Self-evolution semantic classification and dynamic knowledge base construction system
By using a self-evolving semantic classification and dynamic knowledge base construction system, we have achieved fine classification of semantic information and construction of a structured knowledge base. This solves the problems of model adaptability and high cost in existing technologies and provides support for the scalability and practicality of intelligent systems.
Patent Information
- Application Number
- CN202511094401.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-06
- Publication Date
- 2025-11-21
AI Technical Summary
Existing technologies struggle to achieve precise classification of semantic information, adaptive optimization of models, and automatic construction of structured semantic classification knowledge bases. Furthermore, existing methods are costly or lack structured constraints in their output.
The system employs a self-evolving semantic classification and dynamic knowledge base construction system, which includes an information input module, a first-level semantic classification module, a second-level semantic sub-classification module, a classification fusion module, a feedback training module, and a semantic classification knowledge base construction module. It utilizes pre-trained language models and generative language models to identify and fuse main and sub-classes, and combines semantic similarity and confidence calculations to support model self-evolution and new category identification.
It achieves hierarchical classification of information semantics, automatically generates standardized classification results, continuously learns from feedback, and automatically identifies new semantic types, constructs a structured classification knowledge base, provides good scalability and practical value, and supports knowledge graphs, question answering systems, and intelligent recommendations.
Smart Images
Figure CN120994826A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of language processing systems, more particularly, to a self-evolution semantic classification and dynamic knowledge base construction system. BACKGROUND
[0002] With the massive growth of text data, how to identify and organize semantic information has become the core problem of intelligent systems. Existing methods focus on shallow classification of text topics, sentiment, entities, etc., lacking a unified semantic ontology and classification structure.
[0003] At the same time, most classification models (such as BERT) are difficult to adapt to new semantic categories after training, and their outputs are difficult to directly form a structured knowledge system. Traditional classifiers (such as SVM, CNN) rely on fixed classification systems and cannot adapt to the dynamic evolution of semantics; knowledge base construction (such as WordNet) requires a large amount of manual annotation, which is costly; the direct output of the generation model (such as GPT) lacks structured constraints and needs additional post-processing.
[0004] Therefore, there is an urgent need for a system that can finely classify semantic information, optimize model adaptation, automatically construct structured semantic classification knowledge base, and support intelligent semantic expansion and new category identification. SUMMARY
[0005] The purpose of the present application is to solve the problems existing in the prior art and to propose a self-evolution semantic classification and dynamic knowledge base construction system.
[0006] To solve the above problems, the present application adopts the following technical solution:
[0007] A self-evolution semantic classification and dynamic knowledge base construction system, comprising:
[0008] An information input module, which pre-processes raw text data into standardized information units;
[0009] A primary semantic classification module, which identifies main classes based on pre-trained language models for information units;
[0010] A secondary semantic classification module, which uses a generative language model to generate sub-class suggestions based on context and aligns with the main class;
[0011] A classification fusion module, which fuses the main class and sub-class results, determines the final main-sub-class combination through semantic similarity, pragmatic consistency and confidence calculation, and outputs structured classification results;
[0012] A feedback training module, which writes high-confidence classification results into a training pool for model self-maintenance and self-evolution;
[0013] A new semantic type detection module monitors the density and distribution of unattributed expressions and mines potential new categories through semantic clustering.
[0014] A semantic classification knowledge base construction module is used to store structured classification results, form a classification knowledge base, and support external system calls.
[0015] As a further scheme of the present application, the pre-trained language model adopted by the primary semantic classification module is BERT or RoBERTa.
[0016] As a further scheme of the present application, the main categories include factual information, subjective expression category information, logical reasoning category information, axiom / rule category information, hypothesis / guess information, command / intention category information, meta information, and noise / error / misleading information.
[0017] As a further scheme of the present application, the generative language model adopted by the secondary semantic classification module is GPT, Claude or LLaMA.
[0018] As a further scheme of the present application, the classification fusion module includes:
[0019] A confidence score calculation sub-module is used to calculate the confidence score of the primary and secondary category prediction results, providing a quantitative basis for subsequent fusion;
[0020] A semantic similarity calculation sub-module is used to evaluate the semantic compatibility and consistency of the primary and secondary categories;
[0021] A pragmatic consistency verification sub-module is used to evaluate the rationality and consistency of the primary and secondary categories in actual application scenarios;
[0022] A fusion decision sub-module is used to comprehensively consider the results of each sub-module and make a final classification decision;
[0023] A result structured output sub-module formats the final classification result into structured data.
[0024] As a further scheme of the present application, the system further includes a visualization interface and a manual review module, the visualization interface is used to display the classification results and knowledge base content, and the manual review module is used to confirm new semantic categories.
[0025] As a further scheme of the present application, the semantic classification knowledge base is stored in JSON, RDF or SQL format and has an interface that can be called by external APIs.
[0026] As a further scheme of the present application, the system further includes an update management module, which is used to dynamically update semantic categories.
[0027] Compared with the prior art, the present application has the following advantages:
[0028] The scheme realizes hierarchical classification of information semantics, automatic generation of standardized classification results, continuous feedback learning, automatic identification of new semantic types and construction of a structured classification knowledge base, provides standard semantic support for knowledge graphs, question and answer systems, intelligent recommendation and the like, and has good expansibility and practical value. BRIEF DESCRIPTION OF DRAWINGS
[0029] Figure 1 is a system diagram of the present application;
[0030] Figure 2 is a schematic diagram of the classification fusion module of the present application;
[0031] Figure 3 is a flowchart of the present application;
[0032] Figure 4 is a high / low confidence information result processing flowchart of the present application;
[0033] Figure 5 is a flowchart of the present application for unattributed information processing. DETAILED DESCRIPTION
[0034] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application; obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments; based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0035] Please refer to Figures 1-2 A self-evolution semantic classification and dynamic knowledge base construction system comprises an information input module, a primary semantic classification module, a secondary semantic classification module, a classification fusion module, a feedback training module, a new semantic type detection module and a semantic classification knowledge base construction module.
[0036] The information input module pre-processes original text data into standardized information units, that is, receives original text data and performs the following pre-processing: text cleaning (removing irrelevant characters, formatting), sentence and segment division, language recognition and coding unification, generating standardized information units (containing text content and metadata), and completing preliminary semantic segmentation.
[0037] The primary semantic classification module uses a pre-trained language model to identify the main class of the information unit; wherein the main class includes factual information, subjective expression class information, logical reasoning class information, axiom / rule class information, hypothesis / guess information, command / intention class information, meta-information and noise / error / misleading information. The pre-trained language model is BERT or RoBERTa or other equivalent models.
[0038] The secondary semantic sub-classification module generates sub-class suggestions based on the context using a generative language model, such as: factual information → [time, location, characters, events...]; subjective expression → [evaluation, emotion, preference...], and ensures semantic alignment between sub-classes and main classes through attention mechanisms. The generative language model is a large language model such as GPT, Claude, LLaMA, etc.
[0039] The classification fusion module fuses the main class and sub-class results, determines the final main-sub-class combination and outputs structured classification results through semantic similarity, pragmatic consistency and confidence calculation. The output results include the original text, main class label, sub-class label (suggestion or determination), model confidence, timestamp and source marker, optional language and meta-attributes.
[0040] The classification fusion module includes five sub-modules working together:
[0041] The confidence calculation submodule is used to calculate the confidence score of the main class and sub-class prediction results, providing a quantitative basis for subsequent fusion;
[0042] The semantic similarity calculation submodule is used to evaluate the compatibility and consistency of the main class and sub-class in terms of semantics;
[0043] The pragmatic consistency verification submodule is used to evaluate the rationality and consistency of the main class and sub-class in actual application scenarios;
[0044] The fusion decision submodule is used to integrate the results of each submodule and make the final classification decision;
[0045] The result structured output submodule formats the final classification results into structured data.
[0046] The feedback training module writes high-confidence classification results into the training pool for subsequent fine-tuning, achieving model self-maintenance and self-evolution.
[0047] The new semantic type detection module monitors the density and distribution of unattributed expressions, mines potential new categories through semantic clustering, and incorporates them into the category system after artificial confirmation.
[0048] The semantic classification knowledge base construction module is used to store all classification results in a standard structure to form a classification knowledge base, supporting external system (search engine, knowledge graph, question and answer system) calling. The semantic classification knowledge base is stored in JSON, RDF or SQL format, with an interface that can be called by external API.
[0049] The system further comprises a visualization interface, a manual review module, and an update management module. The visualization interface is used to display the classification results and the knowledge base content. The manual review module is used to confirm new semantic categories. The update management module is used to dynamically update the semantic categories.
[0050] Referring to Figures 3-5 Based on the above embodiments, the present application provides a self-evolution semantic classification and dynamic knowledge base construction process, comprising:
[0051] Stage 1, input processing:
[0052] Input module: receives text data to be classified, and performs preprocessing and standardization.
[0053] Stage 2, classification model processing:
[0054] Primary semantic classification (BERT classifier): using a pre-trained BERT model for initial semantic classification, assigning the input text to an existing primary main category;
[0055] Secondary semantic classification (sub-category suggestion generation): generating sub-category suggestions using GPT / Claude / LLaMA generative models based on the primary classification results;
[0056] Fusion judgment and confidence control: comprehensively considering the results of the main category and the sub-category, making a final classification decision and calculating the classification confidence.
[0057] Stage 3, classification result output and knowledge base management:
[0058] Structured classification result output: converting the final classification result into structured data;
[0059] Classification result storage in knowledge base: storing high-confidence results in the classification knowledge base as baseline knowledge;
[0060] External system query: knowledge base open API interface, supporting real-time query of classification results by external systems.
[0061] Stage 4, low-confidence / unattributed information processing:
[0062] Unattributed information detection: identifying low-confidence or unclassifiable data;
[0063] Semantic clustering: clustering analysis of unattributed information;
[0064] New category suggestion: generating new category suggestions based on clustering results.
[0065] Stage 5, manual intervention and confirmation
[0066] Manual confirmation: manual review and labeling of low-confidence classification results and new category suggestions;
[0067] Update category system: expand or adjust the classification system according to manual confirmation results.
[0068] Stage 6, feedback and evolution mechanism
[0069] Classifier feedback and training pool construction: collect annotated classification results, low confidence samples and new class samples confirmed by manual confirmation to the feedback training pool, and construct the training data set;
[0070] Model optimization and update: periodically retrain the model using feedback data to adapt to new categories and boundary cases;
[0071] New semantic type detection and category growth mechanism: automatically detect emerging semantic types and trigger category expansion process.
[0072] Based on the above embodiments, the present application provides the following examples:
[0073] Example 1 (basic classification):
[0074] Input text: "I like red."
[0075] Processing flow:
[0076] 1. Information input module: perform word segmentation and stop word removal on the text to generate standardized information units ["I", "like", "red"].
[0077] 2. Primary semantic classification module: based on a pre-trained language model (such as BERT), identify the main class as "subjective expression information" (confidence 0.91).
[0078] 3. Secondary semantic classification module: generative language model (such as GPT-3) generates subclass suggestions based on context, including "emotional expression" (0.75) and "personal preference expression" (0.88).
[0079] 4. Classification fusion module: calculate semantic similarity (subclass and main class association degree 0.92), pragmatic consistency (context matching degree 0.90), and determine the final output:
[0080] → Main class: subjective expression information → Subclass: personal preference expression → Comprehensive confidence: 0.93.
[0081] 5. Knowledge base construction: store the classification results in a structured form ({main class: subjective expression, subclass: personal preference, text ID: 001}) in the semantic classification knowledge base.
[0082] Example 2 (rules and reasoning mixed)
[0083] Input text: "If it rains, the road will be slippery."
[0084] Processing flow:
[0085] 1. The primary semantic classification module identifies the main class as "logical reasoning information" (confidence 0.87).
[0086] 2. The secondary semantic classification module generates sub-class suggestions: "causal prediction" (0.89), "hypothesis reasoning" (0.82).
[0087] 3. The classification fusion module checks by logical rules ("if... will..." is a typical causal sentence pattern), selects the optimal sub-class:
[0088] → Main class: logical reasoning information → Sub-class: causal prediction → Confidence: 0.89.
[0089] 4. Feedback training: high-confidence results are added to the training pool, strengthening the model's recognition ability for causal text.
[0090] Example 3 (meta-information labeling)
[0091] Input text: "According to Xinhua News Agency..."
[0092] Processing flow:
[0093] 1. The primary semantic classification module identifies the keyword "according to... report" and directly matches the pre-defined meta-information template, with the main class as "meta-information" (confidence 0.97).
[0094] 2. The secondary semantic classification module supplements sub-class suggestions through generative models: "source explanation" (0.95), "reference annotation" (0.82).
[0095] 3. The classification fusion module prioritizes the template matching result:
[0096] → Main class: meta-information → Sub-class: source explanation → Confidence: 0.95.
[0097] Example 4 (new type clustering recognition)
[0098] Scenario: The system monitors high-frequency unclassified expressions in the technology field (such as "machine consciousness" "consciousness upload").
[0099] Processing flow:
[0100] 1. New semantic type detection module:
[0101] Statistical unattributed expression density (more than 500 times / week threshold), triggering clustering analysis.
[0102] Discover potential class features through semantic vector clustering (such as UMAP dimensionality reduction + K-means).
[0103] 2. Recommend new category: generate the suggestion "interdisciplinary hypothesis information" and associate the parent category as "scientific theory information".
[0104] 3. Artificial review and knowledge base update:
[0105] After the administrator confirms, the system adds the new category to the classification system.
[0106] Historical data backtracking classification (e.g., classifying "quantum immortality" into this subclass).
[0107] 4. Model self-evolution: new category samples are added to the training pool, and the secondary classification model is fine-tuned.
[0108] Based on the above, the present application provides a self-evolving semantic classification and dynamic knowledge base construction system and method, which realizes hierarchical classification of information semantics, automatic generation of standardized classification results, continuous feedback learning, automatic identification of new semantic types, and construction of a structured classification knowledge base, providing standard semantic support for knowledge graphs, question and answer systems, intelligent recommendations, and other systems, and has good scalability and practical value.
[0109] The above is only a preferred specific embodiment of the present application; however, the protection scope of the present application is not limited thereto. Any person skilled in the art, within the technical scope disclosed by the present application, can make equivalent replacements or changes to the technical solutions and improvement concepts of the present application, which should be covered within the protection scope of the present application.
Claims
1. A self-evolving semantic classification and dynamic knowledge base construction system, characterized in that: include: The information input module preprocesses raw text data into standardized information units; The first-level semantic classification module identifies the main class of information units based on a pre-trained language model. The secondary semantic fine-classification module uses a generative language model to generate subclass suggestions based on context and align them with the main class; The classification fusion module merges the main class and subclass results, determines the final main-subclass combination through semantic similarity, pragmatic consistency and confidence calculation, and outputs the structured classification result; The feedback training module writes high-confidence classification results into the training pool for model self-maintenance and self-evolution. The new semantic type detection module monitors the density and distribution of unassigned expressions and mines potential new categories through semantic clustering; The semantic classification knowledge base construction module is used to store structured classification results, form a classification knowledge base, and support external system calls.
2. The self-evolving semantic classification and dynamic knowledge base construction system according to claim 1, characterized in that: The pre-trained language model used in the first-level semantic classification module is BERT or RoBERTa.
3. The self-evolving semantic classification and dynamic knowledge base construction system according to claim 1, characterized in that: The main categories include factual information, subjective expression information, logical reasoning information, axiom / rule information, hypothesis / guess information, command / intention information, meta-information, and noise / error / misleading information.
4. The self-evolving semantic classification and dynamic knowledge base construction system according to claim 1, characterized in that: The generative language model used in the secondary semantic sub-classification module is GPT, Claude, or LLaMA.
5. The self-evolving semantic classification and dynamic knowledge base construction system according to claim 1, characterized in that: The classification fusion module includes: The confidence score calculation submodule is used to calculate the confidence scores of the prediction results of the main class and the subclass, providing a quantitative basis for subsequent fusion. The semantic similarity calculation submodule is used to evaluate the semantic compatibility and consistency between the main class and its subclasses; The pragmatic consistency verification submodule is used to evaluate the rationality and consistency of the main class and subclasses in actual application scenarios; The fusion decision submodule is used to integrate the results of each submodule and make a final classification decision. The results structured output submodule formats the final classification results into structured data.
6. The self-evolving semantic classification and dynamic knowledge base construction system according to claim 1, characterized in that: The system also includes a visual interface and a manual review module. The visual interface is used to display classification results and knowledge base content, while the manual review module is used to confirm new semantic categories.
7. The self-evolving semantic classification and dynamic knowledge base construction system according to claim 1, characterized in that: The semantic classification knowledge base is stored in JSON, RDF, or SQL format and has an interface that can be called by external APIs.
8. The self-evolving semantic classification and dynamic knowledge base construction system according to claim 1, characterized in that: The system also includes an update management module, which is used to dynamically update semantic categories.