Intelligent question and answer management system and method based on large model Chinese

Abstract

The invention discloses an intelligent Chinese question and answer management system and method based on a large model. The method comprises the steps that real-time Chinese question data of a user is obtained, a Chinese text is partitioned through recursive character segmentation, a knowledge base is constructed in combination with automatic information extraction, the knowledge base is updated in real time through incremental learning, and a target knowledge base is obtained; acquiring historical dialogue data, and identifying a user intention through a model to generate an initial reply strategy; converting the real-time Chinese question data into vector representation by using a pre-trained RoBERTa semantic embedding model, and calculating a semantic similarity between a user question and a candidate answer through a GNN graph neural network in combination with an entity relationship of a knowledge graph in a target knowledge base; and generating a target reply strategy based on the initial reply strategy and the semantic similarity. And the question-answering accuracy and efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent question-and-answer management technology, and in particular to a large-model-based intelligent Chinese question-and-answer management system and method. Background Technology

[0002] Traditional intelligent question-answering management systems for Chinese language learning lack a deep understanding of Chinese semantics, relying mostly on keyword matching or simple semantic vector calculations. This makes it difficult to capture the complex semantic relationships between user questions and candidate answers. Especially when dealing with in-depth questions involving character relationships, plot development, and rhetorical devices in literary works, semantic biases often occur, resulting in low accuracy and inefficiency in answering questions. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a large-scale model-based intelligent question-and-answer management system and method for Chinese language.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following modules in the above-mentioned intelligent question-and-answer management system for Chinese language based on a large model: The knowledge graph construction module is used to acquire real-time Chinese language question data from users, divide Chinese text into blocks using recursive character segmentation, and build a knowledge base by combining automated information extraction. The knowledge base is updated in real time using incremental learning to obtain the target knowledge base. The user intent recognition module is used to acquire historical dialogue data and generate an initial response strategy by recognizing user intent through the Transformer model. The intelligent question-answering management module is used to convert the real-time Chinese question data into vector representations using a pre-trained RoBERTa semantic embedding model, and calculate the semantic similarity between user questions and candidate answers by combining the entity relationships in the knowledge graph of the target knowledge base and using a GNN graph neural network. The personalized adaptation module is used to generate a target response strategy based on the initial response strategy and the semantic similarity using a Transformer model.

[0005] Furthermore, in the aforementioned large-scale model-based intelligent question-answering management system for Chinese language, the knowledge graph construction module includes the following units: The invalid character filtering unit is used to obtain real-time Chinese question data from users, match and remove invalid characters using regular expressions, and then perform standardization processing to obtain standard question data. An abnormal data processing unit is used to split the standard question data into multiple sub-question fragments and complete them by combining them with the user's historical dialogue context to obtain the initial question data.

[0006] Furthermore, in the aforementioned large-model-based intelligent question-answering management system for Chinese language, the knowledge graph construction module also includes the following units: The text preliminary segmentation unit is used to initially segment the text of the initial query data according to an initial length to obtain text block data; The semantic integrity detection unit is used to call the pre-trained BERT model to calculate the co-occurrence frequency of core semantic words in the text block data, obtain the detected text block data, and combine the detected text block data with automated information extraction to construct a knowledge base. The timed update unit is used to start incremental learning tasks every 30 minutes, scanning newly added question data and processing results within 30 minutes; The knowledge base update unit is used to associate supplementary information of the same entity with the corresponding entity node; for new entities, new nodes are created and relationship edges with related entities are established to obtain the target knowledge base.

[0007] Furthermore, in the aforementioned large-model-based intelligent question-answering management system for Chinese language, the user intent recognition module includes the following units: The model parameter configuration unit is used to perform natural language processing tasks using the Transformer base model. The hidden layer dimension is set to 512, the number of attention heads is 8, and the number of encoder and decoder layers is 6. The semantic association computing unit is used to input the user's historical dialogue data into the trained Transformer model, use the model to encode the question content and extract features, and capture key information and semantic associations in the question through a multi-layer attention mechanism; The user intent recognition unit is used by the model to calculate the probability that the question belongs to each intent label based on the learned knowledge, and to determine the label with the highest probability as the user intent.

[0008] Furthermore, in the aforementioned large-model-based intelligent question-and-answer management system for Chinese language, the intelligent question-and-answer management module includes the following units: The question data input unit is used to input preprocessed real-time Chinese question data into the RoBERTa semantic embedding model; The semantic feature capture unit is used to process the segmented text through a multi-layer Transformer structure, with each layer capturing semantic features at different levels. The vector representation generation unit is used to generate a corresponding word vector for each word, and then combine all word vectors through aggregation to generate the vector representation of the entire real-time question data.

[0009] Furthermore, in the aforementioned large-model-based intelligent question-and-answer management system for Chinese language, the intelligent question-and-answer management module includes the following units: The graph neural network building unit is used to input the vector representation and transformed graph structure of real-time query data into the GNN graph neural network; The feature information calculation unit is used to update the feature representation of each node by propagating and aggregating information on the graph. During the propagation process, each node combines the feature information of its neighboring nodes to understand the semantic relationship between entities. The semantic similarity acquisition unit is used to calculate the degree of semantic association between real-time question data and each candidate answer in the knowledge graph, and obtain the semantic similarity.

[0010] Furthermore, in the aforementioned large-model-based intelligent question-answering management system for Chinese language, the personalized adaptation module includes the following units: The strategy generation unit is used by the decoder to gradually generate the target response strategy based on the encoded vector output by the encoder and its own language generation capabilities. The relation determination unit is used to model the causal, progressive, and parallel relationships between sentences, and to unfold the generated responses in a clear logical order.

[0011] In a large-model-based intelligent question-answering management method for Chinese language, the method includes the following steps: The system acquires real-time Chinese language questions from users, segments the Chinese text into blocks using recursive character segmentation, constructs a knowledge base by combining automated information extraction, and updates the knowledge base in real time using incremental learning to obtain the target knowledge base. Acquire historical dialogue data and use the Transformer model to identify user intent and generate an initial response strategy; The real-time Chinese question data is transformed into vector representation using a pre-trained RoBERTa semantic embedding model. Combined with the entity relationships in the knowledge graph of the target knowledge base, the semantic similarity between the user's question and the candidate answer is calculated using a GNN graph neural network. Based on the initial response strategy and the semantic similarity, a target response strategy is generated using the Transformer model.

[0012] Furthermore, in the aforementioned large-model-based intelligent question-answering management method for Chinese language, the step of using a pre-trained RoBERTa semantic embedding model to convert the real-time Chinese question data into vector representations, combining the entity relationships in the knowledge graph of the target knowledge base, and calculating the semantic similarity between the user's question and candidate answers through a GNN graph neural network includes: The preprocessed real-time Chinese question data is input into the RoBERTa semantic embedding model; The segmented text is processed through a multi-layered Transformer structure, with each layer capturing semantic features at different levels. For each word, a corresponding word vector is generated. Then, all word vectors are combined by aggregation to generate a vector representation of the entire real-time question data.

[0013] Furthermore, in the aforementioned large-model-based intelligent question-answering management method for Chinese language, the step of using a pre-trained RoBERTa semantic embedding model to convert the real-time Chinese question data into vector representations, combining the entity relationships in the knowledge graph of the target knowledge base, and calculating the semantic similarity between the user's question and candidate answers using a GNN graph neural network further includes: The vector representation of the real-time query data and the transformed graph structure are input into the GNN graph neural network; By propagating and aggregating information on the graph, the feature representation of each node is updated. During the propagation process, each node combines the feature information of its neighboring nodes to understand the semantic relationship between entities. The semantic similarity is obtained by calculating the degree of semantic association between real-time question data and each candidate answer in the knowledge graph.

[0014] Its beneficial effects are as follows: 1. It effectively improves the accuracy of semantic matching between user questions and candidate answers, especially when dealing with complex character relationships and plot logic in literary works. 2. It ensures that the generated target response strategies not only conform to the user's questioning intent but are also rich in content, logically coherent, and highly targeted, providing users with precise and effective guidance for Chinese language learning, thus enhancing their learning experience and effectiveness. 3. By analyzing user's historical dialogue data to identify user intent, it can generate personalized response strategies based on the different learning characteristics and needs of users, meeting their diverse and personalized needs in Chinese language learning. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of a large-scale model-based intelligent Chinese question-answering management system according to the present invention; Figure 2 This is a schematic diagram of a second embodiment of a large-model-based intelligent Chinese question-and-answer management system according to the present invention; Figure 3 This is a schematic diagram of an embodiment of a large-model-based intelligent question-and-answer management method for Chinese language. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings, such as... Figure 1 As shown, a large-scale model-based intelligent question-and-answer management system for Chinese language includes the following modules: The knowledge graph construction module is used to acquire real-time Chinese language question data from users, divide Chinese text into blocks using recursive character segmentation, and build a knowledge base by combining automated information extraction. The knowledge base is updated in real time using incremental learning to obtain the target knowledge base. Specifically, this embodiment also includes an invalid character filtering unit, which is used to obtain the user's real-time Chinese question data, match and remove invalid characters through regular expressions, and then perform standardization processing to obtain standard question data; The abnormal data processing unit is used to split the standard question data into multiple sub-question fragments and complete them by combining them with the user's historical dialogue context to obtain the initial question data.

[0020] The initial text segmentation unit is used to initially segment the text of the initial query data according to the initial length to obtain text block data; The semantic integrity detection unit is used to call the pre-trained BERT model to calculate the co-occurrence frequency of core semantic words in the text block data, obtain the detected text block data, and combine the detected text block data with automated information extraction to build a knowledge base. The timed update unit is used to start incremental learning tasks every 30 minutes, scanning newly added question data and processing results within 30 minutes; The knowledge base update unit is used to associate supplementary information of the same entity with the corresponding entity node; for new entities, new nodes are created and relationship edges with related entities are established to obtain the target knowledge base.

[0021] Specifically: I. Acquisition of real-time user questions in Chinese language; (a) Multi-channel data access; Text input channels: Users can directly type Chinese questions through the input box on the web page or the input interface of the mobile APP. The system captures the input field in real time, and temporarily stores it after each character is entered. After the user clicks the send button, the complete text data is sent to the processing module.

[0022] Voice-to-text channel: Integrates mainstream voice recognition interfaces (Baidu Voice, iFlytek Voice, etc.). When users enter voice questions through the microphone, the system converts the voice stream into text data in real time, while retaining the confidence parameter of voice recognition (value range 0-1). When the confidence is higher than 0.85, it is directly processed as valid data; when it is lower than 0.85, a secondary confirmation mechanism is triggered, prompting the user to check or re-enter.

[0023] To calculate the confidence level, first determine the confidence level, which represents the probability that the population parameter falls within the confidence interval. Common values ​​are 0.8, 0.9, etc. For example, a confidence level of 0.9 means there is a 95% certainty that the population parameter value will fall within the confidence interval. Next, find or calculate the Z-value or t-value. The Z-value is suitable for confidence intervals of a normal distribution, while the t-value is used for confidence intervals of non-normal distributions or small samples. These values ​​can be obtained by consulting the standard normal distribution table or the t-distribution table. For example, a Z-value of 1.96 corresponds to a confidence level of 0.95. Then, calculate the standard error, which is the standard deviation of the sample mean, used to measure the accuracy of the sample mean's estimation of the population mean. The formula is the sample standard deviation divided by the square root of the sample size. Finally, calculate the confidence interval using the formula: "Confidence Interval = Sample Mean ± Z-value or t-value * Standard Error". For example, if the sample mean is 50, the sample standard deviation is 10, the sample size is 100, and a confidence level of 0.95 is chosen, then the confidence interval is 50 ± 1.96. * 10 / √100, which is 44.04 to 55.96; Finally, the confidence level is the selected confidence level, such as 0.9, 0.99, etc., which means that in repeated sampling, approximately a certain proportion of the confidence intervals will contain the true population parameter.

[0024] (ii) Data preprocessing workflow; Invalid character filtering: Match and remove the following using regular expressions: consecutive repeated spaces (keeping single spaces), non-Chinese and uncommon punctuation marks, and meaningless strings longer than 5 characters (random characters such as "asdfghj").

[0025] Standardize the format: convert full-width characters to half-width characters, unify punctuation styles, and standardize the case of English characters.

[0026] Abnormal data handling: For excessively long questions exceeding 1000 characters, they are automatically split into multiple sub-question segments (split by paragraph separators); for excessively short questions less than 5 characters ("What does it mean?" "Explain"), they are completed by combining the user's historical dialogue context. If the context information is insufficient, guiding prompts are generated.

[0027] II. Recursive character segmentation for Chinese text partitioning; (a) Recursive segmentation of execution logic; First round of segmentation: The preprocessed text is initially segmented according to its initial length to obtain several text blocks, and each text block is marked with the start and end character positions.

[0028] Semantic integrity detection: A lightweight semantic analysis tool (a simplified version of the pre-trained Chinese BERT model) is called to calculate the co-occurrence frequency of core semantic words (nouns and verbs) within the text block. If the co-occurrence frequency is lower than 0.5 (i.e., the connection between core words is weak), it is judged as semantic incomplete.

[0029] Co-occurrence frequency (COF) is a key indicator for measuring the degree of association between core semantic words (nouns and verbs) in a text. Its calculation is based on the co-occurrence of word pairs within the text. Specifically, firstly, all nouns and verbs are extracted from the text as a core word set. Then, the number of all possible core word pairs is counted (i.e., the number of combinations C(n,2) of randomly selecting two words from n words). Next, using a sliding window (e.g., a window of 3 words) or grammatical dependencies, the number of times each core word pair co-occurs in the text is counted. To eliminate the influence of text length and the total number of word pairs, the COF of each word pair is divided by a normalization factor (e.g., the number of possible co-occurrence opportunities within the window N-1, or directly using the total number of word pairs) to obtain the local COF of that word pair. Finally, the average of the local frequencies of all core word pairs is calculated to obtain the overall COF of the text. If this value is lower than a preset threshold (e.g., 0.5), it indicates a weak association between core words, and the text semantics may be incomplete. This method provides an objective and calculable basis for semantic integrity detection by quantifying the co-occurrence relationship between words.

[0030] Dynamically adjust segmentation: For semantically incomplete text blocks, if the length exceeds 120% of the initial value, the length is reduced by 20% and re-segmented; if the length is less than 80% of the initial value, the length is expanded by 20% and adjacent fragments are merged, and the detection is repeated until the semantic integrity reaches 0.7 or higher (co-occurrence frequency of core words ≥ 0.7).

[0031] Semantic completeness is an important indicator for evaluating the semantic coherence and sufficiency of core information in a text block. Its calculation is based on the co-occurrence frequency of core words, combined with a dynamic adjustment mechanism to achieve precise quantification. The specific calculation logic is as follows: First, based on a pre-trained Chinese semantic analysis tool, all nouns and verbs are extracted from the text block as a core word set. Using a sliding window or grammatical dependencies, the co-occurrence frequency of each core word pair in the text is counted, and its co-occurrence frequency—that is, the average co-occurrence frequency of all core word pairs—is calculated. For example, if there are 5 core words in the text, forming 10 word pairs, and 6 of these pairs have a co-occurrence relationship, then the basic co-occurrence frequency is 6 / 10 = 0.6.

[0032] To better align with semantic integrity requirements, the basic co-occurrence frequency is normalized into a semantic integrity score: when the co-occurrence frequency is ≥0.7, semantic integrity is directly determined, and the score is recorded as 1.0 (full marks); when the co-occurrence frequency is <0.7, the score is equal to the actual co-occurrence frequency value. For example, when the co-occurrence frequency is 0.65, the semantic integrity score is 0.65, indicating that there is a slight semantic break in the text.

[0033] For text blocks with a score below 0.7, the system initiates a dynamic adjustment mechanism: if the text length exceeds 120% of the initial value, it is reduced by 20%, prioritizing the deletion of non-core words or low-relevance segments; if the length is less than 80% of the initial value, it is expanded by 20%, merging adjacent semantically related fragments. After adjustment, the semantic completeness is recalculated until the score is ≥0.7 or the maximum number of iterations is reached. This process, through quantified semantic association and dynamic boundary optimization, ensures that text blocks retain both core information and logical coherence.

[0034] Boundary handling rules: When a text block is close to the beginning or end of the original text, if the remaining characters are less than 50% of the adjusted length, it will be directly merged into the adjacent text block to avoid excessively short fragments.

[0035] III. Automated information extraction to build an initial knowledge base; (a) Entity recognition and classification; Core entity type definition: Clearly define the entity categories to be extracted, including: Literary subjects: writers (Lu Xun, Li Bai), works (Dream of the Red Chamber, Climbing High), literary schools (realism, Misty Poetry), literary characters (Ah Q, Lin Daiyu).

[0036] Language elements: words (idioms, classical Chinese vocabulary), sentence structure (parallel sentences, passive sentences), and rhetorical devices (metaphor, hyperbole).

[0037] Cultural knowledge: historical background (the Tang Dynasty, the New Culture Movement), literary awards (Mao Dun Literature Prize, Nobel Prize in Literature).

[0038] Entity recognition tool: A Chinese NER model based on the RoBERTa-wwm-ext pre-trained model with fine-tuning is used. Trained on labeled corpora, the model achieves an entity recognition accuracy of over 92% on the test set. During the recognition process, ambiguous entities ("Dongpo" could refer to Su Shi or a place name) are disambiguated by combining the context, prioritizing matching common interpretations in the literary field.

[0039] (ii) Relation extraction and storage; Relationship Type System: This system outlines the core relationships within the field of Chinese language knowledge. In addition to the previously listed subordinate, causal, explanatory, and associative relationships, it supplements the following: Creative relationship: such as "Du Fu wrote 'Spring View'".

[0040] Inclusion relationship: such as "The Book of Songs contains 'Feng', 'Ya', and 'Song'".

[0041] Comparative relationship: such as "There are differences in the metrical rules between classical poetry and modern poetry".

[0042] Relation extraction process: The TPLinker-based relation extraction model is used to predict entity pair relations in the segmented text blocks. For implicit relations that are not explicitly stated, supplementary extraction is performed using a rule base (containing 200+ Chinese language knowledge reasoning rules).

[0043] Knowledge base storage structure: The graph database (Neo4j) is used to store triple data. Each entity node contains a unique ID, name, type, and attributes (work nodes contain attributes such as creation year and genre). Relationship edges contain information such as relationship type, confidence (predicted probability given by the extraction model), and source text block ID, which facilitates traceability and updates.

[0044] IV. Incremental learning updates the knowledge base in real time; (a) Update trigger mechanism; Scheduled updates: Incremental learning tasks are initiated every 30 minutes to scan for newly added questions and processing results within the last 30 minutes.

[0045] (ii) Information screening criteria; Duplicate detection: Calculate the similarity between the newly extracted entity-relation triplet and the existing triplet in the knowledge base. The similarity is determined by a combination of entity name similarity (using the edit distance algorithm) and relation type matching. If the similarity is less than 30%, it is considered new information.

[0046] In knowledge graph duplication detection, determining whether a newly extracted entity-relation triple is redundant information in the existing knowledge base hinges on calculating its comprehensive similarity to triples already in the database. This process is achieved by quantifying the character similarity of entity names and the logical matching of relation types. The specific steps are as follows: First, entity name similarity is calculated using the edit distance algorithm. This algorithm calculates the minimum number of character editing operations (including insertion, deletion, and replacement) required to transform one entity name into another, and then normalizes the number of operations to the range of 0 to 1. For example, when comparing "Tang Poetry and Song Lyrics" with "Complete Collection of Tang Poetry, Song Lyrics, and Yuan Drama," 4 characters need to be inserted, resulting in an edit distance of 4. Since the longest possible name length is 8, the entity similarity is 1 - 4 / 8 = 0.5. This method effectively captures character-level differences in names, avoiding misjudgments caused by abbreviations or suffixes.

[0047] Secondly, the relation type matching degree is determined by a binary judgment: if the new triplet has the exact same relation type as the triplet in the database (e.g., "poet" and "poet"), the matching degree is 1; otherwise, it is 0. This strict matching ensures the accuracy of the relation logic.

[0048] Ultimately, the similarity score is calculated by weighted summation of entity similarity and relationship matching, typically with entities accounting for 60% and relationships for 40%. If the overall score is below 30%, it is considered new information; otherwise, it is considered duplicate data. This strategy, by combining character-level details with logical-level rules, effectively filters redundant content in the knowledge base while ensuring detection accuracy.

[0049] (iii) Incremental update execution; Knowledge fusion: For supplementary information of the same entity (adding representative works to existing author nodes), it is directly associated with the corresponding entity node; for new entities, a new node is created and a relationship edge is established with related entities.

[0050] The user intent recognition module is used to acquire historical dialogue data and generate an initial response strategy by recognizing user intent through the Transformer model. Specifically, this embodiment also includes a model parameter configuration unit, which is used to perform natural language processing tasks using the Transformer basic model, with the hidden layer dimension set to 512, the number of attention heads to 8, and the number of encoder and decoder layers to 6. The semantic association computing unit is used to input the user's historical dialogue data into the trained Transformer model, use the model to encode the question content and extract features, and capture key information and semantic associations in the question through a multi-layer attention mechanism; The user intent recognition unit is used by the model to calculate the probability that the question belongs to each intent label based on the learned knowledge, and to determine the label with the highest probability as the user intent.

[0051] Specifically: I. Processing historical dialogue data; (a) Scope and methods of collecting historical dialogue data; Historical dialogue data encompasses all user interactions with the system since its inception, including language-related questions initiated by users through various channels such as web pages, mobile apps, and third-party educational platforms, as well as the system's corresponding responses. This data is automatically collected through the system's backend log recording function, ensuring data integrity and continuity.

[0052] II. Transformer Model Training and User Intent Recognition; (a) Construction of a user intent tagging system; Based on the characteristics of Chinese language question-and-answer scenarios, a comprehensive user intent tagging system is constructed, specifically including: Knowledge Point Inquiry: Users can ask specific Chinese language knowledge points, such as "What is metaphor?" or "Who is the author of the Records of the Grand Historian?".

[0053] Sentence analysis: Request a grammatical analysis and semantic interpretation of a sentence, such as "analyze the usage of the word 'green' in 'The spring breeze has turned the south bank of the Yangtze River green again'".

[0054] Essay-related questions: These involve essay writing, such as "How to write a good opening for an argumentative essay?" or "Please provide an essay topic about family affection."

[0055] Literature knowledge consultation: Ask about general knowledge related to literature, such as "What are the main schools of Tang poetry?" and "What are the selection criteria for the Nobel Prize in Literature?"

[0056] Word explanation: This requires explaining the meaning and usage of words.

[0057] (ii) Preparation of training data for the Transformer model; Data annotation: Experts in the field of Chinese language education manually annotated the cleaned historical dialogue data, assigning corresponding user intent tags to each user question.

[0058] Dataset partitioning: The labeled dataset is divided into training set, validation set and test set in a ratio of 7:2:1.

[0059] (III) Transformer model training process; Model selection and configuration: Select the Transformer base model suitable for natural language processing tasks, set reasonable hyperparameters, such as setting the hidden layer dimension to 512, the number of attention heads to 8, the number of encoder and decoder layers to 6, and the initial learning rate to 5e-5.

[0060] Training process monitoring: During model training, the loss value and intent recognition accuracy on the training and validation sets are monitored in real time. Training is stopped when the accuracy on the validation set no longer improves for 5 consecutive epochs to prevent overfitting.

[0061] (iv) Implementation of user intent recognition; Real-time user questions are input into a trained Transformer model. The model encodes and extracts features from the questions, capturing key information and semantic relationships through a multi-layer attention mechanism. Then, based on the learned knowledge, the model calculates the probability that the question belongs to each intent label, and identifies the label with the highest probability as the user's intent.

[0062] III. Initial Response Strategy Generation; (a) Construction of a response framework based on intent tags; For different user intent tags, design corresponding response frameworks to provide structural guidance for generating initial response strategies: Knowledge point search: The response framework is "definition + characteristics + examples", that is, first give a clear definition of the knowledge point, then explain its main characteristics, and finally illustrate it with specific examples.

[0063] (ii) Initial response strategy generation logic; Based on the identified user intent tags, the corresponding response framework is invoked. Combined with response experience for similar intent questions from historical dialogue data, an initial response strategy is generated. During the generation process, it is ensured that the response strategy conforms to the accuracy and standardization of Chinese language knowledge, and that the language is concise, clear, and easy to understand, providing valuable reference for users.

[0064] The intelligent question answering management module is used to convert real-time Chinese question data into vector representations using a pre-trained RoBERTa semantic embedding model, and combine the entity relationships in the knowledge graph of the target knowledge base to calculate the semantic similarity between user questions and candidate answers through a GNN graph neural network. Specifically, this embodiment also includes a question data input unit, used to input preprocessed real-time Chinese question data into the RoBERTa semantic embedding model; The semantic feature capture unit is used to process the segmented text through a multi-layer Transformer structure, with each layer capturing semantic features at different levels. The vector representation generation unit is used to generate a corresponding word vector for each word, and then combine all word vectors through aggregation to generate the vector representation of the entire real-time question data.

[0065] The graph neural network building unit is used to input the vector representation and transformed graph structure of real-time query data into the GNN graph neural network; The feature information calculation unit is used to update the feature representation of each node by propagating and aggregating information on the graph. During the propagation process, each node combines the feature information of its neighboring nodes to understand the semantic relationship between entities. The semantic similarity acquisition unit is used to calculate the degree of semantic association between real-time question data and each candidate answer in the knowledge graph, and obtain the semantic similarity.

[0066] Specifically: I. Real-time Chinese language question data vector representation transformation; (a) Vector representation generation; The preprocessed real-time Chinese question data is input into the RoBERTa semantic embedding model. The model processes the segmented text through a multi-layer Transformer structure, with each layer capturing semantic features at different levels. For each word, the model generates a corresponding word vector, and then combines all word vectors through specific aggregation methods (averaging, maximizing, etc.) to form a vector representation of the entire real-time question data. This vector can comprehensively reflect the semantic information of the question.

[0067] II. Application of entity relationships in the target knowledge base knowledge graph; (a) Organizing and storing entity relationships in knowledge graphs; The knowledge graph in the target knowledge base is systematically analyzed to clarify all entities and the relationships between them. Entities include authors, works, literary concepts, and words, while entity relationships encompass various types such as subordination, causation, explanation, and association. These entities and relationships are then stored in a structured format. (ii) Extraction and integration of entity relationships; When semantic similarity calculations require combining entity relationships from a knowledge graph, entities and relationships relevant to the real-time question data are extracted from the knowledge graph. By analyzing the vector representation of the real-time question data, key entities are identified. Then, based on these key entities, other related entities and corresponding relationships are searched within the knowledge graph. The extracted entities and relationships are integrated to form a sub-knowledge graph relevant to the current question for centralized processing and computation.

[0068] III. Graph Neural Networks (GNNs) calculate semantic similarity; (a) Sub-knowledge graphs are transformed into graph structures; The integrated sub-knowledge graphs related to real-time questioning are transformed into graph structures that can be processed by GNN (Graph Neural Network). Entities in the knowledge graph are treated as nodes, and relationships between entities are represented as edges connecting the nodes. Each node and edge is assigned corresponding features; node features can be vector representations of the entity (generated using a RoBERTa model), while edge features can include relationship type information, etc.

[0069] (II) GNN Model Selection and Configuration; Choose a suitable GNN model for handling semantic similarity calculations in knowledge graphs, such as GCN (Graph Convolutional Network) or GAT (Graph Attention Network). Set the relevant model parameters according to the actual task requirements, such as the number of hidden layers, the number of neurons in each hidden layer, and the learning rate. For Chinese question-answering scenarios, it is common to set 2-3 hidden layers, with 128 or 256 neurons in each hidden layer, to ensure computational efficiency while fully capturing the semantic information in the graph.

[0070] (III) Semantic similarity calculation process; The vector representation of the real-time query data and the transformed graph structure are input into the GNN (Graph Neural Network). The GNN model continuously updates the feature representation of each node by propagating and aggregating information on the graph.

[0071] (iv) Similarity result processing; The calculated semantic similarity results are organized and sorted. Candidate answers are ranked from highest to lowest semantic similarity, and the top 10 with the highest similarity are selected to provide a strong reference for generating the subsequent target response strategy. Simultaneously, the similarity score of each candidate answer is recorded for comprehensive consideration in subsequent steps in conjunction with the initial response strategy.

[0072] The personalized adaptation module is used to generate a target response strategy based on the initial response strategy and semantic similarity using the Transformer model.

[0073] Specifically, this embodiment also includes a strategy generation unit, which is used by the decoder to gradually generate the target response strategy based on the encoding vector output by the encoder and its own language generation capabilities. The relation determination unit is used to model the causal, progressive, and parallel relationships between sentences, and to unfold the generated responses in a clear logical order.

[0074] Specifically: I. Transformer Model Preparation and Configuration; (a) Model selection and fine-tuning; The Transformer model, suitable for natural language generation tasks, was chosen as the base model. This model has strong capabilities in text generation, capturing contextual semantics and generating coherent text. To better suit the needs of intelligent question answering in the Chinese language domain, the model was fine-tuned using a large amount of dialogue data in the Chinese language field. During the fine-tuning process, the focus was on optimizing the model's use of Chinese language terminology, understanding of complex semantics, and the logic and accuracy of generated responses.

[0075] (ii) Model parameter settings; Based on the characteristics of Chinese language question-and-answer sessions, reasonable model parameters were set. For example, the maximum sequence length was set to 512 to meet the input requirements of most Chinese language questions and related information; appropriate decoding strategies were set, such as using beamsearch with a beam width of 5-10, to improve the quality and diversity of responses while ensuring generation speed; and the temperature parameter was adjusted to 0.7-0.9 to ensure that the generated responses were neither too conservative nor too arbitrary.

[0076] III. The process of generating the target response strategy; (a) Input data encoding; The fused input data is fed into the encoder of the Transformer model, which performs deep semantic understanding and feature extraction. Through a multi-layer self-attention mechanism, the model is able to focus on key information in the input data.

[0077] (ii) Response strategy generation and decoding; The decoder, based on the encoded vectors output by the encoder and its own language generation capabilities, gradually generates the target response strategy. During the decoding process, the model references the framework of the initial response strategy and appropriately utilizes information from candidate answers to enrich and refine the response content.

[0078] (iii) Logical control during the generation process; When generating target response strategies, the model emphasizes the logical coherence and structure of the responses. By modeling causal, progressive, and parallel relationships between sentences, it ensures that the generated responses unfold in a clear logical order.

[0079] IV. Optimization and Evaluation of Target Response Strategies; (a) Optimization of response strategy; Further optimization of the generated target response strategy mainly includes the following aspects: Language expression optimization: Check the reply for issues such as inappropriate word choice, awkward sentences, and grammatical errors, and make modifications and improvements to make the reply language more standardized, fluent, accurate, and in line with the habits of Chinese expression.

[0080] (ii) Evaluation of response strategies; An evaluation index system should be established to assess the target response strategy, mainly including the following indicators: Accuracy: Assess whether the response is accurate, consistent with language knowledge, and correctly answers the user's question.

[0081] Relevance: Determine the degree of correlation between the response and the user's real-time question data, and whether it closely revolves around the user's question.

[0082] V. Output of Target Response Strategy; The optimized and evaluated target response strategies will be presented to users in a natural and user-friendly language. The output format can be adjusted according to the user's usage scenario. Its beneficial effects are as follows: 1. It effectively improves the accuracy of semantic matching between user questions and candidate answers, especially when dealing with complex character relationships and plot logic in literary works, thus improving the accuracy and efficiency of question and answering. 2. It ensures that the generated target response strategies not only conform to the user's questioning intent but are also rich in content, logically coherent, and highly targeted, providing users with precise and effective guidance for Chinese language learning, thereby enhancing the user's learning experience and learning outcomes. 3. By analyzing and identifying user intent through historical dialogue data, it can generate personalized response strategies based on the different learning characteristics and needs of users, meeting their diverse and personalized needs in Chinese language learning.

[0083] Please see Figure 2 In a large-scale model-based intelligent question-answering management system for Chinese language, the knowledge graph construction module includes the following units: The initial text segmentation unit is used to initially segment the text of the initial query data according to the initial length to obtain text block data; The semantic integrity detection unit is used to call the pre-trained BERT model to calculate the co-occurrence frequency of core semantic words in the text block data, obtain the detected text block data, and combine the detected text block data with automated information extraction to build a knowledge base. The timed update unit is used to start incremental learning tasks every 30 minutes, scanning newly added question data and processing results within 30 minutes; The knowledge base update unit is used to associate supplementary information of the same entity with the corresponding entity node; for new entities, new nodes are created and relationship edges with related entities are established to obtain the target knowledge base.

[0084] Please see Figure 3 In a large-scale model-based intelligent question-answering management method for Chinese language, the method includes the following steps: Step 301: Obtain real-time Chinese language question data from users, segment the Chinese text into blocks using recursive character segmentation, construct a knowledge base by combining automated information extraction, and update the knowledge base in real time using incremental learning to obtain the target knowledge base. Step 302: Obtain historical dialogue data and generate an initial response strategy by identifying user intent through the Transformer model; Step 303: Use the pre-trained RoBERTa semantic embedding model to convert real-time Chinese question data into vector representation, combine the entity relationships in the knowledge graph of the target knowledge base, and calculate the semantic similarity between user questions and candidate answers through GNN graph neural network; Step 304: Based on the initial response strategy and semantic similarity, generate the target response strategy using the Transformer model.

[0085] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A large model-based language intelligent question and answer management system, characterized in that, The big model-based Chinese language intelligent question and answer management system comprises the following modules: A knowledge graph construction module is configured to acquire real-time Chinese language question data of a user, divide the Chinese text by using recursive character segmentation, construct a knowledge base in combination with automatic information extraction, update the knowledge base in real time by using incremental learning, and obtain a target knowledge base; A user intention recognition module is configured to acquire historical dialogue data, identify user intention by using a Transformer model, and generate an initial reply strategy; An intelligent question and answer management module is configured to convert the real-time Chinese language question data into a vector representation by using a pre-trained RoBERTa semantic embedding model, calculate semantic similarity between a user question and a candidate answer by using a GNN graph neural network in combination with entity relationships of a knowledge graph in the target knowledge base, and generate a target reply strategy by using a Transformer model based on the initial reply strategy and the semantic similarity. The knowledge graph construction module comprises the following units:

2. The large model-based Chinese intelligent question and answer management system of claim 1, wherein, An invalid character filtering unit is configured to acquire real-time Chinese language question data of a user, remove invalid characters by using regular expression matching, and perform standardization processing to obtain standard question data; An abnormal data processing unit is configured to split the standard question data into multiple sub-question segments, complete the sub-question segments in combination with user historical dialogue context, and obtain initial question data. The knowledge graph construction module further comprises the following units:

3. The large model-based Chinese language intelligent question and answer management system of claim 2, wherein, A text preliminary segmentation unit is configured to preliminarily segment text of the initial question data according to an initial length to obtain text block data; A semantic integrity detection unit is configured to calculate co-occurrence frequency of core semantic words in the text block data by using a pre-trained BERT model, obtain detection text block data, and construct a knowledge base in combination with the detection text block data and automatic information extraction; A timing update unit is configured to start an incremental learning task every 30 minutes, and scan newly added question data and processing results within 30 minutes; A knowledge base updating unit is configured to associate supplementary information of a same entity to a corresponding entity node, create a new node for a new entity, and establish a relationship edge with a related entity to obtain a target knowledge base. The user intention recognition module comprises the following units:

4. The large model-based Chinese intelligent question answering management system of claim 1, wherein, A model parameter configuration unit is configured to use a Transformer basic model for natural language processing tasks, set a hidden layer dimension to 512, set a number of attention heads to 8, and set a number of layers of an encoder and a decoder to 6 layers; A semantic association calculation unit is configured to input historical dialogue data of a user into a trained Transformer model, encode and extract features of question content by using the model, and capture key information and semantic associations in the question by using a multi-layer attention mechanism; A user intention recognition unit is configured to calculate probabilities of questions belonging to various intention labels according to learned knowledge of the model, and determine a label with the highest probability as a user intention. The intelligent question and answer management module comprises the following units:

5. The large model-based Chinese intelligent question answering management system of claim 1, wherein, A question data input unit is configured to input preprocessed real-time Chinese language question data into a RoBERTa semantic embedding model; ​ The semantic feature capturing unit is configured to process the segmented text through a multi-layer Transformer structure, each layer of which can capture semantic features at different levels. The vector representation generating unit is configured to generate a corresponding word vector for each word, and then combine all the word vectors through an aggregation manner to generate a vector representation of the real-time question data.

6. The large model-based Chinese intelligent question answering management system of claim 1, wherein, The intelligent question answering management module comprises the following units: The graph neural network establishing unit is configured to input the vector representation of the real-time question data and the transformed graph structure into a GNN graph neural network. The feature information calculating unit is configured to update the feature representation of each node by performing information propagation and aggregation on the graph, and in the propagation process, each node combines the feature information of its adjacent nodes to understand the semantic association between entities. The semantic similarity obtaining unit is configured to calculate the semantic association degree between the real-time question data and each candidate answer in the knowledge graph to obtain a semantic similarity.

7. The large model-based Chinese intelligent question answering management system of claim 1, wherein, The personalized adaptation module comprises the following units: The strategy generating unit is configured to generate a target reply strategy step by step according to the encoding vector output by the encoder and in combination with the language generation capability of the decoder. The relationship determining unit is configured to model the causal relationship, progressive relationship and parallel relationship between sentences, and expand the generated reply according to a clear logical order.

8. A large model language-based intelligent question and answer management method, characterized in that, The intelligent question answering management method based on a large model Chinese language comprises the following steps: Obtain real-time Chinese question data of a user, perform block segmentation on the Chinese text by using a recursive character segmentation, construct a knowledge base in combination with automatic information extraction, update the knowledge base in real time by using incremental learning, and obtain a target knowledge base; Obtain historical dialogue data, and generate an initial reply strategy by identifying the user intent through a Transformer model; Convert the real-time Chinese question data into a vector representation by using a pre-trained RoBERTa semantic embedding model, combine the entity relationship of the knowledge graph in the target knowledge base, and calculate the semantic similarity between the user question and the candidate answers through a GNN graph neural network; Generate a target reply strategy based on the initial reply strategy and the semantic similarity through a Transformer model.

9. The large model-based Chinese intelligent question answering management method of claim 8, wherein, The method for converting the real-time Chinese question data into a vector representation by using a pre-trained RoBERTa semantic embedding model, combining the entity relationship of the knowledge graph in the target knowledge base, and calculating the semantic similarity between the user question and the candidate answers through a GNN graph neural network comprises the following steps: Input the preprocessed real-time Chinese question data into the RoBERTa semantic embedding model; Process the segmented text through a multi-layer Transformer structure, each layer of which can capture semantic features at different levels; Generate a corresponding word vector for each word, and then combine all the word vectors through an aggregation manner to generate a vector representation of the real-time question data.

10. The large model-based Chinese language intelligent question and answer management method of claim 8, wherein, The method for converting the real-time Chinese question data into a vector representation by using a pre-trained RoBERTa semantic embedding model, combining the entity relationship of the knowledge graph in the target knowledge base, and calculating the semantic similarity between the user question and the candidate answers through a GNN graph neural network further comprises the following steps: The vector representation of the real-time question data and the transformed graph structure are input into a GNN graph neural network; By information propagation and aggregation on the graph, the feature representation of each node is updated. During the propagation process, each node combines the feature information of its adjacent nodes to understand the semantic association between entities. The semantic association degree between the real-time question data and each candidate answer in the knowledge graph is calculated to obtain the semantic similarity.

Images