Intelligent question answering method and device based on large model and medium

Abstract

The invention provides an intelligent question answering method and device based on a large model and a medium, and the method comprises the steps: obtaining a first to-be-consulted event, carrying out the parameter analysis of the first to-be-consulted event based on the large model, obtaining a parameter result, and judging whether the parameter result meets a preset condition or not; if the parameter result does not meet the preset condition, generating a first inquiry verbal skill for modifying the first to-be-consulted event based on the large model, and obtaining a second to-be-consulted event according to the first inquiry verbal skill; in combination with the first to-be-consulted event and the second to-be-consulted event, analyzing first emotion information and a complete first target event based on a large model, and creating a corresponding keyword set, so that a user selects a target keyword from the keyword set; and obtaining a target answer style, and based on the target keyword, explaining the first target event in the target answer style to generate an answer description. Customized intelligent questions and answers can be provided for the user, and the user experience is improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent question answering technology, and in particular to a large model-based intelligent question answering method, device and medium. Background Art

[0002] The current online intelligent question-and-answer service uses semantic understanding of user text and extraction of keywords. However, when answering questions, it is unable to intelligently determine the completeness of the matter to be consulted, nor can it guide users to provide more complete responses, let alone provide users with personalized answers, which leads to a poor user experience. Summary of the invention

[0003] The purpose of this application is to address the deficiencies of the above-mentioned prior art and to propose a large-model-based intelligent question-answering method, device and medium that can generate more personalized answer descriptions, thereby providing users with customized intelligent questions and answers, thereby improving the user experience.

[0004] In a first aspect, the present application provides an intelligent question answering method based on a large model, comprising:

[0005] Obtain a first event to be consulted, perform parameter analysis on the first event to be consulted based on the big model, obtain a parameter result, and determine whether the parameter result meets a preset condition;

[0006] If the parameter result does not meet the preset condition, a first inquiry script is generated based on the large model to modify the first event to be consulted, and a second event to be consulted is obtained according to the first inquiry script;

[0007] Combining the first event to be consulted and the second event to be consulted, parsing the first emotion information and the complete first target event based on the large model, and respectively assigning different weights to the first emotion information and the first target event, creating a corresponding keyword set so that the user can select a target keyword from the keyword set;

[0008] A target answer style is obtained, and based on the target keyword, the first target event is explained in the target answer style to generate an answer description.

[0009] The present application adopts a large model to parse the consultation event, and determines whether the parameter result meets the preset conditions, and generates corresponding inquiry words for the situation that the preset conditions are not met, so as to guide the user to provide a more complete consultation event and facilitate the generation of richer answer descriptions; and, the consultation event is parsed into emotional information and target events, so as to obtain accurate questions to be consulted and take into account the user's current emotional state. By weighting the emotional information and target events based on different weights and combining the target answer style, a more personalized answer description can be generated, thereby providing users with customized intelligent questions and answers, thereby improving the user experience.

[0010] Further, the acquiring of the target answer style, and based on the target keyword, explaining the first target event in the target answer style to generate an answer description, includes:

[0011] Based on the large model, obtaining a target answer style determined by the user;

[0012] Matching the target answer style with a plurality of predefined modal particles respectively to determine the target modal particles;

[0013] The target keyword and the first target event are explained using the target modal particle to generate a response description.

[0014] Further, after generating the answer description, the method further includes:

[0015] Based on the large model, a fourth inquiry script is generated to inquire whether the user is satisfied with the answer description;

[0016] According to the fourth inquiry language, a follow-up event is obtained, the follow-up event is analyzed to obtain third emotional information and a third target event, and a follow-up target keyword is obtained based on the third emotional information and the third target event. If the target answer style is not modified, the follow-up target keyword, the target keyword, the first target event and the follow-up event are comprehensively explained using the target modal particle to generate a comprehensive answer description.

[0017] This application also uses a large model to conduct a survey on user satisfaction. If the user is not satisfied with the generated answer description, an inquiry script is generated to provide follow-up service, thereby realizing an online follow-up link, thereby improving the user experience.

[0018] Furthermore, the keyword set is displayed in a manner of randomly arranging candidate keywords, and the number of candidate keywords in the keyword set is proportional to the size of the event scale of the first event to be consulted; wherein, the size of the event scale is obtained by obtaining the parameter value of the event scale after judging the parameter results.

[0019] Further, the event scale includes: time range, object range and event type; wherein, performing parameter analysis on the first event to be consulted based on the big model to obtain parameter results, and judging whether the parameter results meet preset conditions, includes:

[0020] Performing problem judgment on the first event to be consulted based on the big model;

[0021] If the first event to be consulted is a question type, parsing the event scale of the first event to be consulted to obtain a parameter result;

[0022] It is determined whether the parameter result includes the time range, the object range, and the event type to obtain a result of whether the parameter result satisfies the integrity parameter.

[0023] Further, the step of obtaining whether the parameter result satisfies the result of the integrity parameter includes:

[0024] For the result of the judgment of the question of the first event to be consulted, a question sentence indicator is set, for the result of whether the integrity parameter is met, an integrity parameter indicator is set, and the first event to be consulted is set in the reply parameter;

[0025] A data structure of a preset format is generated according to the question indicator, the integrity parameter indicator and the reply parameter, and based on the large model and according to the data structure, a prompting speech requiring the input of a question is generated or the first inquiry speech is generated until the integrity parameter indicator is true.

[0026] Furthermore, after giving different weights to the first emotion information and the first target event respectively, creating a corresponding keyword set includes:

[0027] Based on the large model, a customized keyword set is constructed for the second emotion information and the second target event given different weights;

[0028] Alternatively, based on the large model, a keyword set matching the event type of the target event is selected from a plurality of preset keyword sets.

[0029] This application is based on a large model. It can not only select the most appropriate keyword set from multiple preset keyword sets to meet the answer description of simple events to be consulted, but also customize the generation of keyword sets to meet the answer description of complex events to be consulted, thereby realizing a variety of business scenarios and providing users with richer keyword sets, thereby improving user experience.

[0030] Furthermore, after explaining the first target event in the target answer style, the method further includes:

[0031] Acquire current time information, and determine the current time type based on the current time information;

[0032] Based on the current time type, extract corresponding description information from a preset event database;

[0033] Generate an answer description based on the current time type, the description information and the explanation information.

[0034] Based on a large model, this application parses the current time information and extracts corresponding descriptive information from a preset event database. It can combine the time type, descriptive information and explanation information to provide a more complete answer description related to the real information, thereby improving the user experience.

[0035] In a second aspect, the present application provides an intelligent question-answering device based on a large model, comprising: a parameter judgment module, a guidance module, a keyword set creation module and an answering module; wherein,

[0036] The parameter judgment module is used to obtain a first event to be consulted, perform parameter analysis on the first event to be consulted based on the big model, obtain a parameter result, and judge whether the parameter result meets a preset condition;

[0037] The guiding module is configured to generate, based on the large model, a first inquiry script for modifying the first event to be consulted if the parameter result does not meet the preset condition, and obtain a second event to be consulted according to the first inquiry script;

[0038] The keyword set creation module is used to combine the first event to be consulted and the second event to be consulted, parse out the first emotion information and the complete first target event based on the large model, and create a corresponding keyword set after giving different weights to the first emotion information and the first target event, so that the user can select a target keyword from the keyword set;

[0039] The answer module is used to obtain a target answer style, and based on the target keyword, explain the first target event in the target answer style to generate an answer description.

[0040] In a third aspect, the present application provides a readable storage medium storing a program or instruction, which, when executed by a processor, implements the steps of the large-model-based intelligent question-answering method as described in the first aspect and further limited to the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flowchart of an intelligent question-answering method based on a large model provided in an embodiment of the present application;

[0042] Figure 2 It is an interactive schematic diagram of another intelligent question-answering method based on a large model provided in an embodiment of the present application;

[0043] Figure 3 It is an interactive schematic diagram of the intelligent questioning link based on a large model provided in an embodiment of the present application;

[0044] Figure 4 It is an interactive schematic diagram of another intelligent question-answering method based on a large model provided in an embodiment of the present application;

[0045] Figure 5 It is an interactive schematic diagram of another intelligent question-answering method based on a large model provided in an embodiment of the present application;

[0046] Figure 6 It is another interactive schematic diagram of an intelligent questioning link based on a large model provided in an embodiment of the present application;

[0047] Figure 7 It is an interactive schematic diagram of another intelligent question-answering method based on a large model provided in an embodiment of the present application;

[0048] Figure 8 It is a structural diagram of an intelligent question-answering device based on a large model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0050] It is worth mentioning that the existing related technologies are either conducted offline when answering questions. If it is inconvenient for the user, the question and answer cannot be carried out immediately. If the question and answer is conducted online, it is impossible to provide users with intelligent and adaptive questions and answers, resulting in the inability to provide users with personalized questions and answers, which in turn leads to a low user experience.

[0051] Based on this, the present application provides a method, device and medium for intelligent question and answer based on a large model, which provides intelligent personalized question and answer through a large model, thereby improving the user experience. In order to more clearly illustrate the solution of the present application, the following embodiments will be described in detail.

[0052] Example 1

[0053] See also Figure 1, is a flow chart of an intelligent question-answering method based on a large model provided in an embodiment of the present application, including: steps S11 to S14, specifically:

[0054] Step S11: obtain a first event to be consulted, perform parameter analysis on the first event to be consulted based on the big model, obtain parameter results, and determine whether the parameter results meet preset conditions.

[0055] It is worth noting that the first, second, third or fourth do not indicate the order of precedence, but only distinguish between the same nouns from different sources. The first, second, third or fourth can be understood as names. It can be understood that the first event to be consulted is the first event to be consulted of the user, and the second event to be consulted is the event to be consulted obtained after the first event to be consulted is guided after the parameter judgment based on the first event to be consulted when the first event to be consulted does not meet the integrity parameters.

[0056] The large model is an abbreviation of a large language model (LLM). In some embodiments, the large model may be a series model of ChatGPT, specifically, the large model may be ChatGPT4.

[0057] In some embodiments, in addition to ChatGPT4, the large model can also be a Chinese-based ERNIE3.0 Titan model, a general language model (General Language Model, GLM)-130B model providing Chinese and English, a Pathways architecture-based PaLM model (Pathways Language Model) providing multiple languages, a BLOOM (Bidirectional and Self-Enforcing Neural Network with Oriented Optimization for Understanding and Generation) model, or an LLaMA (Large Language Model Meta AI) model.

[0058] In some embodiments, determining whether the parameter result satisfies a preset condition includes determining completeness and / or accuracy.

[0059] In some embodiments, the integrity judgment includes: performing problem judgment on the first event to be consulted based on the big model; if the first event to be consulted is a question type, parsing the time range, object range and event type of the first event to be consulted to obtain parameter results, and judging whether the parameter results meet the preset conditions.

[0060] In some embodiments, the integrity judgment is specifically: judging whether the parameter result includes a time range, an object range, and an event type, so as to obtain whether the parameter result satisfies the result of the integrity parameter.

[0061] In some embodiments, the event scale includes: time range, object range and event type; the parameter value of the event scale includes: specific values ​​of the time range, object range and event type.

[0062] In some embodiments, the integrity judgment includes: setting a question indicator for the result of the judgment of the first event to be consulted, setting an integrity parameter indicator for the result of whether the integrity parameter is met, and setting the first event to be consulted in the reply parameter; generating a data structure in a preset format according to the question indicator, the integrity parameter indicator and the reply parameter, and based on the big model and according to the data structure, generating a prompt script that requires the input of a question or generating a first inquiry script until the integrity parameter indicator is true.

[0063] In some embodiments, a problem judgment is performed on the first event to be consulted based on the big model, including: judging the sentence type of the first event to be consulted, if the sentence type of the first event to be consulted is a question type, it indicates that the sentence input by the user has an event to be consulted, and the question indicator is set to true to characterize this; if the sentence type of the first event to be consulted is not a question type, the question indicator is false, and it is necessary to generate a prompt script for inputting a question to guide the user to input an event to be consulted of the question type; if the completeness parameter indicator is false, it is impossible to understand the user's intention, resulting in an inability to answer, and it is necessary to generate a first inquiry script to guide the user to complete the event to be consulted.

[0064] In some embodiments, if the sentence type of the first event to be consulted is a question type, the corresponding question indicator is true, otherwise it is false.

[0065] In some embodiments, the time range includes: within K months, or one year; where K is a positive integer not greater than 12.

[0066] In some embodiments, the object range includes: L individuals; L is a positive integer.

[0067] In some embodiments, if the parameter result does not include at least any one of: time range, object range or event type, the integrity parameter is not satisfied and the corresponding integrity parameter indicator is false; otherwise, the corresponding integrity parameter indicator is true.

[0068] In other embodiments, if the parameter result includes at least one of the time range, object range and event type, the integrity parameter is satisfied and the corresponding integrity parameter indicator is true; otherwise, the corresponding integrity parameter indicator is false.

[0069] In some embodiments, when the integrity parameter indicator is false, that is, when the integrity parameters are not met, it is necessary to output the inquiry words based on the big model to actively guide the user to complete the events to be consulted; when the integrity parameter indicator is true, that is, when the integrity parameters are met, the user is not actively guided, but if the user actively supplements the events to be consulted, it is still necessary to combine the supplemented events to be consulted with the first input events to be consulted to output the answer description.

[0070] In some embodiments, a user inputs an event to be consulted, and based on the big model, a problem judgment is performed on the event to be consulted to identify whether the event to be consulted is a question type. If the event to be consulted is a question type, parameter analysis is performed on the event to be consulted to obtain parameter results, and an integrity parameter judgment is performed on the parameter results to determine whether the event to be consulted meets the consultation conditions. If the integrity parameters are not met, it is necessary to generate an inquiry script based on the big model to guide the user to complete the event to be consulted; if the integrity parameters are met, a confirmation button pops up on the interface, and the user clicks confirm to proceed to the next step; in the process of the user completing the event to be consulted, the user can make multiple supplements or modifications until the confirmation button is clicked.

[0071] In some embodiments, the accuracy judgment is specifically as follows: a semantic similarity threshold is set in advance, and the large model uses a semantic similarity algorithm to judge the similarity between the user input, that is, the first event to be consulted, and the large model, that is, the parameter result. When the similarity meets the preset similarity threshold, it means that the large model’s understanding of the user’s input first event to be consulted is accurate, and the user’s input first event to be consulted can be directly parsed to generate a corresponding answer description; if the preset similarity threshold is not met, then it is necessary to generate a first inquiry script based on the large model to modify the first event to be consulted, so as to obtain the second-generation consultation event input by the user. It should be noted that the similarity algorithm can be, but is not limited to, distance similarity, cosine similarity, Sentence-BERT or Semantic Textual Similarity, etc.

[0072] In some embodiments, accuracy judgment includes: after obtaining the first generation of consulting events, it is necessary to perform word segmentation and part-of-speech tagging on the first event to be consulted to identify key information and vocabulary, and then the big model performs parameter parsing on the first generation of consulting events to obtain parameter results, and judge whether the parameter results contain key information and vocabulary. If it contains key information and vocabulary, the first event to be consulted input by the user can be directly parsed to generate a corresponding answer description; if it does not contain key information and vocabulary, it is necessary to generate a first inquiry script based on the big model to modify the first event to be consulted to obtain the second generation of consulting events input by the user.

[0073] It is worth noting that traditional technology is unable to make precise judgments on user input, including: judging the questions asked by the user and useless modal particles. Traditional technology cannot judge whether the user input is a specific question when a bunch of modal particles and meaningless phrases, or a sentence that does not contain a formed question, nor can it make a very humanized reply. Of course, traditional technology cannot judge the accuracy of the analysis results. The embodiment of the present application can provide humanized services by analyzing and judging the input event to be consulted, correcting the problem and providing problem confirmation, thereby improving the user experience.

[0074] Furthermore, due to the uncontrollable user input, if traditional technology wants to correct the event to be consulted, it can only be achieved by introducing manual customer service, resulting in high labor costs and low service efficiency. It is precisely because the cost-effectiveness of manual customer service is too low that in the question-and-answer scenario, traditional technology will not add the step of correcting the event to be consulted in the pre-step. The embodiment of the present application implements the pre-step of correcting the event to be consulted based on the big model, which can obtain the event to be consulted with more complete information, so that a richer answer description can be given, thereby improving the user experience. Moreover, when correcting the event to be consulted by parsing and judging the event to be consulted, if the user enters some emotional text, the emotional information and target event of the consultation event are separated based on the big model, and a keyword set is created based on different weights, which can also provide users with more personalized services, thereby improving the user experience.

[0075] Step S12: If the parameter result does not meet the preset conditions, a first inquiry script is generated based on the large model to modify the first event to be consulted, and a second event to be consulted is obtained based on the first inquiry script.

[0076] It is worth mentioning that the first inquiry dialogue is an inquiry dialogue generated based on the first event to be consulted when the first event to be consulted does not meet the preset conditions; in the following, the second inquiry dialogue is a dialogue used to ask the user which keyword to select as the target keyword, the third inquiry dialogue is a dialogue used to ask the user which answer style to use, and the fourth inquiry dialogue is a dialogue used to ask the user whether he is satisfied with the answer description.

[0077] In some embodiments, whether the preset conditions are met is determined based on the response parameter results corresponding to all the received events to be consulted. If the preset conditions are still not met, the inquiry script is continued to be generated to guide the user to complete the missing parameters.

[0078] In some embodiments, it is necessary to perform parameter analysis on the event to be consulted that is modified or supplemented by the user, obtain reply parameters, and determine whether the reply parameters are the same as the event that the user wants to consult, that is, determine whether the large model analysis is accurate. The specific method for accurate judgment may be but is not limited to: similarity judgment method and keyword labeling, etc. The specific method has been expanded in step S11 and will not be repeated here.

[0079] In some embodiments, it is determined whether the input event to be consulted is a question sentence to obtain whether there is a question that the user wants to consult in the event to be consulted. If not, the question indicator isQuestion is set to false, otherwise it is set to true.

[0080] In some embodiments, determining whether the input event to be consulted is suitable includes: determining in turn whether the event to be consulted is within the time range, object range and time type. If the integrity parameters are met, it indicates that the event to be consulted is suitable, and the integrity parameter indicator isSuitable is set to true, otherwise it is set to false.

[0081] In some embodiments, if the question indicator isQuestion is true, a button to confirm the question pops up.

[0082] In some embodiments, the question indicator isQuestion is used to make the generated answer description better, so that the AI ​​understands that when the question indicator isQuestion is false, the user can be asked whether to modify or supplement the event to be consulted.

[0083] In some embodiments, the pending consultation event to be replied to the user is placed in the reply parameter reply, and the question in the modified pending consultation event is placed in the question parameter editQuestion.

[0084] In some embodiments, each time the user modifies or supplements the question to be consulted, the generated event to be consulted is placed in the reply parameter, and a data structure in a preset format is generated by combining the completeness parameter index and the question index.

[0085] In some embodiments, the data structure in the preset format may be any one of a JSON format, a JSON5 format, a STOB format, an ENO format, an AXON format, or a BRE format.

[0086] In some embodiments, the data structure generated in JSON format can be represented as:

[0087] {

[0088] "isQuestion":false,

[0089] "isSuitable":false,

[0090] "reply":"It sounds like you don't have a question. You can ask...",

[0091] "editQuestion":""

[0092] }

[0093] Step S13, combining the first event to be consulted and the second event to be consulted, based on the big model, parsing the first emotional information and the complete first target event, and after giving different weights to the first emotional information and the first target event respectively, creating a corresponding keyword set to allow the user to select the target keyword from the keyword set.

[0094] It is worth explaining that the first emotion information and the first target event are the total events to be consulted after combining the user's first event to be consulted and the event to be consulted supplemented by the user after the guidance of the first event to be consulted, and the emotion information and the target event are obtained after analyzing the total events to be consulted; the second emotion information and the second target event are the result of giving different weights to the first emotion information and the first target event respectively.

[0095] In some embodiments, the emotional information includes: at least one of the user's emotions, interjections used by the user, emotional word frequency, interjection word frequency, emotions expressed by images, or image frequency. The emotions expressed by images include: emotional expressions of emoticons or emotional expressions of user-defined pictures.

[0096] In some embodiments, after giving different weights to the first emotion information and the first target event, respectively, corresponding keyword sets are created, including: based on the big model, building a custom keyword set for the second emotion information and the second target event given different weights; or, based on the big model, selecting a keyword set that matches the event type of the target event from a plurality of preset keyword sets.

[0097] It is worth explaining that the keyword set can be understood as the description direction determined by the big model. After the big model determines the keyword set, it will be displayed to the user. On the one hand, the user can select several target keywords from the keyword set to explain the first target event based on the target keywords and the target answer style and generate an answer description. On the other hand, if the user does not make a selection, the big model can directly explain the first target event based on the keyword set and in the target answer style to generate an answer description.

[0098] The embodiment of the present application is based on a big model, which not only allows users to select the most appropriate keyword set from multiple preset keyword sets to meet the answer description of simple events to be consulted, but also allows the big model to custom generate keyword sets to meet the answer description of complex events to be consulted, thereby realizing a variety of business scenarios and providing users with richer keyword sets, thereby improving user experience.

[0099] In some embodiments, the weight of the second emotion information is not greater than the weight of the second target event.

[0100] In some embodiments, the weight of the second emotion information is less than the weight of the second target event.

[0101] In some embodiments, if the event type of the first event to be consulted or the second event to be consulted and the event scale corresponding to the event type are relatively simple, a keyword set matching the event type can be directly selected from multiple preset keyword sets.

[0102] For example, if the target event obtained according to the first event to be consulted and the second event to be consulted is which one to choose between two options, a keyword set with three keywords or other relatively simple keyword sets can be directly selected from the preset multiple keyword sets.

[0103] In some embodiments, if the time range involved in the target event is large, the target event can be regarded as a complex and large problem. From the preset multiple keyword sets, a keyword set of thirteen keywords can be selected based on the large model; wherein, after the user selects the first keyword, he selects the second keyword. When explaining the second keyword, the explanation is based on the answer description of the first keyword explanation, and the final comprehensive answer description is generated.

[0104] In some embodiments, for different target events, a keyword set may select a suitable keyword set of different keywords.

[0105] Step S14: Obtain a target answer style, and based on the target keyword, explain the first target event in the target answer style to generate an answer description.

[0106] In some embodiments, a target answer style is obtained, and based on the target keyword, the first target event is explained in the target answer style to generate an answer description, including: based on a large model, obtaining the target answer style determined by the user, matching the target answer style with a plurality of predefined modal particles respectively, determining the target modal particles, and explaining the target keyword and the first target event in the target modal particles to generate an answer description.

[0107] In some embodiments, by predefining modal particles in Prompt as gentle or harsh, and adjusting the basic content of the target keyword in the database according to the predefined modal particles, the description style and tone bias when explaining each target keyword can be changed, a customized answer style can be generated, and personalized answer descriptions can be provided to users, thereby improving the user experience.

[0108] It is worth mentioning that the target answer style is the overall style of the large model answer description, while the target modal particle is the style and tone bias of the description of each target keyword. The two are different concepts.

[0109] In some embodiments, based on the big model, target keywords and target answer styles are obtained respectively, including: based on the big model, a second inquiry dialogue for inquiry selection and a third inquiry dialogue for inquiry answer style are sequentially generated, and the target keywords and target answer styles are obtained correspondingly according to the second inquiry dialogue and the third inquiry dialogue.

[0110] In some embodiments, based on the target keywords, the first target event is explained in a target answer style to generate an answer description, including: extracting the basic definition of the target keyword from a database, combining the basic definition with the target modal particle based on a large model, explaining the first target event, and generating an answer description.

[0111] In some embodiments, after generating the answer description, a language type of inquiry answer description is generated based on the big model, and based on the language type of inquiry answer description, the target language type is obtained, and the answer description is output according to the target language type.

[0112] In some embodiments, based on the large model of ChatGPT4, basic content in multiple languages ​​is set in the database, including at least three basic languages: Simplified Chinese, Traditional Chinese and English, and may also include: Indian languages, Iranian languages, other Germanic languages ​​except English, foreign language types in the Romance languages ​​or Slavic languages, etc.

[0113] In some embodiments, the database data may affect the answering style, and therefore, there may be multiple different databases, for example, for specific consulting areas, or different language styles, there may be data with a more considerate tone, or data with a more severe and precise tone, and so on.

[0114] In some embodiments, according to data type classification, the data classification in the database includes: the meaning of the keyword and the meaning of the keyword corresponding to the description picture.

[0115] In some embodiments, after generating the answer description, the method includes: based on the big model, generating a fourth inquiry script to inquire whether the answer description is satisfactory; according to the fourth inquiry script, obtaining a follow-up event, parsing the follow-up event to obtain third emotional information and a third target event, and obtaining a follow-up target keyword based on the third emotional information and the third target event; if the target answer style is not modified, comprehensively explaining the follow-up target keyword, target keyword, first target event and follow-up event with the target modal particle to generate a comprehensive answer description.

[0116] It is worth noting that asking whether the user is satisfied with the answer description is actually to guide the user and provide the user with a follow-up questioning link. The user's target event and the answer description generated by the large model are used as pre-information. After the user inputs the follow-up question event, the pre-information and the follow-up question event are used as inputs of the large model, so that the large model outputs a new comprehensive answer description. Although the follow-up questioning link is a link that has always been maintained offline, it is difficult to implement with traditional online technology because the user's input cannot be controlled and traditional technology cannot understand diverse semantics. Therefore, the embodiment of the present application implements this link online, which is conducive to subsequent communication after the user has answered, thereby improving the user experience.

[0117] It is worth noting that, since the user's emotional information is dynamically changing, when the user's follow-up event is obtained in the follow-up link, the user's emotional information and the target event in the follow-up event will be analyzed in real time to determine the user's current emotional value, and the user's current emotional value will be recorded. When generating a comprehensive answer description for the user, the user's current emotional value will be combined to give appropriate comfort. For example, but not limited to, when the user answers the first follow-up event, the big model analyzes the user's current emotional information and scores the user's emotional information. If the user is currently sad and his emotional value is 30 points, then the big model will extract the corresponding comfort description from the preset event library to comfort the user in the answer description. If the user answers the next follow-up question, it is analyzed that the user's emotional information has improved, then the emotional value will increase, and the comfort description will be appropriately modified. Among them, the preset event library is pre-set, containing multiple key-value pairs of emotional states and corresponding emotional values, the emotional state is the key, including but not limited to pleasure, sadness and anger, etc., and the emotional value is the corresponding score.

[0118] In some embodiments, the query target keyword and the target keyword may be extracted from an online virtual keyword set.

[0119] In some embodiments, according to the fourth inquiry script, the user inputs a follow-up event, and a keyword set is given again according to the follow-up event, allowing the user to extract the target keyword again, combining the previous target keyword and the latest extracted target keyword, and then making further answers to the user's follow-up event.

[0120] In some embodiments, it can be specified in the Prompt: "The target keyword is..., the latest question target keyword drawn is..., and the question event is...", so that the big model can combine this information to generate a comprehensive answer description.

[0121] In some embodiments, if the target answer style is not modified, the target modal particles can be used to comprehensively explain the follow-up target keywords, target keywords, first target events and follow-up events to generate a comprehensive answer description; the target modal particles can also be used to comprehensively explain the follow-up target keywords, first target events and follow-up events to generate a comprehensive answer description; the target modal particles can also be used to directly explain the follow-up target keywords and follow-up events to generate a comprehensive answer description.

[0122] In some embodiments, if the user input is obtained to modify the target answer style, the target answer style is changed, and the changed target answer style is re-matched with multiple predefined modal particles to obtain updated target modal particles, and the updated target modal particles are used to comprehensively explain the follow-up target keyword, target keyword, first target event and follow-up event to generate a comprehensive answer description.

[0123] In some embodiments, the updated target modal particles are used to comprehensively explain the follow-up target keywords, the first target event and the follow-up event to generate a comprehensive answer description. The updated target modal particles can also be used to comprehensively explain the follow-up target keywords and the follow-up event to generate a comprehensive answer description.

[0124] In some embodiments, when selecting a target keyword from a randomly arranged keyword set online, the user clicks one or more keywords as the target keyword.

[0125] In some embodiments, the user may also select target keywords from a randomly arranged keyword set offline. The user may select the target keywords offline and then directly provide the keyword set and the extracted target keywords online. Alternatively, the user may select the target keywords offline and upload photos or videos of the extracted keywords so that the uploaded photos or videos can be identified with the target keywords based on the large model.

[0126] It is worth noting that the online target keyword extraction in the embodiment of the present application is not much different from the traditional online target keyword extraction, but in the offline target keyword extraction method, a product equipped with a large model can be used to take photos or videos of user extraction, and perform target detection and recognition, which has multiple usage scenarios and higher applicability.

[0127] It is worth noting that all the above embodiments can be combined and arranged to obtain new embodiments that still fall within the protection scope of the embodiments of the present application.

[0128] In some embodiments, after explaining the first target event in the target answer style, it also includes: obtaining current time information, and determining the current time type based on the current time information; based on the current time type, extracting corresponding description information from a preset event database; based on the current time type, the description information and the explanation information, generating an answer description.

[0129] In some embodiments, after the big model explains the target keyword and the first target event with the target modal particle, it needs to obtain the current time information and determine the current time type based on the current time information, such as "Spring Festival", "Mid-Autumn Festival", "National Day", "Monday", "Sunday" and other types. After determining the current time type, the big model will extract the corresponding description information from the preset event database, wherein the event database contains key-value pairs of pre-set time types and description information, with the time type as the key and the description information as the pair; finally, the big model integrates the current time type and the description information into the explanation information to generate an answer description.

[0130] Based on a large model, this application parses the current time information and extracts corresponding descriptive information from a preset event database. It can combine the time type, descriptive information and explanation information to provide a more complete answer description related to the real information, thereby improving the user experience.

[0131] Exemplarily, the above steps are illustrated.

[0132] In some embodiments, see Figure 2 , is an interactive schematic diagram of another intelligent question-answering method based on a large model provided in an embodiment of the present application. Figure 2 In the process, the question type and integrity parameter of the first event to be consulted input by the user need to be judged multiple times until the event to be consulted input by the user meets the question type and integrity parameter. Figure 2 In the process, users select offline entities and upload corresponding photos or videos, which are then recognized by an AI intelligent question-answering robot based on a large model to obtain target keywords.

[0133] In some embodiments, see Figure 3 , is an interactive schematic diagram of the intelligent questioning link based on a large model provided in an embodiment of the present application, Figure 3 If the user is not satisfied with the generated answer description, he or she can provide feedback for a follow-up question, wherein the question type and completeness parameters of the follow-up question event are judged multiple times until the follow-up question event input by the user meets the question type and completeness parameters, and then a comprehensive explanation is performed to generate a comprehensive answer description.

[0134] In some embodiments, see Figure 4 , is an interactive schematic diagram of another intelligent question-answering method based on a large model provided in an embodiment of the present application. Figure 4 In the process, users virtually input target keywords online.

[0135] In some embodiments, see Figure 5 , is an interactive schematic diagram of another intelligent question-answering method based on a large model provided in an embodiment of the present application. Figure 5 In the example, if the first event to be consulted input by the user meets the question type and completeness parameters, the first event to be consulted is directly parsed, and after the user selects the target keyword through the offline entity, the corresponding photo or video is uploaded, and the AI ​​intelligent question-answering robot based on the large model recognizes it and generates an answer description after obtaining the target keyword.

[0136] In some embodiments, see Figure 6 , is another interactive schematic diagram of an intelligent questioning link based on a large model provided in an embodiment of the present application, and Figure 3 compared to, Figure 6 In the appending stage, if the additional target keywords input by the user directly meet the question type and completeness parameters, a comprehensive explanation will be directly performed to generate a comprehensive answer description.

[0137] In some embodiments, see Figure 7 , is an interactive schematic diagram of another intelligent question-answering method based on a large model provided in an embodiment of the present application, and Figure 5 The difference is, Figure 7 Users in the website select virtual target keywords online.

[0138] The embodiment of the present application adopts a large model to parse the consultation event, and performs integrity parameter judgment on the obtained parameter results, and generates corresponding inquiry words for the situation that does not meet the integrity parameters, so as to guide the user to provide a more complete consultation event and facilitate the generation of richer answer descriptions; and, the consultation event is parsed into emotional information and target events, so as to obtain accurate questions to be consulted, and based on different weights for emotional information and target events, combined with the target answer style, a more personalized answer description can be generated, thereby providing users with customized intelligent questions and answers, thereby improving the user experience.

[0139] Example 2

[0140] See also Figure 8 , is a structural diagram of an intelligent question-answering device based on a large model provided in an embodiment of the present application, comprising: a parameter judgment module 21, a guidance module 22, a keyword set creation module 23 and an answering module 24.

[0141] In some embodiments, the parameter judgment module 21 inputs parameter results to the guidance module 22 and the keyword set creation module 23 based on the first event to be consulted. The guidance module 22 outputs the second event to be consulted based on the input parameter results and the first event to be consulted, and transmits the first event to be consulted and the second event to be consulted to the keyword set creation module 23. The keyword set creation module 23 parses the first target event and creates a keyword set based on the first event to be consulted and the second event to be consulted, outputs the corresponding target keyword based on the keyword set, and inputs the first target event and the target keyword to the answer module 24. The answer module 24 outputs an answer description based on the input target keyword and the first event to be consulted.

[0142] The parameter judgment module 21 is used to obtain the first event to be consulted, perform parameter analysis on the first event to be consulted based on the big model, obtain parameter results, and judge whether the parameter results meet preset conditions.

[0143] In some embodiments, parameters of the first event to be consulted are parsed based on the big model to obtain parameter results, and it is determined whether the parameter results meet preset conditions, including: question judgment is performed on the first event to be consulted based on the big model, if the first event to be consulted is a question type, the event scale is parsed for the first event to be consulted to obtain parameter results, and it is determined whether the parameter results include a time range, an object range and an event type, so as to obtain a result of whether the parameter results meet the integrity parameters.

[0144] In some embodiments, the event scale includes: time range, object range, and event type.

[0145] In some embodiments, to obtain the result of whether the parameter result satisfies the integrity parameter, including: setting a question indicator for the result of judging the first event to be consulted, setting the integrity parameter indicator for the result of whether the integrity parameter is satisfied, and setting the first event to be consulted in the reply parameter; generating a data structure in a preset format according to the question indicator, the integrity parameter indicator and the reply parameter, and based on the big model and according to the data structure, generating a prompting speech that requires input of a question or generating a first inquiry speech until the integrity parameter indicator is true.

[0146] The guiding module 22 is used to generate a first inquiry script to modify the first event to be consulted based on the large model if the parameter result does not meet the preset conditions, and obtain the second event to be consulted according to the first inquiry script.

[0147] A keyword set creation module 23 is used to combine the first event to be consulted and the second event to be consulted, parse out the first emotional information and the complete first target event based on the big model, and give different weights to the first emotional information and the first target event respectively, and then create a corresponding keyword set to allow the user to select the target keyword from the keyword set.

[0148] In some embodiments, after giving different weights to the first emotion information and the first target event, respectively, corresponding keyword sets are created, including: based on the big model, building a custom keyword set for the second emotion information and the second target event given different weights; or, based on the big model, selecting a keyword set that matches the event type of the target event from a plurality of preset keyword sets.

[0149] The answer module 24 is used to obtain a target answer style, and based on the target keyword, explain the first target event in the target answer style to generate an answer description.

[0150] In some embodiments, a target answer style is obtained, and based on the target keyword, the first target event is explained in the target answer style to generate an answer description, including: based on a large model, obtaining the target answer style determined by the user, matching the target answer style with a plurality of predefined modal particles respectively, determining the target modal particles, and explaining the target keyword and the first target event in the target modal particles to generate an answer description.

[0151] In some embodiments, target keywords and target answer styles are obtained respectively, including: based on the large model, sequentially generating a second inquiry dialogue for inquiring about the target keyword and a third inquiry dialogue for inquiring about the answer style, and obtaining the target keywords and target answer styles accordingly according to the second inquiry dialogue and the third inquiry dialogue.

[0152] In some embodiments, after generating the answer description, it includes: after generating the answer description, it includes: based on the big model, generating a fourth inquiry script to inquire whether the answer description is satisfactory, and according to the fourth inquiry script, obtaining a follow-up event, parsing the follow-up event to obtain third emotional information and a third target event, and obtaining a follow-up target keyword based on the third emotional information and the third target event; if the target answer style is not modified, the follow-up target keyword, target keyword, first target event and follow-up event are comprehensively explained with the target modal particles to generate a comprehensive answer description.

[0153] In some embodiments, the keyword set is displayed in a randomly arranged manner with candidate keywords, and the number of candidate keywords in the keyword set is proportional to the size of the event scale of the first event to be consulted; wherein, the size of the event scale is obtained by obtaining the parameter value of the event scale after judging the parameter results.

[0154] The present application adopts a large model to parse the consultation event, and determines whether the parameter result meets the preset conditions, and generates corresponding inquiry words for the situation that the preset conditions are not met, so as to guide the user to provide a more complete consultation event and facilitate the generation of richer answer descriptions; and, the consultation event is parsed into emotional information and target events, so as to obtain accurate questions to be consulted and take into account the user's current emotional state. By weighting the emotional information and target events based on different weights and combining the target answer style, a more personalized answer description can be generated, thereby providing users with customized intelligent questions and answers, thereby improving the user experience.

[0155] Example 3

[0156] It is a readable storage medium provided in an embodiment of the present application; wherein, the readable storage medium stores programs or instructions, and when the programs or instructions are executed by the processor, the steps of the intelligent question-answering method based on a large model as described in Example 1 are implemented.

[0157] When the embodiment of the present application stores the large model-based intelligent question-answering method as a program in a storage medium, it can be run on different systems or different products by running or reading the file processing program in the storage medium. It can be applicable to a variety of systems and electronic products, and has a wider range of application scenarios and stronger scalability.

[0158] Those skilled in the art will appreciate that the embodiments of the present application may also provide computer program products. Therefore, the present application may adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application may adopt the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program codes.

[0159] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0160] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0161] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0162] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. An intelligent question answering method based on a large model, characterized in that: include: Obtain a first event to be consulted, perform parameter analysis on the first event to be consulted based on the big model, obtain a parameter result, and determine whether the parameter result meets a preset condition; If the parameter result does not meet the preset condition, a first inquiry script is generated based on the large model to modify the first event to be consulted, and a second event to be consulted is obtained according to the first inquiry script; Combining the first event to be consulted and the second event to be consulted, parsing the first emotion information and the complete first target event based on the large model, and respectively assigning different weights to the first emotion information and the first target event, creating a corresponding keyword set so that the user can select a target keyword from the keyword set; A target answer style is obtained, and based on the target keyword, the first target event is explained in the target answer style to generate an answer description.

2. The intelligent question-answering method according to claim 1, characterized in that: The step of obtaining a target answer style, and based on the target keyword, explaining the first target event in the target answer style to generate an answer description includes: Based on the large model, obtaining a target answer style determined by the user; Matching the target answer style with a plurality of predefined modal particles respectively to determine the target modal particles; The target keyword and the first target event are explained using the target modal particle to generate a response description.

3. The intelligent question-answering method according to claim 2, characterized in that: After the generated answer description, include: Based on the large model, a fourth inquiry script is generated to inquire whether the user is satisfied with the answer description; According to the fourth inquiry speech, a follow-up event is obtained, the follow-up event is parsed to obtain third emotion information and a third target event, and a follow-up target keyword is obtained based on the third emotion information and the third target event; If the target answer style is not modified, the target interrogation target keyword, the target keyword, the first target event and the interrogation event are comprehensively explained using the target modal particle to generate a comprehensive answer description.

4. The intelligent question-answering method according to claim 1, wherein: The keyword set is displayed in a manner of randomly arranging candidate keywords, and the number of candidate keywords in the keyword set is proportional to the size of the first event to be consulted; wherein, the size of the event scale is obtained by obtaining the parameter value of the event scale after judging the parameter result.

5. The intelligent question-answering method according to claim 4, characterized in that: The event scale includes: time range, object range and event type; wherein, performing parameter analysis on the first event to be consulted based on the big model to obtain parameter results, and judging whether the parameter results meet the preset conditions, includes: Performing problem judgment on the first event to be consulted based on the big model; If the first event to be consulted is a question type, parsing the event scale of the first event to be consulted to obtain a parameter result; It is determined whether the parameter result includes the time range, the object range, and the event type to obtain a result of whether the parameter result satisfies the integrity parameter.

6. The intelligent question-answering method according to claim 5, characterized in that: The step of obtaining whether the parameter result satisfies the result of the integrity parameter includes: For the result of the judgment of the question of the first event to be consulted, a question sentence indicator is set, for the result of whether the integrity parameter is met, an integrity parameter indicator is set, and the first event to be consulted is set in the reply parameter; A data structure of a preset format is generated according to the question indicator, the integrity parameter indicator and the reply parameter, and based on the large model and according to the data structure, a prompting speech requiring the input of a question is generated or the first inquiry speech is generated until the integrity parameter indicator is true.

7. The intelligent question-answering method according to claim 1, characterized in that: After giving different weights to the first emotion information and the first target event respectively, creating a corresponding keyword set includes: Based on the large model, a customized keyword set is constructed for the second emotion information and the second target event given different weights; Alternatively, based on the large model, a keyword set matching the event type of the target event is selected from a plurality of preset keyword sets.

8. The intelligent question-answering method according to claim 1, characterized in that: After explaining the first target event in the target answer style, the method further includes: Acquire current time information, and determine the current time type based on the current time information; Based on the current time type, extract corresponding description information from a preset event database; Generate an answer description based on the current time type, the description information and the explanation information.

9. An intelligent question-answering device based on a large model, characterized in that: include: Parameter judgment module, guidance module, keyword set creation module and answer module; among them, The parameter judgment module is used to obtain a first event to be consulted, perform parameter analysis on the first event to be consulted based on the big model, obtain a parameter result, and judge whether the parameter result meets a preset condition; The guiding module is configured to generate, based on the large model, a first inquiry script for modifying the first event to be consulted if the parameter result does not meet the preset condition, and obtain a second event to be consulted according to the first inquiry script; The keyword set creation module is used to combine the first event to be consulted and the second event to be consulted, parse out the first emotion information and the complete first target event based on the large model, and create a corresponding keyword set after giving different weights to the first emotion information and the first target event, so that the user can select a target keyword from the keyword set; The answer module is used to obtain a target answer style, and based on the target keyword, explain the first target event in the target answer style to generate an answer description.

10. A readable storage medium, characterized in that: The readable storage medium stores programs or instructions, and when the programs or instructions are executed by the processor, the steps of the large model-based intelligent question-answering method as described in any one of claims 1 to 8 are implemented.

Images