Answer generation method, apparatus, device, and medium

Abstract

The application relates to the technical field of natural language processing, and discloses a question answering generation method and device, equipment and a medium, which comprises the following steps: receiving a question of an object to be answered, and obtaining object attribute information of the object to be answered; extracting attribute information of the question to obtain question attribute information; matching the object attribute information and the question attribute information to obtain a matching result; determining a first answer generation strategy based on the matching result, and generating an answer to the question based on the first answer generation strategy. The application can be applied to the question and answer scene of financial technology and medical health, and the appropriateness of the answer content is improved.

Description

Technical Field

[0001] This application relates to the fields of natural language processing and artificial intelligence, and is applied to fintech and healthcare scenarios, particularly to a method, apparatus, device, and medium for generating answers. Background Technology

[0002] With the rapid development of large language model technology, intelligent question answering systems have been widely applied in various fields such as fintech, healthcare, legal services, and online education. Existing intelligent question answering systems directly input questions from users into a large language model to generate and output answers. It is evident that current technology generates answers based on the question itself, failing to provide differentiated responses according to the actual circumstances of different users. In fintech and healthcare scenarios, this can easily lead to answers that exceed the user's risk tolerance or cognitive level, resulting in low appropriateness of the responses. Summary of the Invention

[0003] This application provides an artificial intelligence-based answer generation method, apparatus, computer device, and medium, aiming to address the technical problem that current intelligent question-answering systems cannot provide differentiated answers based on the actual situation of different users. In fintech and healthcare scenarios, the answers are prone to exceeding the user's risk tolerance or cognitive level, resulting in low appropriateness of the answers. This application aims to improve the appropriateness of the answers.

[0004] Firstly, a method for generating answers is provided, including: Receive the question from the object to be answered, and obtain the object attribute information of the object to be answered; Extract the attribute information of the problem to obtain the problem attribute information; The object attribute information is matched with the problem attribute information to obtain a matching result; Based on the matching results, a first answer generation strategy is determined, and an answer to the question is generated based on the first answer generation strategy.

[0005] Secondly, an answer generation apparatus is provided, comprising: The acquisition module is used to receive the question of the object to be answered and to acquire the object attribute information of the object to be answered; The extraction module is used to extract the attribute information of the problem to obtain the problem attribute information; The matching module is used to perform fit matching between the object attribute information and the problem attribute information to obtain a matching result; The generation module is used to determine a first answer generation strategy based on the matching results, and generate an answer to the question based on the first answer generation strategy.

[0006] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described answer generation method.

[0007] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described answer generation method.

[0008] In the above-mentioned solution implemented by the answer generation method, apparatus, computer equipment, and storage medium, the following steps are taken: receiving a question from an object to be answered and obtaining the object attribute information of the object to be answered; extracting the attribute information of the question to obtain question attribute information; performing a fit matching between the object attribute information and the question attribute information to obtain a matching result; determining a first answer generation strategy based on the matching result; and generating an answer to the question based on the first answer generation strategy. In this application, before answering the question, the object attribute information and question attribute information are first matched for fit, and then a first answer generation strategy is determined based on the matching result. Based on the first answer generation strategy, an answer to the question is generated. This can provide differentiated answers according to the actual situation of different users, thereby reducing the situation in financial technology and healthcare scenarios where the answer content exceeds the user's risk tolerance or cognitive level, and improving the appropriateness of the answer content. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of an application environment for the answer generation method in one embodiment of this application; Figure 2 This is a flowchart illustrating a response generation method in one embodiment of this application; Figure 3 yes Figure 1 A schematic diagram of a specific implementation method for step S20; Figure 4 yes Figure 1 A schematic diagram of a specific implementation method for step S30; Figure 5 yes Figure 1 A flowchart illustrating an implementation method following step S40; Figure 6This is a schematic diagram of the structure of a response generation device in one embodiment of this application; Figure 7 This is a schematic diagram of the structure of a computer device according to one embodiment of this application; Figure 8 This is another structural schematic diagram of a computer device in one embodiment of this application. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] The answer generation method provided in this application embodiment can be applied to, for example, Figure 1 In this application environment, the client communicates with the server via a network. The server can receive questions from the object to be answered through the client and obtain the object attribute information of the object to be answered; extract the attribute information of the question to obtain question attribute information; perform fit matching between the object attribute information and the question attribute information to obtain a matching result; determine a first answer generation strategy based on the matching result; and generate an answer to the question based on the first answer generation strategy. In this application, before answering the question, the object attribute information and the question attribute information are first matched for fit, and then a first answer generation strategy is determined based on the matching result. Based on the first answer generation strategy, an answer to the question is generated. This can provide differentiated answers according to the actual situation of different users, thereby reducing the situation in fintech and healthcare scenarios where the answer content exceeds the user's risk tolerance or cognitive level, and improving the appropriateness of the answer content.

[0013] The present application will now be described in detail through specific embodiments.

[0014] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the answer generation method provided in this application embodiment includes the following steps: S10: Receive the question from the object to be answered and obtain the object attribute information of the object to be answered.

[0015] The answer generation method provided in this application can be applied to intelligent question-answering systems in various application scenarios. Intelligent question-answering systems are typically implemented through a server-side architecture, which can receive questions from users in real time. For example, in the fintech field, users may ask financial investment advisory questions, requiring an intelligent question-answering system to respond and provide investment advice. Similarly, in the healthcare field, users may ask various health-related questions based on their health conditions, requiring an intelligent question-answering system to respond and provide appropriate health guidance and medical advice.

[0016] The "object to be answered" refers to the person who submits a question to the intelligent question-answering system and requests an answer.

[0017] Object attribute information refers to various types of information used to characterize the features of the object to be answered, and is the basic data reflecting the actual situation of the object to be answered.

[0018] The input format for questions can include, but is not limited to, various formats such as text and voice; this application does not impose specific limitations on this. Methods for obtaining object attribute information can include, but are not limited to: obtaining it from the registration information of the object to be answered, obtaining it from the historical interaction records of the object to be answered, and obtaining it through real-time collection of relevant information from the object to be answered.

[0019] For example, in fintech applications, the person to be answered submits a question via voice, such as "Is now a good time to buy a financial product?". After receiving the question, the intelligent question-answering system obtains the object attribute information from the object's object account information for subsequent suitability matching calculations.

[0020] For example, in healthcare applications, the person to be answered submits a question via voice, such as "I have a cough recently, what medicine can I take?" After receiving the question, the intelligent question-answering system obtains the person's attribute information from their electronic health record for subsequent fit matching calculations.

[0021] In some embodiments, the question is in the form of an audio question. Before step S10, i.e., before obtaining the object attribute information of the object to be answered, the following steps S50-S60 are also included: S50: Extract the voiceprint features of the problem.

[0022] S60: Obtain the device identifier corresponding to the object to be answered.

[0023] If the voiceprint feature matches the preset voiceprint feature and the device identifier is the same as the preset device identifier, then the step of obtaining the object attribute information of the object to be answered is executed.

[0024] Among them, the problem of phonetic form can be represented by the symbol A, and its mathematical form is A∈R. 1×T Where R represents the set of real numbers, 1 represents mono speech, and T represents the number of speech sampling points; T≥48000 is required, corresponding to a speech signal with a sampling rate of 16kHz and a duration of not less than 3 seconds, in order to ensure the accuracy of voiceprint recognition.

[0025] Once the question is obtained in speech form, a pre-trained voiceprint recognition model can be used to extract the voiceprint embedding of the question, which can be represented by the symbol e. speaker This means that the voiceprint recognition model converts the input speech signal into a real number vector, which can uniquely represent the speaker's voiceprint features.

[0026] The voiceprint recognition model takes a 16kHz sampling rate, mono audio signal with a duration of at least 3 seconds as input and outputs a fixed 192-dimensional voiceprint embedding vector. The model is pre-trained with weight initialization, fine-tuned using contrastive learning on a business scenario speech dataset, employing the Adam optimizer with a learning rate of 1e-3, and trained until the voiceprint recognition accuracy converges. Device identifier refers to the unique hardware identifier assigned to the device at the factory, used to distinguish different physical devices.

[0027] During the registration phase of the subject to be answered, standard speech is collected. Voiceprints are extracted using this voiceprint recognition model, embedded, and stored as preset voiceprint features. Simultaneously, the device identifier of the subject to be answered is bound to it, forming a one-to-one voiceprint-device registration database. The unique identifier of the device currently used by the subject to be answered can be obtained, which can be represented by the symbol d. id This means that Android devices read ANDROID_ID as the device identifier through the system API, while iOS devices read IDFV as the device identifier through the system API.

[0028] Query the registration database to retrieve the preset voiceprint features stored during the registration of the object to be answered. This can be done using the symbol 'e'. reg The symbol d can be used to indicate and preset device identifiers. reg Indicates; calculates the extracted voiceprint features e speaker With preset voiceprint characteristics e reg The cosine similarity between them can be expressed by the symbol cos sim (e speaker ,e reg )express.

[0029] If the cosine similarity is cos sim (e speaker ,e reg If the value is less than 0.75, it means the voiceprint feature does not match the preset voiceprint feature, or the device identifier d... id ≠ d regIf the device identifier is different from the preset device identifier, the request will be rejected and an authentication failure message will be returned.

[0030] If the cosine similarity is cos sim (e speaker ,e reg ) ≥ 0.75, meaning the voiceprint feature matches the preset voiceprint feature, and the device identifier d id = d reg If the device identifier matches the preset device identifier, the verification passes, and the object attribute information is loaded, which can be represented by the symbol P. user express.

[0031] For example, in fintech applications, object attribute information can include risk level, investment period, the date of the most recent risk assessment, and real-time holding information, such as equity position ratio and used margin ratio.

[0032] Among them, the risk level refers to the classification result of the risk tolerance of the respondent obtained through standardized risk assessment questionnaires. It is usually divided into five levels, from C1 (conservative type) to C5 (aggressive type), and is the core basis for judging the suitability of the respondent with investment products.

[0033] The investment period refers to the cumulative time the respondent has been involved in the investment field of fintech, measured in years. It reflects the respondent's familiarity with the financial market and level of investment experience.

[0034] The most recent risk assessment date refers to the specific date on which the subject last completed a standardized risk assessment. It is used to assess the timeliness of risk level information. Regulatory requirements stipulate that the validity period of risk assessment results is usually one year.

[0035] Real-time holdings information refers to the real-time status of various financial assets currently held by the entity being addressed; it is dynamically updated attribute information of the entity.

[0036] Equity position ratio refers to the proportion of the market value of equity assets held by the respondent, such as stocks and equity funds, to their total assets. The value ranges from 0 to 1, with a higher value indicating a higher level of risk exposure for the respondent.

[0037] The used margin ratio refers to the proportion of the margin amount used by the respondent in its margin trading business to its total available margin, and is used to measure the respondent's leverage level and liquidity risk.

[0038] For example, in healthcare applications, object attribute information can include age, gender, medical history, allergy risk factor, current medication, date of the most recent physical examination, and severity of underlying conditions.

[0039] The voiceprint-device two-factor authentication mechanism used in the above steps combines two independent verification dimensions: non-replicable biometric features and unique hardware features. Compared with single password verification or single biometric verification, it can effectively resist most security threats such as account theft and identity impersonation. Moreover, the verification process is completed automatically in the background without requiring any additional operation from the person being answered, and it does not affect the normal voice interaction experience. At the same time, the use of a pre-trained voiceprint recognition model ensures the recognition accuracy in complex environments, providing a solid security guarantee for application scenarios involving highly sensitive information such as financial technology and healthcare.

[0040] It should be noted that the above is only one example of a voiceprint recognition model. In practical applications, other voiceprint recognition models can also be used, and no specific restrictions are made here.

[0041] S20: Extract the attribute information of the problem to obtain the problem attribute information.

[0042] Among them, problem attribute information refers to various types of information used to characterize the characteristics of the problem itself, and is the basic data that reflects the inherent attributes of the problem.

[0043] In this step, features are extracted from the received questions to obtain question attribute information that characterizes the questions themselves. This question attribute information is another important basis for subsequent fit matching, as it reflects the inherent characteristics of the questions themselves.

[0044] For example, in the fintech field, when faced with a question submitted by a user, such as "Is now a good time to buy financial products?", an intelligent question-answering system uses natural language processing technology to analyze the question and extract its attribute information.

[0045] For example, in the field of healthcare, when faced with a question submitted by a user, such as "I've had a cough recently, what medicine can I take?", the intelligent question-answering system uses natural language processing technology to analyze the question and extract its attribute information.

[0046] In some embodiments, such as Figure 3 As shown, the problem attribute information includes first problem attribute information and second problem attribute information. Step S20, which is to extract the problem attribute information and obtain the problem attribute information, includes the following steps S21-S22: S21: By using a semantic analysis model, semantic analysis is performed on the question to obtain the first question attribute information.

[0047] S22: Determine the proportion of the preset vocabulary in the question in the total vocabulary to obtain the second question attribute information.

[0048] For example, in fintech applications, for questions in speech form, a pre-trained speech recognition model can be used to perform automatic speech recognition processing, parsing the question text T; for questions in text form, the original text T can be obtained directly. Then, a semantic analysis model fine-tuned with LoRA is used to perform semantic parsing. The model parameters are set to rank 8 and scaling factor 16. After feature mining of the text T, three types of first-question attribute information are output, namely, risk level r representing the degree of risk of the question. q The risk levels are categorized into three levels: Level 1 (low risk), Level 2 (medium risk), and Level 3 (high risk), representing the cognitive complexity c, which indicates the difficulty of understanding the problem. q The numerical range is between 0 and 1, determined based on a combination of the number of technical terms and the length of the logical chain, as well as the behavioral tendency type b that represents the questioner's intent. q The data is categorized into two types: a neutral (0) type and an aggressive (1) type. Specifically, if aggressive expressions such as "all in," "betting everything," or "bottom fishing" are detected, the data is classified as aggressive; otherwise, it is classified as neutral. A financial dictionary containing 2800 professional terms is used as the reference set D. finance The statistical text T belongs to set D. finance The vocabulary, as the preset vocabulary, is calculated according to the formula ρ. term =∣T∩D finance | / |T|, calculate the proportion of the preset vocabulary in all words in text T, and obtain the term density ρ. term This serves as the attribute information for the second question.

[0049] Among them, the speech recognition model is used to convert speech signals into text. LoRA fine-tuning refers to a lightweight model fine-tuning method that optimizes the model's semantic recognition capabilities with small parameter changes. Risk level is an attribute parameter that quantifies the level of investment risk contained in a question. Cognitive complexity is a numerical indicator that measures the professional difficulty and comprehension threshold of a question. Behavioral tendency type is a classification attribute that determines the investment operation intention behind the question asked by the respondent. Terminology density is a quantitative value that reflects the professional level of the question, representing the proportion of professional terms in the statistical text.

[0050] For example, in healthcare applications, for questions in speech form, a speech recognition model can be used to perform automatic speech recognition processing to parse the question text T; for text-based questions, the original text T can be directly obtained. Then, a lightweight, fine-tuned semantic analysis model is used to perform semantic parsing, mining text features and outputting three categories of first-level question attribute information: risk level (representing the degree of health risk, categorized as low risk 1, medium risk 2, and high risk 3); cognitive complexity (representing the difficulty of understanding the question, ranging from 0 to 1), which is determined by combining the amount of medical vocabulary and the length of the logical chain of the disease description; and behavioral tendency type (representing the behavioral tendency of the person seeking answers, categorized as routine consultation tendency 0 and blind treatment tendency 1). Expressions such as self-medication and unauthorized discontinuation of medication are identified as blind treatment tendency, while others are routine consultation. Using a medical dictionary with a massive number of entries as a reference set, preset medical terms within the text are filtered, and terminology density is calculated according to a proportion formula. The calculation result is used as the second-level question attribute information.

[0051] It should be noted that the above is only one example of a speech recognition model and a semantic analysis model. In practical applications, other speech recognition models and semantic analysis models can also be used, and no specific restrictions are made here.

[0052] The above steps, through fine-tuning the semantic analysis model, can accurately analyze the semantic features of risk level, cognitive complexity, and behavioral tendency type. Combined with the term density quantification statistical method, the attribute dimensions are improved, and the characteristics of the question itself are comprehensively and objectively characterized. This provides a complete and reliable data basis for subsequent matching judgment and answer strategy formulation, effectively improving the accuracy of question-and-answer adaptation and compliance control capabilities.

[0053] S30: Match the object attribute information with the problem attribute information to obtain the matching result.

[0054] In this step, the object attribute information and the problem attribute information are matched to obtain a matching result that reflects the degree of compatibility between the two.

[0055] For example, in the application of fintech, intelligent question-answering systems perform a fit calculation by matching the object's attribute information with the question's attribute information, such as "Is now a good time to buy a financial product?", to obtain a matching result. This matching result reflects the degree of fit between the object's own characteristics and the question related to the financial product.

[0056] In some embodiments, an AI model can be used to perform fit matching between the object feature vector and the problem feature vector to obtain a matching result. For example, Figure 4 As shown, step S30, which involves matching the object attribute information with the problem attribute information to obtain the matching result, includes the following steps S31-S32: S31: Map object attribute information to object feature vectors, and map problem attribute information to problem feature vectors.

[0057] S32: The object feature vector and the problem feature vector are matched using the fitness matching model to obtain the matching result.

[0058] For example, in fintech applications, object attribute information and problem attribute information are transformed into corresponding numerical feature vectors. The specific attribute mapping rules are as follows: risk levels C1-C5 are mapped to values ​​1-5 respectively; the investment period is the actual number of years the user has invested, with any period exceeding 5 years counted as 5 years; the equity position ratio is the user's equity assets / total assets, normalized to the 0-1 range. Object attribute information P user Mapping generates object feature vector x user The vector is filled with risk level codes in sequence. code (C1=1…C5=5)), Investment period (invest) years ), equity position ratio ratio By integrating the various attribute parameters of the problem, a feature vector x is constructed. q The vector is determined by the problem risk level r. q Cognitive complexity c q Technical terminology density ρ term Behavioral tendency type b q The model consists of four types of parameters. Then, a multilayer perceptron (MLP) fitness matching model is constructed with hidden layer dimensions of 64 and 32, respectively, using ReLU as the activation function. The model training dataset is a sample set containing user attributes, question attributes, and matching degree labels, with labels either manually labeled or generated by rules. The model training uses the cross-entropy loss function, the Adam optimizer, a learning rate of 1e-3, 50 training epochs, and a batch size of 32. The final output of the model is normalized to the [0,1] interval using the Sigmoid activation function to obtain the matching degree score s. match This is used as the matching result. MLP is a commonly used deep neural network model for feature fitting and score prediction.

[0059] For example, in healthcare applications, object attribute information and problem attribute information are converted into corresponding numerical feature vectors. Specifically, object attribute information is mapped to generate object feature vectors, which are then filled with age codes, basic symptom levels, and allergy risk coefficients. Problem attribute parameters are integrated to construct problem feature vectors, which are composed of four types of parameters: health risk level, cognitive complexity, medical terminology density, and behavioral tendency type. The two sets of feature vectors are then input into a fit matching model for computation, ultimately outputting a fit score in the range of 0 to 1 as the matching result.

[0060] It should be noted that the specific calculation method for fitness matching can also be implemented using various matching algorithms known in the art. This application does not impose any specific limitations on this.

[0061] It should also be noted that the above is only one example of a fit matching model. In practical applications, other fit matching models can also be used, and no specific restrictions are imposed here.

[0062] It should also be noted that the matching result can be a quantitative value, such as a matching score between 0 and 1, or a qualitative classification result, such as match or no match.

[0063] The above steps transform various heterogeneous attribute data into numerical vectors through standardized mapping, adapt to multi-layer neural network models to complete feature fusion calculations, objectively quantify the degree of fit between the object to be answered and the consultation question, and provide accurate and stable matching results, thus providing a reliable numerical basis for subsequent differentiated answering strategies.

[0064] S40: Based on the matching results, determine the first answer generation strategy, and generate the answer to the question based on the first answer generation strategy.

[0065] In this embodiment of the application, the answer generation strategy, such as the first answer generation strategy or the second answer generation strategy, refers to a set of rules used to guide the answer generation process, which specifies the requirements for the content, style and form of the answer.

[0066] In this step, a first answer generation strategy is determined based on the matching results, and then an answer to the question is generated based on this first answer generation strategy using the answer generation model.

[0067] The answer generation model can be a pre-trained large language model or other models; no specific restrictions are imposed here.

[0068] Different matching results correspond to different first answer generation strategies. The intelligent question answering system automatically selects the corresponding first answer generation strategy based on the matching results and generates answer content that meets the requirements based on the first answer generation strategy.

[0069] For example, if the matching result is a match, the intelligent question-answering system determines that the first answer generation strategy is a normal answer strategy and generates an answer about purchasing financial products according to the first answer generation strategy; if the matching result is a mismatch, the system determines that the first answer generation strategy is a risk warning strategy and generates a warning about the risks of the financial product according to the first answer generation strategy.

[0070] In some embodiments, step S40, namely determining the first answer generation strategy based on the matching results, includes: If the matching result meets the preset matching conditions, or the question attribute information meets the preset attribute conditions, then the first answer generation strategy is determined based on the matching result.

[0071] In this embodiment, when the matching result meets the preset requirements, or when the attributes of the question itself meet the preset requirements, the intelligent question answering system will determine the first answer generation strategy and generate the corresponding answer based on the matching result, which can cover more practical application scenarios and avoid unnecessary answer restrictions.

[0072] Among them, the preset matching conditions are pre-set standards used to determine whether the matching results meet the requirements. Their specific content can be adjusted according to the actual application scenario. For example, the preset matching conditions can be that the matching result is "matched", or that the matching score is greater than or equal to a certain preset threshold.

[0073] Preset attribute conditions are pre-defined standards used to determine whether the attribute information of a question meets the requirements. Their specific content can also be adjusted according to the actual application scenario. For example, preset attribute conditions could be that the question belongs to the basic consultation category, or that the question's behavioral tendency type is a non-preset behavioral tendency type, such as a neutral tendency type.

[0074] Behavioral tendency types are typically defined using binary or multi-class classification methods. This application adopts a binary classification system, dividing them into two categories: pre-defined tendency types and non-pre-defined tendency types. For example, in fintech applications, a predefined tendency type refers to a high-risk behavioral tendency that is predefined and requires special compliance control. This typically corresponds to an aggressive tendency type, manifested as the respondent's expressed intentions in the question to engage in excessive speculation, blindly follow trends, operate with full leverage, or use excessive leverage. In this application, this is represented by the symbol b. q =1 indicates that the non-preset tendency type refers to all normal behavioral tendencies other than the preset tendency type, usually corresponding to the neutral tendency type. This is manifested in the respondent's expression of rational consultation, knowledge acquisition, risk avoidance, etc., in the questions. In this application, this is represented by the symbol b. q =0 indicates that, in healthcare applications, the preset bias type corresponds to the tendency to blindly treat patients, while the non-preset bias type corresponds to the tendency to provide routine consultation.

[0075] For example, assuming the preset matching condition is that the matching score is greater than or equal to the first preset score, and the preset attribute condition is that the behavioral tendency type is a non-preset tendency type, then if the matching score obtained in step S30 is greater than or equal to the first preset score, or the behavioral tendency type included in the question attribute information is a non-preset tendency type, then the first answer generation strategy is determined based on the matching result.

[0076] The higher the matching score, the better the fit between the object to be answered and the question. The first preset score can be set based on actual needs, and there are no specific restrictions here. For example, the first preset score can be 0.6.

[0077] For example, in a healthcare application scenario, the question is "How to manage high blood pressure in daily life?", and the calculated value is s. match =0.83>0.6, satisfying the preset matching condition, b q =0 indicates a non-preset tendency type that meets the preset attribute conditions. If both conditions are met, the first answer generation strategy can be determined based on the matching results.

[0078] For example, in fintech applications, the question might be "Is now a good time to buy a money market fund?", and the result calculated is s. match =0.85>0.6, satisfying the preset matching condition, b q =0 indicates a non-preset tendency type, which satisfies the preset attribute conditions. It can be seen that both conditions are met, and the first answer generation strategy can be determined based on the matching results.

[0079] In some embodiments, when determining the first answer generation strategy based on the matching results, the matching score s obtained in step S30 can be used as a reference. match Different value ranges are used to determine the first answer generation strategy, which instructs the corresponding answer generation constraint prefix to be input into the answer generation model along with the question. This constraint prefix is ​​used to limit the content scope, tone style, and risk warning level of the answer, ensuring that the generated answer content matches the suitability of the object to be answered.

[0080] The above approach can provide clear strategic suggestions for high-adaptability scenarios and necessary risk warnings for medium- and low-adaptability scenarios, while strictly adhering to the regulatory requirements for content appropriateness while meeting the consultation needs of the respondents.

[0081] For example, when s match When ≥0.8, generate the first-level constraint prefix: [Strategy Suggestion]; when 0.6≤s match When s > 0.8, a second-level constraint prefix is ​​generated: [Risk Warning] This suggestion involves high risk; your current risk level is Cx. Please make a careful decision. match When the value is less than 0.6, a third-level constraint prefix is ​​generated: "[Suitability Not Met]". According to regulatory requirements, this type of suggestion is not provided at this time.

[0082] The intelligent question-answering system concatenates the generated constraint prefixes with the question and inputs them into the answer generation model to generate the final answer.

[0083] If the question is "Is now a good time to buy a money market fund?", the calculated value is s. match=0.88>0.8, so the first answer generation strategy is determined to be generating the first-level constraint prefix "[Strategy Recommendation]". The final input model content is "[Strategy Recommendation] Is it suitable to buy a money market fund now?"; the model output is "[Strategy Recommendation] The current market liquidity is relatively loose, and the yield of money market funds remains stable, making them suitable as a short-term cash management tool. You can allocate funds according to your own financial arrangements."

[0084] If the question is "Is now a good time to buy equity funds?", the calculated value is s. match =0.72, ≥0.6, and <0.8, therefore the first answer generation strategy is determined to be generating the second-level constraint prefix "[Risk Warning]". The final input model content is "[Risk Warning] This suggestion involves high risk. Your current risk level is C3. Please make a careful decision. Is it suitable to buy equity funds now?"; the model output is "[Risk Warning] This suggestion involves high risk. Your current risk level is C3. Please make a careful decision. Equity funds mainly invest in the stock market, and their net asset value fluctuates greatly, posing a risk of principal loss. The current market is in a period of fluctuation and adjustment. It is recommended that you control your position and allocate in batches."

[0085] If the question is "Is now a good time to buy stock index futures?", the calculated value is s. match =0.42 < 0.6, therefore the first answer generation strategy is determined to be generating the third-level constraint prefix "[Suitability Ineligible]". The final input model content is "[Suitability Ineligible] According to regulatory requirements, this type of advice is not provided at this time. Is it suitable to buy stock index futures now?"; the model output is "[Suitability Ineligible] According to regulatory requirements, your current risk tolerance does not match this type of investment product, and related investment advice is not provided at this time. You can learn about the basic concepts of stock index futures."

[0086] The above steps adopt a strategy that can be triggered as long as any condition is met, taking into account both the matching suitability and the attributes of the question itself, broadening the scope of reasonable consultation response, strictly controlling the output of high-risk questions and answers, while also meeting the normal consultation needs, thus improving the practicality and compliance balance of the question and answer service.

[0087] In some embodiments, such as Figure 5 As shown, after generating the answer to the question based on the first answer generation strategy, the following steps S70-S90 are also included: S70: Determine the problem type based on object attribute information and problem attribute information.

[0088] S80: If the problem type is a preset problem type, then obtain the content to be filled in corresponding to the preset problem type.

[0089] S90: Add the content to be added to the answer to get the answer after addition.

[0090] The preset question type is the knowledge mismatch type, and its judgment condition is: the cognitive complexity of the question is greater than the preset cognitive threshold, and the number of years of relevant domain experience of the person to be answered is less than the preset experience threshold.

[0091] In this embodiment, the preset cognitive threshold and preset experience threshold can be set based on actual conditions, and no specific restrictions are imposed here. For example, the preset cognitive threshold is set to 0.7, and the preset experience threshold is set to 2 years, that is: if c q >0.7, and invest years If the value is less than 2, it is determined to be a knowledge mismatch type, and a knowledge mismatch flag is set. knowledge_mismatch =1.

[0092] When the system detects the flag knowledge_mismatch When =1, the corresponding content to be added will be automatically retrieved: "It is recommended to understand the basic concepts first: [link to the corresponding investor education video]", and added to the end of the basic answer.

[0093] For example, in fintech applications, if the question is "How to construct a cross-market volatility surface arbitrage strategy?", the calculated value of c is... q =0.92>0.7, invest years =1.2 < 2 years, judged as knowledge mismatch type, flag set. knowledge_mismatch =1; The basic answer is "Cross-market volatility surface arbitrage strategy is a complex quantitative strategy that utilizes the price differences in volatility between different markets to generate profits, involving professional knowledge such as option pricing and volatility modeling."; The final answer after addition is "Cross-market volatility surface arbitrage strategy is a complex quantitative strategy that utilizes the price differences in volatility between different markets to generate profits, involving professional knowledge such as option pricing and volatility modeling. It is recommended to understand the basic concepts first: [link to basic option investment education video]". If the question is "How to construct a cross-market volatility surface arbitrage strategy?", the calculated value is c. q =0.92>0.7, invest years =3.5 > 2 years, so it is judged as a non-knowledge mismatch type, and no knowledge mismatch mark is set; the final answer is directly output as the basic answer, without adding any content.

[0094] For example, in a healthcare application scenario, if the question is "Adjusting the dosage of hypoglycemic drugs to control blood sugar," and the calculated cognitive complexity is 0.88 > 0.7, and the duration of illness is 1.1 years < 2 years, it is judged as a knowledge mismatch type and marked with the corresponding identifier. The basic answer is "Adjusting the dosage of hypoglycemic drugs needs to be combined with blood sugar indicators and physical condition; unauthorized changes pose health risks." After supplementation, the final answer is "Adjusting the dosage of hypoglycemic drugs needs to be combined with blood sugar indicators and physical condition; unauthorized changes pose health risks. It is recommended to first learn about common knowledge of medication for chronic diseases: [Health Science Popularization Link]."

[0095] The above steps enable precise adaptation to the cognitive level of the target audience. For complex questions from novice target audiences, basic investor education content is automatically supplemented. The post-supplementation architecture does not affect the execution efficiency of the pre-response generation process, and the system response speed is unaffected. The supplemented content can be dynamically adjusted according to the question type and supports multiple formats such as links, text, and images, improving the learning experience of the target audience, effectively reducing investment risks caused by insufficient understanding of the target audience, and further strengthening the system's compliance.

[0096] In some embodiments, after adding the content to be added to the answer and obtaining the added answer, the method further includes: It can identify whether the supplemented response includes pre-set statements about increasing asset holdings. If the statement includes a pre-set description of increasing asset holdings, and the real-time equity position ratio of the respondent is greater than the pre-set position threshold, then the corresponding position risk warning content will be obtained. Add the position risk warning to the end of the completed answer to obtain the target answer.

[0097] The preset asset holding descriptions include, but are not limited to, "can be increased," "recommended to increase holdings," "can be bought," and "recommended allocation." The preset position threshold can be set according to actual needs, and there are no specific restrictions here. For example, the preset position threshold can be set to 0.8.

[0098] In this embodiment, if the answer includes statements related to asset increases, and equity ratio If the value is greater than 0.8, the risk warning for adding positions will be: "Your current equity position is high, and we recommend that you operate with caution." For example, in fintech applications, if the basic answer is "The current valuation of the consumer sector is at a historical low, and it is advisable to increase holdings of high-quality leading consumer stocks," then the real-time equity position ratio (equity) can be calculated. ratio =0.87>0.8; the system detected that the answer contained the phrase "can be increased," which is an asset-related statement; therefore, a risk warning was added to the end of the answer; the final answer was: "The current valuation of the consumer sector is at a historical low, and it is advisable to increase holdings of high-quality leading consumer stocks. Your current equity position is high, and we recommend that you operate with caution." If the basic answer is "The current valuation of the consumer sector is at a historical low, so it is advisable to increase holdings of high-quality leading consumer stocks," then the real-time equity position ratio will be calculated. ratio =0.62 < 0.8; Although statements related to increasing asset holdings were identified, the position did not exceed the threshold; The final answer will be the basic answer without any additional content.

[0099] If the basic answer is "The current market is highly volatile, and it is recommended to remain on the sidelines," and no statements regarding increasing asset holdings are identified, no additional position risk warnings will be added regardless of the respondent's current position.

[0100] In some embodiments, after generating an answer to the question based on the first answer generation strategy, the method further includes: Each answer is attached with a unique digital signature, which contains key information such as the identifier of the object to be answered, the question timestamp, the matching score, and the strategy type. Output the final answer with a digital signature.

[0101] The digital signature can be generated using asymmetric encryption algorithms (such as RSA, SM2, etc.). The generation process involves extracting key business information from the response generation process, including but not limited to: the unique identifier of the object to be responded to, "user". id Issue submission timestamp, fit score (s) match The type of answer generation strategy adopted. type (Values ​​can be either the first or second answer generation strategy), the hash value of the answer content. hash The system's private key is used to encrypt the combination of the above key information to generate a unique digital signature; the digital signature is then appended to the end of the response to form the final output.

[0102] During subsequent auditing, the digital signature can be verified using the system's public key: if the verification passes, it proves that the answer content has not been tampered with and that all key business information is true and valid; if the verification fails, it proves that the answer content or key information has been tampered with.

[0103] Through the above steps, each answer carries a digital signature containing complete business information, and all interaction records are traceable and verifiable; the digital signature is generated using an asymmetric encryption algorithm, and any tampering with the answer content will cause the signature verification to fail; the generated digital signature can be directly connected to the regulatory system to meet the compliance reporting requirements of the financial industry.

[0104] In some embodiments, the method further includes the following steps S100-S110: S100: If the matching result does not meet the preset matching conditions, and the question attribute information does not meet the preset attribute conditions, then determine the second answer generation strategy.

[0105] S110: Generate an answer to the question based on the second answer generation strategy.

[0106] The second response generation strategy refers to a set of response generation rules specifically designed for high-risk scenarios. Its core principle is to refuse to provide responses that do not meet the appropriateness requirements and to explain the reasons for the refusal to the person to be responded to, thereby ensuring the compliance of the service and protecting the legitimate rights and interests of the person to be responded to.

[0107] This embodiment is used to handle high-risk scenarios where two triggering conditions are not met simultaneously. When the fit between the object to be answered and the question does not meet the requirements, and the question implies a behavioral tendency that requires special control, the intelligent question answering system will determine a second answer generation strategy and generate corresponding answer content to block the output of inappropriate content at the source.

[0108] In some embodiments, the matching result does not meet the preset matching conditions specifically when the matching score is less than the first preset score, and the problem attribute information does not meet the preset attribute conditions specifically when the behavioral tendency type included in the problem attribute information is a preset tendency type.

[0109] Accordingly, in this embodiment, if the matching score obtained in step S30 is less than the first preset score, and the behavioral tendency type included in the question attribute information is the preset tendency type, then a second answer generation strategy is determined, and an answer to the question is generated based on the second answer generation strategy.

[0110] Using the aforementioned notation, the above judgment logic can be simplified as follows: If s match <0.6, and b q =1, i.e., the radical tendency type, then the second answer generation strategy is determined.

[0111] For example, in fintech applications, if the question is "I want to use all my savings to leverage and buy stocks, can you recommend a few that will double in value?", the calculated value is s. match =0.18 < 0.6, the preset matching condition is not met, b q =1, indicating an aggressive tendency, which does not meet the preset attribute conditions. It can be seen that both conditions are not met simultaneously, so the second answer generation strategy is determined to be a suitability rejection strategy, generating the answer: "[Suitability Ineligibility] According to regulatory requirements, your current risk tolerance is not compatible with this type of investment product, and no relevant advice is provided at this time." If the question is "How do I open a stock index futures trading account? I want to engage in intraday high-frequency trading," s is calculated. match =0.32 < 0.6, which does not meet the preset matching condition, b q =1 indicates an aggressive tendency, which does not meet the preset attribute conditions. It can be seen that both conditions are not met, so the second answer generation strategy is determined to be a suitability rejection strategy, generating the answer: "[Suitability Not Met] According to regulatory requirements, your current risk tolerance is not compatible with this type of investment product, and no relevant advice will be provided at this time." For example, in a healthcare application scenario, the problem is "unauthorized increase in medication dosage to quickly cure chronic diseases," and the calculated value is s. match =0.15 < 0.6, which does not meet the matching criteria. The behavioral tendency type is determined to be the preset tendency type, which does not meet the attribute requirements. Since both conditions are not met, a rejection response strategy is adopted, and the following reply is generated: "[Health Risk Warning] Your idea has safety risks. Based on your physical condition, no relevant advice is provided at this time. Please follow your doctor's advice to take care of your body." In some embodiments, if the matching result does not meet the preset matching conditions and the problem attribute information does not meet the preset attribute conditions, a high-risk flag is generated; for example, this can be achieved by setting a high-risk flag bit. high_risk This flag is implemented using a value of 1. Correspondingly, if the intelligent question-answering system detects a high-risk flag... high_risk If the value is 1, then the second answer generation strategy is determined directly, and the answer to the question is generated based on the second answer generation strategy.

[0112] The above steps employ a strict circuit breaker logic that is triggered only when two conditions are met simultaneously. The rejection mechanism is only activated when the person to be answered is seriously mismatched with the question and has a high-risk behavioral tendency. This ensures compliance while avoiding excessive restrictions on normal consultation needs.

[0113] As can be seen, in the above solution, before answering the question, the object attribute information is first matched with the question attribute information for compatibility. Then, based on the matching result, the first answer generation strategy is determined, and based on the first answer generation strategy, the answer to the question is generated. This can provide differentiated answers according to the actual situation of different objects to be answered, thereby reducing the situation in fintech and healthcare scenarios where the answer content exceeds the user's risk tolerance or cognitive level, and improving the appropriateness of the answer content.

[0114] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0115] In one embodiment, an answer generation apparatus is provided, which corresponds one-to-one with the answer generation method in the above embodiments. For example... Figure 6 As shown, the answer generation device includes an acquisition module 101, an extraction module 102, a matching module 103, and a generation module 104. Detailed descriptions of each functional module are as follows: The acquisition module 101 is used to receive the question of the object to be answered and to obtain the object attribute information of the object to be answered; Extraction module 102 is used to extract the attribute information of the problem and obtain the problem attribute information; Matching module 103 is used to perform fit matching between object attribute information and problem attribute information to obtain matching results; The generation module 104 is used to determine a first answer generation strategy based on the matching results, and generate an answer to the question based on the first answer generation strategy.

[0116] In one embodiment, the generation module 104 is specifically used for: If the matching result meets the preset matching conditions, or the question attribute information meets the preset attribute conditions, then the first answer generation strategy is determined based on the matching result.

[0117] In one embodiment, the generation module 104 is specifically used for: If the matching result does not meet the preset matching conditions, and the question attribute information does not meet the preset attribute conditions, then the second answer generation strategy is determined. The answer to the question is generated based on the second answer generation strategy.

[0118] In one embodiment, the answer generation apparatus further includes a supplementation module, which is specifically used for: The problem type is determined based on object attribute information and problem attribute information; If the question type is a preset question type, then retrieve the content to be filled corresponding to the preset question type; Add the missing content to the answer to get the answer after filling in the missing content.

[0119] In one embodiment, the problem attribute information includes first problem attribute information and second problem attribute information. The extraction module 102 is specifically used for: By using a semantic analysis model, semantic analysis is performed on the question to obtain the first question attribute information; Determine the proportion of the preset vocabulary in the question within its total vocabulary to obtain the second question attribute information.

[0120] In one embodiment, the matching module 103 is specifically used for: Mapping object attribute information to object feature vectors, and mapping problem attribute information to problem feature vectors; The fitness matching model is used to match the feature vectors of the object with the feature vectors of the problem to obtain the matching result.

[0121] In one embodiment, the question is in the form of an audio message, and the acquisition module 101 is specifically used for: Extract the voiceprint features of the question; obtain the device identifier corresponding to the object to be answered; If the voiceprint feature matches the preset voiceprint feature and the device identifier is the same as the preset device identifier, then the step of obtaining the object attribute information of the object to be answered is executed.

[0122] This application provides an answer generation device that, before answering a question, first matches the object attribute information with the question attribute information for compatibility, then determines a first answer generation strategy based on the matching result, and generates an answer to the question based on the first answer generation strategy. This device can provide differentiated answers according to the actual situation of different objects to be answered, thereby reducing the occurrence of answers that exceed the user's risk tolerance or cognitive level in fintech and healthcare scenarios, and improving the appropriateness of the answer content.

[0123] Specific limitations regarding the answer generation device can be found in the limitations on answer generation described above, and will not be repeated here. Each module in the aforementioned answer generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0124] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements a function or step on the server side of an answer generation method.

[0125] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a response generation method on the client side. In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Receive the question from the object to be answered and obtain the object attribute information of the object to be answered; Extract the attribute information of the problem to obtain the problem attribute information; The object attribute information is matched with the problem attribute information to obtain the matching result; Based on the matching results, a first answer generation strategy is determined, and an answer to the question is generated based on the first answer generation strategy.

[0126] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: The acquisition module is used to receive the question from the object to be answered and to obtain the object attribute information of the object to be answered. The extraction module is used to extract the attribute information of the problem and obtain the problem attribute information; The matching module is used to match the object's attribute information with the problem's attribute information to obtain the matching result; The generation module is used to determine the first answer generation strategy based on the matching results, and generate the answer to the question based on the first answer generation strategy.

[0127] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0128] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0129] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0130] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. The personal information of the individuals to be answered in the embodiments of this application is all authorized (aware and consented to) by the relevant individuals or fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0131] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for generating answers, characterized in that, include: Receive the question from the object to be answered, and obtain the object attribute information of the object to be answered; Extract the attribute information of the problem to obtain the problem attribute information; The object attribute information is matched with the problem attribute information to obtain a matching result; Based on the matching results, a first answer generation strategy is determined, and an answer to the question is generated based on the first answer generation strategy.

2. The answer generation method according to claim 1, characterized in that, The step of determining the first answer generation strategy based on the matching results includes: If the matching result meets the preset matching conditions, or the question attribute information meets the preset attribute conditions, then a first answer generation strategy is determined based on the matching result.

3. The answer generation method according to claim 2, characterized in that, The method further includes: If the matching result does not meet the preset matching conditions, and the question attribute information does not meet the preset attribute conditions, then a second answer generation strategy is determined. An answer to the question is generated based on the second answer generation strategy.

4. The answer generation method according to claim 1, characterized in that, After generating the answer to the question based on the first answer generation strategy, the method further includes: Based on the object attribute information and the problem attribute information, the problem type is determined; If the problem type is a preset problem type, then obtain the content to be added corresponding to the preset problem type; The content to be added is added to the answer to obtain the answer after addition.

5. The answer generation method according to claim 1, characterized in that, The problem attribute information includes first problem attribute information and second problem attribute information. Extracting the problem attribute information to obtain the problem attribute information includes: The semantic analysis model is used to perform semantic analysis on the question to obtain the first question attribute information; Determine the proportion of the preset words in the question within its total vocabulary to obtain the second question attribute information.

6. The answer generation method according to claim 1, characterized in that, The step of matching the object attribute information with the problem attribute information to obtain a matching result includes: Mapping the object attribute information to an object feature vector, and mapping the problem attribute information to a problem feature vector; The object feature vector and the problem feature vector are matched using a fitness matching model to obtain a matching result.

7. The answer generation method according to claim 1, characterized in that, The question is in the form of an audio message. Before obtaining the object attribute information of the object to be answered, the process further includes: Extract the voiceprint features of the problem; Obtain the device identifier corresponding to the object to be answered; If the voiceprint feature matches a preset voiceprint feature, and the device identifier is the same as a preset device identifier, then the step of obtaining the object attribute information of the object to be answered is executed.

8. An answer generation device, characterized in that, include: The acquisition module is used to receive the question of the object to be answered and to acquire the object attribute information of the object to be answered; The extraction module is used to extract the attribute information of the problem to obtain the problem attribute information; The matching module is used to perform fit matching between the object attribute information and the problem attribute information to obtain a matching result; The generation module is used to determine a first answer generation strategy based on the matching results, and generate an answer to the question based on the first answer generation strategy.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the answer generation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the answer generation method as described in any one of claims 1 to 7.

Images