary intent confidence meets a pre-set primary intent confidence threshold, a secondary intent knowledge graph of the primary intent label is constructed; According to the securities consultation data of the target user, the intent multi-dimensional features of the target user are analyzed; Based on the multi-dimensional intention feature, a secondary intention score of the secondary intention knowledge graph is analyzed to label a secondary candidate intention of the target user; An intention risk coefficient of the secondary candidate intention is analyzed; Based on the intention risk coefficient, a secondary intention label of the target user is determined.

[0035] The primary intention confidence threshold refers to a pre-set numerical limit for determining whether the confidence of the primary intention label is high enough. The secondary intention knowledge graph refers to a structured knowledge representation constructed around a specific primary intention label. It is not just a simple classification list, but a network-like structure that organizes related secondary intentions, key entities (such as stocks, funds, industry terms), concepts, their relationships (such as inclusion, association, and prerequisite conditions), and related background knowledge, rules or constraints with the primary intention as the core. For example, under the primary intention label "market analysis", the secondary intention knowledge graph connects "market trend query", "industry sector analysis" and other secondary intentions, and associates them with "Shanghai Composite Index", "semiconductor industry", "technical indicators" and other entities and concepts. The multi-dimensional intention feature refers to a series of features extracted from the target user's securities consulting data that can reflect the deep-level information of their intentions. The secondary intention score is a quantitative score calculated by the system based on the matching degree and association strength of the multi-dimensional intention features and each secondary intention in the secondary intention knowledge graph after analyzing the multi-dimensional intention features. The secondary candidate intention refers to those secondary intentions that meet certain criteria (e.g., higher than a pre-set candidate threshold) after calculating the secondary intention score. The intention risk coefficient is a numerical value assigned to each secondary candidate intention to measure its potential risk or sensitivity. The secondary intention label refers to the most specific and accurate secondary intention that the system finally confirms and labels as the target user's current securities consulting data.

[0036] Optionally, the analysis of the intention risk coefficient of the secondary candidate intention can be determined by analyzing the user adaptation risk and content sensitivity of the target user through the risk level label of the dynamic user portrait.

[0037] S3, track the multi-round historical dialogue record of the target user, and output the structured query parameters of the target user according to the primary intention label, the secondary intention label, and the multi-round historical dialogue record.

[0038] The application tracks the multi-round historical dialogue record of the target user, and outputs the structured query parameter system of the target user according to the primary intention label, the secondary intention label and the multi-round historical dialogue record, so that the most relevant information can be quickly and accurately screened from massive data. It should be explained that the multi-round historical dialogue record refers to the continuous interaction data generated by the user and the system in a single session (Session), and the structured query parameter refers to the conversion of the user's natural language intention into a standardized query instruction (entity + operation + constraint condition) executable by a machine, which is used to retrieve accurate results from a financial database or a knowledge graph.

[0039] S4, according to the structured query parameter and the dynamic user portrait, output the double-channel data of the securities consulting data, wherein the double-channel data includes structured channel data and unstructured channel data, and the key word blacklist, implicit commitment and undisclosed risk of the double-channel data are identified to filter the double-channel data to obtain filtered multi-modal knowledge fragments.

[0040] According to the structured query parameter and the dynamic user portrait, the application outputs the double-channel data of the securities consulting data to ensure the comprehensiveness and reliability of the consulting content.

[0041] In detail, according to the structured query parameter and the dynamic user portrait, the double-channel data of the securities consulting data is output, including: According to the structured query parameter, determine the query API of the securities consulting data; Determine the structured channel data of the securities consulting data through the query API; Use the risk level label and knowledge level index of the dynamic user portrait to mark the outputable nodes of the pre-constructed financial graph; According to the structured query parameter, traverse the outputable nodes to obtain unstructured channel data; Combine the structured channel data and the unstructured channel data to determine the double-channel data of the securities consulting data.

[0042] The query API refers to a set of application programming interfaces designed and deployed in advance to obtain securities consulting data, the APIs are standardized channels for the system to interact with external data sources (such as stock exchange data interfaces, financial data providers, internal databases, etc.), the structured channel data refers to data obtained from external data sources or internal databases by calling selected query APIs, and the data is organized in a structured format (such as JSON, XML, database table records, etc.), the financial graph refers to a pre-constructed financial field knowledge base in the form of a graph database or a knowledge graph, which takes entities (such as companies, stocks, funds, industries, concepts, regulations, and persons) in the financial field as nodes and the relationship (such as membership, investment relationship, influence relationship, associated transaction, and concept association) between the entities as edges, and the output node refers to a node set in the financial graph that is marked as being able to be displayed or provided to the user according to the dynamic user portrait information (especially the risk level label and the knowledge level index) of the current target user, the unstructured channel data refers to data extracted by traversing, querying and aggregating the output nodes of the filtered financial graph, and the data usually exists in the form of text, paragraphs, and descriptive information, for example, detailed interpretation of the latest financial report of a company, comments of industry analysts on the industry, and influence analysis of macroeconomic news on the market, and the double-channel data refers to an integrated data set formed by fusing the structured channel data and the unstructured channel data.

[0043] Optionally, the double-channel data of the securities consulting data combined with the structured channel data and the unstructured channel data can be output by determining the proportion of the structured channel data and the unstructured channel data in the double-channel data according to the dynamic user portrait.

[0044] The application identifies the keyword blacklist, implicit commitment and undisclosed risk of the double-channel data, effectively avoids information misdirection and potential legal compliance risks, and provides a safe and reliable information environment for users.

[0045] In detail, the identification of the keyword blacklist, implicit commitment and undisclosed risk of the double-channel data comprises: constructing a multi-level blacklist library of the double-channel data; determining potential extension words and potential transformation forms of the multi-level blacklist library; determining the keyword blacklist of the double-channel data according to the multi-level blacklist library, the potential extension words and the potential transformation forms; determining the commitment semantic pattern of the double-channel data to retrieve the implicit commitment of the double-channel data; detecting risk words in the dual-channel data; determining specific risk points in the dual-channel data; determining undisclosed risks in the dual-channel data by combining the risk words and the specific risk points.

[0046] The multi-level blacklist library refers to a pre-constructed word library or rule set for identifying and filtering inappropriate or high-risk content. The "multi-level" feature is reflected in that it may contain lists of words and phrases of different severity or different application scenarios. For example, the first-level blacklist may contain words that are explicitly illegal, rule-breaking, or extremely sensitive; the second-level blacklist may contain words that, although not completely rule-breaking, may cause misunderstandings, have potential misleadingness, or lack of risk warnings. The potential extension words refer to words that are usually similar or related to the core words in meaning. For example, if "insider trading" is in the blacklist, the extension words may include "unpublished information trading", "message leakage", etc. The potential transformation forms refer to the possible changes in the core words. For example, if "limit up" is in the blacklist, the transformation forms may include "zhangting" (pinyin), "zhangting" (homonym), etc. The keyword blacklist refers to a more comprehensive and dynamic keyword identification list formed by combining potential extension words and potential transformation forms on the basis of the multi-level blacklist library. The commitment semantic pattern refers to a series of language patterns used to identify text containing commitment statements, such as "guarantee profit", "steady profit", "future will definitely rise", "absolutely safe", etc. The implied commitment refers to content in the dual-channel data that indirectly conveys a deterministic expectation or guarantee meaning through tone, wording, analysis logic, or selective display of information, etc. The risk word refers to a specific word related to financial risk, investment risk, operational risk, etc. that is pre-defined or identified through analysis. The specific risk point refers to a specific risk source explicitly pointed to in the specific content of the dual-channel data, such as "major customer loss" and "revenue decline" in "the company faces the risk of revenue decline due to loss of major customers". The undisclosed risk refers to a risk in the dual-channel data that should be mentioned according to the user's query intent and combined with financial knowledge but is not fully disclosed or explained.

[0047] Optionally, the construction of the multi-level blacklist library of the dual-channel data can be achieved through the design of database tables (such as primary table-subtable relationship, field classification) in a relational database (such as MySQL, PostgreSQL).

[0048] Optionally, the determination of the specific risk points in the dual-channel data can be achieved through natural language processing (NLP) technology It should be explained that the filtered multi-modal knowledge fragments refer to a series of knowledge units containing valuable information and meeting specific quality standards obtained after screening, cleaning and refining the original dual-channel data (i.e., a comprehensive data set combining structured channel data and unstructured channel data).

[0049] S5、According to the filtered multi-modal knowledge fragments and the dynamic user portrait, a multi-dimensional answer of the target user is generated, and explicit feedback and implicit feedback of the multi-dimensional answer are obtained to generate multi-strategy adjustment parameters and user portrait update parameters of the multi-strategy model and the dynamic user portrait, and multi-round question and answer of the target user is performed based on the multi-strategy adjustment parameters and user portrait update parameters.

[0050] According to the filtered multi-modal knowledge fragments and the dynamic user portrait, the multi-dimensional answer of the target user is generated, which can improve the experience of the user.

[0051] In detail, the multi-dimensional answer of the target user is generated according to the filtered multi-modal knowledge fragments and the dynamic user portrait, including: According to the knowledge level index of the dynamic user portrait, a generation mode of the filtered multi-modal knowledge fragments is generated, wherein the generation mode includes a professional mode and an explanation mode; When the intention ambiguity score of the dynamic user portrait is greater than a preset intention ambiguity threshold, a counter-question label of the target user is determined; Based on the generation mode and the counter-question label, a multi-dimensional answer of the target user is generated.

[0052] The professional mode refers to an answer generation strategy for high-knowledge-level users (such as institutional investors and experienced individual investors), for example, containing Greek letter indicators such as alpha / Beta, the explanation mode refers to a popularized explanation strategy for novice users, for example, using ConceptNet to generate analogy explanations (such as "PE is like the price tag of stocks, 36 times means..."), dynamically inserting a schematic diagram (ASCII technical indicator legend), the counter-question label refers to a guiding question label generated by the system when the intention is ambiguous or the information is insufficient, and the multi-dimensional answer refers to a comprehensive answer finally generated and presented to the target user, which is around the consultation question and is elaborated from multiple angles or levels.

[0053] Optionally, when the intention ambiguity score of the dynamic user portrait is greater than the preset intention ambiguity threshold, the system can determine the target user's question tag by combining the previous dialogue history through the DST technology. The system can determine which context the user's ambiguous question is directed to or which information that has been mentioned in the previous context but needs to be confirmed this time is missing. For example, if the previous discussion is about multiple companies and only asks "financial report", it needs to ask which company specifically. DST helps the system locate the ambiguity point in the more macro dialogue process.

[0054] It should be explained that the explicit feedback refers to the user's explicit and direct evaluation or indication of the system-generated multi-dimensional answer, such as rating, likes, etc. The implicit feedback refers to the user's non-direct and indirect behavior or pattern exhibited during the interaction with the system, such as interaction duration, subsequent question relevance, etc. The multi-strategy adjustment parameter refers to the specific value used to optimize and adjust the multiple strategies (such as information retrieval strategy, knowledge graph query strategy, generation model strategy, question strategy, answer organization strategy, etc.) used by the system to generate multi-dimensional answers according to the collected explicit and implicit feedback. The user portrait update parameter refers to the specific value used to update and correct each attribute (knowledge level index, risk level label, intention ambiguity) constituting the dynamic user portrait according to the feedback information. For reference, Figure 2 The user portrait update schematic diagram for implementing the multi-round question and answer method based on a large model for securities consultation is provided for an embodiment of the present application.

[0055] Finally, in the subsequent each round of dialogue of the multi-round question and answer guidance system for the target user based on the multi-strategy adjustment parameter and the user portrait update parameter, the system can more accurately understand the user's current real intention and cognitive state, thereby providing more personalized and effective securities consultation answers that meet the user's needs, forming a closed-loop interactive process of continuous learning and continuous optimization.

[0056] Firstly, the introduction of dynamic user portrait greatly improves the personalization and safety of consultation. The accurate characterization of risk level label, knowledge level index and intention ambiguity score enables the system to dynamically adjust the consultation strategy according to the user's risk tolerance, professional knowledge background and current intention clarity. For users with high risk level and low knowledge level, the system can automatically filter high-risk product information and use more popular and easy-to-understand language for explanation, effectively reducing the potential loss caused by user misoperation or understanding deviation, enhancing the safety of consultation service. For users with ambiguous intention, the system can intervene in clarification in time to avoid providing information deviating from the user's real needs, improving the communication efficiency. Secondly, multi-level intention recognition and structured query parameter generation significantly improve the accuracy and efficiency of information retrieval. By accurately positioning the secondary intention through the first-level intention label and confidence, and combining the multi-round historical dialogue record to generate structured query parameters, the system can quickly and accurately filter out the most relevant information from massive data, which avoids the inefficient search caused by information overload or inaccurate keyword matching in traditional consultation, enabling users to obtain the required answers more quickly and improving user experience. Thirdly, the dual-channel data fusion and intelligent filtering mechanism ensures the comprehensiveness and reliability of the consultation content. The combination of structured data channel and unstructured knowledge channel ensures that the answer has accurate data support and rich background knowledge. More importantly, the system can actively identify and filter keyword blacklists, implicit promises and undisclosed risks, effectively avoiding information misdirection and potential legal compliance risks, providing a safe and reliable information environment for users. Finally, the synergistic optimization of multi-strategy model and user portrait realizes the continuous improvement and intelligent evolution of consultation service. By obtaining explicit and implicit feedback of users on multi-dimensional answers, the system dynamically adjusts the multi-strategy model parameters and updates the user portrait, enabling the system to continuously learn the user's preferences and demand changes and optimize the answer strategy. Therefore, the present application can improve the accuracy and personalization of securities consultation.

[0057] Embodiment 2: As Figure 3 shown, it is a function module diagram of a multi-round question and answer system based on a large model for securities consultation according to the present application.

[0058] The multi-round question and answer system based on a large model for securities consultation 300 according to the present application can be installed in an electronic device. According to the realized functions, the multi-round question and answer system based on a large model for securities consultation can include a user portrait construction module 301, a user intention analysis module 302, a query parameter construction module 303, a dual-channel data query module 304 and an intelligent multi-round question and answer module 305. The modules according to the present application can also be called units, which refer to a series of computer program segments that can be executed by an electronic device processor and can complete fixed functions, and are stored in the memory of the electronic device.

[0059] In the embodiments of the present application, the functions of each module / unit are as follows: The user portrait construction module 301 is configured to receive security consulting data of a target user, wherein the security consulting data comprises explicit data and implicit data, so as to generate a dynamic user portrait of the target user, wherein the dynamic user portrait comprises a risk level label, a knowledge level index and an intention ambiguity score. The user intention analysis module 302 is configured to analyze a first-level intention label and a first-level intention confidence of the security consulting data, and determine a second-level intention label of the target user according to the first-level intention label and the first-level intention confidence. The query parameter construction module 303 is configured to track multi-round historical dialogue records of the target user, and output a structured query parameter of the target user according to the first-level intention label, the second-level intention label and the multi-round historical dialogue records. The dual-channel data query module 304 is configured to output dual-channel data of the security consulting data according to the structured query parameter and the dynamic user portrait, wherein the dual-channel data comprises structured channel data and unstructured channel data, identify a keyword blacklist, an implied commitment and an undisclosed risk of the dual-channel data, filter the dual-channel data to obtain a filtered multi-modal knowledge fragment. The intelligent multi-round question and answer module 305 is configured to generate a multi-dimensional answer of the target user according to the filtered multi-modal knowledge fragment and the dynamic user portrait, obtain explicit feedback and implicit feedback of the multi-dimensional answer, generate a multi-strategy model and multi-strategy adjustment parameters and user portrait update parameters of the dynamic user portrait, and perform multi-round question and answer of the target user based on the multi-strategy adjustment parameters and user portrait update parameters.

[0060] In detail, the modules in the multi-round question and answer system 300 for security consulting based on a large model in the embodiments of the present application adopt the same technical means as the multi-round question and answer method for security consulting based on a large model in the above-mentioned Figure 1 , and can produce the same technical effects, which will not be described here.

[0061] It is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0062] Finally, it should be noted that the above examples are merely intended to illustrate the technical solutions of the present application and not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for implementing multi-round question and answer under securities consulting based on a large model, characterized in that, The method comprises: receiving security consultation data of a target user, wherein the security consultation data comprises explicit data and implicit data, to generate a dynamic user portrait of the target user, wherein the dynamic user portrait comprises a risk level label, a knowledge level index, and an intention ambiguity score; analyzing a first-level intention label and a first-level intention confidence of the security consultation data, and determining a second-level intention label of the target user according to the first-level intention label and the first-level intention confidence; tracking multi-round historical dialogue records of the target user, and outputting a structured query parameter of the target user according to the first-level intention label, the second-level intention label, and the multi-round historical dialogue records; outputting double-channel data of the security consultation data according to the structured query parameter and the dynamic user portrait, wherein the double-channel data comprises structured channel data and unstructured channel data, identifying a keyword blacklist, an implied commitment, and an undisclosed risk of the double-channel data to filter the double-channel data to obtain a filtered multi-modal knowledge segment; generating a multi-dimensional answer of the target user according to the filtered multi-modal knowledge segment and the dynamic user portrait, and obtaining explicit feedback and implicit feedback of the multi-dimensional answer to generate a multi-strategy model and multi-strategy adjustment parameters and user portrait update parameters of the dynamic user portrait, and executing multi-round question and answer of the target user based on the multi-strategy adjustment parameters and user portrait update parameters.

2. The method of claim 1, wherein, The generation of the dynamic user portrait of the target user comprises: analyzing a basic risk and a sensitivity of the target user according to the explicit data and the implicit data of the target user; dynamically adjusting the basic risk based on the sensitivity to obtain a risk level label of the target user; analyzing a financial term understanding degree, a question complexity, and a historical interaction depth of the explicit data and the implicit data to determine a knowledge level index of the target user; analyzing a semantic confusion degree and an intention confidence difference of the target user according to the explicit data and the implicit data; determining an intention ambiguity score of the target user in combination with the semantic confusion degree and the intention confidence difference; generating the dynamic user portrait of the target user in combination with the risk level label, the knowledge level index, and the intention ambiguity score.

3. The method of claim 2, wherein, The analysis of the semantic confusion degree and the intention confidence difference of the target user according to the explicit data and the implicit data comprises: determining a cosine similarity score and a sentence coherence weight of the target user according to the explicit data and the implicit data; calculating the semantic confusion degree of the target user based on the cosine similarity score and the sentence coherence weight by using the following formula: ； wherein, represents a semantic confusion degree of the target user, represents a total number of utterances of a current dialogue turn of the target user, represents a context cosine similarity score of the target user, represents an utterance coherence weight, represents the i-th utterance of a current dialogue turn of the target user, utterance text; analyzing an intention probability of the target user according to the explicit data and the implicit data; and determining an intention confidence difference of the target user based on the intention probability.

4. The method of claim 3, wherein, The determination of the intention confidence difference of the target user based on the intention probability comprises: calculating the intention confidence difference of the target user based on the intention probability by using the following formula: ； wherein, denotes an intent confidence difference, denotes a number of candidate intents of the target user, denotes a first candidate intent in the nth round of dialogues of the target user, denotes a first candidate intent in the nth round of dialogues of the target user, denotes a second candidate intent in the nth round of dialogues of the target user, denotes a second candidate intent in the nth round of dialogues of the target user, denotes a probability function of a candidate intent, denotes a max function.

5. The method of claim 4, wherein, The analysis of the first-level intention label and the first-level intention confidence of the securities consultation data comprises: The securities consultation data is preprocessed to obtain processed securities consultation data; Key entities of the processed securities consultation data are labeled; According to the key entities, a trained intention classification model is used to output a first-level intention label of the securities consultation data; The first-level intention label is analyzed in terms of term enhancement confidence and entity consistency to determine the first-level intention confidence of the first-level intention label.

6. The method of claim 5, wherein, The determination of the second-level intention label of the target user according to the first-level intention label and the first-level intention confidence comprises: When the first-level intention confidence meets a preset first-level intention confidence threshold, a second-level intention knowledge graph of the first-level intention label is constructed; The intention multi-dimensional features of the target user are analyzed according to the securities consultation data of the target user; Based on the intention multi-dimensional features, the second-level intention score of the second-level intention knowledge graph is analyzed to label the second-level candidate intention of the target user; The intention risk coefficient of the second-level candidate intention is analyzed; Based on the intention risk coefficient, the second-level intention label of the target user is determined.

7. The method of claim 6, wherein, The output of the double-channel data of the securities consultation data according to the structured query parameter and the dynamic user portrait comprises: According to the structured query parameter, a query API of the securities consultation data is determined; Through the query API, the structured channel data of the securities consultation data is determined; The risk level label and the knowledge level index of the dynamic user portrait are used to label the outputable nodes of the pre-constructed financial graph; According to the structured query parameter, the outputable nodes are traversed to obtain unstructured channel data; The double-channel data of the securities consultation data is determined by combining the structured channel data and the unstructured channel data.

8. The method of claim 7, wherein, The identification of the keyword blacklist, implicit commitment and undisclosed risk of the double-channel data comprises: A multi-level blacklist library of the double-channel data is constructed; The potential extension words and the potential transformation forms of the multi-level blacklist library are determined; According to the multi-level blacklist library, the potential extension words and the potential transformation forms, the keyword blacklist of the double-channel data is determined; The commitment semantic pattern of the double-channel data is determined to retrieve the implicit commitment of the double-channel data; The risk words of the double-channel data are detected; The specific risk points of the double-channel data are determined; The undisclosed risk of the double-channel data is determined by combining the risk words and the specific risk points.

9. The method of claim 8, wherein, The generation of the multi-dimensional answer of the target user according to the filtered multi-modal knowledge segment and the dynamic user portrait comprises: According to the knowledge level index of the dynamic user portrait, a generation mode of the filtered multi-modal knowledge segment is generated, wherein the generation mode comprises a professional mode and an explanation mode; When the intention ambiguity score of the dynamic user portrait is greater than a preset intention ambiguity threshold, a counter-question label of the target user is determined; Based on the generation mode and the counter-question label, the multi-dimensional answer of the target user is generated.

10. A multi-round question-answering system under securities consultation based on a large model, characterized in that, The system comprises: The user portrait construction module is configured to receive security consultation data of a target user, wherein the security consultation data comprises explicit data and implicit data, to generate a dynamic user portrait of the target user, wherein the dynamic user portrait comprises a risk level label, a knowledge level index, and an intention ambiguity score. The user intention analysis module is configured to analyze a first-level intention label and a first-level intention confidence of the security consultation data, and determine a second-level intention label of the target user according to the first-level intention label and the first-level intention confidence. The query parameter construction module is configured to track multi-round historical dialogue records of the target user, and output structured query parameters of the target user according to the first-level intention label, the second-level intention label, and the multi-round historical dialogue records. The dual-channel data query module is configured to output dual-channel data of the security consultation data according to the structured query parameters and the dynamic user portrait, wherein the dual-channel data comprises structured channel data and unstructured channel data, identify a keyword blacklist, an implied commitment, and an undisclosed risk of the dual-channel data, filter the dual-channel data to obtain filtered multi-modal knowledge fragments. The intelligent multi-round question and answer module is configured to generate multi-dimensional answers of the target user according to the filtered multi-modal knowledge fragments and the dynamic user portrait, obtain explicit feedback and implicit feedback of the multi-dimensional answers, generate multi-strategy adjustment parameters and user portrait update parameters of the multi-strategy model and the dynamic user portrait, and perform multi-round question and answer of the target user based on the multi-strategy adjustment parameters and the user portrait update parameters.

Images