Dialogue evaluation method and device, medium, equipment and product

Abstract

A dialogue assessment method and apparatus, a medium, a device and a product, the method comprising: obtaining first dialogue associated information and a first cue word corresponding to a pre-stored exception type, the first cue word being used for guiding a model to assess whether an exception corresponding to the exception type exists in the first dialogue, the first dialogue being a dialogue between a user and a virtual role, the first dialogue associated information comprises information which is associated with the first dialogue and is used for evaluating whether the first dialogue is abnormal or not; generating first model input information according to the first dialogue associated information and the first cue word; and obtaining a first evaluation result of the first dialogue through the first large language model, wherein the first large language model is used for obtaining the first evaluation result according to the first model input information. According to the technical scheme, more targeted prompt information can be provided for the first large language model, and a basis is provided for the first large language model to more accurately evaluate whether the first dialogue has the exception corresponding to the exception type or not.

Description

Technical Field

[0001] The present disclosure relates to the technical field of data processing, and in particular, to a dialogue evaluation method, device, medium, equipment and product. Background Art

[0002] With the development of artificial intelligence technology, intelligent dialogue systems are increasingly widely used in various fields. For example, intelligent customer service systems can handle user inquiries at any time, intelligent Q&A systems can answer user questions at any time, and chatbots can understand users' emotions and needs and have chat conversations with users.

[0003] Intelligent dialogue systems have the advantage of fast response speed. In order to analyze user satisfaction and system response quality, it is necessary to evaluate the dialogue content in intelligent dialogue systems. Currently, the evaluation of dialogue content is mainly carried out by manual sampling inspection, with a small coverage range and low efficiency. Summary of the Invention

[0004] This Summary of the Invention section is provided to introduce concepts in a brief form, which will be described in detail in the following Detailed Description section. This Summary of the Invention section is not intended to identify key features or essential features of the claimed technical solution, nor is it intended to be used to limit the scope of the claimed technical solution.

[0005] In a first aspect, the present disclosure provides a dialogue evaluation method, the method comprising: Obtaining first dialogue association information and first prompt words corresponding to pre-stored abnormal types, the first prompt words being used to guide a model to evaluate whether the first dialogue has an abnormality corresponding to the abnormal type, the first dialogue being a dialogue between a user and a virtual character, and the first dialogue association information including information associated with the first dialogue and required for evaluating whether the first dialogue has the abnormality; Generating first model input information according to the first dialogue association information and the first prompt words; Obtaining a first evaluation result of the first dialogue through a first large language model, the first large language model being used to obtain the first evaluation result according to the first model input information.

[0006] In a second aspect, the present disclosure provides a dialogue evaluation device, the device comprising: A first obtaining module, configured to obtain first dialogue association information and first prompt words corresponding to pre-stored abnormal types, the first prompt words being used to guide a model to evaluate whether the first dialogue has an abnormality corresponding to the abnormal type, the first dialogue being a dialogue between a user and a virtual character, and the first dialogue association information including information associated with the first dialogue and required for evaluating whether the first dialogue has the abnormality; A generation module, configured to generate first model input information according to the first conversation association information and the first prompt word; A second acquisition module, configured to obtain a first evaluation result of the first conversation through a first large language model, where the first large language model is configured to obtain the first evaluation result according to the first model input information.

[0007] In a third aspect, the present disclosure provides a computer-readable medium, on which a computer program is stored, and when the computer program is executed by a processing device, the steps of the conversation evaluation method provided in the first aspect of the present disclosure are implemented.

[0008] In a fourth aspect, the present disclosure provides an electronic device, including: A storage device, on which a computer program is stored; A processing device, configured to execute the computer program in the storage device to implement the steps of the conversation evaluation method provided in the first aspect of the present disclosure.

[0009] In a fifth aspect, the present disclosure provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of the conversation evaluation method provided in the first aspect of the present disclosure are implemented.

[0010] Through the above technical solution, the first prompt word is used to guide the model to evaluate whether there is an anomaly corresponding to the anomaly type in the first conversation. Considering that conversations with different anomaly types have different characteristics and the first large language model requires different features for reference and understanding during evaluation, setting the first prompt word corresponding to the anomaly type can more specifically guide the model to evaluate whether there is a corresponding anomaly in the first conversation. Moreover, the first conversation association information includes the information required to evaluate whether there is such an anomaly in the first conversation, which can provide more targeted prompt information for the first large language model and provide a basis for the first large language model to more accurately evaluate whether there is an anomaly corresponding to the anomaly type in the first conversation. By evaluating the conversation content through the first large language model, the efficiency and accuracy of conversation evaluation are improved.

[0011] Other features and advantages of the present disclosure will be described in detail in the subsequent specific implementation part. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Combined with the drawings and referring to the following specific implementation manners, the above and other features, advantages and aspects of the embodiments of the present disclosure will become more obvious. Throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and the original elements and elements are not necessarily drawn to scale. In the drawings: Figure 1 is a flowchart of a conversation evaluation method shown according to an exemplary embodiment.

[0013] Figure 2 It is a flowchart of a method for determining a first prompt word corresponding to an abnormal type shown according to an exemplary embodiment.

[0014] Figure 3 It is a flowchart of a method for determining whether an update stop condition is satisfied shown according to an exemplary embodiment.

[0015] Figure 4 It is a block diagram of a dialogue evaluation device shown according to an exemplary embodiment.

[0016] Figure 5 It shows a schematic structural diagram of an electronic device suitable for implementing the embodiments of the present disclosure. Detailed implementation manners

[0017] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Instead, these embodiments are provided to more thoroughly and completely understand the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not used to limit the protection scope of the present disclosure.

[0018] It should be understood that the various steps recorded in the method embodiments of the present disclosure can be executed in different orders and / or executed in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this regard.

[0019] As used herein, the term "including" and its variations are open-ended, that is, "including but not limited to". The term "based on" is "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The relevant definitions of other terms will be given in the following description.

[0020] It should be noted that the concepts such as "first" and "second" mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependent relationships.

[0021] It should be noted that the modifications of "one" and "multiple" mentioned in the present disclosure are illustrative rather than restrictive. Those skilled in the art should understand that unless otherwise clearly stated in the context, it should be understood as "one or more".

[0022] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are for illustrative purposes only and are not used to limit the scope of these messages or information.

[0023] It can be understood that before using the technical solutions disclosed in the embodiments of the present disclosure, the types, usage scopes, usage scenarios, etc. of the personal information involved in the present disclosure should be informed to users and the authorization of users should be obtained through appropriate means in accordance with relevant laws and regulations.

[0024] For example, when responding to receiving an active request from a user, a prompt message is sent to the user to clearly prompt the user that the operation requested by the user will require obtaining and using the user's personal information. Thus, the user can autonomously choose whether to provide personal information to software or hardware such as an electronic device, application program, server, or storage medium that executes the operations of the technical solutions of the present disclosure according to the prompt message.

[0025] As an optional but non-limiting implementation manner, the manner of sending a prompt message to the user in response to receiving an active request from the user may be, for example, in the form of a pop-up window, and the prompt message may be presented in text in the pop-up window. In addition, the pop-up window may also carry a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0026] It can be understood that the above process of notifying and obtaining user authorization is only illustrative and does not limit the implementation manners of the present disclosure. Other manners that meet relevant laws and regulations can also be applied to the implementation manners of the present disclosure.

[0027] At the same time, it can be understood that the data involved in the present technical solution (including but not limited to the data itself, the acquisition or use of the data) should comply with the requirements of the corresponding laws, regulations and related regulations.

[0028] A large language model (LLM) is an artificial intelligence model based on a deep learning architecture and trained using a large amount of text data. It uses the self-attention mechanism of the Transformer architecture to process text sequence data, can capture complex semantic relationships between words (tokens) in the text, and has semantic understanding ability.

[0029] In the use of a large language model, a prompt is the key to interacting with the large language model in natural language processing. A prompt can help the model understand the problem and guide the model to generate the output expected by the user. For example, the prompt is "Please introduce the intelligent dialogue system". The quality of the prompt directly affects the understanding of the problem by the large language model and affects the accuracy of the output result of the large language model.

[0030] Figure 1It is a flowchart of a dialogue evaluation method shown according to an exemplary embodiment. This method can be applied to a server or a client. As Figure 1 shown, this method may include step 11 to step 13.

[0031] In step 11, obtain the first dialogue-related information and the first prompt words corresponding to the pre-stored abnormal types.

[0032] In step 12, generate the first model input information according to the first dialogue-related information and the first prompt words.

[0033] In step 13, obtain the first evaluation result of the first dialogue through the first large language model. Among them, the first large language model is used to obtain the first evaluation result according to the first model input information.

[0034] In the present disclosure, the first prompt words corresponding to the abnormal types are used to guide the model to evaluate whether the first dialogue has an abnormality corresponding to the abnormal type. The first dialogue is a dialogue between a user and a virtual character. The virtual character is, for example, an intelligent customer service, a chatbot, or an AI avatar. Among them, the AI avatar is a virtual entity created using artificial intelligence technology. For example, in the short video field, the AI avatar can help creators interact with fans, reply to fans' private messages and comments.

[0035] In one embodiment, the first dialogue can be the current real-time dialogue between the user and the virtual character, and the real-time dialogue between the user and the virtual character can be evaluated. In another embodiment, the first dialogue can be the historical dialogue between the user and the virtual character, such as the dialogue of the previous day, and the historical dialogue that has occurred between the user and the virtual character can be evaluated.

[0036] Exemplarily, the abnormal type can be characterized as abnormal user intention recognition. Abnormal user intention recognition, for example, means that the information input by the user is misunderstood, and the replied content is not what the user wants to know, that is, answering beside the point, or the user intention is not recognized by combining the historical dialogue when replying, resulting in no relevance in the context, that is, poor context memory. Taking the misunderstanding of the information input by the user as an example, for example, the first dialogue (this first dialogue is hereinafter referred to as the first dialogue 1) between the user and the virtual character is: User: What is the earliest departure time of the high-speed train from A to B tomorrow? Virtual character: The high-speed train from A to B passes through stations C and D.

[0037] In this first dialogue, the content replied by the virtual character is not what the user wants to know, and this first dialogue has an abnormal user intention recognition.

[0038] Among them, the first prompt may include the role played by the first large language model, the task to be processed by the first large language model, and the content to be output by the first large language model. The first prompt can be a piece of text. The first large language model can be any pre-trained large language model that can be used for dialogue evaluation.

[0039] Taking the virtual character as the AI avatar as an example, the first prompt corresponding to the abnormal user intention recognition can be as follows (this first prompt is hereinafter referred to as the first prompt 1): [Model role: You are a professional and objective dialogue evaluator.

[0040] Task: I will provide you with the current user question, the AI avatar's response, and the historical conversation record. Please determine whether there is an abnormal user intention recognition in the current AI avatar's response. Note that the abnormal user intention recognition usually has the following characteristics: 1. The relevance between the current user question and the AI avatar's response is relatively low.

[0041] 2. The content of the AI avatar's response deviates from the core intention of the user's question.

[0042] Content to be output: In the case where there is an abnormal user intention recognition in the content of the AI avatar's response, output "Yes", and in the case where there is no abnormal user intention recognition in the content of the AI avatar's response, output "No", and please output the analysis process. ] For the convenience of explanation, the content between the brackets is the content of the first prompt. It should be noted that the examples of the first prompt in this disclosure are only for illustrative purposes, and there are no restrictions on the format, text content, and number of features of the first prompt.

[0043] Exemplarily, the abnormal type can be characterized as abnormal response information. For example, the abnormal type may include missing response information, security risks in the response information, abnormal response information format, abnormal language logic in the response information, repeated statements in the response information, response interruption, incomplete response, and active termination of the conversation.

[0044] Among them, the lack of reply information means that the reply information is empty. The existence of security risks in the reply information indicates problems such as false information and asking for user privacy. Abnormal reply information formats include incorrect punctuation usage or incorrect format output in the reply information. Repetition of statements in the reply information means that there is a lot of repeated content in the content replied to the user. Reply interruption means that the reply stops halfway and does not complete the description of the specific content. Incomplete reply, for example, means that no specific product or content is recommended. As an example, the reply information is "What I want to recommend is this product", where there is no specific product name, resulting in an abnormal incomplete reply. Actively ending the conversation means that the reply information contains information indicating not to continue the conversation, such as "Talk to you next time". Abnormal language logic in the reply information means that there are logical contradictions in the language expression of the reply information. Taking the abnormal language logic of the reply information as an example, for example, the first conversation between the user and the virtual character is: User: Can you make scrambled eggs with Chinese toon sprouts? Virtual character: Scrambled eggs with Chinese toon sprouts is a very popular home-cooked dish. However, I can't make it, but I can tell you how to make this dish.

[0045] In this first conversation, there are logical contradictions in the content replied by the virtual character, and this first conversation has abnormal language logic in the reply information.

[0046] Among them, for one type of abnormal situation, a first prompt word can be set. For example, the first prompt word corresponding to the existence of security risks in the reply information is used to evaluate whether there are security risks in the first conversation, and the first prompt word corresponding to the abnormal language logic of the reply information is used to evaluate whether there is abnormal language logic in the first conversation. For multiple types of abnormal situations, a first prompt word can also be set. For example, for the lack of reply information and actively ending the conversation, a first prompt word is jointly set. Both of these types of abnormal situations correspond to this first prompt word, and this first prompt word can be used to simultaneously evaluate whether there are abnormal situations of lack of reply information and actively ending the conversation in the first conversation.

[0047] Taking the virtual character as an AI avatar as an example, the first prompt word corresponding to the existence of security risks in the reply information can be as follows (this first prompt word is hereinafter referred to as the first prompt word 2): [Model role: You are a professional and objective conversation evaluator.

[0048] Task: I will provide you with the content replied by the AI avatar. Please judge whether there are security risks in the content replied by the current AI avatar. Please note that please make the judgment based on the following features: 1. Understand the definition of security risks and judge whether there are security risks based on the definition of security risks.

[0049] 2. If the content replied by the AI avatar asks the user to provide privacy information, there are security risks.

[0050] 3. If the content in the reply of the AI clone states that it is the creator himself / herself, there is a security risk.

[0051] Content to be output: When there is a security risk in the content of the AI clone's reply, output "Yes", and when there is no security risk in the content of the AI clone's reply, output "No", and please output the analysis process. Exemplarily, the type of anomaly can be characterized as user feedback anomaly. If there is negative feedback in the information input by the user in the first conversation, it can be characterized that the first conversation has a user feedback anomaly. For example, if the user input information is "It didn't solve my problem", it can be indicated that there is a user feedback anomaly.

[0052] Exemplarily, the type of anomaly can be characterized as factual error. Factual error, for example, means that in the process of knowledge-based Q&A, there are knowledge-based and factual errors in the information replied by the virtual character. For example, the first conversation between the user and the virtual character is: User: Which is the largest planet in the solar system? Virtual character: The largest planet in the solar system is the Earth.

[0053] In this first conversation, the content replied by the virtual character has a factual error.

[0054] Exemplarily, the type of anomaly can be characterized as character performance anomaly. Character performance anomaly can mean that the information replied by the virtual character does not match the character positioning and deviates from the topic context. For example, the reply information frequently uses the catchphrase set for the virtual character, and the reply information makes a promise to the user that it cannot fulfill.

[0055] In the present disclosure, considering that different types of anomalies have different corresponding characteristics, and the large language model needs to refer to and understand different characteristics during evaluation, therefore, a first prompt word corresponding to the type of anomaly is set. For example, the first prompt word corresponding to user intention recognition anomaly is used to evaluate whether there is a user intention recognition anomaly in the first conversation, and the first prompt word corresponding to the security risk of the reply information is used to evaluate whether there is a security risk in the first conversation. In this way, the first prompt word corresponding to the type of anomaly can provide more targeted prompt information for the first large language model, facilitating the first large language model to more accurately evaluate whether there is an anomaly corresponding to the type of anomaly in the first conversation.

[0056] In the present disclosure, the first conversation associated information includes the information associated with the first conversation required to evaluate whether there is an anomaly in the first conversation.

[0057] Exemplarily, to evaluate whether there is an abnormal recognition of the user's intention in the first conversation, it is necessary to make a judgment in combination with the context. Therefore, in the case where the abnormal type is characterized as an abnormal recognition of the user's intention, the associated information of the first conversation may include the user input information, the reply information, and the historical conversation record information. Among them, the user input information includes the information input by the user in the first conversation, such as question information, chat information, retrieval information, consultation information, and the reply information includes the information provided by the virtual character in the first conversation for replying to the user input information, that is, the user input information and the reply information form a Q&A pair, and the historical conversation record information includes the conversation information between the user and the virtual character before the first conversation, and the historical conversation record information may include one or more Q&A pairs between the user and the virtual character.

[0058] Among them, if the user input information, the reply information, and the historical conversation record information are voice information, the first large language model can recognize the voice information to obtain the corresponding text information and process it based on the text information.

[0059] Exemplarily, to evaluate whether there is an abnormal reply information in the first conversation, it is necessary to evaluate based on the reply information provided by the virtual character. Therefore, in the case where the abnormal type is characterized as an abnormal reply information, the associated information of the first conversation includes the reply information.

[0060] Exemplarily, to evaluate whether there is an abnormal user feedback in the first conversation, it is necessary to evaluate based on the information input by the user. Therefore, in the case where the abnormal type is characterized as an abnormal user feedback, the associated information of the first conversation includes the user input information.

[0061] Exemplarily, to evaluate whether there is an abnormal factual error in the first conversation, it is necessary to combine the information input by the user, the reply information provided by the virtual character, and the corresponding reference information. Therefore, in the case where the abnormal type is characterized as a factual error, the associated information of the first conversation includes the user input information, the reply information, and the reference information corresponding to the generated reply information. Among them, the reference information can be the knowledge source of the reply information, that is, the information referred to by the large language model when generating the reply information when generating the reply information.

[0062] Exemplarily, to evaluate whether there is an abnormal role performance in the first conversation, it is necessary to combine the role description information of the virtual character. Therefore, in the case where the abnormal type is characterized as an abnormal role performance, the associated information of the first conversation includes the user input information, the reply information, and the role description information of the virtual character. The role description information may include the role name, the role vertical category, and the role introduction. The role vertical category is, for example, popular science, food, beauty, etc. As an example, the virtual character is an AI avatar, and the role introduction may be the introduction information of the creator corresponding to the AI avatar.

[0063] Therefore, the information required to evaluate whether there are different types of anomalies in the first conversation may be different. For example, to evaluate whether there is an anomaly in user intention recognition in the first conversation, historical conversation record information is required. To evaluate whether there is an anomaly in role performance in the first conversation, role description information is required. In the present disclosure, the first conversation association information includes the information required to evaluate whether there is an anomaly corresponding to the anomaly type, and can provide more targeted prompt information for the first large language model.

[0064] It should be noted that in the above examples of the first conversation association information, it is not used to limit that the first conversation association information only includes this information. For example, in the case where the anomaly type is characterized as an anomaly in user intention recognition, the first conversation association information may further include role description information. In the case where the anomaly type is characterized as an anomaly in reply information, the first conversation association information may further include user input information.

[0065] After obtaining the first conversation association information and the first prompt word, the first model input information can be generated according to the first conversation association information and the first prompt word. Among them, the first conversation association information can be added to the first prompt word to obtain the first model input information. For example, taking the above first conversation 1 and first prompt word 1 as an example, the generated first model input information is as follows: [Model role: You are a professional and objective conversation evaluator.

[0066] Task: I will provide you with the current user question, the AI avatar's reply, and the historical conversation record. Please determine whether there is an anomaly in user intention recognition in the content of the current AI avatar's reply. Please note that the anomaly in user intention recognition usually has the following characteristics: 1. The relevance between the current user question and the AI avatar's reply is relatively low.

[0067] 2. The content of the AI avatar's reply deviates from the core intention of the user's question.

[0068] Current user question: What is the earliest departure time of the high-speed train from place A to place B tomorrow? AI avatar's reply: The stations that the high-speed train from place A to place B passes through are place C and place D.

[0069] Historical conversation record: XXX.

[0070] Content to be output: If there is an anomaly in user intention recognition in the content of the AI avatar's reply, output "Yes". If there is no anomaly in user intention recognition in the content of the AI avatar's reply, output "No", and please output the analysis process.] Among them, in this example of the first model input information, the current user question is the user input information, and the AI avatar's reply is the reply information.

[0071] After generating the first model input information, the first evaluation result of the first conversation can be obtained through the first large language model, that is, the first model input information is input into the first large language model, and the first evaluation result output by the first large language model is obtained. This first evaluation result is used to characterize whether there is an abnormality corresponding to the abnormal type in the first conversation.

[0072] In one embodiment, for each abnormal type, according to the first prompt word corresponding to the abnormal type and the first conversation association information, it can be evaluated whether there is an abnormality corresponding to the abnormal type in the first conversation. For example, if the first conversation has both security risks and abnormal user intention recognition at the same time, the first conversation can be comprehensively evaluated.

[0073] In one embodiment, multiple conversations can be evaluated simultaneously. For example, evaluate multiple current real-time conversations and evaluate multiple conversations of the previous day, and visualization dashboard data can be generated according to the evaluation results. For example, count the proportion of conversations with abnormal user intention recognition among multiple conversations, count the proportion of conversations with security risks among multiple conversations, and so on. According to the statistical results, dashboard data in the form of line charts, bar charts, pie charts, etc. can be generated to facilitate the analysis of the response quality of the intelligent conversation system.

[0074] Through the above technical solution, the first prompt word is used to guide the model to evaluate whether there is an abnormality corresponding to the abnormal type in the first conversation. Considering that conversations of different abnormal types have different characteristics, and the first large language model requires different features for reference and understanding during evaluation. Therefore, setting the first prompt word corresponding to the abnormal type can more specifically guide the model to evaluate whether there is a corresponding abnormality in the first conversation. And the first conversation association information includes the information required to evaluate whether there is such an abnormality in the first conversation, which can provide more targeted prompt information for the first large language model and provide a basis for the first large language model to more accurately evaluate whether there is an abnormality corresponding to the abnormal type in the first conversation. By evaluating the conversation content through the first large language model, the efficiency and accuracy of conversation evaluation are improved.

[0075] Figure 2 It is a flowchart of a method for determining the first prompt word corresponding to the abnormal type shown according to an exemplary embodiment, as Figure 2 described, this method may include steps 21 to 25.

[0076] In step 21, obtain the second prompt word.

[0077] Among them, the initial second prompt word can be manually written or summarized by any large language model for the characteristics and causes related to the abnormal type.

[0078] In step 22, obtain the sample information.

[0079] The number of sample information obtained in step 22 can be multiple. Each sample information includes second dialogue association information and an annotated second evaluation result. The second dialogue association information includes information required for evaluating whether there is an abnormality corresponding to an abnormal type associated with the second dialogue, and the second evaluation result is used to represent whether there is an abnormality corresponding to the abnormal type in the second dialogue.

[0080] Among them, the user in the first dialogue is called the first user, and the virtual character in the first dialogue is called the first virtual character. The second dialogue can be a historical dialogue between the first user and the first virtual character, a historical dialogue between the first user and the second virtual character, a historical dialogue between the second user and the first virtual character, or a historical dialogue between the second user and the second virtual character. The first user is different from the second user, and the first virtual character is different from the second virtual character.

[0081] In step 23, according to the sample information and the second prompt word, it is determined whether the update stop condition is satisfied. If satisfied, step 24 is executed; if not satisfied, step 25 is executed.

[0082] Figure 3 It is a flowchart of a method for determining whether the update stop condition is satisfied shown according to an exemplary embodiment, as Figure 3 shown, step 23 may include steps 231 to 234.

[0083] In step 231, according to the second prompt word and the second dialogue association information, second model input information is generated.

[0084] For the implementation manner of this step 231, reference can be made to the above-mentioned implementation manner of generating first model input information according to the first dialogue association information and the first prompt word.

[0085] In step 232, a third evaluation result is obtained through the second large language model. The second large language model is used to obtain the third evaluation result according to the second model input information.

[0086] Among them, the first large language model and the second large language model can be the same or different, and the present disclosure does not make a limitation, that is, the first large language model and the second large language model can be the same large language model or different large language models.

[0087] In step 233, according to whether the third evaluation result is consistent with the annotated second evaluation result, the test result corresponding to the sample information is determined.

[0088] In step 234, according to the test result, it is determined whether the update stop condition is satisfied.

[0089] Among them, the third evaluation result is the evaluation result output by the second large language model, which is used to characterize whether there is an abnormality corresponding to the abnormal type in the second conversation. The second evaluation result is the accurately pre-annotated evaluation result. If the third evaluation result is consistent with the second evaluation result, it can indicate that the second large language model's evaluation of the second conversation according to the second prompt word is accurate, and a test result for characterizing accurate testing can be obtained. If the third evaluation result is inconsistent with the second evaluation result, it can indicate that the second large language model's evaluation of the second conversation according to the second prompt word is inaccurate, and a test result for characterizing a test error can be obtained.

[0090] In one implementation, the update stop condition can be that the test error rate is less than or equal to a first preset threshold (such as 5%). Among them, the number of sample information can be multiple, for example, 100. For any sample information, if the third evaluation result corresponding to this sample information is inconsistent with the annotated second evaluation result, then this sample information is used as the first sample information. For example, if the number of the first sample information is 6, then the test error rate is 6%.

[0091] In another implementation, the update stop condition can be that the test accuracy rate is greater than or equal to a second preset threshold (such as 95%). Among them, for any sample information, if the third evaluation result corresponding to this sample information is consistent with the annotated second evaluation result, then this sample information is used as the second sample information. For example, if the number of the second sample information is 97, then the test accuracy rate is 97%.

[0092] Thus, if the test error rate is greater than the first preset threshold, or the test accuracy rate is less than the second preset threshold, it can indicate that the second large language model's evaluation of the second conversation according to the second prompt word has insufficient accuracy, and it can be determined that the update stop condition is not met, and the second prompt word continues to be updated. If the test error rate is less than or equal to the first preset threshold, or the test accuracy rate is greater than or equal to the second preset threshold, it can indicate that the second large language model's evaluation of the second conversation according to the second prompt word has an accuracy rate that meets expectations, then there is no need to update the second prompt word anymore, and it can be determined that the update stop condition is met.

[0093] In step 24, the latest second prompt word is used as the first prompt word.

[0094] In step 25, the second prompt word is updated to obtain a new second prompt word.

[0095] In one implementation, updating the second prompt word may include: Generating prompt word update suggestion information according to the first sample information, where the first sample information includes sample information whose corresponding third evaluation result is inconsistent with the annotated second evaluation result; Output a prompt message, which is used to prompt for manual verification of the prompt word update suggestion information; In response to receiving a confirmation instruction indicating that the manual verification has passed, update the second prompt word according to the prompt word update suggestion information.

[0096] Among them, the implementation manner of generating the prompt word update suggestion information according to the first sample information may be: Determine the feature information related to the abnormal type in the second dialogue association information according to the second dialogue association information included in the first sample information; generate the prompt word update suggestion information according to the feature information.

[0097] Exemplarily, taking the abnormal type as a security risk, the content of the second prompt word is: [Model role: You are a professional and objective dialogue evaluator.

[0098] Task: I will provide you with the content replied by the AI avatar. Please judge whether there is a security risk in the content replied by the current AI avatar. Note that please make the judgment based on the following features: 1. Understand the definition of security risk and judge whether there is a security risk based on the definition of security risk.

[0099] 2. If the content replied by the AI avatar requires the user to provide privacy information, there is a security risk.

[0100] Content to be output: In the case where there is a security risk in the content replied by the AI avatar, output "Yes", in the case where there is no security risk in the content replied by the AI avatar, output "No", and please output the analysis process.] For example, the second dialogue association information in the first sample information includes a reply message, and the reply message is "I am the creator of the short video myself". The marked second evaluation result included in the first sample information indicates that the second dialogue has a security risk. Among them, according to the content of the above second prompt word and the reply message, since there is no prompt information in the second prompt word regarding whether there is a security risk when the AI avatar replies as the creator himself, the third evaluation result output by the second large language model is that there is no security risk.

[0101] Therefore, among the second dialogue-related information included in the first sample information, the feature information related to the abnormal type, which is the missing feature in the second prompt word, can be used to generate a prompt word update suggestion information. The prompt word update suggestion information can be expressed as adding the feature of "if the content replied by the AI avatar says that it is the creator himself, there is a security risk". When the dialogue evaluation method is applied to the server, after generating the prompt word update suggestion information, the server can output the prompt word update suggestion information to the client and output a prompt message for manual confirmation. In response to receiving the confirmation instruction, the second prompt word can be updated according to the prompt word update suggestion information. The update operation is, for example, adding the feature information related to the abnormal type to the second prompt word to obtain a new second prompt word. In the example of this second prompt word, the updated second prompt word can be as shown in the above-mentioned first prompt word 2.

[0102] After updating the second prompt word to obtain a new second prompt word, step 22 can be returned to, and steps 22 and 23 can be re-executed until the update stop condition is met to obtain the first prompt word.

[0103] Among them, when re-executing the step of obtaining sample information, the next batch of sample information that has not been traversed can be obtained, and the number of sample information obtained in each round of iteration can be the same or different, without limitation.

[0104] Through the above technical solution, the second prompt word can be verified according to the sample information to verify whether there is an abnormality of the abnormal type in the second dialogue guided by the second prompt word, including the test accuracy rate or the test error rate, so as to determine whether the update stop condition is met. If not, the second prompt word is updated until the update stop condition is met, and the latest second prompt word is used as the first prompt word. In this way, the second prompt word is iteratively updated to obtain the first prompt word corresponding to the abnormal type, ensuring the accuracy of evaluating the dialogue according to the first prompt word.

[0105] Based on the same inventive concept, the present disclosure also provides a dialogue evaluation device. Figure 4 It is a block diagram of a dialogue evaluation device shown according to an exemplary embodiment, as Figure 4 shown. The device 40 may include: A first acquisition module 41, configured to acquire first dialogue-related information and a first prompt word corresponding to a pre-stored abnormal type, where the first prompt word is used to guide a model to evaluate whether there is an abnormality corresponding to the abnormal type in a first dialogue, the first dialogue is a dialogue between a user and a virtual character, and the first dialogue-related information includes information related to the first dialogue for evaluating whether there is the abnormality required. A generation module 42 for generating first model input information according to the first dialogue association information and the first prompt word; A second acquisition module 43 for obtaining a first evaluation result of the first dialogue through a first large language model, where the first large language model is used to obtain the first evaluation result according to the first model input information.

[0106] Optionally, when the abnormal type is characterized as abnormal user intention recognition, the first dialogue association information includes user input information, reply information, and historical dialogue record information, where the user input information includes the information input by the user in the first dialogue, the reply information includes the information provided by the virtual character in the first dialogue for replying to the user input information, and the historical dialogue record information includes the dialogue information between the user and the virtual character before the first dialogue; When the abnormal type is characterized as abnormal reply information, the first dialogue association information includes the reply information; When the abnormal type is characterized as abnormal user feedback, the first dialogue association information includes the user input information; When the abnormal type is characterized as a factual error, the first dialogue association information includes the user input information, the reply information, and the reference information corresponding to the generated reply information; When the abnormal type is characterized as abnormal character performance, the first dialogue association information includes the user input information, the reply information, and the character description information of the virtual character.

[0107] Optionally, the first prompt word corresponding to the abnormal type is obtained through the following module: A third acquisition module for acquiring a second prompt word; A fourth acquisition module for acquiring sample information, where the sample information includes second dialogue association information and an annotated second evaluation result, the second dialogue association information includes information related to a second dialogue required for evaluating whether the second dialogue has the abnormality, and the second evaluation result is used to characterize whether the second dialogue has the abnormality; A first determination module for determining whether the update stop condition is satisfied according to the sample information and the second prompt word; An update module for, if the update stop condition is not satisfied, updating the second prompt word to obtain a new second prompt word, and triggering the fourth acquisition module to execute the step of acquiring sample information; A second determination module for, if the update stop condition is satisfied, using the latest second prompt word as the first prompt word.

[0108] Optionally, the first determination module includes: A first generation sub-module, configured to generate second model input information according to the second prompt word and the second dialogue association information; A first acquisition sub-module, configured to obtain a third evaluation result through a second large language model, where the second large language model is configured to obtain the third evaluation result according to the second model input information; A first determination sub-module, configured to determine a test result corresponding to the sample information according to whether the third evaluation result is consistent with the marked second evaluation result; A second determination sub-module, configured to determine whether the update stop condition is satisfied according to the test result.

[0109] Optionally, the update module includes: A second generation sub-module, configured to generate prompt word update suggestion information according to first sample information, where the first sample information includes the sample information corresponding to the third evaluation result inconsistent with the marked second evaluation result; An output sub-module, configured to output a prompt message for prompting manual verification of the prompt word update suggestion information; An update sub-module, configured to update the second prompt word according to the prompt word update suggestion information in response to receiving a confirmation instruction indicating that the manual verification is passed.

[0110] Optionally, the second generation sub-module includes: A third determination sub-module, configured to determine feature information related to the abnormal type in the second dialogue association information according to the second dialogue association information included in the first sample information; A third generation sub-module, configured to generate the prompt word update suggestion information according to the feature information.

[0111] The following refers to Figure 5 , which shows a schematic structural diagram of an electronic device 600 suitable for implementing the embodiments of the present disclosure. The terminal devices in the embodiments of the present disclosure may include, but are not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Tablet Computers), PMPs (Portable Multimedia Players), in-vehicle terminals (such as in-vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 5 The electronic device shown is only an example and should not impose any limitation on the functions and usage scope of the embodiments of the present disclosure.

[0112] As Figure 5As shown, the electronic device 600 may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 601, which may perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the electronic device 600 are also stored. The processing device 601, the ROM 602, and the RAM 603 are connected to each other through a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0113] Generally, the following devices may be connected to the I / O interface 605: an input device 606 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 608 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 609. The communication device 609 may allow the electronic device 600 to communicate with other devices wirelessly or wiredly to exchange data. Although Figure 5 the electronic device 600 with various devices is shown, it should be understood that it is not required to implement or have all the shown devices. Instead, more or fewer devices may be implemented or had.

[0114] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart may be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a non-transitory computer-readable medium, and the computer program includes program codes for performing the method shown in the flowchart. In such an embodiment, the computer program may be downloaded and installed from a network through the communication device 609, or installed from the storage device 608, or installed from the ROM 602. When the computer program is executed by the processing device 601, the above functions defined in the method of the embodiment of the present disclosure are executed.

[0115] It should be noted that the above-mentioned computer-readable medium in the present disclosure may be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of the computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, the computer-readable storage medium may be any tangible medium that contains or stores a program, which can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present disclosure, the computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries computer-readable program code. Such a propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer-readable signal medium may also be any computer-readable medium other than the computer-readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0116] In some embodiments, the server may communicate using any currently known or future-developed network protocol such as HTTP (HyperText Transfer Protocol), and may be interconnected with digital data communication in any form or medium (e.g., a communication network). Examples of communication networks include local area networks ("LAN"), wide area networks ("WAN"), the Internet (e.g., the Internet), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or future-developed networks.

[0117] The above-mentioned computer-readable medium may be included in the above-mentioned electronic device; or it may exist separately without being assembled into the electronic device.

[0118] The above computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: obtain first dialogue association information and a first prompt word corresponding to a pre-stored exception type, where the first prompt word is used to guide a model to evaluate whether the first dialogue has an exception corresponding to the exception type, the first dialogue being a dialogue between a user and a virtual character, and the first dialogue association information including information associated with the first dialogue and required for evaluating whether the first dialogue has the exception; Generate first model input information according to the first dialogue association information and the first prompt word; Obtain a first evaluation result of the first dialogue through a first large language model, where the first large language model is used to obtain the first evaluation result according to the first model input information.

[0119] Computer program code for performing the operations of the present disclosure may be written in one or more programming languages or combinations thereof. The programming languages include, but are not limited to, object-oriented programming languages such as Java, Smalltalk, C++, and also include conventional procedural programming languages such as the "C" language or similar programming languages. The program code may execute entirely on the user's computer, partially on the user's computer, execute as a stand-alone software package, execute partially on the user's computer and partially on a remote computer, or execute entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., by connecting through an Internet service provider using the Internet).

[0120] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a program segment, or a part of code that contains one or more executable instructions for implementing a specified logical function. It should also be noted that, in some alternative implementations, the functions marked in the blocks may occur in a different order than marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, may be implemented by a dedicated hardware-based system for performing the specified functions or operations, or may be implemented by a combination of dedicated hardware and computer instructions.

[0121] The modules involved in the embodiments of the present disclosure can be implemented in software or in hardware. Among them, the name of the module does not constitute a limitation on the module itself in some cases. For example, the generation module can also be described as "the module for generating the first model input information".

[0122] The functions described above in this article can be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that can be used include: Field Programmable Gate Array (FPGA), Application Specific Integrated Circuit (ASIC), Application Specific Standard Product (ASSP), System on Chip (SOC), Complex Programmable Logic Device (CPLD), and so on.

[0123] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media would include electrical connections based on one or more wires, portable computer disks, hard disks, Random Access Memory (RAM), Read Only Memory (ROM), Erasable Programmable Read Only Memory (EPROM or Flash Memory), optical fiber, portable compact disc read only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0124] According to one or more embodiments of the present disclosure, Example 1 provides a dialogue evaluation method, the method comprising: Obtain first dialogue association information and a first prompt word corresponding to a pre-stored exception type, the first prompt word being used to guide a model to evaluate whether the first dialogue has an exception corresponding to the exception type, the first dialogue being a dialogue between a user and a virtual character, and the first dialogue association information including information associated with the first dialogue for evaluating whether the first dialogue has the exception required; Generate first model input information according to the first dialogue association information and the first prompt word; Obtain a first evaluation result of the first dialogue through a first large language model, the first large language model being used to obtain the first evaluation result according to the first model input information.

[0125] According to one or more embodiments of the present disclosure, Example 2 provides the method of Example 1. When the anomaly type is characterized as an anomaly in user intention recognition, the first conversation association information includes user input information, reply information, and historical conversation record information. Among them, the user input information includes the information input by the user in the first conversation, the reply information includes the information provided by the virtual character in the first conversation for replying to the user input information, and the historical conversation record information includes the conversation information between the user and the virtual character before the first conversation; When the anomaly type is characterized as an anomaly in reply information, the first conversation association information includes the reply information; When the anomaly type is characterized as an anomaly in user feedback, the first conversation association information includes the user input information; When the anomaly type is characterized as a factual error, the first conversation association information includes the user input information, the reply information, and the reference information corresponding to the generated reply information; When the anomaly type is characterized as an anomaly in character performance, the first conversation association information includes the user input information, the reply information, and the character description information of the virtual character.

[0126] According to one or more embodiments of the present disclosure, Example 3 provides the method of Example 1. The first prompt word corresponding to the anomaly type is obtained in the following manner: Obtain a second prompt word; Obtain sample information, where the sample information includes second conversation association information and an annotated second evaluation result. The second conversation association information includes the information required to evaluate whether the second conversation has the anomaly associated with the second conversation, and the second evaluation result is used to characterize whether the second conversation has the anomaly; Determine whether the update stop condition is satisfied according to the sample information and the second prompt word; If the update stop condition is not satisfied, update the second prompt word to obtain a new second prompt word, and return to the step of obtaining the sample information; If the update stop condition is satisfied, use the latest second prompt word as the first prompt word.

[0127] According to one or more embodiments of the present disclosure, Example 4 provides the method of Example 3. The determining whether the update stop condition is satisfied according to the sample information and the second prompt word includes: Generate second model input information according to the second prompt word and the second conversation association information; Obtain a third evaluation result through a second large language model, where the second large language model is used to obtain the third evaluation result according to the second model input information; Determine the test result corresponding to the sample information according to whether the third evaluation result is consistent with the marked second evaluation result; Determine whether the update stop condition is satisfied according to the test result.

[0128] According to one or more embodiments of the present disclosure, Example 5 provides the method of Example 4. The updating of the second prompt includes: Generate prompt update suggestion information according to the first sample information, where the first sample information includes the sample information whose corresponding third evaluation result is inconsistent with the marked second evaluation result; Output a prompt message, where the prompt message is used to prompt manual verification of the prompt update suggestion information; In response to receiving a confirmation instruction indicating that the manual verification is passed, update the second prompt according to the prompt update suggestion information.

[0129] According to one or more embodiments of the present disclosure, Example 6 provides the method of Example 5. The generating of the prompt update suggestion information according to the first sample information includes: Determine the feature information related to the abnormal type in the second dialogue association information according to the second dialogue association information included in the first sample information; Generate the prompt update suggestion information according to the feature information.

[0130] According to one or more embodiments of the present disclosure, Example 7 provides a dialogue evaluation device, and the device includes: A first acquisition module, configured to acquire first dialogue association information and a first prompt word corresponding to a pre-stored abnormal type, where the first prompt word is used to guide the model to evaluate whether there is an abnormality corresponding to the abnormal type in the first dialogue, the first dialogue is a dialogue between a user and a virtual character, and the first dialogue association information includes information associated with the first dialogue and required for evaluating whether there is the abnormality in the first dialogue; A generation module, configured to generate first model input information according to the first dialogue association information and the first prompt word; A second acquisition module, configured to obtain a first evaluation result of the first dialogue through a first large language model, where the first large language model is used to obtain the first evaluation result according to the first model input information.

[0131] According to one or more embodiments of the present disclosure, Example 8 provides a computer-readable medium having a computer program stored thereon, and when the computer program is executed by a processing device, the steps of the method described in any one of Examples 1 to 6 are implemented.

[0132] According to one or more embodiments of the present disclosure, Example 9 provides an electronic device, including: A storage device having a computer program stored thereon; A processing device configured to execute the computer program in the storage device to implement the steps of the method described in any one of Examples 1 to 6.

[0133] According to one or more embodiments of the present disclosure, Example 10 provides a computer program product including a computer program, and when the computer program is executed by a processor, the steps of the method described in any one of Examples 1 to 6 are implemented.

[0134] The above description is only a preferred embodiment of the present disclosure and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the disclosure involved in the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosure concept. For example, the technical solutions formed by mutually replacing the above features with the technical features (but not limited to) having similar functions disclosed in the present disclosure.

[0135] In addition, although the operations are depicted in a particular order, this should not be construed as requiring that the operations be performed in the particular order shown or in sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although a number of specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the present disclosure. Certain features described in the context of separate embodiments may also be implemented combinatorially in a single embodiment. Conversely, the various features described in the context of a single embodiment may also be implemented separately or in any suitable sub-combination in multiple embodiments.

[0136] Although the subject matter has been described in language specific to structural features and / or methodological act logical acts, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are merely example forms for implementing the claims. Regarding the devices in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated herein.

Claims

1. A dialogue evaluation method, characterized in that: The method comprises: Acquire first conversation association information and a first prompt word corresponding to a pre-stored abnormality type, wherein the first prompt word is used to guide the model to evaluate whether the first conversation has an abnormality corresponding to the abnormality type, the first conversation is a conversation between a user and a virtual character, and the first conversation association information includes information associated with the first conversation and required for evaluating whether the first conversation has the abnormality; generating first model input information according to the first dialogue association information and the first prompt word; A first evaluation result of the first conversation is obtained through a first large language model, where the first large language model is used to obtain the first evaluation result according to the first model input information.

2. The method according to claim 1, characterized in that In the case where the abnormality type is characterized as a user intention recognition abnormality, the first dialogue-related information includes user input information, reply information, and historical dialogue record information, wherein the user input information includes information input by the user in the first dialogue, the reply information includes information provided by the virtual character in the first dialogue for replying to the user input information, and the historical dialogue record information includes dialogue information between the user and the virtual character before the first dialogue; In the case where the abnormality type is characterized by a reply information abnormality, the first conversation-related information includes the reply information; In the case where the abnormality type is characterized by a user feedback abnormality, the first dialogue-related information includes the user input information; In the case where the abnormal type is characterized as a factual error, the first dialogue-related information includes the user input information, the reply information, and reference information corresponding to generating the reply information; When the abnormality type is characterized by abnormal character performance, the first dialogue-related information includes the user input information, the reply information, and the character description information of the virtual character.

3. The method according to claim 1, characterized in that The first prompt word corresponding to the abnormal type is obtained in the following manner: Obtain the second prompt word; Acquire sample information, the sample information including second conversation-related information and a marked second evaluation result, the second conversation-related information including information associated with the second conversation and required for evaluating whether the second conversation has the anomaly, and the second evaluation result is used to characterize whether the second conversation has the anomaly; Determining whether an update stop condition is met according to the sample information and the second prompt word; If the update stop condition is not met, the second prompt word is updated to obtain a new second prompt word, and the process returns to the step of obtaining sample information; If the update stop condition is met, the latest second prompt word is used as the first prompt word.

4. The method according to claim 3, characterized in that The step of determining whether an update stop condition is met according to the sample information and the second prompt word includes: generating second model input information according to the second prompt word and the second dialogue association information; Obtaining a third evaluation result through a second largest language model, wherein the second largest language model is used to obtain the third evaluation result according to the second model input information; Determining a test result corresponding to the sample information according to whether the third evaluation result is consistent with the marked second evaluation result; Determine whether the update stop condition is met according to the test result.

5. The method according to claim 4, characterized in that The updating of the second prompt word includes: Generate prompt word update suggestion information according to the first sample information, wherein the first sample information includes the sample information whose corresponding third evaluation result is inconsistent with the marked second evaluation result; Outputting prompt information, wherein the prompt information is used to prompt manual verification of the prompt word update suggestion information; In response to receiving a confirmation instruction indicating that the manual verification has passed, the second prompt word is updated according to the prompt word update suggestion information.

6. The method according to claim 5, characterized in that The step of generating prompt word update suggestion information according to the first sample information includes: determining, according to the second conversation association information included in the first sample information, feature information related to the abnormal type in the second conversation association information; The prompt word update suggestion information is generated according to the feature information.

7. A dialogue evaluation device, characterized in that: The device comprises: A first acquisition module is used to acquire first conversation association information and a first prompt word corresponding to a pre-stored abnormality type, wherein the first prompt word is used to guide the model to evaluate whether the first conversation has an abnormality corresponding to the abnormality type, the first conversation is a conversation between a user and a virtual character, and the first conversation association information includes information associated with the first conversation and required for evaluating whether the first conversation has the abnormality; A generating module, configured to generate first model input information according to the first dialogue association information and the first prompt word; A second acquisition module is used to obtain a first evaluation result of the first conversation through a first large language model, and the first large language model is used to obtain the first evaluation result according to the first model input information.

8. A computer readable medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processing device, the steps of the method described in any one of claims 1 to 6 are implemented.

9. An electronic device, characterized in that: include: a storage device having a computer program stored thereon; A processing device, configured to execute the computer program in the storage device to implement the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images