Information detection method and device, equipment, storage medium and product

Abstract

The invention discloses an information detection method and device, equipment, a storage medium and a product, relates to the technical field of data security, and discloses a method for performing knowledge retrieval based on to-be-detected information corresponding to an information detection instruction in response to the information detection instruction to obtain at least one reference knowledge point associated with the to-be-detected information; performing conflict detection on the to-be-detected information based on each reference knowledge point to obtain a conflict detection result corresponding to each reference knowledge point; and generating an attribute detection result of the to-be-detected information based on the conflict detection result corresponding to each reference knowledge point. The method can effectively improve the accuracy and interpretability of the detection result.

Description

Technical Field

[0001] The present application relates to the field of data security technology, and in particular to an information detection method, device, equipment, storage medium and product. Background Art

[0002] In the era of information explosion, attribute detection of massive amounts of information and determining whether it contains specific attributes such as rumors, confidential information, etc. has become a key link in information security management and content governance.

[0003] At present, the use of large language models for information attribute detection has become a mainstream technical means. By inputting the information to be detected into the model, based on its powerful semantic understanding and generation capabilities, the detection results can be quickly output. However, this method has significant defects: the detection logic of the large language model belongs to the "black box" mechanism, and its internal decision-making process is difficult to be observed and explained by the outside world. Users can only obtain the final judgment conclusion, but cannot know what knowledge the model is based on and what reasoning process it uses to reach the conclusion, resulting in a lack of explainability of the detection logic and results. This unexplainability not only hinders the verification of the accuracy of the test results, but also makes it difficult for users to trust the test conclusions.

[0004] Therefore, providing an information attribute detection method that is both accurate and interpretable has become an important technical issue that needs to be solved urgently.

[0005] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention

[0006] The main purpose of this application is to provide an information detection method, device, equipment, storage medium and product that can effectively improve the accuracy and interpretability of detection results.

[0007] To achieve the above purpose, the present application proposes an information detection method, which comprises: In response to the information detection instruction, performing knowledge retrieval based on the information to be detected corresponding to the information detection instruction to obtain at least one reference knowledge point associated with the information to be detected; Based on each reference knowledge point, conflict detection is performed on the information to be detected to obtain a conflict detection result corresponding to each reference knowledge point; Based on the conflict detection results corresponding to the reference knowledge points, an attribute detection result of the information to be detected is generated.

[0008] Optionally, the generating the attribute detection result of the information to be detected based on the conflict detection result corresponding to each reference knowledge point includes: Generate a comprehensive detection result based on the conflict detection results corresponding to each reference knowledge point; Determine the business scenario to which the information detection instruction belongs; Based on the conversion rules corresponding to the business scenario, convert the content of the comprehensive detection result to obtain the attribute detection result.

[0009] Optionally, the conflict detection result indicates whether the content of each reference knowledge point conflicts with the content of the information to be detected; the attribute detection result indicates whether the information to be detected belongs to the target attribute information, and includes an explanation of the reason why the information to be detected belongs to the target attribute information.

[0010] Optionally, generating the attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point includes: When there is a target knowledge point among the at least one reference knowledge point, indicate in the attribute detection result that the information to be detected belongs to the target attribute information, where the target knowledge point is a knowledge point that conflicts with the content of the information to be detected; Based on the target knowledge point, generate an explanation of the reason why the information to be detected belongs to the target attribute information.

[0011] Optionally, generating the explanation of the reason why the information to be detected belongs to the target attribute information based on the target knowledge point includes: Obtain the metadata of the target knowledge point; Based on the target knowledge point and the metadata of the target knowledge point, generate an explanation of the reason why the information to be detected belongs to the target attribute information.

[0012] Optionally, the metadata of the target knowledge point includes at least one of the following: The original document to which the target knowledge point belongs, the knowledge point serial number corresponding in the original document, the attribute label, the release date, the knowledge point summary, the knowledge source, and the uniform resource locator.

[0013] Optionally, generating the attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point includes: Obtain a set of constraint rules, where the set of constraint rules is used to restrict the content and format of the attribute detection result; Through a first large language model, based on the conflict detection results corresponding to each reference knowledge point, generate an attribute detection result that meets the requirements of the set of constraint rules.

[0014] Optionally, generating the attribute detection result that meets the requirements of the set of constraint rules through the first large language model based on the conflict detection results corresponding to each reference knowledge point includes: Generate multiple candidate words through the first large language model based on the context sequence formed by the conflict detection results corresponding to each reference knowledge point; Filter out the candidate words that do not meet the requirements of the set of constraint rules from the multiple candidate words, and select candidate words from the remaining candidate words and add them to the context sequence to obtain an updated context sequence; Regenerate the next candidate word based on the updated context sequence until the attribute detection result of the information to be detected is obtained.

[0015] Optionally, the responding to the information detection instruction and performing knowledge retrieval based on the information to be detected corresponding to the information detection instruction to obtain at least one reference knowledge point associated with the information to be detected includes: Respond to the information detection instruction, perform attribute detection on the information to be detected through the second large language model to obtain an initial detection result, where the initial detection result indicates whether the information to be detected belongs to the target attribute information; In the case where the initial detection result indicates that the information to be detected belongs to the target attribute information, perform knowledge retrieval based on the information to be detected to obtain at least one reference knowledge point associated with the information to be detected.

[0016] Optionally, before the responding to the information detection instruction and performing attribute detection on the information to be detected through the second large language model to obtain an initial detection result, the method further includes: Obtain a sample data set, where the sample data set includes sample information and sample labels, and the sample labels indicate whether the sample information belongs to the target attribute information; Based on the sample data set, fine-tune the third large language model so that the initial detection result generated by the second large language model obtained by fine-tuning is a binary classification label, where the binary classification label includes a first label or a second label, the first label indicates that the information to be detected belongs to the target attribute information, and the second label indicates that the information to be detected does not belong to the target attribute information.

[0017] Optionally, the method further includes: Obtain text information from multiple sources; Extract knowledge points from the text information through the first large language model to obtain multiple knowledge points; Store the multiple knowledge points in a knowledge base for knowledge retrieval.

[0018] Optionally, the method further includes: Classify multiple pieces of information to be detected detected within a historical time period according to their semantic topics to obtain multiple topic data sets; Obtain the detection performance parameters corresponding to each topic dataset; When the detection performance parameter corresponding to any semantic topic dataset is less than the performance threshold, update the knowledge belonging to the semantic topic in the knowledge base, where the knowledge base is used for knowledge retrieval.

[0019] In addition, to achieve the above object, the present application also proposes an information detection device, the device includes: A knowledge retrieval module, configured to respond to an information detection instruction, perform knowledge retrieval based on the information to be detected corresponding to the information detection instruction, and obtain at least one reference knowledge point associated with the information to be detected; A conflict detection module, configured to perform conflict detection on the information to be detected based on each reference knowledge point, and obtain a conflict detection result corresponding to each reference knowledge point; An attribute detection module, configured to generate an attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point.

[0020] Optionally, the attribute detection module is configured to generate a comprehensive detection result based on the conflict detection results corresponding to each reference knowledge point; determine the business scenario to which the information detection instruction belongs; and convert the content of the comprehensive detection result based on the conversion rule corresponding to the business scenario to obtain the attribute detection result.

[0021] Optionally, the conflict detection result indicates whether the content of each reference knowledge point conflicts with the information to be detected; the attribute detection result indicates whether the information to be detected belongs to the target attribute information, and includes a reason explanation for whether the information to be detected belongs to the target attribute information.

[0022] Optionally, the attribute detection module includes: An attribute detection unit, configured to indicate in the attribute detection result that the information to be detected belongs to the target attribute information when there is a target knowledge point in the at least one reference knowledge point, where the target knowledge point is a knowledge point that conflicts with the content of the information to be detected; A reason generation unit, configured to generate a reason explanation for the information to be detected belonging to the target attribute information based on the target knowledge point.

[0023] Optionally, the reason generation unit is configured to obtain the metadata of the target knowledge point; and generate a reason explanation for the information to be detected belonging to the target attribute information based on the target knowledge point and the metadata of the target knowledge point.

[0024] Optionally, the metadata of the target knowledge point includes at least one of the following: The original document to which the target knowledge point belongs, the corresponding knowledge point serial number within the original document, the attribute label, the release date, the knowledge point summary, the knowledge source, and the uniform resource locator.

[0025] Optionally, the attribute detection module includes: A rule acquisition unit for acquiring a set of constraint rules, where the set of constraint rules is used to restrict the content and format of the attribute detection result; A result generation unit for generating an attribute detection result that meets the requirements of the set of constraint rules based on the conflict detection results corresponding to each reference knowledge point through a first large language model.

[0026] Optionally, the result generation unit is used to generate a plurality of candidate words through the first large language model based on the context sequence formed by the conflict detection results corresponding to each reference knowledge point; filter out the candidate words that do not meet the requirements of the set of constraint rules from the plurality of candidate words, and select candidate words from the remaining candidate words and add them to the context sequence to obtain an updated context sequence; regenerate the next candidate word based on the updated context sequence until the attribute detection result of the information to be detected is obtained.

[0027] Optionally, the knowledge retrieval module is used to, in response to the information detection instruction, perform attribute detection on the information to be detected through a second large language model to obtain an initial detection result, where the initial detection result indicates whether the information to be detected belongs to the target attribute information; in the case where the initial detection result indicates that the information to be detected belongs to the target attribute information, perform knowledge retrieval based on the information to be detected to obtain at least one reference knowledge point associated with the information to be detected.

[0028] Optionally, the device further includes: A model training module for obtaining a sample data set, where the sample data set includes sample information and sample labels, and the sample labels indicate whether the sample information belongs to the target attribute information; based on the sample data set, fine-tune a third large language model so that the initial detection result generated by the fine-tuned second large language model is a binary classification label, where the binary classification label includes a first label or a second label, the first label indicates that the information to be detected belongs to the target attribute information, and the second label indicates that the information to be detected does not belong to the target attribute information.

[0029] Optionally, the device further includes: A knowledge storage module for obtaining text information from multiple sources; extracting knowledge points from the text information through a first large language model to obtain a plurality of knowledge points; storing the plurality of knowledge points in a knowledge base, where the knowledge base is used for knowledge retrieval.

[0030] Optionally, the device further includes: A knowledge update module, configured to classify multiple pieces of information to be detected detected within a historical time period according to their semantic topics, to obtain multiple topic data sets; acquire detection performance parameters corresponding to each topic data set; and update the knowledge belonging to the semantic topic in a knowledge base when the detection performance parameter corresponding to any semantic topic data set is less than a performance threshold, where the knowledge base is used for knowledge retrieval.

[0031] In addition, to achieve the above object, the present application further provides an information detection device, where the device includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the computer program is configured to implement the steps of the information detection method as described above.

[0032] In addition, to achieve the above object, the present application further provides a storage medium, where the storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium, and when the computer program is executed by a processor, it implements the steps of the information detection method as described above.

[0033] In addition, to achieve the above object, the present application further provides a computer program product, where the computer program product includes a computer program, and when the computer program is executed by a processor, it implements the steps of the information detection method as described above.

[0034] One or more technical solutions proposed by the present application have at least the following technical effects: The information detection solution provided by the present application provides a clear basis for detection by retrieving reference knowledge points associated with the information to be detected, avoiding the drawbacks of solely relying on the subjective judgment of large language models. By performing conflict detection based on each reference knowledge point respectively to obtain the conflict detection results of each reference knowledge point, and then generating the attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point, the complex information detection is disassembled into multiple traceable sub-processes, making the detection logic transparent, and enabling users to clearly understand the source and basis of the judgment. Thus, the dilemma of the unexplainable detection result is broken, and the user's trust in the detection result is greatly improved. In addition, this detection solution with multi-knowledge point cross-verification also enhances the accuracy and reliability of the detection result, enabling the solution to obtain trustworthy attribute detection results in various detection scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings here are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present application, and are used together with the specification to explain the principles of the present application.

[0036] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0037] Figure 1 Schematic diagram of an implementation environment for the information detection method of the present application; Figure 2 Flow chart provided by the first embodiment of the information detection method of the present application; Figure 3 Flow chart provided by the second embodiment of the information detection method of the present application; Figure 4 Flow chart provided by the third embodiment of the information detection method of the present application; Figure 5 Flow chart provided by the fourth embodiment of the information detection method of the present application; Figure 6 Schematic diagram of an information detection system provided by the embodiments of the present application; Figure 7 Module structure diagram of the information detection device in the embodiments of the present application; Figure 8 Device structure diagram of the hardware operating environment involved in the information detection method in the embodiments of the present application.

[0038] The realization of the purpose, functional features, and advantages of the present application will be further described with reference to the embodiments and the drawings. Specific implementation manners

[0039] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0040] To better understand the technical solutions of the present application, the following will be described in detail with reference to the drawings in the specification and the specific implementation manners.

[0041] Figure 1 Schematic diagram of an implementation environment provided by an embodiment of the present disclosure. Refer to Figure 1 , this implementation environment includes a terminal 101 and a server 102. The terminal 101 and the server 102 are connected through a wireless or wired network. Exemplarily, a target application provided by the server 102 is installed on the terminal 101, and the terminal 101 can implement functions such as data transmission and message interaction through this target application.

[0042] Exemplarily, the terminal 101 is a computer, a mobile phone, a tablet computer or other terminals. Exemplarily, the target application is the target application in the operating system of the terminal 101, or the target application provided by a third party. For example, the target application is a search application, an information detection application or a social application, etc. This application can determine whether the information to be detected belongs to the target attribute information. For example, it can determine whether the information to be detected belongs to a rumor, whether it is confidential, or whether it is hallucination information generated by a model. The server 102 is the background server corresponding to the target application.

[0043] In this application, the terminal 101 is used to send an information detection instruction to the server in response to the user's information detection operation. For example, after the user inputs the information to be detected or selects the information to be detected in the terminal 101 and clicks the information detection control, the terminal 101 sends an information detection instruction carrying the information to be detected to the server 102. Correspondingly, the server 102 responds to the information detection instruction, performs knowledge retrieval based on the information to be detected corresponding to the information detection instruction, and obtains at least one reference knowledge point associated with the information to be detected; performs conflict detection on the information to be detected based on each reference knowledge point, and obtains the conflict detection results corresponding to each reference knowledge point; generates an attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point. Then, the server 102 returns the attribute detection result of the information to be detected to the terminal 101. The terminal 101 can display the attribute detection result of the information to be detected on the interface.

[0044] Alternatively, the above information detection process can also be completed by the terminal 101 alone. Or, the terminal 101 completes it through the installed target application. The embodiments of this application do not limit this.

[0045] The information detection method provided by this application is applicable to various scenarios. Optionally, a rumor detection scenario. For example, the user inputs the information to be detected in the terminal and clicks the rumor detection button. The terminal sends a rumor detection instruction carrying the information to be detected to the server. The server retrieves the factual information associated with the information to be detected as a reference knowledge point, and determines whether the information to be detected belongs to a rumor according to whether the content of each reference knowledge point conflicts with the information to be detected, and then returns the rumor detection result to the terminal.

[0046] Optionally, a classified information detection scenario. For example, when an enterprise employee sends a specific file through the office system, the system automatically triggers classified information detection. That is, the office system sends an information detection instruction containing the file content to the server. The server retrieves the enterprise confidentiality system and national confidentiality regulations associated with the file content from the knowledge base as reference knowledge points. It detects whether there are rule conflicts between each reference knowledge point and the file content. For example, the file content contains confidential information that is not allowed to be privately disseminated by the reference knowledge point. Then, based on the conflict detection results of each reference knowledge point, a classified information detection result is generated. Thus, according to the classified information detection result, data protection measures such as intercepting the sending of this file are determined.

[0047] Optionally, a hallucination detection scenario. For example, after a large language model generates an answer to a question in a question-and-answer system, an information detection instruction carrying the answer content is sent to the server for information detection. The server retrieves the factual information associated with the answer content as a reference knowledge point, detects whether there are conflicts between each reference knowledge point and the answer content, and thus generates a hallucination detection result to indicate whether the answer content is hallucination information generated by the model, that is, information that does not conform to logic or violates facts. Thus, according to the hallucination detection result, it is determined whether to change the answer strategy of the model so that the model regenerates the answer.

[0048] Figure 2 This is a schematic flowchart of the first embodiment of the information detection method of this application. Refer to Figure 2 , taking the server as the execution subject as an example, the information detection method includes the following steps S10 to S30: Step S10, in response to the information detection instruction, perform knowledge retrieval based on the information to be detected corresponding to the information detection instruction, and obtain at least one reference knowledge point associated with the information to be detected.

[0049] The information detection instruction is an operation instruction triggered by a user or a system, requesting to detect the attributes of specific information. For example, a user clicks the "Detect whether this information is a rumor" button on a social platform; a classified information detection request triggered when an enterprise office system automatically intercepts an externally sent file, etc. The information detection instruction contains the content of the information to be detected. In addition, the information detection instruction can also contain the metadata of the information to be detected, such as the information source, the business scenario it belongs to, the information type, the timestamp, the detection type, etc. Among them, the detection type can be rumor detection, classified information detection, hallucination detection, legality detection, etc.

[0050] The information to be detected is the original content of the property to be verified, which can be presented in various forms. For example, text, pictures, files, AI (Artificial Intelligence) generated content, etc. Knowledge retrieval refers to the process of screening relevant reference bases from an authoritative knowledge base based on the information to be detected. The reference knowledge points are the authoritative bases directly related to the information to be detected retrieved from the knowledge base. It can be various forms of data such as original documents, knowledge points, knowledge point summaries, text paragraphs, etc.

[0051] Among them, the knowledge base can cover various types. For example, a fact base that stores government open data, scientific literature, etc. for rumor detection or hallucination detection; a rule base that stores provisions of the Network Security Law, enterprise confidentiality systems, etc. for compliance or confidentiality detection; a case base that stores adjudicated similar cases for assisting in decision-making. The knowledge base can also cover data from various sources, such as Baidu Encyclopedia, Weibo, WeChat, official websites, etc. The knowledge base can also cover data with various attributes, such as rumors, popular science, confidentiality, etc. The knowledge base can also have various data storage forms. For example, in the knowledge base for storing reliable source manuscripts and the content of Baidu Encyclopedia, the data is stored in a segmented form. In the knowledge base for storing rumor information, the data is stored in the form of rumor knowledge points. In the knowledge base for storing popular science information, the data is stored in the form of popular science knowledge points.

[0052] Optionally, the construction method of the knowledge base is: obtaining text information from multiple sources; extracting knowledge points from the text information through a first large language model to obtain multiple knowledge points; storing the multiple knowledge points in the knowledge base, and the knowledge base is used for knowledge retrieval.

[0053] Among them, the data sources can include open data, industry norms, internal systems, etc. For example, for the rumor knowledge base and the popular science knowledge base, integrating the authoritative data of 231 national and provincial-level rumor refutation sites provides dynamic knowledge support for rumor detection. Through web crawler technology, adapting a customized crawler mechanism for the data interfaces of platforms such as Weibo and WeChat official accounts and the structure of official rumor refutation websites to achieve dynamic data collection and real-time update.

[0054] The large language model is a natural language processing model based on deep learning, which has the ability to understand, generate, and process large-scale text. In the construction of the knowledge base, it is used to automatically parse text information, extract key knowledge points, and improve the efficiency of knowledge extraction. For example, the large language model can be DeepSeek-V3 (model name).

[0055] Knowledge point extraction is the process of identifying and extracting knowledge units with independent semantics, such as facts, rules, concepts, clauses, etc. from unstructured text information and converting them into structured data. For example, when performing knowledge point extraction on the collected news text, a single news article is split into multiple independent rumor or science popularization knowledge points. In this way, the minimum storage unit of the knowledge base is upgraded from "news" to "rumor / science popularization knowledge point". On the one hand, it makes knowledge retrieval more accurate and efficient, and on the other hand, it facilitates refined management such as rumor quantity statistics and science popularization knowledge classification.

[0056] It should be noted that in order to improve data timeliness, an hourly data disk storage strategy can be adopted. In order to improve data processing efficiency, a distributed storage method can be used. In order to achieve the generality of the solution, the module for building the knowledge base can be decoupled from the data source, so that the knowledge base building module can adapt to data from various sources.

[0057] Exemplarily, in addition to performing knowledge point extraction, metadata of each knowledge point can also be obtained. For example, the original document to which the knowledge point belongs, the knowledge point serial number corresponding in the original document, attribute tags, release date, knowledge point summary, knowledge source, and uniform resource locator, etc. Among them, the knowledge point serial number corresponding in the original document can also be called the knowledge point index. The attribute tag indicates the attribute of the knowledge point, such as rumor, science popularization, confidential, etc. These metadata can be used to assist knowledge point retrieval. For example, when performing knowledge retrieval based on the information to be detected, when any knowledge point is successfully recalled, the knowledge point summary of this knowledge point can be recalled together for use as a reference knowledge point. Conversely, when any knowledge point summary is successfully recalled, the corresponding knowledge point can be recalled together for use as a reference knowledge point. This combined indexing method of the summary and the original content can reduce the retrieval time consumption and improve the retrieval recall rate.

[0058] Exemplarily, a large language model and a rule engine can also be combined to build a knowledge base. For example, for a science popularization knowledge base and a rumor knowledge base, a large language model is used to extract knowledge points, and the structured knowledge points are stored in the database. For a Baidu Encyclopedia knowledge base and a reliable manuscript source knowledge base, a rule engine is used to segment the encyclopedia content or manuscript content, and the obtained text paragraphs are stored in the database. In this way, on the one hand, it can support complex and diverse data types, and on the other hand, the constructed knowledge base can support rich attribute detection scenarios.

[0059] In the embodiments of the present application, by collecting multi-source text information, the knowledge base covers a wider range of knowledge fields, avoiding the limitations of a single data source. Utilizing the powerful semantic understanding ability of the large language model to extract knowledge points instead of manual or static rules not only improves the efficiency of knowledge base construction but also converts unstructured text into structured knowledge points, making knowledge retrieval more accurate and efficient. Subsequently, retrieving reference knowledge points from this knowledge base as a basis for attribute detection can improve the accuracy of attribute detection.

[0060] Step S20: Based on each reference knowledge point, perform conflict detection on the information to be detected, and obtain the conflict detection results corresponding to each reference knowledge point.

[0061] Conflict detection refers to the process of comparing the information to be detected with each reference knowledge point to identify the conflicting content between the two. Conflict detection includes multiple detection dimensions. For example, factual conflict, that is, the information to be detected is contradictory to the factual data in the reference knowledge point. For example, the information to be detected states that "someone died in 2023", but the reference knowledge point shows that the person is still alive; rule conflict, that is, the information to be detected violates the rules defined in the knowledge point. For example, the outgoing document contains the customer's ID number, violating the rule that "personal sensitive information needs to be desensitized"; logical conflict, that is, the internal logic of the information to be detected is contradictory or does not conform to the logical rules of the knowledge point. For example, the AI answers that "all even numbers are not prime numbers", which is contradictory to the mathematical definition.

[0062] The conflict detection result is the comparison result obtained for each reference knowledge point with the information to be detected, indicating whether the content of each reference knowledge point conflicts with the information to be detected. For example, the conflict detection result can be conflict or support. It should be noted that for the information to be detected for which no relevant reference knowledge point is retrieved or the number of retrieved relevant reference knowledge points is insufficient, the conflict detection result can also be set to unverifiable. In addition, the conflict detection result can also include the specific reason for the conflict. For example, the reason for the conflict is that the expression "vaccine side effects" in the information to be detected contradicts "no serious adverse reactions" in the authoritative report.

[0063] Step S30: Based on the conflict detection results corresponding to each reference knowledge point, generate the attribute detection result of the information to be detected.

[0064] The attribute detection result is the final determination of the attribute of the information to be detected by comprehensively considering all conflict detection results, indicating whether the information to be detected belongs to the target attribute information. In addition, the attribute detection result can also include an explanation of the reason why the information to be detected belongs or does not belong to the target attribute information. For example, in the rumor detection scenario, the attribute detection result is rumor or normal, along with a specific reason explanation; in the hallucination detection scenario, the attribute detection result is hallucination or normal, along with a specific reason explanation; in the classified information detection scenario, the attribute detection result is classified or normal, along with a specific reason explanation.

[0065] Optionally, when the target knowledge point exists in at least one reference knowledge point, it is indicated in the attribute detection result that the information to be detected belongs to the target attribute information, where the target knowledge point is a knowledge point that conflicts with the content of the information to be detected; based on the target knowledge point, a reason explanation for the information to be detected belonging to the target attribute information is generated. When the target knowledge point does not exist in at least one reference knowledge point, it is indicated in the attribute detection result that the information to be detected does not belong to the target attribute information. Of course, a reason explanation for the information to be detected not belonging to the target attribute information can also be generated, or a preset reason explanation for the information to be detected not belonging to the target attribute information can be directly obtained. For example, "The known target attribute standard of the information to be detected does not match at all."

[0066] Among them, generating a reason explanation for the information to be detected belonging to the target attribute information based on the target knowledge point includes: generating a reason explanation for the information to be detected belonging to the target attribute information based on the conflict detection result corresponding to the target knowledge point. For example, the conflict detection result contains the conflict point or conflict reason between the target knowledge point and the information to be detected, so a reason explanation for the information to be detected belonging to the target attribute information can be generated based on the conflict point or conflict reason. For example, the conflict point or conflict reason in the conflict detection result is: the target knowledge point shows that "there is no earthquake record above magnitude 3 in area A in 2024", which has a factual conflict in time and magnitude with the information to be detected "an earthquake of magnitude 5.0 occurred in area A on X month X day in 2024". Accordingly, the reason explanation in the generated target attribute information is: According to the information released by the seismic network, no earthquake of magnitude 5.0 was monitored in area A on X month X day in 2024, so this information is determined to be a rumor. The embodiment of the present application generates a reason explanation based on the target knowledge point that conflicts with the content of the information to be detected, provides a specific explanation for the information to be detected being determined to belong to the target attribute information, increases the transparency of the decision-making process, and makes the result easier to understand and trust.

[0067] Optionally, generating a reason explanation for the information to be detected belonging to the target attribute information based on the target knowledge point includes: obtaining the metadata of the target knowledge point; generating a reason explanation for the information to be detected belonging to the target attribute information based on the target knowledge point and the metadata of the target knowledge point. Among them, the metadata of the target knowledge point includes at least one of the following: the original document to which the target knowledge point belongs, the knowledge point serial number corresponding in the original document, the attribute label, the release date, the knowledge point summary, the knowledge point type, the knowledge source, and the uniform resource locator.

[0068] For example, in the scenario of rumor detection, the information to be detected is "On a certain day in a certain month of 2023, Mall B was sealed up due to fire hazards". The target knowledge point is the "Circular on the Fire Safety Inspection of Mall B" issued by the local fire department. The metadata of the knowledge point includes: source: the official website of the Municipal Fire and Rescue Detachment; release time: a certain day in a certain month of 2023; type: official circular; knowledge point summary: The fire inspection result of Mall B is qualified. Since the information to be detected states that Mall B was sealed up, but the content of the target knowledge point is "The inspection is qualified and it continues to operate", there is a direct factual conflict. Moreover, the metadata shows that the source of the knowledge point is the official website of the fire department, which means high authority, and the release time coincides with the event time, which means high credibility. Therefore, the reason for the generated information to be detected being a rumor is: "According to the authoritative circular issued by the official website of the Municipal Fire and Rescue Detachment on a certain day in a certain month of 2023, the result of Mall B's fire safety inspection is qualified and there is no situation of being sealed up. There is a conflict between the detected information and the factual description in the official circular, and the source of the knowledge point has authority and timeliness. Therefore, this information is determined to be a rumor." In the embodiment of the present application, since the original document to which the knowledge point belongs in the metadata, the knowledge point serial number corresponding in the original document, and the uniform resource locator can accurately trace the specific source of the conflicting content, the knowledge source can prove the reliability of the knowledge point, the release date can illustrate the timeliness of the knowledge point, and the knowledge point summary, attribute label, and knowledge point type can help quickly locate the nature of the knowledge point. Therefore, combining these information can not only quickly determine the information attribute, but also improve the accuracy of the reason explanation in the attribute detection result, and enhance the traceability and transparency of the conclusion.

[0069] Exemplarily, a large language model can be used for attribute detection. For example, the information to be detected and at least one reference knowledge point are input into the large language model, and the large language model is instructed to generate conflict detection results corresponding to each reference knowledge point, and the attribute detection result is generated by synthesizing the conflict detection results corresponding to each reference knowledge point. To improve the accuracy of attribute detection, the metadata of each reference knowledge point can also be input into the large language model to assist the large language model in making attribute judgments and generating the reason explanations for the judgment results.

[0070] Optionally, multiple pieces of information to be detected detected within a historical time period are classified according to the semantic topics they belong to, and multiple topic data sets are obtained; the detection performance parameters corresponding to each topic data set are obtained; in the case where the detection performance parameter corresponding to any semantic topic data set is less than the performance threshold, the knowledge belonging to the semantic topic is updated in the knowledge base, and the knowledge base is used for knowledge retrieval. This knowledge base update mechanism can accurately fill the knowledge gap, improve the accuracy, timeliness, and coverage of the knowledge in the knowledge base, thereby improving the cross-topic detection performance and optimizing the global detection performance.

[0071] Exemplarily, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is used to cluster all the information to be detected during the historical time period, obtaining multiple information clusters with similar semantic features. Each information cluster is used as a topic dataset, and a semantic topic is generated for each topic dataset. It should be noted that the user can set clustering parameters according to needs, thereby determining the semantic similarity degree of the information to be detected in each information cluster.

[0072] Exemplarily, obtaining the detection performance parameters corresponding to each topic dataset includes: statistically calculating the accuracy rate, recall rate, etc. of each topic dataset. The performance threshold is used to determine whether the knowledge under each semantic topic in the knowledge base needs to be updated in a timely manner. For example, the performance threshold is the recall rate threshold. If the recall rate corresponding to any semantic topic dataset is less than the recall rate threshold, the knowledge belonging to this semantic topic is updated in the knowledge base. Among them, knowledge update can include adding knowledge of the corresponding semantic topic in the knowledge base, and replacing the knowledge of the corresponding semantic topic in the knowledge base, etc.

[0073] The information detection solution provided by this application provides a clear basis for detection by retrieving the reference knowledge points associated with the information to be detected, avoiding the drawbacks of simply relying on the subjective judgment of large language models. By performing conflict detection based on each reference knowledge point respectively to obtain the conflict detection results of each reference knowledge point, and then generating the attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point, the complex information detection is decomposed into multiple traceable sub-processes, making the detection logic transparent, and enabling users to clearly understand the source and basis of the judgment. Thus, it breaks the dilemma that the detection result is unexplainable, greatly improving the user's trust in the detection result. In addition, this detection solution with cross-verification of multiple knowledge points also enhances the accuracy and reliability of the detection result, enabling the solution to obtain trustworthy attribute detection results in various detection scenarios.

[0074] Based on the above first embodiment, the second embodiment of this application is proposed. For the same or similar content as the first embodiment, reference can be made to the above introduction and will not be repeated hereinafter. Refer to Figure 3 , in the second embodiment, the above step S30 includes steps S301 to S303: Step S301, generating a comprehensive detection result based on the conflict detection results corresponding to each reference knowledge point.

[0075] Among them, the comprehensive detection result is a comprehensive and systematic detection conclusion formed by integrating the conflict detection results corresponding to all reference knowledge points. For example, if there is a conflict detection result with the result of "conflict" among the multiple conflict detection results obtained from the conflict detection, the comprehensive detection result is "conflict"; if all the multiple conflict detection results obtained from the conflict detection are "supported", the comprehensive detection result is "supported". In addition, the comprehensive detection result can also integrate the conflict points or conflict reasons in multiple conflict detection results.

[0076] Exemplarily, the comprehensive detection result is in the form of a list. It can include the information to be detected, all reference knowledge points, the conflict detection results corresponding to each reference knowledge point, and the final conclusion, that is, "conflict" or "supported". It can also include the final conflict reason description obtained by integrating the conflict reasons in each conflict detection result.

[0077] Step S302, determine the business scenario to which the information detection instruction belongs.

[0078] Exemplarily, the information detection instruction not only carries the information to be detected, but also carries a business scenario identifier. Correspondingly, the business scenario identifier can be extracted from the information detection instruction, so as to determine the business scenario to which the information detection instruction belongs according to the business scenario identifier.

[0079] Step S303, convert the content of the comprehensive detection result based on the conversion rule corresponding to the business scenario to obtain an attribute detection result.

[0080] Exemplarily, the comprehensive detection result includes two result labels: "conflict" or "supported". If the result label of the comprehensive detection result is "conflict", the result label in the attribute detection result is "belongs to the target attribute"; if the result label of the comprehensive detection result is "supported", the result label in the attribute detection result is "does not belong to the target attribute" or "normal".

[0081] Exemplarily, in the case where the business scenario is a rumor detection scenario, the content of the comprehensive detection result is converted based on the conversion rule corresponding to the rumor detection scenario, and the obtained rumor detection result indicates whether the information to be detected belongs to a rumor, and includes a reason description of whether the information to be detected belongs to a rumor. Exemplarily, the two result labels of "conflict" and "supported" in the comprehensive detection result respectively correspond to "rumor" and "normal" in the rumor detection result.

[0082] In the case where the business scenario is a classified information detection scenario, the content of the comprehensive detection result is converted based on the conversion rules corresponding to the classified information detection scenario, and the obtained classified information detection result indicates whether the information to be detected is classified and includes an explanation of the reason why the information to be detected is classified or not. Exemplarily, the two result tags of "conflict" and "support" in the comprehensive detection result respectively correspond to "classified" and "normal" in the classified information detection result.

[0083] In the case where the business scenario is a hallucination detection scenario, the content of the comprehensive detection result is converted based on the conversion rules corresponding to the hallucination detection scenario, and the obtained hallucination detection result indicates whether the information to be detected belongs to a hallucination and includes an explanation of the reason why the information to be detected belongs to a hallucination or not. Exemplarily, the two result tags of "conflict" and "support" in the comprehensive detection result respectively correspond to "hallucination" and "normal" in the hallucination detection result.

[0084] It should be noted that if the reference knowledge points of the information to be detected are not retrieved, or the number of reference knowledge points is insufficient, such as lower than the quantity threshold, or the content relevance with the information to be detected is relatively low, such as lower than the relevance threshold, the result tag of the comprehensive detection result can be "unverifiable". In this case, knowledge retrieval can be performed again until a sufficient number of reference knowledge points with relatively high content relevance are obtained. And conflict detection is performed again until a comprehensive detection result with the content of "conflict" or "support" is obtained for subsequent conversion into an attribute detection result. Or, in the case where the comprehensive detection result is "unverifiable", the corresponding attribute detection result can also be defined as "normal". This means that in the case where it is impossible to verify whether the content of the information to be detected conflicts with the content of the existing factual information or rules, the information to be detected is considered normal information.

[0085] In the embodiments of the present application, regardless of which business scenario the information to be detected belongs to, the detection process is to obtain relevant reference knowledge points, compare the content of each reference knowledge point with the information to be detected to obtain the conflict detection result of each reference knowledge point and the information to be detected. Then, based on the conflict detection results of each reference knowledge point and the information to be detected, a comprehensive detection result is generated. Furthermore, according to the business scenario, the comprehensive detection result is converted into an attribute detection result adapted to the business scenario. This scenario-based result conversion mechanism decouples the general detection ability from the business scenario and, based on the conversion rules of each business scenario, realizes "one system, multi-scenario customized output", solves the adaptation problem of the information detection system in multiple business scenarios, makes the detection result more in line with the actual application requirements, and is applicable to the information detection system that needs to support diverse services. In addition, when a new business scenario needs to be added, only the conversion rules corresponding to the business scenario need to be added, without changing the core detection module of the information detection system, reducing the deployment and R & D costs of the information detection system that supports cross-domain and cross-business.

[0086] Based on the above first embodiment, the third embodiment of the present application is proposed. For the same or similar content as the first embodiment, reference can be made to the above introduction and will not be repeated hereinafter. Referring to Figure 4 , in the third embodiment, the above step S30 includes steps S304 to S305: Step S304, obtain a set of constraint rules, which is used to restrict the content and format of the attribute detection results.

[0087] The set of constraint rules may include content constraint rules, format constraint rules, and logical constraint rules, which are used to define the content scope, format specification, and logical relationship of the attribute detection results.

[0088] For example, the content constraint rules specify the information items that must be included or prohibited in the attribute detection results. The format constraint rules stipulate the presentation form of the attribute detection results, such as JSON format (JavaScript Object Notation, a lightweight data exchange format), specific field naming rules, numerical precision requirements, etc. The logical constraint rules define the logical relationship within the attribute detection results. For example, the comprehensive detection results must be generated based on the conflict detection results corresponding to each reference knowledge point, and then the comprehensive detection results are converted into attribute detection results according to the business scenario.

[0089] Step S305, through the first large language model, generate attribute detection results that meet the requirements of the set of constraint rules based on the conflict detection results corresponding to each reference knowledge point.

[0090] The first large language model is used to structurally output content that meets the requirements according to the set of constraint rules, such as automatically generating attribute detection results in JSON format.

[0091] Optionally, generating attribute detection results that meet the requirements of the set of constraint rules through the first large language model based on the conflict detection results corresponding to each reference knowledge point includes: generating multiple candidate words through the first large language model based on the context sequence formed by the conflict detection results corresponding to each reference knowledge point; filtering out the candidate words that do not meet the requirements of the set of constraint rules from the multiple candidate words, and selecting candidate words from the remaining candidate words and adding them to the context sequence to obtain an updated context sequence; regenerating the next candidate word based on the updated context sequence until the attribute detection results of the information to be detected are obtained.

[0092] Exemplarily, through the first large language model, multiple candidate words and the probabilities of each candidate word are generated based on the context sequence composed of the conflict detection results corresponding to each reference knowledge point. Check whether each candidate word meets all the constraint rules in the constraint rule set. For any candidate word, if it meets all the constraint rules, set its corresponding mask to 1; otherwise, set its corresponding mask to 0. Then, multiply the probability of each candidate word by its mask to obtain the corrected probability of each candidate word. Select the target candidate word based on the corrected probabilities of each candidate word, and add the target candidate word to the context sequence to obtain the updated context sequence. Repeat this process until the termination condition is met, and use the currently obtained context sequence as the attribute detection result. Exemplarily, the termination condition is that the length of the generated sequence reaches the preset maximum length, or the first large language model outputs a termination symbol, or the constraint rules force termination, for example, when no candidate word that meets the constraint rule set can be generated.

[0093] It should be noted that the generation method of the conflict detection results corresponding to each reference knowledge point is the same as that of the attribute detection result. That is, based on each reference knowledge point, conflict detection is performed on the information to be detected to obtain the conflict detection results corresponding to each reference knowledge point, including: through the first large language model, based on each reference knowledge point, conflict detection is performed on the information to be detected to generate conflict detection results that meet the requirements of the constraint rule set. Specifically, the constraint rule set is also used to restrict the content and format of the conflict detection results. For example, the content constraint rules in the constraint rule set specify the information items that the conflict detection results must contain or prohibit. The format constraint rules stipulate the presentation form of the conflict detection results, such as JSON format, specific field naming rules, numerical precision requirements, etc. The logical constraint rules define the internal logical relationships of the conflict detection results. For example, in the case where the conflict detection result corresponding to any reference knowledge point is "conflict", the reason for the conflict must be generated.

[0094] Exemplarily, through the first large language model, based on each reference knowledge point, conflict detection is performed on the information to be detected to generate conflict detection results that meet the requirements of the constraint rule set, including: through the first large language model, multiple candidate words are generated based on the context sequence composed of the information to be detected and each reference knowledge point; the candidate words that do not meet the requirements of the constraint rule set among the multiple candidate words are filtered out, and candidate words are selected from the remaining candidate words and added to the context sequence to obtain the updated context sequence; based on the updated context sequence, the next candidate word is regenerated until the conflict detection results corresponding to each reference knowledge point are obtained.

[0095] It should be noted that the above process of generating the conflict detection result and the process of generating the attribute detection result can be continuously executed by the first large language model. That is to say, the generation of the conflict detection result and the attribute detection result can be completed by the first large language model in a single continuous inference, rather than being independently executed in stages. Specifically, the context sequence composed of the information to be detected and each reference knowledge point is used as the original context sequence, and the first large language model continuously iteratively generates the next candidate word based on the constraint rule set and the original context sequence, thereby continuously growing the context sequence until the final attribute detection result is generated. During the iteration process, a conflict detection result will be generated and used to generate the final attribute detection result.

[0096] In the embodiment of the present application, the content and format of the attribute detection result generated by the large language model are restricted through a constraint rule set to obtain a structured attribute detection result. On the one hand, it can improve the standardization and compliance of the attribute detection result. On the other hand, it enables the detection result to present key information in a clearer and more intuitive manner, greatly reducing the understanding cost, facilitating technicians to quickly locate data anomalies, and also enabling business personnel and management personnel to quickly obtain core conclusions, thereby realizing the efficient flow and communication of the detection result among different levels and functions and improving the cross-departmental collaboration efficiency.

[0097] Based on the above first embodiment, the fourth embodiment of the present application is proposed. For the same or similar content as the first embodiment, reference can be made to the above introduction and will not be repeated hereinafter. Referring to Figure 5 , in the fourth embodiment, the above step S10 includes steps S101 to S102: Step S101, in response to the information detection instruction, perform attribute detection on the information to be detected corresponding to the information detection instruction through the second large language model to obtain an initial detection result, where the initial detection result indicates whether the information to be detected belongs to the target attribute information.

[0098] The second large language model is a large language model specifically used to perform attribute pre-detection in the information detection process, with a different division of labor from the first large language model mentioned above. It is used to quickly determine whether the information to be detected belongs to the target attribute information, such as whether it belongs to a rumor, whether it is confidential, and whether it belongs to an illusion. The initial detection result output by the second large language model provides a decision basis for whether to start knowledge retrieval subsequently.

[0099] Exemplarily, the second large language model is a model obtained by fine-tuning a general large language model. The difference from the general large language model lies in its output result tags, such as rumor or normal. Compared with the general large language model that needs to output a relatively long text content, the second large language model compresses the length of the model output content. Therefore, the speed of attribute detection is increased by 20 times compared with the original general large language model. In addition, because it has been fine-tuned through a selected dataset, the accuracy of attribute detection of the second large language model has also been greatly improved compared with the original general large language model.

[0100] The initial detection result is the conclusion output after the second large language model preliminarily analyzes the information to be detected, indicating whether the information to be detected belongs to the target attribute information. Exemplarily, in the rumor detection scenario, the initial detection result is a binary classification label, such as "rumor" or "normal".

[0101] Step S102, in the case where the initial detection result indicates that the information to be detected belongs to the target attribute information, perform knowledge retrieval based on the information to be detected to obtain at least one reference knowledge point associated with the information to be detected.

[0102] Exemplarily, in the case where the initial detection result indicates that the information to be detected does not belong to the target attribute information, directly use the initial detection result as the final attribute detection result.

[0103] Optionally, in response to the information detection instruction, before obtaining the initial detection result by performing attribute detection on the information to be detected through the second large language model, the method further includes: obtaining a sample dataset, where the sample dataset includes sample information and sample labels, and the sample labels indicate whether the sample information belongs to the target attribute information; based on the sample dataset, fine-tuning the third large language model so that the initial detection result generated by the fine-tuned second large language model is a binary classification label, and the binary classification label includes a first label or a second label, the first label indicates that the information to be detected belongs to the target attribute information, and the second label indicates that the information to be detected does not belong to the target attribute information.

[0104] The third large language model is a general large language model that has not been optimized for the current attribute detection task. For example, Qwen / Qwen3-0.6B, which is used as the initial model for subsequent fine-tuning and is the basic architecture of the second large language model.

[0105] Exemplarily, in the rumor detection scenario, the sample dataset may include a rumor dataset and a normal dataset. For example, the rumor dataset may contain rumor datasets from multiple sources, such as a Weibo rumor dataset, a rumor article dataset, etc. The normal dataset may contain real and objective datasets such as authoritative news and government announcements.

[0106] Fine-tuning is a process of further training on the basis of the third largest pre-trained language model using a sample dataset to adapt it to the label generation task, that is, to generate binary classification labels for attribute detection. Exemplarily, the LoRA (Low-Rank Adaptation) method can be used for fine-tuning to improve the training efficiency and reduce the training cost.

[0107] Exemplarily, the attribute detection result of the second largest language model can be reflected by the following formula (1): Formula (1):

[0108] Where token is the attribute detection result output by the second largest language model, 1 represents that the information to be detected belongs to the target attribute information, 0 represents that the information to be detected does not belong to the target attribute information, query is the information to be detected input into the second largest language model, and LLM is the second largest language model.

[0109] In the solution provided in the embodiments of the present application, since the second largest language model does not rely on knowledge base retrieval and reduces external interface calls, the attribute detection speed is extremely fast, and the initial detection result can be quickly output within a millisecond response time. Then, the solution only performs knowledge retrieval when the initial detection result indicates that the information to be detected belongs to the target attribute information, avoiding full-scale retrieval of all information to be detected. On the one hand, it can reduce the load on the knowledge base, and on the other hand, it can improve the overall detection efficiency. This hierarchical strategy of "pre-screening - fine detection" can significantly improve the efficiency on the premise of ensuring the detection quality, so as to support large data volume and high-concurrency data detection scenarios.

[0110] Figure 6 It is a schematic diagram of an information detection system provided in the embodiments of the present application. Refer to Figure 6, the information detection system has a three-layer architecture, including a business application layer, a core detection layer, and a knowledge base. Among them, the information to be detected for each business interface (including rumor detection business, classified information detection business, hallucination detection business, etc.) is uniformly sent to the public detection entrance of the business application layer. The public detection entrance forwards the information to be detected to the unified detection entrance of the core detection layer. Then, the retrieval module of the core detection layer retrieves reference knowledge points from each knowledge base based on the information to be detected. In the case of retrieving reference knowledge points, the conflict detection module performs conflict detection based on the information to be detected and each reference knowledge point to obtain a conflict detection result (including three cases: conflict, support, and unable to verify); in the case of not retrieving reference knowledge points, an unable-to-verify conflict detection result is directly obtained. Then, the conflict detection result is sent to the public detection entrance of the business application layer. The public detection entrance converts the conflict detection result into an attribute detection result adapted to the business scenario based on the business scenario to which the information to be detected belongs, and sends the attribute detection result to the corresponding business interface. It should be noted that the above conflict detection result refers to the comprehensive detection result obtained by comprehensively considering the conflict detection results of each reference knowledge point. In addition, during the process of the information retrieval system performing the information detection task, each knowledge base (including rumor knowledge base, popular science knowledge base, reliable source knowledge base, encyclopedia knowledge base, etc.) is continuously updated with the knowledge crawled from the network to ensure the timeliness of the knowledge.

[0111] The following Table 1, Table 2, and Table 3 record the performance of the information detection system in performing the rumor detection task. Table 1 contains the test results of 6 datasets, where the query dataset is a dataset constructed from the queries (information to be detected) actually received by the system. The category refers to the categories covered by the sample data in the dataset. For example, the article dataset contains both normal data samples and rumor data samples. Precision, recall, and f1-score are three evaluation metrics of the information detection system.

[0112] Table 1

[0113] The following Table 2 records the performance of the information detection system for the information to be detected with 6 different semantic themes.

[0114] Table 2

[0115] Table III below records the performance differences between the third largest language model adopted by the information detection system and the second largest language model after fine-tuning when performing the rumor detection task. Among them, Qwen / Qwen3-0.6B is the third largest language model, and qwen3-yaoyan:0.6B is the second largest language model obtained after fine-tuning. It can be seen that the rumor detection performance of the second largest language model after fine-tuning is significantly better than that of the third largest language model.

[0116] Table III

[0117] Another point that needs to be explained is that the above examples are only for understanding this application and do not constitute a limitation on the information detection method of this application. Based on this technical concept, more forms of simple transformations are within the protection scope of this application.

[0118] This application also provides an information detection device. Please refer to Figure 7 , the information detection device includes: A knowledge retrieval module 10, configured to perform knowledge retrieval based on the information to be detected corresponding to the information detection instruction in response to the information detection instruction, and obtain at least one reference knowledge point associated with the information to be detected; A conflict detection module 20, configured to perform conflict detection on the information to be detected based on each reference knowledge point, and obtain a conflict detection result corresponding to each reference knowledge point; An attribute detection module 30, configured to generate an attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point.

[0119] Optionally, the attribute detection module 30 is configured to generate a comprehensive detection result based on the conflict detection results corresponding to each reference knowledge point; determine the business scenario to which the information detection instruction belongs; and convert the content of the comprehensive detection result based on the conversion rule corresponding to the business scenario to obtain the attribute detection result.

[0120] Optionally, the conflict detection result indicates whether the content of each reference knowledge point conflicts with the information to be detected; the attribute detection result indicates whether the information to be detected belongs to the target attribute information, and includes an explanation of the reason why the information to be detected belongs to the target attribute information.

[0121] Optionally, the attribute detection module 30 includes: An attribute detection unit, configured to indicate in the attribute detection result that the information to be detected belongs to the target attribute information when there is a target knowledge point in at least one reference knowledge point, where the target knowledge point is a knowledge point that conflicts with the content of the information to be detected; A reason generation unit, configured to generate an explanation of the reason why the information to be detected belongs to the target attribute information based on the target knowledge point.

[0122] Optionally, a reason generation unit is configured to obtain the metadata of the target knowledge point; based on the target knowledge point and the metadata of the target knowledge point, generate an explanation for the reason why the information to be detected belongs to the target attribute information.

[0123] Optionally, the metadata of the target knowledge point includes at least one of the following: The original document to which the target knowledge point belongs, the knowledge point serial number corresponding to the target knowledge point in the original document, the attribute label, the release date, the knowledge point summary, the knowledge source, and the uniform resource locator.

[0124] Optionally, the attribute detection module 30 includes: A rule acquisition unit is configured to obtain a set of constraint rules, and the set of constraint rules is used to restrict the content and format of the attribute detection result; A result generation unit is configured to generate an attribute detection result that meets the requirements of the set of constraint rules based on the conflict detection results corresponding to each reference knowledge point through the first large language model.

[0125] Optionally, the result generation unit is configured to generate multiple candidate words through the first large language model based on the context sequence formed by the conflict detection results corresponding to each reference knowledge point; filter out the candidate words that do not meet the requirements of the set of constraint rules from the multiple candidate words, and select candidate words from the remaining candidate words and add them to the context sequence to obtain an updated context sequence; regenerate the next candidate word based on the updated context sequence until the attribute detection result of the information to be detected is obtained.

[0126] Optionally, the knowledge retrieval module 10 is configured to, in response to an information detection instruction, perform attribute detection on the information to be detected through the second large language model to obtain an initial detection result, where the initial detection result indicates whether the information to be detected belongs to the target attribute information; in the case where the initial detection result indicates that the information to be detected belongs to the target attribute information, perform knowledge retrieval based on the information to be detected to obtain at least one reference knowledge point associated with the information to be detected.

[0127] Optionally, the apparatus further includes: A model training module is configured to obtain a sample data set, where the sample data set includes sample information and sample labels, and the sample labels indicate whether the sample information belongs to the target attribute information; based on the sample data set, fine-tune the third large language model so that the initial detection result generated by the second large language model obtained by fine-tuning is a binary classification label, and the binary classification label includes a first label or a second label, where the first label indicates that the information to be detected belongs to the target attribute information, and the second label indicates that the information to be detected does not belong to the target attribute information.

[0128] Optionally, the apparatus further includes: A knowledge storage module for obtaining text information from multiple sources; extracting knowledge points from the text information through a first large language model to obtain multiple knowledge points; and storing the multiple knowledge points in a knowledge base for knowledge retrieval.

[0129] Optionally, the apparatus further includes: A knowledge update module for classifying multiple pieces of information to be detected detected within a historical time period according to their semantic topics to obtain multiple topic datasets; obtaining detection performance parameters corresponding to each topic dataset; and updating the knowledge belonging to the semantic topic in the knowledge base when the detection performance parameter corresponding to any semantic topic dataset is less than a performance threshold, where the knowledge base is for knowledge retrieval.

[0130] The information detection apparatus provided in this application adopts the information detection method in the above embodiment, and can solve the technical problems that the information detection results in the related art cannot be verified for accuracy and are unexplainable. Compared with the prior art, the beneficial effects of the information detection apparatus provided in this application are the same as those of the information detection method provided in the above embodiment, and other technical features in the information detection apparatus are the same as those disclosed in the method of the above embodiment, which will not be elaborated here.

[0131] This application provides an information detection device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the information detection method in the above embodiment.

[0132] Next, refer to Figure 8 , which shows a schematic structural diagram of an information detection device suitable for implementing the embodiments of this application. The information detection device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistant), PADs (Portable Application Description: tablet computers), PMPs (Portable Media Player: portable multimedia players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 8 The information detection device shown is only an example and should not impose any limitation on the functions and usage scope of the embodiments of this application.

[0133] As Figure 8As shown, the information detection device may include a processing device 1001 (such as a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM: Random Access Memory) 1004. In the RAM 1004, various programs and data required for the operation of the information detection device are also stored. The processing device 1001, the ROM 1002, and the RAM 1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the information detection device to communicate with other devices wirelessly or wiredly to exchange data. Although the figure shows an information detection device having various systems, it should be understood that it is not required to implement or have all the shown systems. Instead, more or fewer systems can be implemented or had.

[0134] In particular, according to the embodiments disclosed in the present application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program codes for performing the methods shown in the flowcharts. In such an embodiment, the computer program can be downloaded and installed from a network through the communication device, or installed from the storage device 1003, or installed from the ROM 1002. When the computer program is executed by the processing device 1001, the above functions defined in the methods of the embodiments disclosed in the present application are executed.

[0135] The information detection device provided by the present application adopts the information detection method in the above embodiments, and can solve the technical problems that the information detection results in the related art cannot be accurately verified and are not interpretable. Compared with the prior art, the beneficial effects of the information detection device provided by the present application are the same as those of the information detection method provided by the above embodiments, and other technical features in the information detection device are the same as those disclosed in the method of the previous embodiment, and will not be elaborated here.

[0136] It should be understood that each part disclosed in this application can be implemented by hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in a suitable manner in any one or more embodiments or examples.

[0137] As described above, the above are only specific embodiments of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in this application, and all should be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

[0138] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., computer programs) stored thereon, and the computer-readable program instructions are used to execute the information detection method in the above embodiments.

[0139] The computer-readable storage medium provided by this application can be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM: Random Access Memory), read-only memory (ROM: Read Only Memory), erasable programmable read-only memory (EPROM: Erasable Programmable Read Only Memory or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM: CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the above. In this embodiment, the computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, device, or device. The program code contained on the computer-readable storage medium can be transmitted by any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0140] The above computer-readable storage medium can be included in the information detection device; or it can exist separately without being assembled into the information detection device.

[0141] The above computer-readable storage medium carries one or more programs, which, when executed by an information detection device, cause the information detection device to: in response to an information detection instruction, perform knowledge retrieval based on the information to be detected corresponding to the information detection instruction, and obtain at least one reference knowledge point associated with the information to be detected; perform conflict detection on the information to be detected based on each reference knowledge point respectively, and obtain a conflict detection result corresponding to each reference knowledge point; and generate an attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point.

[0142] Computer program code for performing the operations of the present application may be written in one or more programming languages or combinations thereof. The programming languages include 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 a 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).

[0143] 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 application. 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 the 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.

[0144] The modules described in the embodiments of the present application may be implemented in software or in hardware. Wherein, the name of the module does not constitute a limitation to the unit itself in some cases.

[0145] The readable storage medium provided by this application is a computer-readable storage medium. The computer-readable storage medium stores computer-readable program instructions (i.e., computer programs) for executing the above information detection method, which can solve the technical problems that the information detection results in the related art cannot be accurately verified and are unexplainable. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by this application are the same as those of the information detection method provided by the above embodiments, and will not be elaborated here.

[0146] This application also provides a computer program product, including a computer program, and the steps of the above information detection method are implemented when the computer program is executed by a processor.

[0147] The computer program product provided by this application can solve the technical problems that the information detection results in the related art cannot be accurately verified and are unexplainable. Compared with the prior art, the beneficial effects of the computer program product provided by this application are the same as those of the information detection method provided by the above embodiments, and will not be elaborated here.

[0148] The above are only partial embodiments of this application, and do not limit the patent scope of this application accordingly. Any equivalent structural transformation made by using the content of the specification and drawings of this application under the technical concept of this application, or any direct / indirect application in other related technical fields, is included in the patent protection scope of this application.

Claims

1. An information detection method, characterized in that, The method includes: In response to an information detection instruction, perform knowledge retrieval based on the information to be detected corresponding to the information detection instruction, and obtain at least one reference knowledge point associated with the information to be detected; Based on each reference knowledge point, perform conflict detection on the information to be detected to obtain a conflict detection result corresponding to each reference knowledge point; Based on the conflict detection results corresponding to each reference knowledge point, generate an attribute detection result of the information to be detected.

2. The method according to claim 1, wherein The generating the attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point includes: Based on the conflict detection results corresponding to each reference knowledge point, generate a comprehensive detection result; Determine the business scenario to which the information detection instruction belongs; Based on the conversion rule corresponding to the business scenario, convert the content of the comprehensive detection result to obtain the attribute detection result.

3. The method according to claim 1, wherein The conflict detection result indicates whether the content of each reference knowledge point conflicts with the information to be detected; the attribute detection result indicates whether the information to be detected belongs to the target attribute information, and includes an explanation of the reason why the information to be detected belongs to the target attribute information.

4. The method according to claim 3, wherein The generating the attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point includes: In the case where there is a target knowledge point among the at least one reference knowledge point, indicate in the attribute detection result that the information to be detected belongs to the target attribute information, where the target knowledge point is a knowledge point that conflicts with the content of the information to be detected; Based on the target knowledge point, generate an explanation of the reason why the information to be detected belongs to the target attribute information.

5. The method according to claim 4, wherein The generating the explanation of the reason why the information to be detected belongs to the target attribute information based on the target knowledge point includes: Obtain the metadata of the target knowledge point; Based on the target knowledge point and the metadata of the target knowledge point, generate an explanation of the reason why the information to be detected belongs to the target attribute information.

6. The method according to claim 1, wherein The generating the attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point includes: Obtain a set of constraint rules, where the set of constraint rules is used to restrict the content and format of the attribute detection result; Through a first large language model, based on the conflict detection results corresponding to each reference knowledge point, generate an attribute detection result that meets the requirements of the set of constraint rules.

7. An information detection device, characterized in that The device includes: A knowledge retrieval module, configured to, in response to an information detection instruction, perform knowledge retrieval based on the information to be detected corresponding to the information detection instruction, and obtain at least one reference knowledge point associated with the information to be detected; A conflict detection module, configured to perform conflict detection on the information to be detected based on each reference knowledge point to obtain a conflict detection result corresponding to each reference knowledge point; An attribute detection module, configured to generate an attribute detection result of the information to be detected based on the conflict detection results corresponding to each reference knowledge point.

8. An information detection device, characterized in that, The device includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, where the computer program is configured to implement the steps of the information detection method according to any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the information detection method according to any one of claims 1 to 6 are implemented.

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

Images