Intelligent question answering system and method based on generative AI

Abstract

The invention discloses an intelligent question answering system and method based on generative AI, and belongs to the technical field of natural language processing. The system comprises a problem input interface, an extensible classification matching module (including a dynamic threshold decision tree and semantic extension verification), a strategic output engine, a strategy cooperation controller (a main-auxiliary strategy hard binding mechanism) and an output control layer. According to the method, through semantic feature analysis, adaptive classification mapping, strategy cooperation control and conflict solving processes, the problems of fuzzy classification, strategy cooperation deficiency and classification system stiffness in the prior art are solved. The innovation points comprise a dynamic threshold decision tree (reference value = alpha * (1-beta * ln (classification number)), alpha belongs to [0.48, 0.52], beta belongs to [0.08, 0.12]), a strategy collaboration mechanism bound with RBAC authority, and a JSON template classification expansion interface. Through 100,000-level real data tests, the key action missing rate is reduced by 29.8%, the executable path coverage rate is improved by 37.1%, and the method is suitable for multi-strategy collaborative scenes such as law compliance, supply chain management and medical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an intelligent question-answering system and method based on generative artificial intelligence (AI), which is applicable to professional decision support scenarios that require multi-strategy collaboration, including but not limited to business decision-making, legal compliance consulting, supply chain risk management, and medical diagnostic assistance. Background Technology

[0002] Existing intelligent question-answering systems have the following technical shortcomings in professional decision-making scenarios: (1) Classification ambiguity problem: Using a fixed classification threshold leads to a high misclassification rate for boundary cases. For example, in medical diagnosis scenarios, the IBM Watson system has a false trigger rate of over 25% because the fixed threshold cannot adapt to the classification complexity of different diseases; (2) Lack of strategy coordination: There is a lack of coordination mechanism across strategy engines. For example, the question-answering system disclosed in patent US2019030345A1 only implements a single strategy output and does not design a strategy conflict detection and scheduling module, resulting in a lack of information under multi-dimensional decision-making needs; (3) Rigid classification system: Adding new classification labels requires retraining the entire model, resulting in a long deployment cycle. Existing systems such as IBM Watson require end-to-end retraining of the model when expanding to new classifications, with a deployment cycle of more than 72 hours, making it difficult to adapt to rapidly changing business scenarios.

[0003] The comparative analysis of the shortcomings of existing technologies is shown in the table below: Comparison System Fuzzy classification problem Lack of strategic coordination rigid classification system IBM Watson Medical diagnosis false alarm rate >25% Legal compliance and risk-free collaboration Adding new tags requires 72 hours of training. US2019030345A1 Single threshold causes boundary error Cross-policy scheduling not designed No expansion interface Invention Solution Dynamic threshold + semantic extension Hard binding of primary and secondary strategies JSO template interface deployment in <3 hours Summary of the Invention

[0004] Technical problems to be solved This invention aims to solve three major problems existing in the prior art: (1) The problem of policy error triggering caused by classification ambiguity; (2) Decision-making blind spots caused by the lack of multi-strategy coordination mechanisms; (3) Difficulty in expanding the classification system leads to poor scene adaptability.

[0005] System Architecture The intelligent question-answering system based on generative AI provided by this invention includes the following core modules: (1) Question input interface: Receives natural language questions from users, extracts entities (such as “overseas mergers and acquisitions” and “lung ground-glass nodules”) and constraints (such as “listing regulations” and “biopsy indications”) through pre-trained semantic analysis models (such as BERT-base) to provide basic features for subsequent classification and matching.

[0006] (2) Expandable classification matching module: Semantic Vector Generation Unit: Encodes user questions into 768-dimensional semantic feature vectors using the BERT-base model, capturing the deep semantic information of the questions; Adaptive decision tree model: Employs a dynamic threshold splitting strategy. The threshold formula is "benchmark value = α × (1 - β × ln(number of categories))", where α ∈ [0.48, 0.52] (preferably 0.5), β ∈ [0.08, 0.12] (preferably 0.1), and the number of categories is the total number of category labels in the current system (≥2). This dynamic threshold can adaptively decrease as the number of categories increases, avoiding boundary blurring when the number of categories increases. Boundary processing unit: When the cosine similarity between the semantic feature vector and the classification label is in the fuzzy range of [0.6, 0.7), the knowledge graph is called to perform K-hop (K∈[3,5], preferably 5) semantic expansion verification, which improves the classification accuracy by expanding the entity association relationship.

[0007] (3) Strategy-based output engine: Rule reasoning unit: Generates operation instruction chains with regulatory numbers or guideline clauses based on real-time updated domain knowledge graphs (such as legal provisions and medical guidelines) to ensure the professionalism and traceability of the output; Multi-path strategy unit: Establish a quantitative mapping model between resource input (such as cost and time) and expected returns (such as risk reduction rate and accuracy rate) to provide multi-dimensional reference for complex decision-making.

[0008] (4) Strategy Coordination Controller: Main strategy activator: Prioritizes calling the strategy engine that strictly matches the category label, as the core solution; Auxiliary policy scheduler: When a policy conflict is detected (defined as a mutual exclusion rate of multiple policy engine output actions triggered by the same issue > 40%) or when an explicit user request is received (such as "risk analysis" or "prognosis analysis"), the cross-category policy engine is invoked to supplement the output.

[0009] The primary and secondary strategies are hard-bound through a role-based access control (RBAC) model: the primary strategy engine has full access to all fields of the knowledge graph, while the secondary strategy engine can only read nodes marked with the @public tag, thus avoiding data permission conflicts.

[0010] (5) Output control layer: The output mode is dynamically switched according to the entity density (>3 / sentence) in the question text: when the entity density is high, a visual output (such as a relationship diagram or matrix table) is used, and when the density is low, a plain text output is used to improve the readability of the information.

[0011] Method and Flow The intelligent question-answering method using the above system includes the following steps: Step S101: Analyze the user's question and generate a 768-dimensional semantic feature vector using the BERT-base model; Step S102: Map the semantic feature vector to the classification label using an adaptive decision tree model. The decision tree splitting threshold is set by the formula "benchmark value = α × (1 - β × ln(number of categories))", where β = 0.1. Step S103: Activate the main strategy engine that matches the category label and generate an initial solution (such as an operation instruction with a regulatory number). Step S104: Detect whether there is a policy conflict (output action mutual exclusion rate > 40%) or an explicit user request: If yes, schedule the auxiliary policy engine to supplement the output; if no, proceed directly to step S105. Step S105: Based on the entity density (>3 entities / sentence) in the problem text, dynamically generate and output a plain text or visualization solution.

[0012] Beneficial effects Compared with the prior art, the present invention has the following beneficial effects: (1) Solving the problem of classification ambiguity: By using dynamic threshold decision trees and K-hop semantic expansion verification, the misclassification rate of boundary cases is significantly reduced. Testing showed that the classification error rate was reduced to 12.5%, a 62.3% reduction compared to existing systems (such as IBM Watson). (2) Improve strategy coordination efficiency: Based on the hard binding mechanism of main and auxiliary strategies of RBAC, the permission conflict and scheduling problem of multiple strategy engines is solved, the strategy conflict resolution rate is increased to 92.7%, and the key action missing rate is reduced by 29.8 percentage points; (3) Enhance the scalability of the classification system: The classification labels can be quickly expanded through the JSON template interface. The deployment cycle of new labels is less than 3 hours, and the generalization ability is ensured by verification with noisy samples (accuracy ≥ 95%). (4) Strong cross-scenario adaptability: In scenarios such as legal compliance, supply chain management, and medical diagnosis, the coverage of executable paths increased by 37.1 percentage points, reaching over 93%. Attached Figure Description

[0013] Figure 1: System architecture diagram, showing the connection relationships of the problem input interface, the scalable classification and matching module, the strategy output engine, the strategy coordination controller and the output control layer. The main strategy data flow is represented by red solid arrows, and the auxiliary strategy data flow is represented by blue dashed arrows. Figure 2: Threshold curve of dynamic decision tree, with the X-axis representing the number of categories and the Y-axis representing the baseline threshold, showing the decay trend of the threshold as the number of categories increases; Figure 3: Policy coordination logic diagram, including conflict detection module (mutual exclusion rate > 40% triggers auxiliary policy) and RBAC permission control process; Figure 4: RBAC hard binding pseudocode diagram, showing the implementation of access control between the main policy and the auxiliary policy; Figure 5: β parameter optimization curve, X-axis represents β value (0.05-0.15), Y-axis represents classification error rate, with the lowest error rate when β=0.1; Figure 6: Experimental diagram of semantic expansion hop number selection. The X-axis represents the number of hops K (3-7), and the Y-axis represents accuracy and response latency. K=5 is marked as the optimal value. Figure 7: Bar chart comparing ablation experiments, showing the performance differences between the complete scheme and the schemes lacking dynamic threshold, strategy binding mechanism, and template extension interface. Detailed Implementation

[0014] Example 1: Category Expansion (Supply Chain Risk Early Warning Scenario) When adding a new "Supply Chain Risk Warning" category tag, the strategy engine is bound via a JSON template interface. An example template is shown below: {"category":"supply_risk","strategy":{"class":"RiskEngine","params":{"axis":["probability","impact"],"threshold":0.7}},"validation":{"noise_ratio":0.1,"accuracy_threshold":0.95}} The system automatically adds branch rules to the decision tree: If the semantic vector contains ["supply disruption", "inventory warning"] and information gain ≥ 0.5 × (1 - 0.1 × ln(N)): category = "supply chain risk warning category" Validation process: Inject 10% of the noisy samples based on the MNLI dataset (e.g., replace "supply interruption" with "supply disruption"), test the classification accuracy of 96.3% (≥95%), activate the classification label, and the deployment takes 2.5 hours.

[0015] Example 2: Strategy Collaboration (Legal Compliance Scenario) A user entered the question: "Do overseas mergers and acquisitions comply with listing regulations?" Step S101: Extract the entities “overseas mergers and acquisitions” and “listing regulations” to generate a 768-dimensional semantic vector; Step S102: The category label is matched as "Cross-border M&A Compliance Category", and the decision tree threshold is 0.5×(1-0.1×ln(12))≈0.42 (current number of categories N=12). Step S103: Output of the main policy engine (compliance verification engine): Compliance requirements for the "Guidelines on the Supervision of Cross-border Mergers and Acquisitions": 1. Filing with the State Administration of Foreign Exchange (Regulatory No.: SAFE-2023-Art8) 2. Antitrust review passed (Regulation No.: AML-2022-Art12) Step S104: The user clicks "Risk Analysis" (explicit request), which dispatches the auxiliary strategy engine (risk assessment engine), which only accesses risk nodes tagged with @public in the knowledge graph, and outputs: High-risk point analysis: 1. Foreign exchange control risk (probability 0.8, impact $2M) → Countermeasure: Reserve 15% margin 2. Policy change risk (probability 0.65, impact $1.2M) → Countermeasure: Phased delivery Step S105: Problem entity density = 2 per sentence (<3), output plain text solution.

[0016] Example 3: Application in Medical Diagnosis Scenarios A user entered the question: "Does a biopsy of ground-glass nodules in the lungs shown on a CT scan require?" Step S101: Extract the entities “lung ground-glass nodule” and “biopsy”, and generate semantic vectors; Step S102: The classification label is matched as "pulmonary nodule diagnosis and treatment", and the decision tree threshold is 0.5×(1-0.1×ln(8))≈0.45 (current number of categories N=8). Step S103: Output of the main strategy engine (treatment guideline engine): According to the "Guidelines for the Diagnosis and Treatment of Lung Cancer (2024)": 1. Diameter > 8mm: CT-guided biopsy is recommended. 2. Combined high-risk factors (smoking history / family history): Biopsy is strongly recommended. Step S104: No policy conflict, no explicit request from the user, proceed directly to step S105; Step S105: Problem entity density = 3 per sentence (=3), output visualization solution (including a graph showing the relationship between nodule size and biopsy indications).

[0017] Example 4: Derivation of the dynamic threshold formula (β parameter optimization) Test the classification error rate with different β values ​​on the LegalBench dataset (containing 100,000 legal question and answer samples): β_list=[0.05,0.08,0.1,0.12,0.15] error_rates=[ for β in β_list: model=DecisionTree(threshold formula=0.5×(1-β×ln(N))) error_rate=model.evaluate(LegalBench_testset) error_rates.append(error_rate) #Results: The lowest error rate (18.2%) was observed when β=0.1. Figure 5 Lowest point of the curve

[0018] Example 5: Semantic Expansion Hop Count Selection Experiment Performance tests were conducted on a medical knowledge graph (containing 500,000 entity nodes) with different K values: Number of jumps K accuracy Average response time (ms) 3 82.3% 120 4 90.5% 220 5 95.1% 310 6 95.2% 520 7 95.3% 820 Choosing K=5 (accuracy > 95% and latency < 350ms) as the optimal value corresponds to... Figure 6 Experimental results.

[0019] Example 6: Effect Verification Experiment (1) Ablation experiment (to verify the necessity of the innovation) Test results based on 100,000 pieces of commercial bank consultation data: Technical solution Classification error rate Strategy conflict resolution rate No dynamic threshold 38.7% 70.1% No strategy binding mechanism 25.3% 65.8% No template extension interface 20.1% 85.3% Complete solution 12.5% 92.7% The results show that dynamic threshold, policy binding mechanism, and template extension interface are all key innovations for improving performance.

[0020] (2) Cross-domain performance verification Application scenarios Key motion missing rate Executable path coverage Legal compliance 12.5% 96.2% Supply chain risk warning 14.2% 94.7% Medical diagnostic aid 15.8% 93.1% Test results show that the invention maintains excellent performance in multiple fields, verifying its versatility.

[0021] The above specific embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the protection scope of the present invention.

Claims

1. An intelligent question-answering system based on generative AI, characterized in that, include: Question input interface: Configured to receive user questions and extract entities and constraints from the questions using a pre-trained semantic analysis model; An extensible classification matching module includes: The semantic vector generation unit transforms user questions into 768-dimensional semantic feature vectors using the BERT-base model. The adaptive decision tree model adopts a dynamic threshold splitting strategy, and its threshold setting satisfies the formula: baseline value = α × (1 - β × ln(number of categories)), where the number of categories is the total number of category labels in the current system and ≥ 2, α ∈ [0.48, 0.52], β ∈ [0.08, 0.12]; The boundary processing unit calls the knowledge graph to perform K-hop semantic expansion verification when the cosine similarity between the semantic feature vector and the classification label is in the range of [0.6, 0.7), where K ∈ [3, 5]. A strategy-based output engine, including: The rule-based reasoning unit generates a chain of operational instructions with regulatory numbers based on a real-time updated knowledge graph. Multi-path strategy unit, establishes a quantitative mapping model between resource input parameters and expected return indicators; The policy coordination controller includes: The main strategy activator prioritizes calling the strategy engine that strictly matches the category label; The auxiliary policy scheduler invokes the cross-category policy engine when a policy conflict or explicit user request is detected. A policy conflict is defined as the same user issue triggering two or more policy engines with a mutual exclusion rate of output actions > 40%. The output control layer dynamically switches between plain text output mode and visualization output mode based on the condition that the entity density threshold in the question text is >3 per sentence.

2. The intelligent question-answering system based on generative AI according to claim 1, characterized in that, The policy coordination controller achieves a hard binding between the main policy engine and the auxiliary policy engine through a role-based access control (RBAC) model, wherein: The main strategy engine has full field access permissions to the knowledge graph; The auxiliary strategy engine can only read nodes in the knowledge graph that are marked with the @public tag; The mutual exclusion rate threshold for the policy conflict is set between 35% and 45%.

3. The intelligent question-answering system based on generative AI according to claim 1, characterized in that, When adding a new category tag, the strategy engine is bound via a JSON template interface. This template includes the following required fields: {"category":"category name string","strategy":{"class":"strategy engine class name string","params":{"parameter key":"parameter value"}}} The verification process includes: injecting 10% of the total samples with synonyms to replace noisy samples, which are generated based on the MNLI dataset, and activating the new classification label when the classification accuracy is ≥95%.

4. An intelligent question-answering method applying the system described in any one of claims 1 to 3, characterized in that, Includes the following steps: Step S101: Analyze the user's question and generate a 768-dimensional semantic feature vector using the BERT-base model; Step S102: Map the semantic feature vector to the classification label using an adaptive decision tree model, where the decision tree splitting threshold is calculated using the formula: baseline value = α × (1 - β × ln(number of categories)), and β = 0.1; Step S103: Activate the main strategy engine to generate an initial solution; Step S104: When a policy conflict is detected, i.e., the output action mutual exclusion rate is >40%, or when an explicit request from the user is received, the scheduling auxiliary policy engine supplements the output; Step S105: Based on the condition that the entity density in the problem text is >3 per sentence, dynamically generate a visual solution.

Images