An intelligent robot based on data processing and having autonomous learning function and a method of using the same

Abstract

The present invention belongs to the field of intelligent robots, and provides an intelligent robot based on data processing and having an autonomous learning function and a method of using the intelligent robot, wherein the intelligent robot is used for legal consultation and includes a legal knowledge base; the intelligent robot receives a first question input by a first user, and obtains a first answer to the first question according to a first model; if the first answer cannot be obtained or the confidence of the first answer is less than a first threshold, at least one second answer output by at least one large model is obtained; the first legal knowledge in the at least one second answer is extracted, the first legal knowledge is added to the legal knowledge base, and marked as unverified knowledge; the background information of the second user and the second question input by the second user are obtained, and the newly added knowledge is verified based on the feedback of the second user, and the first legal knowledge is marked as verified if the verification is passed, and the first model is updated according to the legal knowledge base. The above scheme can increase the ability of autonomous learning for the intelligent robot.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent robots, and in particular relates to an intelligent robot based on data processing and having an autonomous learning function and a use method thereof. Background Art

[0002] Traditional legal consultation requires users to go to a law firm in person or make an appointment with a lawyer by phone, which is time-consuming and laborious. Especially in remote areas, it is difficult for users to obtain timely legal help. The complexity and professionalism of legal issues have further increased the difficulty for ordinary users to obtain legal support. Faced with high consultation fees and the complexity of obtaining information, many users often feel at a loss when facing legal issues.

[0003] At present, some intelligent legal consulting robot products and large models are available for legal consultation. These systems use natural language processing technology to initially understand and answer legal questions raised by users. However, the existing large pre-trained language models, despite their powerful natural language processing capabilities, are unable to provide highly professional and accurate legal answers due to the lack of specialized training in the legal field. These models perform well in handling general problems, but are often unable to cope with complex and professional legal consultations.

[0004] In addition, some intelligent legal consulting systems have deficiencies in autonomous learning and optimization. Information such as laws, regulations, and precedents is a process of constant change and dynamic revision. Existing intelligent legal consulting systems usually require manual maintenance and updating of the knowledge base to ensure the timeliness and accuracy of their legal information. This reliance on manual updates not only increases the cost and complexity of system maintenance, but also limits the system's autonomous learning ability and degree of automation. Summary of the invention

[0005] In order to solve the problems in the prior art, the present invention provides an intelligent robot based on data processing and having an autonomous learning function, the intelligent robot being used for legal consultation, the intelligent robot comprising a legal knowledge base; the intelligent robot comprising a first model, the first model being trained according to the legal knowledge base; the intelligent robot receiving a first question input by a first user, and obtaining a first answer to the first question according to the first model; if the first answer cannot be obtained or the confidence level of the first answer is less than a first threshold, the first question is organized into a second question including at least the background of the first question according to the context of the conversation with the first user, and the second question is input into at least one large model to obtain at least one second answer output by the at least one large model; the at least one second answer is extracted from the at least one large model; The method comprises the following steps: determining the first legal knowledge in a second answer, adding the first legal knowledge to the legal knowledge base, and marking it as unverified knowledge; obtaining the background information of the second user and the second question input by the second user, and if the correlation between the background information of the second user and the first legal knowledge is greater than a second threshold, and the correlation between the first legal knowledge and the second question is not less than a third threshold, presenting the third answer including the first legal knowledge to the user, and if the second user does not question the third answer, improving the confidence of the first legal knowledge according to the background information of the second user, and if the confidence of the first legal knowledge is greater than a fourth threshold, marking the first legal knowledge as verified and solidifying it to the legal knowledge base; and updating the first model according to the legal knowledge base.

[0006] Furthermore, determining the first answer confidence of the first answer to the first question includes: pre-establishing multiple confidence assessment models, wherein the input of the confidence assessment model includes the first question input by the first user and the first answer generated by the first model combined with relevant knowledge in the legal knowledge base, and the output of the confidence assessment model is the single-point prediction confidence of the first answer; and calculating the confidence of the first answer based on the single-point prediction confidence of multiple confidence assessment models.

[0007] Furthermore, extracting the first legal knowledge from the at least one second answer includes: calculating a matching value of the second answer; and selecting a second answer whose matching value is greater than a preset matching value threshold as the selected second answer.

[0008] Furthermore, the calculation of the matching value of the second answer includes: determining the single-point prediction confidence of the second answer through multiple confidence assessment models; calculating the second answer confidence based on the single-point prediction confidence of the second answer output by multiple confidence assessment models; and evaluating the relevance of the second answer to the contextual information related to the current question based on multiple relevance assessment models, wherein the input of the relevance assessment model includes the extracted contextual information related to the current question and one or more second answers generated by the large model based on the reorganized second question, and the output of the relevance assessment model includes the relevance of the second answer to the contextual information related to the current question.

[0009] Furthermore, when determining the relevance of the background information of the second user to the first legal knowledge, keywords are extracted from the user background information, the keywords are converted into vector representations, feature vectors are generated, and the relevance of the background information of the second user to the first legal knowledge is calculated using the feature vectors.

[0010] Furthermore, the use of the feature vector to calculate the relevance of the second user's background information to the first legal knowledge includes: calculating the vector similarity between the feature vector of the second user's background information and the feature vector of the user background information of the sample user; taking the sample user whose vector similarity is greater than a preset vector similarity threshold as the target sample user; calculating the relevance of the second user's background information to the first legal knowledge based on the first legal knowledge relevance between multiple first legal knowledge and multiple sample users and the vector similarity between the feature vector of the second user's background information and the feature vector of the user background information of the sample user.

[0011] Furthermore, when the user is an ordinary user, verified legal knowledge is called first.

[0012] Furthermore, when the knowledge is questioned by practitioners with a legal background, the confidence level of the relevant knowledge is lowered.

[0013] The present invention also discloses a method for using an intelligent robot based on data processing and having an autonomous learning function, and uses the intelligent robot disclosed in the present invention to provide legal consultation.

[0014] The present invention can produce the following beneficial effects through the above technical solution:

[0015] The intelligent robot of the present invention has the function of self-learning, and can continuously optimize itself according to user feedback and new legal knowledge, and improve its answering ability and accuracy. Combining the advantages of the first model and the big model, the big model is used to generate more comprehensive answers to complex questions, ensuring the accuracy and professionalism of the answers.

[0016] The system can adjust the confidence level of legal knowledge and optimize the knowledge base content based on user feedback, especially feedback from professional users. By presenting the legal knowledge to be verified to professional users (such as lawyers), the system can verify the correctness of the knowledge faster and more accurately. The system can dynamically adjust the confidence level of legal knowledge based on user qualifications and feedback to ensure the authority and reliability of the knowledge base content.

[0017] High-confidence legal knowledge is marked as verified and solidified in the knowledge base to ensure that the information in the knowledge base is authoritative and reliable. When answering questions from ordinary users, the system gives priority to using verified legal knowledge to ensure the reliability and accuracy of the answers.

[0018] By combining legal knowledge base, machine learning models, big models and user feedback mechanisms, we provide highly intelligent and personalized legal consulting services. The system can continuously learn and optimize to ensure the accuracy and professionalism of answers, which can significantly improve user experience and system reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0021] Below, the invention is preferably described in conjunction with the accompanying drawings and specific implementation methods.

[0022] This embodiment solves the above problem through the following steps:

[0023] In one embodiment, reference Figure 1 The present invention provides an intelligent robot based on data processing and having autonomous learning function.

[0024] An intelligent robot is a machine device with artificial intelligence technology that can simulate certain cognitive functions of humans, such as understanding natural language, answering questions, learning new knowledge, etc. In this solution, the intelligent robot is mainly used for legal consultation.

[0025] Legal consultation refers to the process of providing legal knowledge and advice to users. In this solution, the intelligent robot analyzes the legal questions raised by users and provides corresponding legal answers.

[0026] The intelligent robot in this embodiment is used for legal consultation. The intelligent robot receives legal questions input by the user and provides accurate legal answers in combination with the built-in legal knowledge base and machine learning model. If the existing model cannot provide a high-confidence answer, the system will use a large pre-trained language model to further process the problem, extract new knowledge from it, and continuously optimize itself. The implementation of the intelligent robot will be described in detail below.

[0027] The intelligent robot includes a legal knowledge base.

[0028] The legal knowledge base is one of the core components of the intelligent robot system, carrying the functions of storing, organizing and managing legal information. It not only provides the basic data for legal consultation for the intelligent robot, but also continuously improves the depth and breadth of the system's knowledge through continuous updating and expansion. The construction and maintenance of the legal knowledge base is the key to ensuring that the intelligent robot can provide accurate, timely and professional legal consultation services.

[0029] In this embodiment, the initial sources of the legal knowledge base include:

[0030] Legal provisions: including various laws, regulations, judicial interpretations and other official documents. These provisions are the basis of the legal knowledge base and cover a wide range of legal fields, such as civil law, criminal law, commercial law, labor law, etc.

[0031] Case Library: contains a large number of legal cases such as court judgments and arbitration awards. These cases provide important references for the practical application of legal provisions.

[0032] Academic literature: including legal academic papers, reviews, monographs, etc. These documents provide in-depth theoretical support and cutting-edge research results for the system.

[0033] Regulatory interpretation: Interpretations and explanations of legal provisions issued by relevant legal bodies help clarify the meaning and scope of application of legal provisions.

[0034] The intelligent robot includes a first model, which is trained according to a legal knowledge base; the intelligent robot receives a first question input by a first user, and obtains a first answer to the first question according to the first model.

[0035] In this solution, the first model is a machine learning model trained based on the legal knowledge base, which is mainly used to understand and answer legal questions raised by users. It is the core computing and reasoning component in the intelligent robot system. After receiving the first question input by the first user, the intelligent robot generates a first answer to the question by calling the first model and combining the relevant knowledge in the legal knowledge base.

[0036] According to the complexity and diversity of legal issues, the first model can select a suitable machine learning model type, such as support vector machine (SVM), decision tree, random forest, deep neural network, etc. The specific model selection is not the focus of the present invention. Those skilled in the art can comprehensively consider factors such as model performance, the scale of training data and computing resources to select a suitable model.

[0037] During the training process, training data is extracted from the legal knowledge base, including legal provisions, cases, regulatory interpretations, etc. These data are preprocessed, including word segmentation, stop word removal, stemming, etc., to improve the efficiency and effectiveness of model training.

[0038] Extract features from the preprocessed data, such as keywords, word frequency, legal terms, etc. These features will serve as input for model training.

[0039] The model is trained using the processed training data to enable it to understand and handle legal issues. Hyperparameter tuning and cross-validation are required during the training process to ensure the generalization ability and accuracy of the model.

[0040] The performance of the model is evaluated through the test set, including indicators such as accuracy, recall, and F1 value, to ensure that the model can provide reliable answers in practical applications.

[0041] The intelligent robot receives and understands the first question input by the first user through natural language processing technology. The user can input the legal question through text or voice.

[0042] Parse the questions input by users to extract keywords, legal terms and background information in the questions. The parsing process can use technologies such as word segmentation, named entity recognition (NER), and syntactic analysis.

[0043] The parsed question is passed as input to the first model, which generates a preliminary answer to the question based on its training results and the information in the legal knowledge base.

[0044] The first model combines relevant knowledge in the legal knowledge base to generate one or more possible answers. Depending on the complexity of the question, the answer may include citations of legal provisions, case analysis, regulatory interpretation, etc.

[0045] The first user can comment on the answer or ask further questions. The system adjusts the expression of the answer or provides more reference information based on the user's feedback.

[0046] If the first answer cannot be obtained or the confidence of the first answer is less than a first threshold, then based on the context of the conversation with the first user, the first question is organized into a second question that at least includes the background of the first question, and the second question is input into at least one large model to obtain at least one second answer output by the at least one large model.

[0047] In this step, the big model refers to the pre-trained large language models, such as Wenxinyiyan and ChatGPT, which have strong natural language understanding and generation capabilities. These models are trained on a large amount of data and can handle complex natural language tasks. However, the big models have not been specifically trained in the legal field, so it is difficult to give better answers in legal formats and terminology.

[0048] When the first model generates an answer, the system will evaluate the confidence of the answer. If the confidence is higher than the set first threshold, the answer will be directly fed back to the user.

[0049] Specifically, multiple confidence assessment models can be established in advance, where the confidence assessment model can be a convolutional neural network (CNN) model, a recurrent neural network (RNN) model, a Transformer model, etc. The input of the confidence assessment model can include a first question input by a first user and a first answer generated by the first model combined with relevant knowledge in the legal knowledge base. The output of the confidence assessment model can be a single-point prediction confidence of the first answer. The confidence of the first answer is calculated based on the single-point prediction confidence of multiple confidence assessment models.

[0050] Preferably, the confidence of the first answer can be calculated based on the following formula:

[0051]

[0052] Among them, C1 is the confidence of the first answer, σ0 is the variance of the preset single-point prediction confidence, σ1 is the variance of the single-point prediction confidence of multiple confidence assessment models, μ is the mean of the single-point prediction confidence of multiple confidence assessment models, K is the total number of confidence assessment models, and C k Single-point prediction confidence of the model for the kth confidence evaluation.

[0053] Using multiple different confidence assessment models can capture different aspects and features of the first question input by the first user and the first answer generated by the first model combined with relevant knowledge in the legal knowledge base. This diversity helps to evaluate the confidence of the answer more comprehensively and accurately. A single confidence assessment model may be biased due to data bias, algorithm limitations, or randomness in the training process. Multiple confidence assessment models can average or combine these deviations, thereby reducing the bias of the overall assessment.

[0054] If the confidence level of the answer generated by the first model is lower than the first threshold or the answer cannot be output at all, the system will determine that the answer may not be accurate or comprehensive enough and requires further processing.

[0055] The system extracts contextual information related to the current question, including the user’s previous conversation content, relevant legal provisions, case analysis, etc. This information helps to better understand the user’s consultation background and specific needs.

[0056] Integrate contextual information into the current question and reorganize it into a second question that includes the question background to make the question expression clearer and more specific.

[0057] Specifically, extract contextual information related to the current question (conversation history, legal provisions, case analysis) and represent it in a structured manner, such as using key-value pairs or nested JSON objects. Filter and sort the extracted information according to its relevance and importance to the current question. Ensure that the most relevant information is processed first. Based on the filtered and sorted information, construct a brief background story or situation description. This description should clearly reflect the specific context and needs of the user's consultation. Parse the original question to identify the keywords, topics, and questions. Merge the constructed background story or situation description with the original question to form a new question with complete background information. Optimize the reorganized question, including adjusting the word order, deleting redundant information, adding necessary conjunctions, etc., to improve the clarity and readability of the question.

[0058] Based on the contextual information, the system reorganizes the original question to ensure that the question contains all relevant background information. For example, if the user asks about the specific application of a certain legal provision, the system will integrate the specific content of the provision and the user's specific situation into the question.

[0059] By reorganizing and optimizing questions, the big model can better understand and process user needs and improve the accuracy and relevance of answers.

[0060] For example, the user's first question is: My company wants to terminate my labor contract, what should I do?

[0061] The following information was learned from the context of the conversation with the user: the user had worked for the company for 5 years and there was no clear termination clause in the contract.

[0062] Organize the user's conversation process into the second question:

[0063] I have worked for the company for 5 years, but there is no clear termination clause in the contract. The company now wants to terminate my labor contract, what should I do?

[0064] In this way, the system can provide more context, allowing the large model to more accurately understand the user's specific situation and generate more detailed and accurate answers.

[0065] The reformulated second question is fed into one or more large models. The large models have strong natural language understanding and generation capabilities, can handle complex questions and provide high-quality answers.

[0066] The big model generates one or more possible second answers based on the second question, which may include more comprehensive legal interpretations, relevant case analysis, etc.

[0067] The first legal knowledge in at least one second answer is extracted, the first legal knowledge is added to the legal knowledge base, and is marked as unverified knowledge.

[0068] The large model generates one or more second answers based on the reorganized second question, which may contain detailed legal interpretations, relevant case analysis, and regulatory references.

[0069] The system conducts a preliminary evaluation of the second answer generated by the large model and selects the answer that best meets user needs and has the most reference value.

[0070] Specifically, the single-point prediction confidence of the second answer can be determined through multiple confidence assessment models, the second answer confidence is calculated based on the single-point prediction confidence of the second answer output by the multiple confidence assessment models, and then the correlation between the second answer and the context information related to the current question is evaluated based on multiple correlation assessment models, wherein the correlation assessment model can be a convolutional neural network (CNN) model, a recurrent neural network (RNN) model, a Transformer model, etc., the input of the correlation assessment model can include the extracted context information related to the current question and one or more second answers generated by the large model according to the reorganized second question, the output of the correlation assessment model can include the correlation between the second answer and the context information related to the current question, the matching value of the second answer is calculated based on the second answer confidence and the correlation between the second answer and the context information related to the current question, and the second answer whose matching value is greater than a preset matching value threshold is selected as the second answer.

[0071] Preferably, the matching value of the second answer is calculated based on the following formula:

[0072]

[0073] Where M2 is the matching value of the second answer, a1 and a2 are preset weights, a1 and a2 are greater than 0, a1+a2=1, C2 is the confidence of the second answer, R m is the relevance between the second answer output by the mth relevance evaluation model and the context information related to the current question, where M is the total number of relevance evaluation models.

[0074] Using multiple confidence assessment models can comprehensively consider the prediction results of different models, reduce the deviation that may be caused by a single model, and thus improve the accuracy and robustness of the overall prediction. Calculating the single-point prediction confidence for the second answer can more accurately evaluate the reliability of the answer and ensure that the selected answer has a high degree of credibility. Through the relevance assessment model, the relevance between the second answer and the contextual information related to the current question can be deeply analyzed. This analysis helps to screen out answers that better match the background and context of the question, and improve the pertinence and practicality of the answer. Combining the confidence and relevance of the second answer, the matching value is calculated. This step provides a quantitative basis for the selection of the answer. By setting a preset matching value threshold, high-quality answers can be automatically screened out, reducing the impact of manual intervention and subjective judgment.

[0075] Extract key legal knowledge from the selected second answer. The extracted content may include specific legal provisions, clause interpretation, case analysis and legal opinions, etc.

[0076] Example: Assuming that the second answer mentions "According to Article 40 of the Labor Contract Law, the company can terminate the labor contract under certain circumstances", the system will extract this legal knowledge.

[0077] The extracted legal knowledge is formatted to meet the storage requirements of the legal knowledge base, for example, by terms, cases, interpretations, etc.

[0078] The formatted legal knowledge is added to the legal knowledge base and stored as a new legal entry. Each new legal knowledge will have an independent record, including information such as source, content, and related tags.

[0079] Example: Add the extracted Article 40 of the Labor Contract Law and its interpretation to the legal knowledge base, and record the source as the answer generated by the large model.

[0080] Newly added legal knowledge is managed by labeling, and multiple tags are added according to its content, applicable fields and relevance to facilitate subsequent retrieval and reference.

[0081] Since the newly extracted legal knowledge has not yet been subject to user feedback and professional review, in order to ensure the authority and accuracy of the knowledge base, this knowledge will be initially marked as unverified knowledge.

[0082] Unverified knowledge refers to legal knowledge that has not been fully verified. The system has reservations about its accuracy and reliability and requires further verification.

[0083] The background information of the second user and the second question input by the second user are obtained. If the relevance of the background information of the second user to the first legal knowledge is greater than a second threshold, and the relevance of the first legal knowledge to the second question is not less than a third threshold, a third answer including the first legal knowledge is presented to the user. If the second user does not question the third answer, the confidence of the first legal knowledge is increased based on the background information of the second user.

[0084] The intelligent robot system not only serves ordinary users, but also provides assistance to legal professionals (such as helping to write documents, etc.), such as lawyers who specialize in a certain legal field. When the second user (such as a labor lawyer) enters a question, the system will match the user's background information and the content of the question with the existing legal knowledge to be verified. If the user's background information is highly relevant to a certain legal knowledge, and the knowledge has a certain degree of relevance to the user's question (in order to actively expose new knowledge to be verified, the relevance here can be set relatively low), the system will present the answer generated by the legal knowledge to the user as one of the optional answers. Through user feedback, the system continuously adjusts and improves the confidence of legal knowledge, thereby optimizing the knowledge base and answering capabilities.

[0085] Specifically:

[0086] The system obtains the background information of the second user from the user registration information, usage history, professional field, years of work, etc. For example, if the second user is a lawyer specializing in labor law, the system will record his / her occupation, professional field and other information.

[0087] The system uses natural language processing technology to parse new legal consultation questions input by users and extract keywords and important information.

[0088] When determining the relevance of the second user's background information to the first legal knowledge, extract keywords from the user's background information, such as "labor law", "lawyer", "labor contract", etc. Convert keywords into vector representations, using methods such as word embedding (such as Word2Vec, G loVe) or TF-IDF. Use feature vectors to calculate the relevance of the second user's background information to the first legal knowledge.

[0089] By matching user background information and legal knowledge, the user's professionalism can be ensured, new knowledge can be embedded in problem cases, and the authenticity of the knowledge can be verified by professionals.

[0090] When the second user (such as a legal professional) enters the second question, the system will calculate the relevance between the user's background information and the legal knowledge, match the legal knowledge with a higher relevance as part of the third answer, and present it to the user. If the relevance between the first legal knowledge and the second question is not less than the third threshold, the third answer containing the legal knowledge will be presented to the user. The third threshold set here is relatively low in order to intentionally present the first legal knowledge to be verified to professional users for the purpose of verifying the correctness and reliability of such legal knowledge.

[0091] Specifically, an association graph may be pre-established, wherein the association graph may be used to characterize the relevance of the first legal knowledge between multiple first legal knowledge and multiple sample users. Keywords may be extracted from the user background information of the sample users, such as "labor law", "lawyer", "labor contract", etc. To convert keywords into vector representations, word embedding (such as Word2Vec, Globe) or TF-IDF and other methods may be used.

[0092] The vector similarity between the feature vector of the background information of the second user and the feature vector of the user background information of the sample user can be calculated, and the sample users whose vector similarity is greater than a preset vector similarity threshold are taken as target sample users. Based on the first legal knowledge relevance between multiple first legal knowledge and multiple sample users and the vector similarity between the feature vector of the background information of the second user and the feature vector of the user background information of the sample user, the relevance of the background information of the second user to the first legal knowledge is calculated.

[0093] The relevance of the second user's background information to the first legal knowledge can be calculated based on the following formula:

[0094]

[0095] Among them, D1 is the relevance between the background information of the second user and the first legal knowledge, S i is the vector similarity between the feature vector of the background information of the second user and the feature vector of the user background information of the i-th target sample user, C i is the relevance between the i-th target sample user and the first legal knowledge, I is the total number of target sample users, σ0 is the preset mean, and D0 is the preset relevance of the first legal knowledge.

[0096] A lower threshold (third threshold) is set, for example, the third threshold is set to be lower than the threshold of the correct answer. Specifically, when the relevance between knowledge and the question reaches 95%, the knowledge is considered to be the answer to the question. However, in order to increase the exposure of the knowledge to be verified, the third threshold can be set to 80%, so that more legal knowledge to be verified can be included in the potential answer range. This is done to deliberately present this legal knowledge to professional users so that the correctness of the knowledge can be verified through their feedback.

[0097] If the second user does not question the third answer, the system will adjust the confidence of the first legal knowledge based on the user's background information. Specifically, if the second user's background information shows that he has high professional qualifications (such as a senior lawyer), the system will significantly increase the confidence of the legal knowledge; if the user's qualifications are low (such as an ordinary user), the system will only increase the confidence slightly.

[0098] For senior lawyers, the system gives greater weight to their feedback. If these users do not question the third answer, the system will significantly increase the confidence of the legal knowledge. This is because the professional judgment of highly qualified users is more authoritative and accurate.

[0099] For others, such as ordinary legal consultants, the system gives less weight to their feedback. If these users do not question the third answer, the system will only increase the confidence by a small amount. This is because the professional ability and experience of ordinary users are relatively low, and the weight of their feedback needs to be carefully considered.

[0100] The system adjusts the confidence level of the first legal knowledge based on the user's background information and qualification level.

[0101] Preferably, the confidence of the first legal knowledge may be adjusted based on the following formula:

[0102] η t+1 =η t+△η×τ

[0103] Among them, η t+1 is the confidence of the adjusted first legal knowledge, η t is the confidence of the first legal knowledge before adjustment, △η is the preset basic adjustment amount of the confidence of the first legal knowledge, and τ is the adjustment weight corresponding to the user.

[0104] Among them, Base Increment is the basic confidence increment, and User Weight is the weight adjusted according to the user's qualification level.

[0105] User weight setting:

[0106] Senior lawyer: A higher weight, such as a weight of 2 or higher, significantly increases confidence.

[0107] Ordinary users: The weight is low, for example, the weight value is 0.5 or lower, and only a small confidence increase is performed.

[0108] The feedback from ordinary users is given a lower weight to avoid unnecessary confidence increases due to their less expertise and experience, and to maintain the stability and accuracy of the system.

[0109] Feedback from different user backgrounds optimizes the confidence adjustment strategy, enabling the system to more accurately evaluate and adjust the confidence of legal knowledge.

[0110] If the confidence level of the first legal knowledge is greater than a fourth threshold, the first legal knowledge is marked as verified and solidified into the legal knowledge base.

[0111] The fourth threshold is the minimum confidence requirement for the system to mark legal knowledge as verified knowledge. The purpose of setting this threshold is to ensure that only legal knowledge that has been fully verified and has high reliability can be solidified into the knowledge base.

[0112] Assume that the fourth threshold is set to 0.8, that is, only when the confidence of a piece of legal knowledge is greater than 0.8, the system will mark it as verified knowledge.

[0113] When the confidence level of a piece of legal knowledge exceeds the fourth threshold, the system marks it as verified knowledge, which indicates that the knowledge has been fully verified and has high reliability.

[0114] Verified legal knowledge will be solidified into the legal knowledge base as authoritative and credible reference information. The solidified knowledge will be more widely used in user consultation, improving the system's service quality and user trust.

[0115] Verified legal knowledge can be more widely used in users' legal consulting services to provide accurate and reliable answers.

[0116] Furthermore, when processing user questions, the system will give priority to calling verified legal knowledge to ensure that the answers provided are highly reliable. In particular, when the user is an ordinary user (non-legal practitioner), the system will give priority to calling verified legal knowledge.

[0117] Furthermore, the solidified knowledge will still be monitored and updated by the system to ensure that it is adjusted and optimized in a timely manner as legal provisions and practical applications change. When knowledge is questioned by practitioners with a legal background, the confidence level of the relevant knowledge is lowered and the evaluation process is re-entered.

[0118] The first model is updated according to the legal knowledge base.

[0119] With user feedback and verification, the legal knowledge base will be continuously updated to maintain its timeliness and accuracy. In order to ensure that the first model always has the latest and most authoritative legal knowledge, the system needs to retrain and optimize the first model regularly or in real time based on the updated content of the legal knowledge base. This process can improve the model's answering ability and accuracy and enhance the overall intelligence level of the system.

[0120] In the above solution, the intelligent robot has the function of self-learning, and can continuously optimize itself according to user feedback and new legal knowledge, and improve its answering ability and accuracy. Combining the advantages of the first model and the big model, the big model is used to generate more comprehensive answers to complex questions, ensuring the accuracy and professionalism of the answers.

[0121] The system can adjust the confidence level of legal knowledge and optimize the content of the knowledge base based on user feedback, especially feedback from professional users. By presenting the legal knowledge to be verified to professional users (such as lawyers), the system can verify the correctness of the knowledge faster and more accurately.

[0122] The system can dynamically adjust the confidence level of legal knowledge based on user qualifications and feedback to ensure the authority and reliability of the knowledge base content.

[0123] The system marks highly confident legal knowledge as verified and solidifies it into the knowledge base, ensuring that the information in the knowledge base is authoritative and reliable. When answering questions from ordinary users, the system gives priority to using verified legal knowledge to ensure the reliability and accuracy of the answers.

[0124] The above solution provides highly intelligent and personalized legal consulting services by combining legal knowledge base, machine learning models, big models and user feedback mechanism. The system can continuously learn and optimize to ensure the accuracy and professionalism of answers, which can significantly improve user experience and system reliability.

[0125] Furthermore, the present invention also provides a method for using an intelligent robot based on data processing and having autonomous learning capabilities, and provides legal consultation based on at least one intelligent robot in the aforementioned schemes.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

[0127] The present invention does not particularly specify the structure of some modules, which shall be subject to the contents recorded in the prior art. The prior art mentioned in the above background technology section and the specific embodiment section of the present invention can be used as part of the present invention to understand the meaning of some technical features or parameters. The scope of protection of the present invention shall be subject to the contents actually recorded in the claims.

Claims

1. An intelligent robot based on data processing and having autonomous learning function, characterized in that: The intelligent robot is used for legal consultation The intelligent robot includes a legal knowledge base; The intelligent robot includes a first model, and the first model is trained according to the legal knowledge base; The intelligent robot receives a first question input by a first user, and obtains a first answer to the first question according to the first model; If the first answer cannot be obtained or the confidence of the first answer is less than a first threshold, organizing the first question into a second question including at least the background of the first question according to the context of the conversation with the first user, inputting the second question into at least one large model, and obtaining at least one second answer output by the at least one large model; Extracting the first legal knowledge from the at least one second answer, adding the first legal knowledge to the legal knowledge base, and marking the first legal knowledge as unverified knowledge; Obtaining background information of a second user and a second question input by the second user, if the relevance between the background information of the second user and the first legal knowledge is greater than a second threshold, and the relevance between the first legal knowledge and the second question is not less than a third threshold, presenting a third answer including the first legal knowledge to the second user, if the second user does not question the third answer, improving the confidence of the first legal knowledge based on the background information of the second user, and if the confidence of the first legal knowledge is greater than a fourth threshold, marking the first legal knowledge as verified and solidifying it into the legal knowledge base; The first model is updated according to the legal knowledge base.

2. The intelligent robot based on data processing and having autonomous learning function according to claim 1, characterized in that: Determining a first answer confidence level of a first answer to the first question includes: Pre-establishing multiple confidence assessment models, wherein the input of the confidence assessment model includes a first question input by a first user and a first answer generated by a first model combined with relevant knowledge in a legal knowledge base, and the output of the confidence assessment model is a single-point prediction confidence of the first answer; Calculate the first answer confidence based on the single-point prediction confidence of multiple confidence assessment models.

3. The intelligent robot based on data processing and having autonomous learning function according to claim 1, characterized in that: Extracting the first legal knowledge from the at least one second answer includes: Calculate the matching value of the second answer; The second answer whose matching value is greater than the preset matching value threshold is taken as the selected second answer.

4. The intelligent robot based on data processing and having autonomous learning function according to claim 3 is characterized in that: The calculating the matching value of the second answer comprises: Determining a single-point prediction confidence of the second answer through multiple confidence assessment models; Calculating the confidence of the second answer based on the single-point prediction confidence of the second answer output by the multiple confidence assessment models; The relevance of the second answer to the context information related to the current question is evaluated based on multiple relevance evaluation models, wherein the input of the relevance evaluation model includes the extracted context information related to the current question and one or more second answers generated by the large model according to the reorganized second question, and the output of the relevance evaluation model includes the relevance of the second answer to the context information related to the current question; based on the second answer confidence of the second answer and the relevance of the context information related to the second answer to the current question, the matching value of the second answer is calculated.

5. The intelligent robot based on data processing and having autonomous learning function according to claim 1, characterized in that: When determining the relevance of the background information of the second user to the first legal knowledge, keywords are extracted from the user background information, the keywords are converted into vector representations, feature vectors are generated, and the relevance of the background information of the second user to the first legal knowledge is calculated using the feature vectors.

6. The intelligent robot based on data processing and having autonomous learning function according to claim 5, characterized in that: The step of using the feature vector to calculate the relevance between the background information of the second user and the first legal knowledge includes: Calculating vector similarity between a feature vector of the background information of the second user and a feature vector of the user background information of the sample user; The sample users whose vector similarity is greater than a preset vector similarity threshold are taken as target sample users; Based on the first legal knowledge relevance between multiple first legal knowledge and multiple target sample users and the vector similarity between the feature vector of the second user's background information and the feature vector of the user background information of the target sample users, the relevance between the second user's background information and the first legal knowledge is calculated.

7. The intelligent robot based on data processing and having autonomous learning function according to claim 1, characterized in that: When the user is an ordinary user, verified legal knowledge is called first.

8. The intelligent robot based on data processing and having autonomous learning function according to claim 1, characterized in that: When knowledge is questioned by practitioners with a legal background, lower the confidence level of the relevant knowledge.

9. A method for using an intelligent robot based on data processing and having autonomous learning function, characterized in that: Use the intelligent robot described in any one of claims 1-8 to provide legal consultation.

Images