Data detection method and device based on artificial intelligence, computer equipment and medium

Abstract

The invention belongs to the field of artificial intelligence and the field of digital medical treatment, and relates to a data detection method and device based on artificial intelligence, computer equipment and a storage medium, and the method comprises the steps: obtaining a target reply which is output by a target large model and corresponds to a target prompt; segmenting the target reply to obtain a plurality of clauses; based on the generation strategy, processing the target prompt by using a target large model to obtain a specified number of first replies, and generating a reply set based on the first replies; constructing a detection prompt based on each clause and the reply set; processing each detection prompt based on the third-party large model to obtain a second reply; scoring all the second replies to obtain a plurality of scores, and screening out a target score with the highest value from all the scores; and generating an illusion detection result of the target reply based on the target score. In addition, illusion detection results may be stored in a blockchain. According to the hallucination detection method based on the large model, the applicability and accuracy of hallucination detection are improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence development technology and the field of digital medicine, and in particular to artificial intelligence-based data detection methods, devices, computer equipment and storage media. Background Art

[0002] With the rapid development of artificial intelligence technology, especially the breakthrough in deep learning, big models have shown great application potential in multiple industries with their excellent semantic understanding and text generation capabilities, especially in the medical industry. In the medical field, big models are designed to assist doctors in disease diagnosis, treatment plan recommendations, and patient education, aiming to improve the quality and efficiency of medical services. However, despite the remarkable achievements of big models in performance, the text content they generate is not completely flawless, and there is a phenomenon called "big model hallucination". "Big model hallucination" refers to the fact that when the model generates text, it may produce content that is inconsistent with the facts, logically contradictory, or beyond the reasonable scope of the context. In medical scenarios, this illusion phenomenon is particularly dangerous because it may mislead doctors to make wrong diagnoses or treatment decisions, or even cause irreversible harm to patients, thereby causing medical safety accidents. Therefore, ensuring the accuracy and reliability of the content generated by medical big models has become the key to promoting its further application.

[0003] In order to deal with the risks brought by large model hallucinations, hallucination detection technology came into being. At present, the mainstream hallucination detection technology mainly adopts methods based on external knowledge retrieval. This method relies on a reliable external knowledge base. By comparing the text to be detected with the information in the knowledge base, it is judged whether the text content is consistent with the known facts, thereby identifying potential hallucination content. However, this method faces many challenges in practical applications. First, it is not easy to obtain a comprehensive, accurate and timely external knowledge base, especially in the medical field, where knowledge is updated very quickly and is highly professional, making the construction and maintenance of the knowledge base expensive. Secondly, even with a knowledge base, its quality and integrity are difficult to guarantee, and the detection results may be biased due to missing or incorrect information. Therefore, the hallucination detection method based on external knowledge retrieval has certain limitations in applicability and accuracy. Summary of the invention

[0004] The purpose of the embodiments of the present application is to propose a data detection method, device, computer equipment and storage medium based on artificial intelligence to solve the technical problem that the existing hallucination detection method based on external knowledge retrieval has certain limitations in applicability and accuracy.

[0005] In order to solve the above technical problems, the embodiment of the present application provides a data detection method based on artificial intelligence, which adopts the following technical solutions:

[0006] Obtaining a target response corresponding to the target prompt output by a preset target large model;

[0007] Segmenting the target response to obtain a plurality of corresponding clauses;

[0008] Based on a preset generation strategy, the target prompt is processed using the target macro model to obtain a specified number of first replies, and a corresponding reply set is generated based on all the first replies;

[0009] Constructing a plurality of corresponding detection prompts based on each of the clauses and the response set;

[0010] Processing each of the detection prompts based on a preset third-party large model to obtain corresponding multiple second replies;

[0011] Scoring all the second replies to obtain a plurality of corresponding scores, and selecting a target score with the highest value from all the scores;

[0012] A hallucination detection result corresponding to the target response is generated based on the target score.

[0013] Furthermore, the step of segmenting the target response to obtain a plurality of corresponding clauses specifically includes:

[0014] Get a library of preset natural language processing tools;

[0015] Calling a target component from the natural language processing tool library;

[0016] The target response is segmented based on the target component to obtain a plurality of clauses corresponding to the target response.

[0017] Furthermore, the step of constructing a plurality of corresponding detection prompts based on each of the clauses and the response set specifically includes:

[0018] Get the preset hallucination detection prompt template;

[0019] Obtain a specified clause; wherein the specified clause is any one of all the clauses;

[0020] Filling the specified clause into the clause area in the hallucination detection prompt template, and filling the reply set into the reply text area in the detection prompt template, to obtain a filled hallucination detection prompt template;

[0021] The filled hallucination detection prompt template is used as the designated detection prompt corresponding to the designated clause.

[0022] Furthermore, the step of scoring all the second replies to obtain corresponding multiple scores specifically includes:

[0023] Call the preset scoring strategy;

[0024] Analyze the designated reply based on the scoring strategy to obtain a corresponding reply analysis result; wherein the designated reply is any one of all the second replies;

[0025] A designated score corresponding to the designated reply is generated based on the reply analysis result.

[0026] Furthermore, the step of generating a hallucination detection result corresponding to the target response based on the target score specifically includes:

[0027] Get the preset hallucination score threshold;

[0028] Determining whether the target score is greater than the hallucination score threshold;

[0029] If the target score is greater than the hallucination score threshold, generating a first hallucination detection result including hallucination in the target response;

[0030] If the target score is not greater than the hallucination score threshold, a second hallucination detection result is generated in which the target response does not contain hallucinations.

[0031] Further, after the step of generating a hallucination detection result corresponding to the target response based on the target score, the method further includes:

[0032] Performing content analysis on the hallucination detection result, and if it is detected that the hallucination detection result indicates that the target response contains a hallucination, performing correction processing on the target response to obtain a corresponding corrected response;

[0033] Get preset reminder information;

[0034] A corresponding reply process is performed based on the revised reply and the reminder information.

[0035] Further, after the step of generating a hallucination detection result corresponding to the target response based on the target score, the method further includes:

[0036] Acquiring hallucination detection process data corresponding to the target response;

[0037] Determine the target storage method;

[0038] The target response, the hallucination detection result and the hallucination detection process data are stored and processed based on the target storage method.

[0039] In order to solve the above technical problems, the embodiment of the present application further provides a data detection device based on artificial intelligence, which adopts the following technical solution:

[0040] A first acquisition module is used to acquire a target response corresponding to the target prompt output by a preset target large model;

[0041] A segmentation module, used for segmenting the target response to obtain a plurality of corresponding clauses;

[0042] A first processing module, configured to process the target prompt using the target macro model based on a preset generation strategy to obtain a specified number of first replies, and generate a corresponding reply set based on all the first replies;

[0043] A construction module, configured to construct a plurality of corresponding detection prompts based on each of the clauses and the response set;

[0044] A second processing module, used for processing each of the detection prompts based on a preset third-party large model to obtain a corresponding plurality of second replies;

[0045] A scoring module, used for scoring all the second replies to obtain corresponding multiple scores, and selecting a target score with the highest value from all the scores;

[0046] A generation module is used to generate a hallucination detection result corresponding to the target response based on the target score.

[0047] In order to solve the above technical problems, the embodiment of the present application further provides a computer device, which adopts the following technical solution:

[0048] Obtaining a target response corresponding to the target prompt output by a preset target large model;

[0049] Segmenting the target response to obtain a plurality of corresponding clauses;

[0050] Based on a preset generation strategy, the target prompt is processed using the target macro model to obtain a specified number of first replies, and a corresponding reply set is generated based on all the first replies;

[0051] Constructing a plurality of corresponding detection prompts based on each of the clauses and the response set;

[0052] Processing each of the detection prompts based on a preset third-party large model to obtain corresponding multiple second replies;

[0053] Scoring all the second replies to obtain a plurality of corresponding scores, and selecting a target score with the highest value from all the scores;

[0054] A hallucination detection result corresponding to the target response is generated based on the target score.

[0055] In order to solve the above technical problems, the embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:

[0056] Obtaining a target response corresponding to the target prompt output by a preset target large model;

[0057] Segmenting the target response to obtain a plurality of corresponding clauses;

[0058] Based on a preset generation strategy, the target prompt is processed using the target macro model to obtain a specified number of first replies, and a corresponding reply set is generated based on all the first replies;

[0059] Constructing a plurality of corresponding detection prompts based on each of the clauses and the response set;

[0060] Processing each of the detection prompts based on a preset third-party large model to obtain corresponding multiple second replies;

[0061] Scoring all the second replies to obtain a plurality of corresponding scores, and selecting a target score with the highest value from all the scores;

[0062] A hallucination detection result corresponding to the target response is generated based on the target score.

[0063] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0064] The present application first obtains the target response corresponding to the target prompt output by the preset target large model; and divides the target response to obtain the corresponding multiple clauses; then based on the preset generation strategy, the target large model is used to process the target prompt to obtain a specified number of first replies, and a corresponding reply set is generated based on all the first replies; then, multiple corresponding detection prompts are constructed based on each of the clauses and the reply set; subsequently, each of the detection prompts is processed based on the preset third-party large model to obtain multiple corresponding second replies; further, all the second replies are scored to obtain multiple corresponding scores, and the target score with the highest value is selected from all the scores; finally, the hallucination detection result corresponding to the target reply is generated based on the target score. The detection method for large model hallucinations provided by the present application does not require the support of additional data such as an external knowledge base, and only needs to rely on an additional, existing third-party large model to quickly and accurately complete the hallucination detection processing of the target reply output by the target large model, which effectively improves the processing efficiency and applicability of the hallucination detection processing, and ensures the accuracy of the obtained hallucination detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the scheme in the present application, a brief introduction is given below to the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0066] Figure 1 is an exemplary system architecture diagram to which the present application may be applied;

[0067] Figure 2 A flowchart of an embodiment of an artificial intelligence-based data detection method according to the present application;

[0068] Figure 3 is a structural schematic diagram of an embodiment of an artificial intelligence-based data detection device according to the present application;

[0069] Figure 4 It is a structural diagram of an embodiment of a computer device according to the present application. DETAILED DESCRIPTION

[0070] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0071] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0072] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0073] like Figure 1 As shown, the system architecture 100 may include a terminal device 101, a network 102 and a server 103. The terminal device 101 may be a laptop 1011, a tablet computer 1012 or a mobile phone 1013. The network 102 is used to provide a medium for a communication link between the terminal device 101 and the server 103. The network 102 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0074] The user can use the terminal device 101 to interact with the server 103 through the network 102 to receive or send messages, etc. Various communication client applications can be installed on the terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0075] The terminal device 101 can be any electronic device with a display screen and supporting web browsing. In addition to the laptop computer 1011, tablet computer 1012 or mobile phone 1013, the terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 (Moving Picture Experts Group Audio Layer IV), a laptop computer, a desktop computer, etc.

[0076] The server 103 may be a server that provides various services, such as a background server that provides support for a web page displayed on the terminal device 101 .

[0077] It should be noted that the artificial intelligence-based data detection method provided in the embodiment of the present application is generally executed by a server / terminal device, and accordingly, the artificial intelligence-based data detection device is generally set in the server / terminal device.

[0078] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0079] Continue to refer Figure 2 , showing a flow chart of an embodiment of an artificial intelligence-based data detection method according to the present application. According to different requirements, the order of the steps in the flow chart can be changed, and some steps can be omitted. The artificial intelligence-based data detection method provided in the embodiment of the present application can be applied to any scenario where hallucination detection of responses to large model outputs is required, and the artificial intelligence-based data detection method can be applied to products in these scenarios, for example, hallucination detection of responses to large model outputs in the field of digital medicine. The artificial intelligence-based data detection method comprises the following steps:

[0080] Step S201, obtaining a target response corresponding to the target prompt output by a preset target macro model.

[0081] In this embodiment, the data detection method based on artificial intelligence is executed on the electronic device (e.g. Figure 1The server / terminal device shown in the figure) can obtain the target reply corresponding to the target prompt through a wired connection or a wireless connection. It should be noted that the above-mentioned wireless connection method may include but is not limited to 3G / 4G / 5G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other wireless connection methods currently known or to be developed in the future. The present application can be applied to the medical industry, and is specifically used for business scenarios in which hallucination detection is performed on the replies output by the medical big model. Among them, the above-mentioned target big model can specifically be a medical big model for providing medical advice to doctors and patients. There is no specific limitation on the selection of the above-mentioned medical big model, for example, a big model such as GPT-4 can be used. In addition, the above-mentioned target prompt can be a prompt (prompt) pre-built according to the patient's query. By inputting the prompt into the target big model, a reply output by the target big model, namely the above-mentioned target reply, will be obtained. The goal of this application is to detect whether there is a hallucination in the target reply (whether it conforms to the facts) and obtain the corresponding hallucination detection result.

[0082] Specifically, the present application proposes a method for hallucination detection based on multiple sampling. If for a prompt, the large model is very clear about the meaning of the prompt and knows how to respond, then using this prompt to let the large model generate multiple responses, the consistency of these responses will be very high, and the possibility of hallucinations in the generated content will be smaller; conversely, if the large model cannot understand or is not sure how to respond to this prompt, then the multiple responses generated will be different, the consistency will be lower, and the possibility of hallucinations in the generated content will be greater. Based on the above theory, the present application uses a large model to generate multiple results under the same conditions (same prompt, same generation strategy), and judges whether the generated text has hallucinations by the consistency of the results.

[0083] Step S202, segmenting the target response to obtain a plurality of corresponding clauses.

[0084] In this embodiment, the specific implementation process of segmenting the target response to obtain the corresponding multiple clauses will be described in further detail in subsequent specific embodiments of the present application and will not be elaborated on here.

[0085] Step S203, based on a preset generation strategy, use the target macro model to process the target prompt to obtain a specified number of first replies, and generate a corresponding reply set based on all the first replies.

[0086] In this embodiment, the strategy content of the above generation strategy may include calling the API of the target large model using selected parameters (such as temperature=1.0), then passing in the target prompt, and requesting to generate a specified number (N times) of replies. There is no specific limitation on the selection of the above parameters, and they can be set according to actual business usage requirements, as long as the generated multiple replies are obtained under the same conditions.

[0087] In addition, the number of times to generate replies, i.e., the above-mentioned specified number, can be determined by comprehensively considering computing resources, time limits, and the expected detection effect. Specifically, the process of determining the specified number includes: 1. Evaluate computing resources: First, understand the available computing resources, including the performance of the CPU, GPU, and memory size. This will directly affect the number of tasks that can be processed in parallel and the execution speed of each task. 2. Consider time limits: Determine the time limit for task completion. If time is tight, you may need to choose a smaller N value to reduce waiting time. 3. Weigh the detection effect: In theory, the larger the N value, the higher the diversity of the generated replies, which may improve the accuracy of hallucination detection. However, this will also significantly increase the computational cost. Therefore, it is necessary to weigh the detection effect and the computational cost and select an N value that can meet the detection requirements without over-consuming resources. 4. Preliminary test: After selecting an N value, you can conduct preliminary tests to evaluate its impact on the detection effect. If you find that the detection effect is not ideal, you can adjust the N value appropriately and retest. 5. Final determination: Based on the test results and actual conditions, a suitable N value is finally determined and used as the above-mentioned specified number.

[0088] Step S204: construct corresponding multiple detection prompts based on each of the clauses and the response set.

[0089] In this embodiment, the specific implementation process of constructing the corresponding multiple detection prompts based on each of the clauses and the response set will be further described in detail in subsequent specific embodiments of the present application and will not be elaborated on here.

[0090] Step S205: Process each of the detection prompts based on a preset third-party large model to obtain corresponding multiple second replies.

[0091] In this embodiment, the third-party large model is specifically a model different from the target large model, for example, the open source Llama-2, the closed source GPT-3.5-turbo, etc. can be selected. Each of the detection prompts is input into the third-party large model, and each detection prompt is processed by the third-party large model, and then the second response corresponding to each detection prompt is output. The number of the second responses is the same as the number of the detection prompts.

[0092] Step S206, scoring all the second replies to obtain corresponding multiple scores, and selecting the target score with the highest value from all the scores.

[0093] In this embodiment, the specific implementation process of scoring all the second replies to obtain the corresponding multiple scores will be described in further detail in subsequent specific embodiments of the present application and will not be elaborated on here.

[0094] Step S207: Generate a hallucination detection result corresponding to the target response based on the target score.

[0095] In this embodiment, the specific implementation process of generating the hallucination detection result corresponding to the target response based on the target score will be further described in detail in subsequent specific embodiments of the present application and will not be elaborated on here.

[0096] The present application obtains the target reply corresponding to the target prompt output by the preset target large model, and divides the target reply to obtain the corresponding multiple clauses. Then, based on the preset generation strategy, the target large model is used to process the target prompt to obtain a specified number of first replies, and a corresponding reply set is generated based on all the first replies. Then, based on each of the clauses and the reply set, a corresponding multiple detection prompts are constructed, and each of the detection prompts is processed respectively based on the preset third-party large model to obtain a corresponding multiple second replies. Subsequently, all the second replies are scored to obtain a corresponding multiple scores, and the target score with the highest value is selected from all the scores. Finally, the hallucination detection result corresponding to the target reply is generated based on the target score. The detection method for the large model hallucination provided by the present application does not require the support of additional data such as an external knowledge base, and only needs to rely on an additional, existing third-party large model to quickly and accurately complete the hallucination detection processing of the target reply output by the target large model, effectively improving the processing efficiency and applicability of the hallucination detection processing, and ensuring the accuracy of the obtained hallucination detection result.

[0097] In some optional implementations, step S202 includes the following steps:

[0098] Get a library of preset natural language processing tools.

[0099] In this embodiment, the natural language processing tool library can specifically adopt an open source natural language processing library. The natural language processing tool library provides a series of efficient and easy-to-use tools for tasks such as text preprocessing, text parsing, named entity recognition, part-of-speech tagging, syntactic analysis and text classification.

[0100] The target component is called from the natural language processing tool library.

[0101] In this embodiment, the target component may specifically be a component having a text segmentation function.

[0102] The target response is segmented based on the target component to obtain a plurality of clauses corresponding to the target response.

[0103] In this embodiment, the target reply is segmented by using the target component to segment the target reply into a corresponding plurality of clauses.

[0104] The present application obtains a preset natural language processing tool library; then calls a target component from the natural language processing tool library; and then segments the target reply based on the target component to obtain multiple clauses corresponding to the target reply. The present application obtains a natural language processing tool library, and then based on the use of the target component in the natural language processing tool library, it can quickly and accurately segment the target reply, thereby obtaining multiple clauses corresponding to the target reply, thereby improving the efficiency of clause generation and ensuring the accuracy of the obtained clauses.

[0105] In some optional implementations of this embodiment, step S204 includes the following steps:

[0106] Get the preset hallucination detection hint templates.

[0107] In this embodiment, the content of the hallucination detection prompt template may specifically include:

[0108] Context:{};

[0109] Text:{};

[0110] Is the text supported by the context above? Answer Yes orNo:.

[0111] Among them, the above Context:{} corresponds to the reply text area, and the above Text:{} corresponds to the clause area.

[0112] Gets the specified clause.

[0113] In this embodiment, the above-mentioned designated clause is any one of all clauses.

[0114] The designated clause is filled into the clause area in the hallucination detection prompt template, and the reply set is filled into the reply text area in the detection prompt template to obtain a filled hallucination detection prompt template.

[0115] In this embodiment, by determining the clause area and the reply text area in the above-mentioned hallucination detection prompt template, the specified clause is filled into the clause area in the hallucination detection prompt template, and the reply set is filled into the reply text area in the detection prompt template, so as to obtain the filled hallucination detection prompt template.

[0116] The filled hallucination detection prompt template is used as the designated detection prompt corresponding to the designated clause.

[0117] The present application obtains a preset hallucination detection prompt template; then obtains a designated clause; wherein the designated clause is any one of all the clauses; then fills the designated clause into the clause area in the hallucination detection prompt template, and fills the reply set into the reply text area in the detection prompt template, to obtain the filled hallucination detection prompt template; subsequently, the filled hallucination detection prompt template is used as the designated detection prompt corresponding to the designated clause. The present application uses the hallucination detection prompt template to fill the acquired designated clause into the clause area in the hallucination detection prompt template, and fills the reply set into the reply text area in the detection prompt template, so as to realize the rapid and accurate completion of the construction process of the designated detection prompt corresponding to the designated clause, and effectively improves the construction efficiency and construction intelligence of the detection prompt.

[0118] In some optional implementations, step S206 includes the following steps:

[0119] Invoke the preset scoring strategy.

[0120] In this embodiment, the strategy content of the above-mentioned scoring strategy may specifically include: inputting the detection prompt corresponding to each of the above-mentioned clauses into the above-mentioned third-party large model to obtain a response, if the response of the third-party large model clearly indicates "Yes" or a similar meaning, then the corresponding clause is scored 0 points, indicating that there is no hallucination. If the response of the third-party large model clearly indicates "No" or a similar meaning, then the corresponding clause is scored 1 point, indicating that there is a hallucination. If the response of the third-party large model is ambiguous or unclear (such as "Maybe", "Unsure", etc.), then the corresponding clause is scored 0.5 points, indicating that it is impossible to determine whether there is a hallucination.

[0121] The designated reply is analyzed based on the scoring strategy to obtain a corresponding reply analysis result; wherein the designated reply is any one of all the second replies.

[0122] In this embodiment, by analyzing the specified reply, a corresponding reply analysis result is obtained. The reply analysis result is a result of the reply type corresponding to the specified reply, and the reply analysis result may include a reply that clearly indicates "Yes" or a similar meaning, or a reply that clearly indicates "No" or a similar meaning, or a reply that is ambiguous or unclear.

[0123] A designated score corresponding to the designated reply is generated based on the reply analysis result.

[0124] In this embodiment, the reply analysis result of the designated reply may be score mapped according to the policy content of the above-mentioned scoring policy, so as to obtain a designated score corresponding to the designated reply.

[0125] The present application calls a preset scoring strategy; then analyzes the designated reply based on the scoring strategy to obtain a corresponding reply analysis result; wherein the designated reply is any one of all the second replies; and subsequently generates a designated score corresponding to the designated reply based on the reply analysis result. The present application analyzes the designated reply using a scoring strategy to obtain a corresponding reply analysis result, and then generates a designated score corresponding to the designated reply based on the reply analysis result, thereby achieving fast and intelligent completion of the scoring process for the second reply, improving the processing efficiency of the scoring process, and ensuring the accuracy of the obtained score.

[0126] In some optional implementations, step S207 includes the following steps:

[0127] Gets the preset hallucination score threshold.

[0128] In this embodiment, the above-mentioned hallucination score threshold will be used to determine whether the generated reply contains a hallucination. There is no specific limitation on the value of the above-mentioned hallucination score threshold, which can be determined based on experimental data or empirical judgment. Specifically, the process of determining the hallucination score threshold may include: 1. Collecting data: First, it is necessary to collect a set of data sets containing known hallucinations and non-hallucination replies. These data can be used to train or calibrate the detection model and help determine the appropriate threshold. 2. Experimental testing: Use a third-party large model to perform hallucination detection on the replies in the data set and record the score of each reply. Then, analyze the relationship between the score and the hallucination label to find the threshold that can best distinguish between hallucinations and non-hallucination replies. 3. Adjust the threshold: According to the experimental results, gradually adjust the threshold to optimize the detection performance. The detection effect under different thresholds can be evaluated by calculating indicators such as accuracy, recall, and F1 score. 4. Consider the application scenario: When setting the threshold, the specific needs of the application scenario also need to be considered. For example, in some applications, it may be more advantageous to reduce false positives (i.e. misclassifying non-hallucination responses as hallucinations), while in other applications, it may be more advantageous to reduce false negatives (i.e. misclassifying hallucination responses as non-hallucinations). 5. Finalization: Based on the experimental results and application requirements, a suitable threshold is finally determined as the hallucination score threshold. This threshold will be used in subsequent hallucination detection tasks to determine whether the generated responses contain hallucinations.

[0129] It is determined whether the target score is greater than the hallucination score threshold.

[0130] In this embodiment, the target score is numerically compared with the hallucination score threshold, and a corresponding numerical comparison result is obtained, wherein the numerical comparison result may include that the target score is greater than the hallucination score threshold, or that the target score is not greater than the hallucination score threshold.

[0131] If the target score is greater than the hallucination score threshold, a first hallucination detection result containing hallucinations in the target response is generated.

[0132] In this embodiment, if the target score is greater than the hallucination score threshold, it is determined that the target response contains hallucination, that is, at least one clause in the target response is highly suspected of containing hallucination, and a first hallucination detection result indicating that the target response contains hallucination is generated.

[0133] If the target score is not greater than the hallucination score threshold, a second hallucination detection result is generated in which the target response does not contain hallucinations.

[0134] In this embodiment, if the target score is greater than the hallucination score threshold, it is determined that the target response does not contain hallucinations or is less likely to contain hallucinations, that is, all clauses contained in the target response are considered unlikely to contain hallucinations, or the degree of hallucinations that exist is not enough to exceed the threshold, and a second hallucination detection result is generated that the target response does not contain hallucinations.

[0135] The present application obtains a preset hallucination score threshold; then determines whether the target score is greater than the hallucination score threshold; if the target score is greater than the hallucination score threshold, a first hallucination detection result containing hallucinations in the target response is generated; and if the target score is not greater than the hallucination score threshold, a second hallucination detection result not containing hallucinations in the target response is generated. The present application obtains a preset hallucination score threshold, and then compares the target score with the hallucination score threshold numerically, and then can quickly and accurately generate a hallucination detection result corresponding to the target response based on the obtained numerical comparison result, thereby ensuring the data accuracy of the obtained hallucination detection result.

[0136] In some optional implementations of this embodiment, after step S207, the electronic device may further perform the following steps:

[0137] Content analysis is performed on the hallucination detection result, and if it is detected that the hallucination detection result indicates that the target reply contains a hallucination, correction processing is performed on the target reply to obtain a corresponding corrected reply.

[0138] In this embodiment, the target response can be corrected using a preset correction strategy to obtain a corresponding corrected response. The correction strategy may include: regenerating the response, using manual review to modify the erroneous or hallucinatory parts of the response, or adjusting the parameters or inputs of the generation model according to the context and expected output.

[0139] Get the preset reminder information.

[0140] In this embodiment, the reminder information is pre-constructed to send a warning to the user and point out the inaccurate or illusory information that may exist in the reply. In addition, additional information or suggestions can be provided in the reminder information to help the user better understand the problem and make a decision.

[0141] A corresponding reply process is performed based on the revised reply and the reminder information.

[0142] In this embodiment, the revised reply and the reminder information can be sent to the relevant users through various communication channels to complete the corresponding reply processing. The communication channels can include text messages, pop-up windows, emails, SMS, etc.

[0143] The present application performs content analysis on the hallucination detection result. If it is detected that the hallucination detection result is that the target reply contains hallucinations, the target reply is corrected to obtain a corresponding corrected reply; then a preset reminder message is obtained; and a corresponding reply is subsequently processed based on the corrected reply and the reminder message. After the present application generates an hallucination detection result corresponding to the target reply based on the target score, if it is detected that the hallucination detection result is that the target reply contains hallucinations, the target reply is intelligently corrected to obtain a corresponding corrected reply, and preset reminder messages are obtained, and then a corresponding reply is processed based on the corrected reply and the reminder message. By correcting the target reply to obtain a corresponding corrected reply, the hallucination problem in the target reply can be corrected, thereby ensuring the data accuracy of the corrected reply. In addition, the reminder message and the corrected reply are intelligently returned to the corresponding user at the same time, which improves the intelligence of the reply processing and is conducive to improving the user's experience.

[0144] In some optional implementations of this embodiment, after step S207, the electronic device may further perform the following steps:

[0145] Acquire hallucination detection process data corresponding to the target response.

[0146] In this embodiment, the hallucination detection process data may include at least a plurality of clauses corresponding to the target response generated during the hallucination detection process corresponding to the target response, and a score of hallucination detection. In addition, if the target response is corrected, the hallucination detection process data may also include data such as the corrected response content, the time and method of reminding the user, etc.

[0147] Determine the target storage method.

[0148] In this embodiment, the selection of the above target storage method is not specifically limited and can be determined according to actual business usage requirements. Exemplarily, the target storage method can adopt any one of local database storage, cloud server storage, blockchain storage, etc. Among them, blockchain storage is preferably used as the above target storage method. By using blockchain to store and manage data, the security and immutability of data can be effectively guaranteed.

[0149] The target response, the hallucination detection result and the hallucination detection process data are stored and processed based on the target storage method.

[0150] In this embodiment, the target storage medium corresponding to the above-mentioned target storage method is called, and then the target response, the hallucination detection result and the hallucination detection process data are stored in the target storage medium to complete the corresponding storage processing.

[0151] The present application obtains the hallucination detection process data corresponding to the target response; then determines the target storage method; and subsequently stores and processes the target response, the hallucination detection result, and the hallucination detection process data based on the target storage method. After generating the hallucination detection result corresponding to the target response based on the target score, the present application also intelligently obtains the hallucination detection process data corresponding to the target response, and further uses the target storage method to store and process the target response, the hallucination detection result, and the hallucination detection process data, effectively ensuring the data security of the target response, the hallucination detection result, and the hallucination detection process data. In addition, by recording the target response, the hallucination detection result, and the hallucination detection process data, it can help to provide corresponding analysis and auditing when tracing the source and accuracy of specific responses later.

[0152] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0153] It should be emphasized that in order to further ensure the privacy and security of the above-mentioned hallucination detection results, the above-mentioned hallucination detection results can also be stored in a node of a blockchain.

[0154] The blockchain referred to in this application is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, platform product service layer, and application service layer.

[0155] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0156] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0157] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0158] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0159] Further references Figure 3 , as a response to the above Figure 2 The present application provides an embodiment of a data detection device based on artificial intelligence, and the device embodiment is similar to Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0160] like Figure 3 As shown, the artificial intelligence-based data detection device 300 described in this embodiment includes: a first acquisition module 301, a segmentation module 302, a first processing module 303, a construction module 304, a second processing module 305, a scoring module 306 and a generation module 307. Among them:

[0161] The first acquisition module 301 is used to acquire a target response corresponding to the target prompt output by a preset target large model;

[0162] A segmentation module 302, configured to segment the target response to obtain a plurality of corresponding clauses;

[0163] A first processing module 303 is used to process the target prompt using the target macro model based on a preset generation strategy to obtain a specified number of first replies, and generate a corresponding reply set based on all the first replies;

[0164] A construction module 304, configured to construct a plurality of corresponding detection prompts based on each of the clauses and the response set;

[0165] The second processing module 305 is used to process each of the detection prompts based on a preset third-party large model to obtain a corresponding plurality of second replies;

[0166] A scoring module 306 is used to score all the second replies to obtain corresponding multiple scores, and select a target score with the highest value from all the scores;

[0167] The generating module 307 is used to generate a hallucination detection result corresponding to the target response based on the target score.

[0168] In this embodiment, the operations performed by the above modules or units correspond one-to-one to the steps of the artificial intelligence-based data detection method in the aforementioned implementation mode, and will not be repeated here.

[0169] In some optional implementations of this embodiment, the segmentation module 302 includes:

[0170] A first acquisition submodule is used to acquire a preset natural language processing tool library;

[0171] A first calling submodule, used to call a target component from the natural language processing tool library;

[0172] The segmentation submodule is used to segment the target response based on the target component to obtain multiple clauses corresponding to the target response.

[0173] In this embodiment, the operations performed by the above modules or units correspond one-to-one to the steps of the artificial intelligence-based data detection method in the aforementioned implementation mode, and will not be repeated here.

[0174] In some optional implementations of this embodiment, the construction module 304 includes:

[0175] The second acquisition submodule is used to acquire a preset hallucination detection prompt template;

[0176] A third acquisition submodule is used to acquire a specified clause; wherein the specified clause is any one of all the clauses;

[0177] A filling submodule, used to fill the specified clause into the clause area in the hallucination detection prompt template, and fill the reply set into the reply text area in the detection prompt template, to obtain a filled hallucination detection prompt template;

[0178] A determination submodule is used to use the filled hallucination detection prompt template as a designated detection prompt corresponding to the designated clause.

[0179] In this embodiment, the operations performed by the above modules or units correspond one-to-one to the steps of the artificial intelligence-based data detection method in the aforementioned implementation mode, and will not be repeated here.

[0180] In some optional implementations of this embodiment, the scoring module 306 includes:

[0181] The second calling submodule is used to call the preset scoring strategy;

[0182] An analysis submodule, configured to analyze a designated reply based on the scoring strategy to obtain a corresponding reply analysis result; wherein the designated reply is any one of all the second replies;

[0183] The first generating submodule is used to generate a specified score corresponding to the specified reply based on the reply analysis result.

[0184] In this embodiment, the operations performed by the above modules or units correspond one-to-one to the steps of the artificial intelligence-based data detection method in the aforementioned implementation mode, and will not be repeated here.

[0185] In some optional implementations of this embodiment, the generating module 307 includes:

[0186] A fourth acquisition submodule is used to acquire a preset hallucination score threshold;

[0187] A judging submodule, used to judge whether the target score is greater than the hallucination score threshold;

[0188] A second generating submodule, configured to generate a first hallucination detection result containing hallucination in the target response if the target score is greater than the hallucination score threshold;

[0189] The third generating submodule is used for generating a second hallucination detection result in which the target response does not contain hallucinations if the target score is not greater than the hallucination score threshold.

[0190] In this embodiment, the operations performed by the above modules or units correspond one-to-one to the steps of the artificial intelligence-based data detection method in the aforementioned implementation mode, and will not be repeated here.

[0191] In some optional implementations of this embodiment, the data detection device based on artificial intelligence further includes:

[0192] A correction module is used to perform content analysis on the hallucination detection result, and if it is detected that the hallucination detection result is that the target reply contains a hallucination, then the target reply is corrected to obtain a corresponding corrected reply;

[0193] The second acquisition module is used to obtain preset reminder information;

[0194] A reply module is used to perform corresponding reply processing based on the revised reply and the reminder information.

[0195] In this embodiment, the operations performed by the above modules or units correspond one-to-one to the steps of the artificial intelligence-based data detection method in the aforementioned implementation mode, and will not be repeated here.

[0196] In some optional implementations of this embodiment, the data detection device based on artificial intelligence further includes:

[0197] A third acquisition module is used to acquire hallucination detection process data corresponding to the target response;

[0198] A determination module, used for determining a target storage mode;

[0199] A storage module is used to store and process the target response, the hallucination detection result and the hallucination detection process data based on the target storage method.

[0200] In this embodiment, the operations performed by the above modules or units correspond one-to-one to the steps of the artificial intelligence-based data detection method in the aforementioned implementation mode, and will not be repeated here.

[0201] To solve the above technical problems, the present application also provides a computer device. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0202] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected through a system bus. It should be noted that the figure only shows a computer device 4 with components 41-43, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), programmable gate arrays (Field-Programmable Gate Array, FPGA), digital processors (Digital Signal Processor, DSP), embedded devices, etc.

[0203] The computer device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with a user through a keyboard, a mouse, a remote controller, a touch pad, or a voice control device.

[0204] The memory 41 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 41 can be an internal storage unit of the computer device 4, such as a hard disk or memory of the computer device 4. In other embodiments, the memory 41 can also be an external storage device of the computer device 4, such as a plug-in hard disk equipped on the computer device 4, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), etc. Of course, the memory 41 can also include both the internal storage unit of the computer device 4 and its external storage device. In this embodiment, the memory 41 is generally used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions of a data detection method based on artificial intelligence, etc. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or are to be output.

[0205] The processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. The processor 42 is generally used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to run computer-readable instructions stored in the memory 41 or process data, such as computer-readable instructions for running the artificial intelligence-based data detection method.

[0206] The network interface 43 may include a wireless network interface or a wired network interface. The network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0207] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0208] In the embodiment of the present application, firstly, the target reply corresponding to the target prompt output by the preset target large model is obtained; and the target reply is segmented to obtain the corresponding multiple clauses; then, based on the preset generation strategy, the target large model is used to process the target prompt to obtain a specified number of first replies, and a corresponding reply set is generated based on all the first replies; then, corresponding multiple detection prompts are constructed based on each of the clauses and the reply set; subsequently, each of the detection prompts is processed based on the preset third-party large model to obtain the corresponding multiple second replies; further, all the second replies are scored to obtain the corresponding multiple scores, and the target score with the highest value is selected from all the scores; finally, the hallucination detection result corresponding to the target reply is generated based on the target score. The detection method for large model hallucination provided by the present application does not require the support of additional data such as an external knowledge base, and only needs to rely on an additional, existing third-party large model to quickly and accurately complete the hallucination detection processing of the target reply output by the target large model, which effectively improves the processing efficiency and applicability of the hallucination detection processing and ensures the accuracy of the obtained hallucination detection result.

[0209] The present application also provides another embodiment, namely, providing a computer-readable storage medium, wherein the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to enable the at least one processor to perform the steps of the artificial intelligence-based data detection method as described above.

[0210] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0211] In the embodiment of the present application, firstly, the target reply corresponding to the target prompt output by the preset target large model is obtained; and the target reply is segmented to obtain the corresponding multiple clauses; then, based on the preset generation strategy, the target large model is used to process the target prompt to obtain a specified number of first replies, and a corresponding reply set is generated based on all the first replies; then, corresponding multiple detection prompts are constructed based on each of the clauses and the reply set; subsequently, each of the detection prompts is processed based on the preset third-party large model to obtain the corresponding multiple second replies; further, all the second replies are scored to obtain the corresponding multiple scores, and the target score with the highest value is selected from all the scores; finally, the hallucination detection result corresponding to the target reply is generated based on the target score. The detection method for large model hallucination provided by the present application does not require the support of additional data such as an external knowledge base, and only needs to rely on an additional, existing third-party large model to quickly and accurately complete the hallucination detection processing of the target reply output by the target large model, which effectively improves the processing efficiency and applicability of the hallucination detection processing and ensures the accuracy of the obtained hallucination detection result.

[0212] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0213] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A data detection method based on artificial intelligence, characterized in that: The steps include: Obtaining a target response corresponding to the target prompt output by a preset target large model; Segmenting the target response to obtain a plurality of corresponding clauses; Based on a preset generation strategy, the target prompt is processed using the target macro model to obtain a specified number of first replies, and a corresponding reply set is generated based on all the first replies; Constructing a plurality of corresponding detection prompts based on each of the clauses and the response set; Processing each of the detection prompts based on a preset third-party large model to obtain corresponding multiple second replies; Scoring all the second replies to obtain a plurality of corresponding scores, and selecting a target score with the highest value from all the scores; A hallucination detection result corresponding to the target response is generated based on the target score.

2. The artificial intelligence-based data detection method according to claim 1, characterized in that: The step of segmenting the target response to obtain a plurality of corresponding clauses specifically includes: Get a library of preset natural language processing tools; Calling a target component from the natural language processing tool library; The target response is segmented based on the target component to obtain a plurality of clauses corresponding to the target response.

3. The artificial intelligence-based data detection method according to claim 1, characterized in that: The step of constructing a plurality of corresponding detection prompts based on each of the clauses and the response set specifically includes: Get the preset hallucination detection prompt template; Obtain a specified clause; wherein the specified clause is any one of all the clauses; Filling the specified clause into the clause area in the hallucination detection prompt template, and filling the reply set into the reply text area in the detection prompt template, to obtain a filled hallucination detection prompt template; The filled hallucination detection prompt template is used as the designated detection prompt corresponding to the designated clause.

4. The artificial intelligence-based data detection method according to claim 1, characterized in that: The step of scoring all the second replies to obtain corresponding multiple scores specifically includes: Call the preset scoring strategy; Analyze the designated reply based on the scoring strategy to obtain a corresponding reply analysis result; wherein the designated reply is any one of all the second replies; A designated score corresponding to the designated reply is generated based on the reply analysis result.

5. The artificial intelligence-based data detection method according to claim 1, characterized in that: The step of generating a hallucination detection result corresponding to the target response based on the target score specifically includes: Get the preset hallucination score threshold; Determining whether the target score is greater than the hallucination score threshold; If the target score is greater than the hallucination score threshold, generating a first hallucination detection result including hallucination in the target response; If the target score is not greater than the hallucination score threshold, a second hallucination detection result is generated in which the target response does not contain hallucinations.

6. The artificial intelligence-based data detection method according to claim 1, characterized in that: After the step of generating a hallucination detection result corresponding to the target response based on the target score, the method further includes: Performing content analysis on the hallucination detection result, and if it is detected that the hallucination detection result indicates that the target response contains a hallucination, performing correction processing on the target response to obtain a corresponding corrected response; Get preset reminder information; A corresponding reply process is performed based on the revised reply and the reminder information.

7. The artificial intelligence-based data detection method according to claim 1, characterized in that: After the step of generating a hallucination detection result corresponding to the target response based on the target score, the method further includes: Acquiring hallucination detection process data corresponding to the target response; Determine the target storage method; The target response, the hallucination detection result and the hallucination detection process data are stored and processed based on the target storage method.

8. A data detection device based on artificial intelligence, characterized in that: include: A first acquisition module is used to acquire a target response corresponding to the target prompt output by a preset target large model; A segmentation module, used for segmenting the target response to obtain a plurality of corresponding clauses; A first processing module, configured to process the target prompt using the target macro model based on a preset generation strategy to obtain a specified number of first replies, and generate a corresponding reply set based on all the first replies; A construction module, configured to construct a plurality of corresponding detection prompts based on each of the clauses and the response set; A second processing module, used for processing each of the detection prompts based on a preset third-party large model to obtain a corresponding plurality of second replies; A scoring module, used for scoring all the second replies to obtain corresponding multiple scores, and selecting a target score with the highest value from all the scores; A generation module is used to generate a hallucination detection result corresponding to the target response based on the target score.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the steps of the artificial intelligence-based data detection method as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the artificial intelligence-based data detection method according to any one of claims 1 to 7.

Images