Intelligent computing platform abnormal root cause positioning method fused with large language model and related equipment

Abstract

The invention provides an intelligent computing platform abnormal root cause positioning method fused with a large language model and related equipment, and relates to the technical field of anomaly detection. The method comprises the steps of firstly obtaining semantic expression features of interaction between a target student and a large language model, and judging whether the interaction obstacle degree exceeds a preset value or not; if yes, the feedback type of the obstacle is determined firstly, then dialogues in a preset window before the obstacle appears are searched, the starting point of the obstacle is positioned, and key semantics of the content of the obstacle in the period are extracted. And then based on key semantics, searching a target learning group and a matching dialogue of users in a set area, identifying a plurality of error types and determining a basic score, matching the error types with an obstacle feedback type to obtain a matching score, combining the error types with the obstacle feedback type to determine a basic error type, and finally modifying and generating a reply and feeding back the reply to students. By implementing the method, a closed loop from obstacle identification to root cause positioning to generation of targeted reply is realized, and the accuracy and efficiency of abnormal root cause positioning are greatly improved.

Description

Technical Field

[0001] This application relates to the field of anomaly detection technology, and in particular to an anomaly root cause localization method and related equipment for an intelligent computing platform that integrates a large language model. Background Technology

[0002] With the rapid development of artificial intelligence technology, large language models have been widely used in education, scientific research, and industry. However, in subjects that require precise knowledge, such as history, science, and mathematics, large language models may provide incorrect concepts, formulas, or facts, and this misleading content can have a serious impact on learners.

[0003] To address the aforementioned issues, the industry has developed a knowledge graph-based large language model verification system. This system constructs a domain-specific knowledge graph and compares the model's output with a validated knowledge base to identify potential errors. When a user questions the model's answer, the output is verified based on the user's question, thus correcting the error. This approach improves the reliability of large language models in precise knowledge domains to some extent.

[0004] However, with existing technologies, when the model's output contains anomalies or logical problems, the user needs to explicitly point out the potential errors before the verification and correction process can be initiated. But in real-world educational scenarios, students often struggle to accurately pinpoint the problems. For example, they might only express confusion or anger, but cannot clearly describe where the large language model's output is wrong or where the logical breaks are, making it difficult to meet the educational demand for rapid and accurate error correction of large language models. Summary of the Invention

[0005] This application provides a method and related equipment for locating the root causes of anomalies in an intelligent computing platform that integrates a large language model. This method is used to accurately identify and locate the root causes of interaction obstacles encountered by students when using the large language model, and then generate targeted responses.

[0006] Firstly, this application provides a method for locating the root cause of anomalies in an intelligent computing platform that integrates a large language model. Applied to an intelligent computing platform, the method includes: acquiring semantic expression features during the interaction between a target student and a large language model; determining whether the target student's interaction obstacle feedback exceeds a preset obstacle level based on these semantic expression features; if it exceeds, judging the type of interaction obstacle feedback for the target student; searching for dialogue content within a preset length window before the target student exhibits interaction obstacle feedback, and locating the starting point of the interaction obstacle feedback; acquiring the interaction obstacle content from the starting point until the target student exhibits the interaction obstacle feedback, and extracting the content related to the interaction obstacle. The key semantics of the interaction barrier content are identified; based on these key semantics, a second user is found whose dialogue content matches the key semantics. This second user is a target learning group and users using the large language model within a defined region. The target learning group includes a network group of classmates using the large language model. Semantic recognition is performed on the dialogue content to be identified to determine multiple error types. A base score for each error type is determined. Each error type is matched with the interaction barrier feedback type to obtain a matching score. The root error type is determined based on the base score and the matching score. The content is modified according to the root error type, and a response is generated and fed back to the target student.

[0007] By employing the aforementioned technical solution, the degree of student interaction barriers is first identified through semantic expression features, ensuring that only cases with truly significant barriers are processed in depth, reducing invalid analysis. Next, the starting point of the barrier is located and key semantics are extracted, defining the precise scope for root cause analysis. By combining matched dialogues with users within the target learning group and region, the reference sample is expanded using group data. Through error type matching and scoring mechanisms, the root cause error is accurately identified from multiple dimensions. Each step progresses progressively, achieving a closed loop from barrier identification to root cause localization and then to generating targeted responses, significantly improving the accuracy and efficiency of abnormal root cause localization and ensuring that students receive effective solutions to their problems. In some embodiments of the first aspect, the step of acquiring semantic expression features during the interaction between the target student and the large language model, and determining whether the target student's interaction barrier feedback exceeds a preset barrier level based on the semantic expression features, specifically includes: determining multiple topic boundaries between the target student and the large language model based on the semantic expression features; generating a unique topic identifier for each topic based on the topic boundaries, and recording the total number of interaction rounds of the target student on the target topic; determining the number of occurrences of the same or similar questions in the target topic based on the total number of interaction rounds; if the number of occurrences exceeds a preset threshold, determining that the target student is suspected of having interaction barrier feedback on the target topic; determining the student's interaction temperature tone based on the student's semantic expression features in the target topic and a preset interaction temperature tone model, wherein the interaction temperature tone model is pre-trained by machine learning from student interaction data labeled with warm tones with non-negative emotions and student interaction data labeled with cold tones with negative emotions; if the student's interaction temperature tone is cold, determining that there is interaction barrier feedback; and determining the level of interaction barrier feedback based on the number of occurrences.

[0008] By employing the aforementioned technical solution, topic boundaries are first defined and unique identifiers are generated, allowing interaction analysis to focus on specific topics and avoid cross-topic interference. The total number of interaction rounds and the frequency of identical questions are statistically analyzed to initially assess the likelihood of obstacles from a quantitative perspective. Then, combined with an interaction temperature and tone model, the presence of obstacles is qualitatively confirmed through the assessment of cold tones in the emotional dimension. This dual-dimensional approach of frequency statistics and emotional analysis avoids misjudging obstacles due to accidental repetitive questions and prevents overlooking low-frequency but emotionally negative real obstacles, significantly improving the accuracy and reliability of interaction obstacle assessment. In conjunction with some embodiments of the first aspect, in some embodiments, the step of determining the current topic boundaries between the target student and the large language model based on the semantic expression features specifically includes: real-time acquisition of continuous interactive text between the target student and the large language model, generating a dialogue sequence in timestamp order, and marking the student input content and the large model's response content as units to be analyzed; generating corresponding semantic vectors based on the units to be analyzed; calculating the cosine similarity between the semantic vector of the current student input text unit and the topic semantic vectors in the previous preset window to obtain a real-time similarity value; if the real-time similarity value is lower than a preset similarity threshold, and multiple consecutive student input text units meet the conditions, then it is initially determined that a new topic has been triggered; marking the first student input text unit that triggers the new topic determination as the starting point of the new topic, and simultaneously recording the ending point of the previous topic to form a topic boundary division.

[0009] By employing the above technical solution, continuous interactive text is generated into a sequence based on timestamps and labeled with units, providing a structured foundation for semantic analysis. Semantic vectors are generated and cosine similarity is calculated, using quantitative indicators to determine topic relevance. A judgment rule of "multiple consecutive units below a threshold satisfying the condition" is established to avoid misjudgments caused by a single low-similarity unit. This segmentation method based on semantic vectors and continuous conditions can accurately identify the starting point of a new topic and the ending point of a previous topic, ensuring the objectivity and accuracy of topic boundary division and providing a precise topic range for subsequent interaction obstacle analysis. In conjunction with some embodiments of the first aspect, in some embodiments, the step of finding the dialogue content within a preset length window before the target student exhibits interaction obstacle feedback and locating the starting point of the interaction obstacle feedback specifically includes: extracting the feedback content of the dialogue content within the preset length window using natural language processing technology; inputting the feedback content into an interaction obstacle recognition learning model to determine whether interaction obstacle feedback exists, wherein the machine learning model is trained on language pattern annotations with different interaction obstacles; and if interaction obstacle feedback is detected, immediately marking the corresponding time point as the starting point of the interaction obstacle feedback.

[0010] By employing the above technical solution, natural language processing technology is used to extract feedback content within a preset window, enabling efficient filtering of dialogue information. The feedback content is then input into an interaction obstacle recognition model trained on labeled data. Leveraging machine learning's pattern recognition capabilities, the model accurately detects the presence of interaction obstacles. Once an obstacle is detected, the starting point is immediately marked, achieving rapid location from the dialogue content to the obstacle's origin. In conjunction with some embodiments of the first aspect, in some embodiments, the step of performing semantic recognition on the dialogue content to be identified and determining multiple error types specifically includes: obtaining error correction dialogue content from the dialogue content to be identified that contains the user's error correction intention in a large language model; performing semantic recognition on the error correction dialogue content and matching it with a preset error type correction vocabulary to determine the error type.

[0011] By employing the aforementioned technical solution, the content containing error-correction intent in the second user's dialogue is first extracted, focusing on error correction information that is truly valuable. This content is then semantically recognized and matched against a pre-defined error type lexicon. Through standardized lexicon comparison, error type judgment has a clear basis. This approach avoids ineffective analysis of non-error-correction content, while unifying error type standards through lexicon matching, reducing subjective judgment differences, and ensuring the consistency and accuracy of error type classification. This lays a reliable foundation for subsequent error scoring and root cause determination. In conjunction with some embodiments of the first aspect, in some embodiments, the step of determining the basic score for each error type specifically includes: statistically analyzing error type data in the dialogue content to be identified, the error type data including the error type and the corresponding associated user identifier; statistically analyzing the cumulative occurrence count of each error type; determining the corresponding number of unique users based on each error type and the associated user identifier, the number of unique users being the total number of non-repeating users who have experienced type errors; determining the severity coefficient of each error type based on the cumulative occurrence count of each error type and the corresponding number of unique users; sorting each error type according to the severity coefficient, and determining the basic score for each error type based on the sorting result.

[0012] By employing the above technical solutions, the cumulative frequency of error types is statistically analyzed to reflect the prevalence of errors; the number of unique users is statistically analyzed to reflect the user scope affected by errors. Combining these two methods to calculate a severity coefficient avoids misjudgments caused by a single dimension (such as looking only at the frequency) and provides a comprehensive assessment of the actual impact of errors. A base score is determined by ranking the coefficients, ensuring that the score objectively reflects the priority of each error type. This multi-dimensional statistical and quantitative scoring method makes the base score more convincing, providing a scientific basis for subsequent matching with interaction obstacle types to determine the root cause, and improving the rationality of root cause localization. In conjunction with some embodiments of the first aspect, in some embodiments, after the step of finding the dialogue content to be identified by the second user that matches the key semantics, the method further includes: sending a help request message to the second user's big language model user terminal, the help request message including the key semantics in the interaction obstacle content; receiving feedback content sent by the big language model user terminal, adjusting the output content of the big language model according to the feedback content; and sending the adjusted output content to the target student's big language model user terminal for display.

[0013] By employing the above technical solution, after finding a matching dialogue, a help request containing key semantics is sent to a second user, supplementing the output of the large language model with the practical experience of similar users. Upon receiving feedback, the output content is adjusted to integrate collective wisdom with the model's capabilities. The adjusted content is then sent to the target student, ensuring that the response is not only based on model analysis but also incorporates the effective experience of actual users. This approach compensates for potential limitations of the model, making the response more relevant to students' actual needs, improving problem-solving efficiency, and enhancing the practicality and effectiveness of the interaction. In conjunction with some embodiments of the first aspect, in some embodiments, the server includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory being used to store computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the server to perform the method as claimed in any one of claims 1-7.

[0014] In a second aspect, this application provides a computer-readable storage medium including instructions, characterized in that, when the instructions are executed on a server, the server causes the server to perform the method described in the first aspect and any possible implementation thereof.

[0015] Thirdly, this application provides a computer program product that, when run on a server, causes the server to perform the method described in the first aspect and any possible implementation thereof.

[0016] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By employing techniques such as semantic expression feature analysis that integrates large language models, root cause mining by combining matching dialogue data of target learning groups and regional users, and error type matching and scoring mechanisms, this technology effectively solves the technical problem of inaccurate root cause localization of student-large language model interaction barriers in existing technologies, thereby achieving the technical effect of accurate localization and efficient analysis of the root causes of interaction barriers.

[0017] 2. By employing cosine similarity calculation based on semantic vectors and combining it with similarity threshold judgment rules for multiple consecutive text units to delineate topic boundaries, the technology effectively solves the problems of inaccurate topic boundary delineation and strong subjectivity in existing technologies. This enables objective and accurate delineation of topic boundaries during student-large language model interaction, providing a reliable scope definition for subsequent interaction obstacle analysis.

[0018] 3. By employing the technique of extracting dialogue content containing error correction intent and semantically matching it with a pre-defined error type lexicon, the technology effectively solves the problems of inaccurate error type identification and inconsistent standards in existing technologies. This achieves accurate and standardized identification of error types, providing a reliable foundation for subsequent error scoring and determination of the root cause of the error. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an anomaly root cause localization method for an intelligent computing platform that integrates a large language model, as described in this application embodiment. Figure 2This is another flowchart illustrating the method for locating the root cause of anomalies in an intelligent computing platform that integrates a large language model, as described in this application. Figure 3 This is a schematic diagram of the physical device structure of an intelligent computing platform in the embodiments of this application. Detailed Implementation

[0020] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0021] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more. For ease of understanding, the method provided in this implementation is described in process below. Please refer to [link / reference]. Figure 1 This is a flowchart illustrating a method for locating the root cause of anomalies in an intelligent computing platform that integrates a large language model, as described in this application.

[0022] S101. Obtain the semantic expression features of the target student during the interaction process with the large language model, and determine whether the feedback of the interaction barrier level of the target student exceeds the preset barrier level based on the semantic expression features. The target student refers to a specific individual student interacting with the learning platform using a large language model. The large language model refers to an AI model with natural language understanding and generation capabilities, used for dialogue interaction with students, such as an AI assistant that can answer subject-related questions and provide learning guidance. Semantic expression features represent characteristic information extracted from the dialogue content between the student and the large language model, reflecting the student's expressive intentions, emotional tendencies, and linguistic logic, such as keywords, sentence structures, and emotional vocabulary in the student's questions. The feedback on the degree of interaction obstacle refers to the severity of difficulties encountered by the student during interaction with the large language model, such as the frequency of repeated questioning and the degree of confusion in expression. The preset obstacle level refers to the threshold standard pre-set by the intelligent computing platform to determine whether a student has significant interaction obstacles; for example, asking the same question repeatedly five times or more is considered exceeding the preset level.

[0023] During real-time interaction between the target student and the large language model, the intelligent computing platform needs to continuously monitor and analyze the interaction to determine if the student is experiencing any interaction obstacles requiring special attention. In actual interaction, if the large language model initially provides an incorrect answer, or an answer irrelevant to the student's question, the student may continue interacting based on this incorrect content. When the initial incorrect answer exceeds the student's cognitive range, or the error is relatively simple, the student often has difficulty recognizing it, which in turn affects subsequent question-and-answer sessions. During this process, students may find the model's content seemingly correct yet uncertain, until subsequent errors exceed their cognitive range, at which point they become clearly aware of the mistakes. During this time, students will experience considerable confusion and questions, and may even feel anger, all of which will be expressed through language or speech. For example, if a student repeatedly asks "why" about a question, it indicates that the logic of the large language model is difficult for the student to understand and may contain contradictions, thus creating an obstacle in the interaction between the student and the model. When determining whether the level of interaction obstacle exceeds a preset obstacle level, the intelligent computing platform will compare and analyze the extracted semantic expression features with preset evaluation standards. The preset obstacle level is set by the platform based on a large amount of historical interaction data and the experience of educational experts, covering multiple dimensions. For example, in the dimension of repeated questioning, the preset standard might be "the same or similar questions appear 3 or more times within 10 rounds of dialogue"; in the dimension of expression clarity, it might be set as "grammatical errors exist in 3 consecutive rounds of dialogue and the core intent cannot be accurately identified"; in the dimension of emotional reaction, it might stipulate "5 or more instances of clearly negative emotional expressions (such as 'I don't understand at all' 'It's too complicated')", etc. The intelligent computing platform comprehensively evaluates these dimensional indicators and calculates a quantitative value for the degree of interaction barrier. When this quantitative value exceeds the threshold corresponding to the preset barrier level, it is determined that the student's interaction barrier feedback exceeds the preset level. The specific implementation process will be described in detail in steps S201 to S207, and will not be repeated here.

[0024] S102. If the number of cases exceeds the limit, then the type of interaction barrier feedback for the target student will be determined. Among them, the interaction barrier feedback type refers to the different forms of barriers that students experience when interacting with a large language model due to problems in the model's output (such as incorrect formulas or incorrect explanations of principles).

[0025] Once it is determined that the target student's feedback on the level of interaction barrier exceeds the preset barrier level, it is necessary to further clarify the specific type of barrier to provide direction for subsequent root cause localization. The intelligent computing platform's classification process for interaction barrier feedback types involves several stages: text preprocessing, feature extraction, dimensional analysis, and type matching. More specifically, in the text preprocessing stage, the platform cleans and standardizes the interaction text between the student and the large language model, including removing meaningless stop words (such as "um" and "oh"), correcting typos (such as correcting "public test" to "formula"), and supplementing omitted components (such as completing "this is not right" to "this formula is not right").

[0026] After extracting key features, the platform compares these features with a pre-defined feature library of interaction obstacle feedback types. This feature library contains typical features for various interaction obstacle types. For example, typical features of formula confusion include frequently mentioning formula errors and abnormal calculation results; typical features of principle misunderstanding include questioning the logical relationships of principles and confusing different principles. The platform calculates the similarity between the target student's interaction obstacle features and the features of each category in the feature library. The category with the highest similarity is the initially determined interaction obstacle feedback type. Simultaneously, the platform also incorporates historical interaction data for supplementary judgment. Reviewing the target student's past interaction records, if they have frequently encountered formula-related obstacles in the past, the likelihood of being classified as formula confusion is higher; if the student has more problems understanding principles in learning other knowledge points, the initial judgment needs to be re-examined to ensure accuracy.

[0027] In some embodiments, the determination of the interaction obstacle feedback type of the target student can be achieved in several ways: Optionally, firstly, an interaction obstacle type determination rule base is constructed, containing the conditions for determining each category. For example, when the frequency of keywords related to "formula" exceeds 30% and there are expressions questioning the calculation results, it is determined to be a "formula confusion" type. Then, the preprocessed student interaction text is matched with the conditions in the rule base, and the corresponding type is determined based on which condition is met. For example, if the word "formula" appears 5 times and "incorrect calculation" appears 3 times in the student interaction text, it meets the determination conditions for the "formula confusion" type, and is therefore determined to be a "formula confusion" type. Optionally, a machine learning model is used for determination. Firstly, a large amount of student interaction data labeled with interaction obstacle types is selected as a training set to train a classification model. Then, the interaction features of the target student are input into the model, and the model outputs the interaction obstacle feedback type based on the learned classification rules.

[0028] S103. Locate the dialogue content within a preset length window before the target student experiences interactive obstacle feedback, and pinpoint the starting point of the interactive obstacle feedback. The preset length window refers to the time or round range of dialogue content pre-set by the intelligent computing platform for finding feedback on interaction barriers. For example, it could be set to "the 10 rounds of dialogue before the interaction barrier feedback appeared" or "dialogue content within the past 5 minutes." Its length can be dynamically adjusted according to the subject difficulty and the student's grade level; for example, it could be set to the first 15 rounds of dialogue for high school physics. Dialogue content refers to all text information generated during the interaction between the target student and the large language model. The starting point of interaction barrier feedback refers to the time point or round of dialogue when the student first exhibits interaction barrier characteristics on the same topic.

[0029] The target student has been identified as having an interaction barrier exceeding a preset level. At this point, it's necessary to trace back the relevant dialogue before the barrier occurred to pinpoint its origin. When the intelligent computing platform performs this step, it first uses natural language processing technology to process the dialogue content within a preset length window to extract the feedback. In this process, the platform first cleans the dialogue content, removing irrelevant symbols, spaces, etc., then performs word segmentation and part-of-speech tagging to filter out the parts belonging to the student's feedback and exclude responses from the large language model.

[0030] After extracting the feedback content, the platform inputs it into the interaction barrier identification learning model. During training, the model's input data consists of the feedback text generated when students interact with the large language model, along with language pattern features extracted from the text. These features include keyword sequences (such as word sequences indicating barriers like "don't understand," "incorrect," and "repetition"), sentence structure features (such as the frequency of consecutive interrogative sentences and the number of times negative sentences are used), and interaction round association features (such as differences in the expression of the same question in different rounds). The output data is the interaction barrier pattern category to which the feedback content belongs, such as "questioning the accuracy of the content," "expressing confusion," "repeatedly asking questions," and "expressing dissatisfaction." If no interaction barrier exists, the output is "no interaction barrier." The labeled data used during training consists of a large amount of student feedback text labeled with clear interaction barrier language patterns. For example, "Your answer is different from the textbook, is it wrong?" is marked as "questioning the accuracy of the content"; "I still don't understand this derivation, can you explain it again?" is marked as "expressing confusion"; "I've asked this question three times, and I still haven't gotten a clear answer" is marked as "repeated questioning"; "You've explained so much, but it's still not clear, it's so frustrating" is marked as "expressing dissatisfaction"; and "Your explanation is very clear, I understand" is marked as "no interaction barriers". During training, the labeled data is first divided into training and test sets in a 7:3 ratio. For the base model (such as the BERT fine-tuning model, random forest model, etc.), training is performed: taking the BERT fine-tuning model as an example, the feedback text is first converted into word embedding vectors and input into the pre-trained BERT model. By fine-tuning the hidden layer parameters of the model, the model learns the mapping relationship between interaction barrier patterns and text features. During training, the cross-entropy loss function is used to calculate the difference between the model's output class probability and the labeled class, and the optimizer adjusts the model parameters to minimize the loss value. After training, the model can extract features and perform pattern matching on new input feedback to accurately determine whether there are interaction barriers and their corresponding patterns. For example, when the input is "I've asked three times, but you still haven't explained it clearly," the model will identify the "repeated questioning" feature of "three times" and the "expression of dissatisfaction" feature of "not explained clearly," and thus output the corresponding barrier pattern category.

[0031] If the model detects a feedback obstacle, the platform will immediately mark the time point corresponding to that feedback as the starting point of the feedback obstacle. This time point is usually the timestamp when the feedback was recorded. By accurately marking the starting point, the platform can help trace the development process of the feedback obstacle and better analyze its causes. For example, if the feedback "This formula is different from the one in my textbook" is detected as a feedback obstacle with a timestamp of "14:35:22", then that time point will be marked as the starting point of the feedback obstacle.

[0032] S104. Obtain the content of the interaction obstacle from the starting point to the time when the target student receives the feedback of the interaction obstacle, and extract the key semantics of the interaction obstacle content; The interaction obstacle feedback period refers to the time or dialogue interval from the starting point until the target student's interaction obstacle level is determined to exceed a preset level, such as from round 15 to round 25. Interaction obstacle content refers to all dialogue text related to interaction obstacles generated during this period when the target student interacts with the large language model. Key semantics refers to semantic information extracted from the interaction obstacle content that reflects the core essence of the obstacle.

[0033] After identifying the starting point of the interaction barrier feedback, it is necessary to focus on the complete process from the emergence to the escalation of the interaction barrier, extracting core information to support subsequent root cause analysis. When the intelligent computing platform performs this step, it first accurately retrieves all interaction content within this period from the dialogue database based on the starting point and the interaction barrier feedback point (i.e., the time point when the barrier level exceeds the preset value), forming a set of interaction barrier content. Next, this content is preprocessed. Subsequently, the platform uses multi-layer semantic analysis technology to extract key semantics. The first layer is topic identification, which determines the core knowledge points involved in the interaction barrier by calculating the TF-IDF value of words, such as "quadratic function" and "Newton's second law"; the second layer is question extraction, which identifies the student's core questions through dependency parsing, such as "the meaning of formula parameters" and "the conflict between the principle and the textbook"; the third layer is emotion and behavior analysis, which counts the frequency of students' questions during the period (e.g., repeated questions 5 times), the number of challenges (e.g., challenges 3 times), and the frequency of negative emotion words (e.g., "don't understand" and "confused" appear 4 times) as supplementary information. During the extraction process, the platform optimizes its analysis logic for different types of interaction barriers: For barriers related to formula confusion, it focuses on extracting key entities such as formula names, parameters, and derivation steps, as well as students' questions about these entities. For example, "In rounds 15-25 of the dialogue, students questioned the writing of '4ac-b²' in the quadratic function vertex formula three times, believing it conflicted with 'b²-4ac' in the textbook." For barriers related to misunderstanding principles, it focuses on the core elements of the principle, students' misunderstandings, and the sources of conflict. For example, "Students' understanding of the inertia principle is limited to 'related to velocity,' which conflicts with the model explanation of 'related to mass,' and they expressed 'not understanding' twice." Ultimately, the platform integrates this information into concise and precise key semantics, providing a clear direction for subsequent analysis steps.

[0034] S105. Based on the key semantics, find the dialogue content to be identified that matches the key semantics of the second user. The second user is the target learning group and users using the large language model in the set area. The target learning group includes the network group of classmates using the large language model. The second user refers to the reference user group other than the target student, including the target learning group (such as the large language model usage network group composed of classmates) and users in the designated area (such as students in the same school, city, or specific online learning community using the large language model), for example, the group of classmates in the second grade (1) class where the target student is located and students in other classes of the junior high school. The dialogue content to be identified refers to the dialogue text related to key semantics generated during the interaction between the second user and the large language model, including the second user's questions (such as "Is there two ways to write the quadratic function vertex formula?"), corrections (such as "The model wrote 'b²-4ac' in the vertex formula instead of '4ac-b²', which is different from the textbook") and the large language model's responses, etc.

[0035] After acquiring the key semantics, it is necessary to leverage the similar interaction experiences of second users to expand the analysis sample and provide a reference for pinpointing the root causes of the target student's interaction barriers. When the intelligent computing platform performs this step, it first converts the key semantics into standardized retrieval vectors. A pre-trained language model is then used to encode the key semantics, resulting in vector representations that reflect their semantic features. Next, the scope of the second users is determined: target learning groups are associated through class IDs or group identifiers. For example, if the target student's class ID is "Grade 8, Class 1," then the interaction records of all users under that ID are selected. Users within a defined region are filtered through geographical information (such as the school and city corresponding to the IP address) or community identifiers. For example, if the defined region is "XX Middle School," then the interaction data of all students from that school are selected. Subsequently, the platform performs semantic matching retrieval in the second user's interaction database. The first layer of retrieval uses vector similarity calculation, performing cosine similarity calculation between the key semantic vector and the second user's dialogue content vector, filtering out dialogue content with a similarity ≥ 0.6. For example, a dialogue "The parameters of the quadratic function vertex formula don't match the textbook, is the model wrong?" has a similarity of 0.75 with the key semantics and is initially selected. The second layer of retrieval is strengthened through keyword matching, extracting core entities from the key semantics (such as "quadratic function", "vertex formula", "textbook"), and filtering out content containing these entities from the initially selected dialogues to ensure high topic relevance. The third layer of retrieval considers time factors, prioritizing dialogue content within the last 3 months, as newer content is more likely to reflect the current learning scenario and model status, such as excluding dialogues from 1 year ago such as "the expression of the vertex formula in the old textbook".

[0036] For the selected dialogue content to be identified, the platform will calculate a comprehensive score according to the formula "vector similarity × 0.7 + keyword matching degree × 0.2 + time decay factor × 0.1" (time decay factor: 1.0 for the most recent month and 0.8 for the last 1-3 months). The top 100 scores will be used as the final dialogue content to be identified, providing sufficient reference samples for subsequent error type identification.

[0037] This step involves obtaining interaction information from users of the target learning group and the designated area who use the large language model. This is because classmates often have identical assignments, leading to a higher likelihood of students asking the same questions or content. Furthermore, high-achieving classmates may immediately spot errors in the model's responses. If multiple students point out errors in the model's answers, the model is likely flawed or logically incorrect. The designated area likely encompasses students from the same school, district, or even city. These students often share similar textbook versions, teaching schedules, and exam syllabi, and face common knowledge difficulties and learning challenges. When a significant number of students within the designated area question or point out errors in the same response from the large language model, it indicates that the content may deviate from local teaching requirements and knowledge systems, or have limitations in applicability.

[0038] It should be noted that the intelligent computing platform can obtain the interaction content between the second user and the large language model with the second user's prior authorization. For example, when the second user registers and uses the large language model for the first time, the platform will pop up an authorization agreement window, clearly informing the user that the interaction content between the user and the model may be used to assist other users in solving interaction obstacles and optimizing model responses after anonymization, and providing "agree" and "disagree" options. Only after the user clicks "agree" will the platform obtain the relevant content in subsequent interactions.

[0039] S106. Perform semantic recognition on the dialogue content to be identified to determine multiple error types; The system can acquire error-correction dialogue content from the second user's input to the large language model, containing the user's intention to correct errors. After semantic recognition of this error-correction dialogue content, it is matched against a pre-defined error-type correction vocabulary library to determine the error type. The user's intention to correct errors refers to the second user's expressed willingness in the dialogue to point out errors in the large language model's response, typically expressed through words such as "incorrect," "wrong," or "should be." Error-correction dialogue content refers to the dialogue text from the second user with this intention, such as "The physics formula you provided is wrong; the correct one should be F=ma." The pre-defined error-type correction vocabulary library is a database pre-built by the intelligent computing platform that stores feature words and examples categorized by error type. For example, the "formula error" category includes feature words such as "formula written incorrectly" and "parameter error," along with related examples.

[0040] After obtaining the conversation content to be recognized, when the intelligent computing platform executes this step, it first needs to obtain the error correction conversation content containing the user's error correction intention. The platform will analyze all the conversation content to be recognized one by one, and judge whether each conversation contains the user's error correction intention through the intention recognition algorithm in natural language processing technology. The intention recognition algorithm will pay attention to whether error correction feature words appear in the conversation, such as "error", "wrong", "not like this", "should", etc., and at the same time judge in combination with the semantic logic of the sentence. For example, for the conversation "The balancing of this chemical equation you mentioned is wrong, and the correct balancing coefficients should be 2, 1, 2", the platform determines that it contains the user's error correction intention by recognizing words such as "wrong" and "correct", and incorporates it into the error correction conversation content set; while for a conversation like "This knowledge point is so difficult. How can I remember it", since it does not show the error correction intention, it will be excluded. After obtaining the error correction conversation content, the platform will perform semantic recognition on it. This process includes multiple links: first, word segmentation and part-of-speech tagging are performed to split the text into words and tag their part-of-speech, such as splitting "You wrote the parameters of the vertex formula of the quadratic function backwards" into "You / pronoun gave / verb of / auxiliary word quadratic function / noun vertex formula / noun parameters / noun wrote backwards / verb le / auxiliary word"; then named entity recognition is performed to extract the key entities, such as "quadratic function vertex formula" and "parameters"; then through dependency syntactic analysis, the grammatical structure and semantic relationship of the sentence are clarified, and the object of the error (such as "the parameters of the quadratic function vertex formula") and the manifestation of the error (such as "wrote backwards") are determined; finally, in combination with the context, the correct content considered by the second user is extracted (such as "It should be b² - 4ac instead of 4ac - b²"). After completing the semantic recognition, the platform will match the recognition result with the preset error type error correction word library. Each error type in the word library has clear feature words and examples. For example, the feature words of "formula error" include "formula written wrong", "parameter error", "symbol error", etc., and the examples are "writing a² as a" and "the parameter order of the vertex formula is reversed". When matching, the platform will calculate the similarity between the semantic recognition result and the features of each error type in the word library. The type with the highest similarity and exceeding the preset threshold (such as 0.7) is the error type corresponding to the error correction conversation content. For example, after semantic recognition, the core of a certain error correction conversation content is "the parameter order of the quadratic function vertex formula is reversed", and the feature similarity with "formula error" in the word library is 0.85, exceeding the threshold, so it is determined that its error type is formula error.

[0041] S107. Determine the basic score of each such error type; This step needs to quantify the severity of different error types through the statistics and analysis of error type-related data, providing a basis for determining the root error type later.

[0042] First, the platform needs to collect data on each error type in the dialogue content to be identified. It iterates through all the dialogue content, extracts each error type and its corresponding associated user identifier, and organizes this information into a structured data table for subsequent statistical analysis. For example, in the dialogue, if user "stu001" points out a formula error, and user "stu002" points out both a formula error and a principle error, the error type data table will record the user identifiers corresponding to the formula error as "stu001" and "stu002," and the user identifier corresponding to the principle error as "stu002." Next, the platform calculates the cumulative occurrences of each error type. For each error type, the platform counts the total number of times it appears in the error type data. For example, if a formula error appears 3 times and a principle error appears 2 times, then their cumulative occurrences are 3 and 2 respectively. Then, the number of unique users is determined based on each error type and its associated user identifier. For each error type, the platform deduplicates its associated user identifiers and counts the number of unique user identifiers. This number is the number of unique users for that error type. For example, if the associated user identifiers for a formula error are "stu001", "stu002", and "stu001", after deduplication, we get "stu001" and "stu002", so the number of unique users for that error type is 2. Next, the severity coefficient of each error type is determined based on the cumulative occurrence count of each error type and the corresponding number of unique users. The platform will use a preset calculation formula to calculate it, such as severity coefficient = (cumulative occurrence count / total cumulative occurrence count) × 0.5 + (number of unique users / total number of unique users) × 0.5, where the total cumulative occurrence count is the sum of the cumulative occurrence counts of all error types, and the total number of unique users is the sum of the number of unique users for all error types.

[0043] Finally, the error types are sorted according to their severity coefficients, and a base score is determined based on the sorting results. The sorting is done from highest to lowest coefficient, with higher-ranked errors receiving higher base scores. For example, an error type with a severity coefficient of 0.5 ranks first and has a base score of 100; a coefficient of 0.4 ranks second and has a base score of 80, and so on. Specific score values ​​can be adjusted based on actual circumstances.

[0044] S108. Match each error type with the interaction barrier feedback type to obtain a matching score, and determine the root error type based on the base score and the matching score; Here, "Error Type" refers to the problem category identified in step S106 of the large language model, such as formula errors, principle errors, and expression errors. "Interaction Barrier Feedback Type" refers to the specific type of interaction barrier identified in step S102, such as formula confusion, principle misunderstanding, and application barrier. "Matching Score" is a quantitative score indicating the degree of match between the error type and the interaction barrier feedback type; a higher score indicates a stronger correlation. For example, a formula error might have a matching score of 90 for formula confusion and 30 for principle misunderstanding. "Root Error Type" refers to the most significant error type identified as causing the target student's interaction barrier after considering both the basic score and the matching score; this type requires priority correction.

[0045] After determining the base score for each error type, it's necessary to combine the correlation between error types and interaction obstacle feedback types to accurately pinpoint the root cause of the target student's interaction obstacle from multiple error types. The intelligent computing platform first establishes a matching rule base between error types and interaction obstacle feedback types. This rule base predefines the correlation between different error types and interaction obstacle feedback types, along with their corresponding matching scores. These rules are developed based on extensive historical interaction data and the experience of educational experts, and are regularly updated and optimized. Next, the platform will match each error type with the interaction obstacle feedback type one by one, and calculate the matching score according to the rules in the rule base. For example, if the target student's interaction obstacle feedback type is "formula confusion type", and the error types include "formula error", "principle error", and "expression error", then the rule base will be queried to obtain the following matching scores: formula error matches formula confusion type with 90 points, principle error matches formula confusion type with 30 points, and expression error matches formula confusion type with 60 points.

[0046] The platform then calculates a comprehensive score for each error type. The formula for calculating the comprehensive score is typically: Comprehensive Score = Base Score × 0.6 + Matching Score × 0.4 (the weights can be adjusted according to the actual scenario). For example, if the base score for a formula error is 100 and the matching score is 90, then its comprehensive score = 100 × 0.6 + 90 × 0.4 = 60 + 36 = 96 points; if the base score for a principle error is 80 and the matching score is 30, then the comprehensive score = 80 × 0.6 + 30 × 0.4 = 48 + 12 = 60 points; and if the base score for a description error is 70 and the matching score is 60, then the comprehensive score = 70 × 0.6 + 60 × 0.4 = 42 + 24 = 66 points. Finally, the platform sorts the overall scores of each error type from highest to lowest, and the error type with the highest overall score is the fundamental error type. For example, in the example above, the formula error has the highest overall score (96 points), so it is determined to be the fundamental error type. If there are cases with the same overall score, the error type with the higher matching score is selected first, because it is more closely related to the interaction obstacle feedback type of the target student.

[0047] S109. Modify the code according to the fundamental error type, generate the response content, and provide feedback to the target student.

[0048] In this context, "modification" refers to correcting and optimizing the output of the large language model to address fundamental error types, such as correcting erroneous formulas or supplementing key information about the underlying principles. "Response content" refers to the textual information generated after modification to address the target student's concerns, and it must accurately and clearly resolve the student's confusion. After identifying the root cause of the error, it's necessary to correct it, generating accurate and easily understandable responses to resolve the student's interaction barriers and restore normal interaction. This involves calling the corresponding correction strategy library based on the root cause of the error. The correction strategy library contains specific correction methods and templates for different error types. For example, the correction strategy for "formula errors" is to "first clearly point out the error in the formula, provide the correct formula, and then explain the application of the formula using examples"; the correction strategy for "principle errors" is to "first explain the correct content of the principle, compare the differences with the incorrect statements, and then verify the correctness of the principle through examples." Next, the platform will generate initial response content based on the correction strategy. For example, if the fundamental error type is "formula error" (specifically, the quadratic function vertex formula is written incorrectly), the initial response content might be: "We are very sorry, the quadratic function vertex formula given earlier was incorrect. The correct vertex formula is y=a(xh)²+k (a≠0), where (h,k) represents the vertex coordinates of the parabola. For example, when the vertex of the parabola is (2,3) and a=1, the function expression is y=(x-2)²+3, which expands to y=x²-4x+7." The platform then optimizes the initial responses. On one hand, it adjusts the wording based on the type of interaction barriers encountered by the target students. For example, for students struggling with formulas, it adds derivation steps and memorization techniques; for students misunderstanding principles, it strengthens the connection between principles and real-life situations, explaining them in more accessible language. On the other hand, it checks the accuracy, completeness, and comprehensibility of the responses, ensuring no new errors, highlighting key information (such as formulas and key principles) (e.g., bolding, numbering), and presenting clear logic and steps.

[0049] Finally, the platform sends the optimized response to the target student's large language model application (such as the student's computer, tablet, or mobile client) and records the response time and content for subsequent follow-up on student feedback. Simultaneously, the platform monitors the student's interaction after receiving the response to determine if any interaction barriers remain. If barriers persist, the above steps may need to be repeated for further analysis and correction.

[0050] Existing large language models typically have the following limitations: Fixed training mode: Large language models are generally pre-trained on large-scale data before deployment, rather than continuously learning in real-time interaction. Lack of memory mechanism: Unless specifically designed, most model deployment methods do not save and analyze user correction information as a source of new knowledge. Privacy and security considerations: Automatic learning of user input may introduce erroneous information or inappropriate content, requiring strict review mechanisms.

[0051] This application constructs a closed-loop mechanism of "obstacle identification - root cause localization - precise correction" through full-process interactive analysis. This mechanism accurately captures interactive obstacles caused by errors in the output of the large language model and, combined with group interaction data, pinpoints the root cause error type. Therefore, it can specifically correct the model's output, effectively solving the problems of delayed response to student interactive obstacles, vague error localization, and lack of targeted correction in traditional intelligent computing platforms. This achieves an intelligent upgrade of the interaction between the large language model and students, improving the accuracy and efficiency of learning assistance and ensuring students' learning experience and effectiveness. The advantage of this method is that it can utilize collective knowledge to improve the quality of responses without waiting for the model itself to be updated or retrained, and it does not rely on the model's own real-time learning capabilities, making it particularly suitable for educational scenarios requiring precise knowledge. Based on the above, the following is a more detailed description of the process provided in this implementation. Please refer to [link / reference]. Figure 2 This is another flowchart illustrating the method for locating the root cause of anomalies in an intelligent computing platform that integrates a large language model, as described in this application.

[0052] S201. Based on the semantic expression features, determine the current topic boundaries between the target student and the large language model; Among them, multiple topic boundaries refer to the dividing points between different topics during the interaction process, used to distinguish different discussion topics. For example, when switching from the topic of "the graph of a quadratic function" to the topic of "the properties of a quadratic function", there is a topic boundary between the two.

[0053] After acquiring the semantic expression features of the target student's interaction with the large language model, it is necessary to clarify the multiple topics involved in the interaction and their scope, laying the foundation for subsequent analysis of the student's interaction under specific topics. The intelligent computing platform first performs real-time data collection and sequence generation. The platform acquires the text input by the target student on the client and the response text of the large language model in real time through the interface. Each piece of content is automatically added with a corresponding timestamp, and then these texts are arranged into a dialogue sequence according to the timestamps from earliest to latest. For example, if the student sends "S1: What is a quadratic function?" at 10:30:00, the model replies "M1: A quadratic function is a function of the form y=ax²+bx+c (a≠0)" at 10:30:10, and the student sends "S2: What are the characteristics of its graph?" at 10:30:20, the model replies "M2: The graph is a parabola...", and these contents are arranged into a dialogue sequence according to the timestamps. Meanwhile, the platform marks the "S1, S2" input by students and the "M1, M2" responses from the model as units to be analyzed.

[0054] Next, semantic vectors are generated. The platform uses a pre-trained language model (such as BERT) to process each unit to be analyzed, converting the text into a fixed-dimensional semantic vector. For example, "S1: What is a quadratic function?" is converted into a 768-dimensional numerical vector containing the semantic information of this sentence. The same method is used to generate semantic vectors for both student input units and model response units to ensure the comparability of the vectors. Then, real-time similarity values ​​are calculated. Based on the semantic vector of the current student input text unit (such as S3), the platform determines a pre-defined window. Assuming the pre-defined window is the first three student input units, if the current input is S3, then the window is S1 and S2. Then, the average of the semantic vectors of all student input units within the window is calculated to obtain the pre-defined topic semantic vector. The cosine similarity formula is then used to calculate the similarity between the semantic vector of S3 and this average vector to obtain the real-time similarity value. For example, the cosine similarity between the semantic vector of S3 and the average vectors of S1 and S2 is 0.3.

[0055] Next, a new topic is determined. The platform compares the real-time similarity value with a preset similarity threshold (e.g., 0.5). If the value is lower than the threshold (0.3 < 0.5), the platform enters the continuous detection phase. The platform will check subsequent student input text units (e.g., S4, S5). If the real-time similarity value of S4 is 0.2 and that of S5 is 0.3, both lower than the threshold, and the number of consecutive similarity units reaches the preset three (S3, S4, S5), then a new topic is preliminarily determined to have been triggered. Finally, the topic boundaries are defined. The platform marks the first student input text unit (e.g., S3) that triggers the new topic determination as the starting point of the new topic, and records the previous unit to be analyzed (e.g., M2) as the ending point of the preceding topic, thus forming the topic boundaries. For example, the scope of the preceding topic is from the first text unit to M2, and the new topic starts from S3, clearly defining the boundaries between the two topics.

[0056] S202. Generate a unique topic identifier for each topic by combining the topic boundary, and record the total number of interaction rounds of the target student on the target topic; Here, a unique topic identifier is a unique identifier assigned to each individual topic to accurately distinguish different topics; it can use a combination of letters and numbers. The target topic refers to the topic the target student is currently interacting with. The total number of interaction rounds refers to the number of complete interactions the target student and the large language model complete within the scope of the target topic.

[0057] First, the platform performs topic segmentation based on the established topic boundaries. It traverses the entire dialogue sequence, dividing the continuous interactive text into multiple independent topic fragments according to the topic boundaries. Each fragment contains all the units to be analyzed from the topic's start point to its end point. Next, a unique identifier is generated for each topic. The platform can use a composite encoding rule of "subject classification + timestamp + sequence number" to ensure the uniqueness and readability of the identifier. For example, the first topic generated in the mathematics subject at 10:30 AM on August 16, 2025, is identified as "MATH_202508161030_001". After generation, the identifier is associated with and stored along with metadata such as the topic fragment's start point, end point, and core keywords, forming a topic index library that supports quick retrieval of the complete interactive content corresponding to a specific identifier. Finally, the target topic is determined. The platform monitors the student's interaction status in real time and identifies the topic the student is currently interacting with as the target topic.

[0058] Finally, the total number of interaction rounds under the target topic is counted. The platform traverses the units to be analyzed starting from the topic's starting point. Whenever a complete combination of "student input unit + model response unit" is detected, the counter is incremented by 1. If the model only responds to M2 after the student sends multiple inputs consecutively (such as S2-1, S2-2), it is considered as one interaction round; if the student does not ask any further questions after the model responds, that response unit is not counted in the round.

[0059] S203. Determine the number of times the same or similar questions appear in the target topic based on the total number of interaction rounds; Among them, "identical questions" refers to questions with completely identical text content asked by target students in different interaction rounds. For example, asking "What is the quadratic formula?" in both round 2 and round 5. "Similar questions" refers to questions with different textual expressions but the same core meaning. For example, "How to solve equations using the quadratic formula?" and "What are the steps of the quadratic formula?", both of which revolve around the application of the quadratic formula. "Number of occurrences" refers to the cumulative number of times identical and similar questions appear in the total number of interaction rounds of the target topic.

[0060] After obtaining the total number of interaction rounds for the target topic, it is necessary to analyze the question repetition pattern to determine whether the student is repeatedly asking questions in the current topic due to not understanding the model's response or because there are logical errors in the large language model. This provides a basis for understanding the student's learning status.

[0061] S204. If the number of occurrences exceeds the preset threshold, it is determined that the target student is suspected of having interactive obstacles in the target topic. Among them, the feedback of suspected interaction barriers refers to the preliminary judgment result based on the number of times the question occurs exceeding the threshold, indicating that students may have communication barriers due to unclear model responses, misunderstandings of their own understanding, or other reasons.

[0062] The intelligent computing platform first obtains the occurrence count and the preset threshold. It retrieves the total occurrence count (e.g., 6 times) from step S203, and simultaneously retrieves the corresponding preset threshold (e.g., 4 times) from the threshold configuration library based on the subject type (e.g., mathematics) and student grade (e.g., ninth grade). The threshold configuration library uses a dynamic update mechanism, periodically optimized based on historical interaction data. For example, if it is found that the same question occurs ≤4 times in 80% of normal interactions under a certain topic, the threshold for that topic is set to 4 times.

[0063] The platform compares the number of occurrences with a preset threshold: if the number of occurrences is less than or equal to the threshold (e.g., 6 times > 4 times is not valid), the student's questioning pattern on the target topic is considered normal, the question repetition does not reach the abnormal level, and the "no obvious abnormal feedback" result is recorded before ending this step; if the number of occurrences is greater than the threshold (e.g., 6 times > 4 times is valid), the suspected judgment process is initiated.

[0064] In the suspected question determination process, the platform will use the interaction characteristics of the target topic for auxiliary verification: First, it checks the time distribution of repeated questions. If the same or similar questions are concentrated in a short period of time (e.g., appearing 4 times within 10 minutes), the suspected question determination is strengthened. Second, it analyzes the relevance of the model's responses. If the similarity of the model's responses to repeated questions is ≥0.8 (indicating that no targeted optimization has been performed), the suspected question determination is further supported. For example, if a student asks similar questions 4 times in a row within 5 minutes, and the model's response similarity reaches 0.9, the platform will increase the confidence level of the suspected question determination. Ultimately, the platform generates a judgment result of "the target student is suspected of having interaction barriers under the target topic", which includes the number of occurrences (6 times), the threshold (4 times), and examples of problem groups (such as "problems related to the quadratic formula"), stores it in the interaction barrier log, and triggers subsequent steps for further verification.

[0065] S205. Determine the student interaction temperature and tone based on the semantic expression features of students in the target topic and the preset interaction temperature and tone model. The interaction temperature and tone model is trained by machine learning from student interaction data labeled with warm tones with non-negative emotions and student interaction data labeled with cold tones with negative emotions. Among them, student interaction data labeled with warm tones (non-negative emotions) refers to student interaction text data marked with warm tones that contain positive, peaceful, and other non-negative emotions, such as "I understand, thank you" or "This explanation is very clear." Student interaction data labeled with cool tones (negative emotions) refers to student interaction text data marked with cool tones that contain negative emotions such as confusion, irritability, and dissatisfaction, such as "I still don't understand" or "What is this about?" Student interaction temperature tone refers to the emotional tendency classification obtained after analyzing the semantic expression characteristics of students through the interaction temperature tone model, which is divided into warm and cool tones. Warm tones indicate that students' emotions are positive or peaceful during interaction, while cool tones indicate that students' emotions are negative during interaction.

[0066] Because the number of questions exceeds a preset threshold, it's possible that the target student is interested in a particular topic and therefore asks more questions. Therefore, after determining that the target student may be experiencing interaction barriers on the target topic, it's necessary to further analyze the student's emotional characteristics to assess their interaction status and provide a basis for confirming whether genuine interaction barriers exist. First, the platform extracts the student's semantic expression characteristics from the target topic. It comprehensively analyzes all input text from the target student on the target topic, extracting lexical features such as the presence of emotional words like "happy," "confused," and "angry"; sentence structure features such as the use of rhetorical questions and exclamatory sentences to express strong emotions; and the overall semantic tendency of the text, such as whether it expresses affirmation, gratitude, doubt, or dissatisfaction.

[0067] Next, the platform invokes a pre-defined interaction temperature and tone model. This model is trained on a large amount of labeled data. During training, the input data consists of the text content of the student's interaction with the large language model and the semantic expression features extracted from it. These features include lexical vectors (numerical vectors converted from text words using tools), sentence structure features (such as binary identifiers for whether a sentence is exclamatory or rhetorical), and the frequency of emotion words (such as the percentage of times emotion words like "happy" and "confused" appear in the text). The output data is the corresponding temperature and tone category, i.e., warm or cool tone. The labeled data used during training is divided into two categories: one is data labeled with warm tones, which contains text in which students express positive, peaceful, understanding, and other non-negative emotions during the interaction; the other is data labeled with cool tones, which contains text in which students express negative emotions such as confusion, irritability, and dissatisfaction during the interaction. During training, the labeled data is first divided into training and validation sets in an 8:2 ratio. The training set is then used to train basic machine learning models (such as support vector machines and neural networks). For support vector machine models, kernel functions are used to map the high-dimensional semantic features of the input to a higher-dimensional space to find the optimal classification hyperplane that maximizes the margin between warm and cool-toned data, thereby learning the feature patterns that distinguish the two classes of data. For neural network models, forward propagation is used to calculate the loss value between the output result and the labeled class (such as cross-entropy loss), and then backpropagation is used to adjust the weights and biases of neurons in each layer to continuously reduce the loss value until the model's classification accuracy on the validation set reaches a preset threshold (such as 90%).

[0068] The platform then inputs the extracted student semantic expression features into the interaction temperature color tone model. The model analyzes and calculates the input features based on its learned patterns. For example, when the input features contain many negative words such as "don't understand" and "confused," and use sentence structures expressing strong emotions such as exclamations, the model will identify that these features match the feature patterns of the cool-toned labeled data. Finally, the model outputs the student interaction temperature color tone result, i.e., a warm or cool tone. If the student's semantic expression features do not show obvious negative emotional characteristics, but rather reflect more positive or peaceful emotions, the model will output a warm tone; if the student's semantic expression features show obvious negative emotional characteristics, the model will output a cool tone.

[0069] S206. If the student's interaction temperature color is a cool color, then it is determined that there is an interaction obstacle feedback. The platform will assess the temperature and color tone results. If the result is a cool color tone, it indicates that the student experienced negative emotions during the interaction. These negative emotions are likely due to obstacles encountered in interacting with the large language model, such as not understanding the model's responses or the model's responses failing to solve their problem. Therefore, the platform will determine that the target student has feedback indicating an interaction obstacle. For example, if a student's interaction temperature and color tone is cool, it suggests that they may have experienced confusion and frustration due to not understanding responses related to "solving quadratic equations," thus the platform will determine that there is feedback indicating an interaction obstacle. If the student's interaction temperature is warm, it indicates that the student is in a positive or peaceful mood during the interaction and has not shown any negative emotions due to interaction barriers. Therefore, the platform will determine that there is no interaction barrier feedback. After determining that interaction barrier feedback exists, the platform will record this determination and associate it with information such as the target student and target topic to provide a basis for subsequent steps.

[0070] S207. Determine the level of interaction barrier based on the number of occurrences.

[0071] The frequency of occurrence refers to the total number of times the same or similar questions appear in the target topic identified in step S203. The feedback on the degree of interaction obstacle refers to a grading result indicating the severity of obstacles encountered by the target student during the interaction process, such as mild, moderate, or severe. Different grades reflect the magnitude of the interaction difficulties encountered by the student.

[0072] The platform will invoke preset rules that correlate the frequency of occurrence with the degree of interaction impairment. These rules are developed based on a large amount of historical interaction data and the experience of education experts, with different ranges of occurrence corresponding to different degrees of interaction impairment. For example, the rules may define an occurrence of 3-5 times as mild interaction impairment, 5-8 times as moderate interaction impairment, and more than 8 times as severe interaction impairment. Next, the platform compares the frequency of occurrence with these rules to determine the corresponding level of interaction barrier. For example, if the frequency is 7 times, falling within the range of 5-8 times, the corresponding level of interaction barrier is moderate. After determining the level of interaction barrier, the platform generates interaction barrier level feedback, which includes the frequency of occurrence, the corresponding level of barrier, and a brief explanation, such as "The same or similar questions appeared 7 times, the level of interaction barrier is moderate, and students may have some difficulty understanding the content related to this topic."

[0073] In this embodiment, by defining the scope of interaction by delineating topic boundaries, initially identifying suspected obstacles by statistically analyzing the frequency of identical or similar questions, and accurately determining the interaction obstacle and its degree by combining the interaction temperature tone model and the frequency of occurrence, a full-chain interaction obstacle identification mechanism of "topic focusing - suspected judgment - emotion verification - degree quantification" is constructed. This mechanism can capture students' interaction anomalies from multiple dimensions such as topic dimension, question pattern, and emotional characteristics. Therefore, it can achieve accurate identification and degree assessment of interaction obstacles, effectively solving the problems of traditional intelligent computing platforms that judge interaction obstacles by only a single indicator, have low identification accuracy, and cannot quantify the degree of obstacles. This enables refined perception of the interaction state between students and the large language model, providing a reliable basis for subsequent root cause localization and accurate response. The intelligent computing platform in the embodiments of this invention is described below from the perspective of hardware processing. Please refer to [link / reference]. Figure 3 This is a schematic diagram of the physical device structure of an intelligent computing platform in an embodiment of this application.

[0074] It should be noted that, Figure 3 The structure of the intelligent computing platform shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0075] like Figure 3 As shown, the intelligent computing platform includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) 302 or programs loaded from storage portion 308 into Random Access Memory (RAM) 303, such as performing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0076] The following components are connected to I / O interface 305: input section 306 including audio input devices, push-button switches, etc.; output section 307 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 308 including a hard disk, etc.; and communication section 309 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0077] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the various functions defined in the present invention.

[0078] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0079] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0080] Specifically, the intelligent computing platform in this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the intelligent computing platform anomaly root cause localization method that integrates a large language model provided in the above embodiment.

[0081] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the intelligent computing platform described in the above embodiments; or it may exist independently and not assembled into the intelligent computing platform. The storage medium carries one or more computer programs, which, when executed by a processor of the intelligent computing platform, enable the intelligent computing platform to implement the intelligent computing platform anomaly root cause localization method based on a fused large language model provided in the above embodiments.

[0082] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0083] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0084] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for locating the root cause of anomalies in an intelligent computing platform by integrating a large language model, applied to an intelligent computing platform, characterized in that, The method includes: Obtain semantic expression features during the interaction between the target student and the large language model, and determine whether the interaction barrier level feedback of the target student exceeds the preset barrier level based on the semantic expression features; If the number exceeds the limit, then the interaction barrier feedback type of the target student will be determined. Find the dialogue content within a preset length window before the target student experiences interactive obstacle feedback, and locate the starting point of the interactive obstacle feedback; Obtain the interactive obstacle content from the starting point to the time when the target student provides feedback on the interactive obstacle, and extract the key semantics of the interactive obstacle content; Based on the key semantics, find the dialogue content to be identified that matches the key semantics of the second user. The second user is the target learning group and users using the large language model in a set area. The target learning group includes the network group of classmates using the large language model. Semantic recognition is performed on the dialogue content to be identified to determine multiple error types; Determine the base score for each of the aforementioned error types; Each error type is matched with the interaction barrier feedback type to obtain a matching score, and the root error type is determined based on the base score and the matching score. Modify the code based on the identified fundamental error type, generate a response, and send it back to the target student.

2. The method according to claim 1, characterized in that, The step of acquiring semantic expression features during the interaction between the target student and the large language model, and determining whether the target student's interaction barrier level feedback exceeds a preset barrier level based on the semantic expression features, specifically includes: Based on the semantic expression features, determine the current topic boundaries between the target student and the large language model; A unique topic identifier is generated for each topic based on the topic boundaries, and the total number of interaction rounds of the target student on the target topic is recorded. The number of times the same or similar questions appear in the target topic is determined based on the total number of interaction rounds. If the number of occurrences exceeds a preset threshold, it is determined that the target student is suspected of having interaction barriers in the target topic. The student interaction temperature color is determined based on the semantic expression features of students in the target topic and the preset interaction temperature color model. The interaction temperature color model is trained by machine learning from student interaction data labeled with warm colors with non-negative emotions and student interaction data labeled with cold colors with negative emotions. If the student interaction temperature color tone is a cool color tone, then it is determined that there is feedback of interaction obstacle; The degree of interaction barrier is determined based on the number of occurrences.

3. The method according to claim 2, characterized in that, The steps for determining the current topic boundaries between the target student and the large language model based on the semantic expression features specifically include: Real-time collection of continuous interactive text between the target student and the large language model; generation of dialogue sequences according to timestamp order; marking the student input and the large model's response as units to be analyzed. Generate the corresponding semantic vector based on the unit to be analyzed; Based on the semantic vector of the current student input text unit, cosine similarity is calculated with the topic semantic vector in the previous preset window to obtain the real-time similarity value; If the real-time similarity value is lower than the preset similarity threshold, and multiple consecutive text units input by students meet the conditions, it is initially determined that a new topic has been triggered. The first student input text unit that triggers the new topic determination is marked as the starting point of the new topic, and the ending point of the previous topic is recorded to form the topic boundary division.

4. The method according to claim 1, characterized in that, The step of finding the dialogue content within a preset length window before the target student experiences interactive obstacle feedback, and locating the starting point of the interactive obstacle feedback, specifically includes: Use natural language processing technology to extract feedback content from the dialogue within a window of preset length; The feedback content is input into the interaction barrier recognition learning model to determine whether there is interaction barrier feedback. The machine learning model is trained on language pattern annotations with different interaction barriers. If an interaction obstacle feedback is detected, the corresponding time point is immediately marked as the starting point of the interaction obstacle feedback.

5. The method according to claim 1, characterized in that, The step of performing semantic recognition on the dialogue content to be identified and determining multiple error types specifically includes: Obtain the error correction dialogue content of the second user to the large language model containing the user's error correction intention from the dialogue content to be identified; After semantic recognition of the error correction dialogue content, it is matched with a preset error type correction lexicon to determine the error type.

6. The method according to claim 1, characterized in that, The step of determining the base score for each of the error types specifically includes: Collect data on each error type in the dialogue content to be identified. The error type data includes the error type and the corresponding associated user identifier. Count the cumulative number of occurrences for each error type; The number of unique users is determined based on each error type and the associated user identifier, where the number of unique users is the total number of non-repeating users who have experienced type errors. The severity coefficient of each error type is determined based on the cumulative number of occurrences of each error type and the corresponding number of unique users. The error types are sorted according to their severity coefficients, and a base score for each error type is determined based on the sorting results.

7. The method according to claim 1, characterized in that, After the step of finding the dialogue content to be identified by the second user that matches the key semantics, the method further includes: Send a help request message to the second user's large language model terminal, the help request message including key semantics in the content of the interaction obstacle; Upon receiving feedback from the user of the large language model, the output of the large language model is adjusted according to the feedback. The adjusted output is sent to the target student's large language model client for display.

8. An intelligent computing platform, characterized in that, The server includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the server to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are executed on the server, the server causes the server to perform the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the server, the server performs the method as described in any one of claims 1-7.

Images