Efficient knowledge retrieval system and method based on user intention

Abstract

The invention provides an efficient knowledge retrieval method and system based on user intentions. The method comprises the following steps: receiving a natural language query of a user, and analyzing and extracting an initial known condition; constructing a conditional decision diagram containing conditional nodes and a dependency relationship in combination with the domain knowledge base; a candidate reasoning path is generated in the graph, interactive optimization is carried out by using a strategy based on beam search, a clarification problem is generated by calculating information gain of a missing condition, and a graph structure and a path score are dynamically updated according to user feedback until an optimal path is locked; and finally, generating a structured query based on the optimal path, executing mixed retrieval and synthesizing a final answer. According to the method, through structured intention modeling and an active interaction clarification mechanism, the problems of intention fuzziness, semantic gaps and inference ability deficiency in vertical field retrieval are solved, and the retrieval accuracy and the user experience are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and information retrieval technology, and in particular to an intelligent question-answering and knowledge retrieval system and method based on conditional decision graphs and interactive reasoning in vertical fields (such as natural resources, medical care, finance, etc.), and especially to an efficient knowledge retrieval method and system based on user intent. Background Technology

[0002] In specialized fields such as natural resource surveys and monitoring, medical diagnosis, and legal consulting, building intelligent question-answering systems capable of accurately answering professional questions is a current research hotspot. The core of such systems lies in quickly and accurately retrieving the most relevant knowledge fragments to the user's question from a massive domain knowledge base (such as policies and regulations, technical procedures, case reports, etc.).

[0003] Currently, existing retrieval technologies are mainly divided into two categories:

[0004] 1. Keyword-based retrieval techniques (such as BM25): This method relies on keyword frequency and suffers from a significant semantic gap. For example, if a user queries "land type for strawberry cultivation," the system may not be able to associate it with professional terms such as "perennial herbaceous crops" or "farmland identification standards." Furthermore, when user queries lack information (such as "How is this land approved?" without specifying the land's attributes), the system cannot proactively identify and complete the missing key constraints.

[0005] 2. Vector semantic similarity-based retrieval techniques (RAG's foundational retrieval): This method calculates vector similarity using pre-trained models (such as BERT). While it alleviates the semantic matching problem, it suffers from "intent drift" and "static retrieval" drawbacks. For complex multi-intent queries or fuzzy queries, a single vector representation struggles to capture all logical conditions, and the retrieval process is typically one-off, lacking the ability to perform multi-step reasoning and self-correction based on initial results.

[0006] In summary, existing technologies lack the ability to structurally model users' deep intentions and lack proactive interaction and progressive reasoning mechanisms when facing fuzzy queries, making it difficult to meet the requirements for retrieval accuracy and recall in complex business scenarios. Summary of the Invention

[0007] This invention aims to solve the aforementioned technical problems by providing an efficient knowledge retrieval system and method based on user intent. The invention constructs a structured conditional decision graph (CDG) and utilizes a beam search-based active interaction mechanism to simulate the expert reasoning process, guiding users to complete fuzzy information, thereby achieving accurate knowledge retrieval.

[0008] The first aspect of this invention discloses an efficient knowledge retrieval method based on user intent; the method includes:

[0009] Step S1: Receive the user's initial natural language query, perform semantic parsing on the initial natural language query, and extract the initial set of known conditions;

[0010] Step S2: Based on the initial set of known conditions and combined with a preset domain knowledge base, construct or expand a conditional decision graph; the conditional decision graph is a directed graph that contains nodes representing business conditions and edges representing dependencies between conditions;

[0011] Step S3: Generate multiple candidate reasoning paths in the conditional decision graph, and perform iterative interactive optimization of the candidate reasoning paths based on the bundle search strategy;

[0012] The iterative interactive optimization includes: calculating the information gain of missing conditions in the current candidate inference path, selecting conditions with high information gain to generate clarification questions and interact with the user, updating the comprehensive score of the condition decision graph and candidate inference path based on user feedback, until a preset termination condition is met, and determining the optimal inference path.

[0013] Step S4: Generate the final retrieval query based on the verified condition set in the optimal reasoning path, perform the retrieval in the domain knowledge base, and synthesize the final answer to return to the user.

[0014] Preferably, in step S1, the semantic parsing adopts a sequence labeling and relation extraction model based on a pre-trained language model; the conditions in the initial known condition set include entity, attribute, or relation triplet;

[0015] In step S2, the construction process of the conditional decision graph includes:

[0016] The initial set of known conditions is used as the seed node;

[0017] Breadth-first search is used to match relevant conditions from the domain knowledge base as expansion nodes;

[0018] The semantic similarity between nodes is calculated to perform deduplication and fusion, and directed edges and initial edge weights are established based on logical dependencies or statistical co-occurrence probabilities in domain knowledge.

[0019] Preferably, in step S3, the formula for calculating the comprehensive score Total(p,t) is:

[0020]

[0021] Where p represents the path and t represents the current round;

[0022] S prior(p) represents the path prior score, calculated as the weights λ of all edges on the path. ij The product;

[0023] S coverage (p,t) is the coverage score, representing the proportion of verified conditions out of the total conditions for the path;

[0024] S specificity (p,t) is the specificity score, calculated based on the inverse document frequency (IDF) of the path conditions;

[0025] |C missing (p,t)∣ represents the number of missing conditions on the path;

[0026] α is the missing condition penalty factor, and β is the balance coefficient.

[0027] Preferably, in step S3, the information gain is calculated based on the expected change in path scores; for path p and missing condition c, the information gain IG(p,c) is calculated as follows:

[0028] IG(p,c)=∑ v∈{True,False} P(c=v)·∣Stotal(p∣c=v)-Stotal(p)∣;

[0029] Where P(c=v) is the prior probability that condition c takes the value v, and Total(p|c=v) is the estimated comprehensive score of path p after assuming condition c takes the value v;

[0030] The system selects the condition with the highest weighted information gain to generate a natural language question to ask the user.

[0031] Preferably, in step S3, updating the conditional decision graph based on user feedback includes:

[0032] If the user's feedback is positive, mark the corresponding condition node as verified;

[0033] If the user feedback is negative, perform soft pruning: retain the condition node, but multiply the weight of all edges connected to the condition node by a preset penalty factor γ, thereby reducing the prior score of the path passing through the node.

[0034] If the user provides a specific value, create a new condition node and associate it with the current node.

[0035] Preferably, in step S4, the retrieval process employs a hybrid retrieval and fusion ranking strategy:

[0036] The optimal reasoning path is reconstructed into a structured query statement using a large language model;

[0037] Semantic retrieval based on vector embedding and inverted index retrieval based on keyword matching are performed respectively to obtain two sets of candidate documents;

[0038] The Reciprocal Ranking (RRF) algorithm is used to fuse the two sets of candidate documents.

[0039] The fused documents are finely sorted using the Cross-Encoder model, and the Top-N documents are selected as reference materials to be input into the large language model to generate the answer.

[0040] Preferably, the edge weight λ ij Dynamic updates are performed based on Bayesian smoothing, and the formula is as follows:

[0041]

[0042] Where Count(c i ∧c j ) represents condition c in historical interactions. i and c j The number of times the numbers co-occur on the successful path, Count(c i ) is condition c i The total number of occurrences, μ is the smoothing hyperparameter, λ prior This represents the prior probability.

[0043] A second aspect of this invention discloses an efficient knowledge retrieval system based on user intent; the system employs the method described in the first aspect above, and the system includes:

[0044] The intent parsing module is configured to receive the user's initial natural language query, use a pre-trained language model to perform entity recognition and relation extraction, and output an initial set of known conditions.

[0045] The graph construction and maintenance module is configured to store and manage conditional decision graphs, and dynamically expand the graph structure based on the initial set of known conditions and the domain knowledge base.

[0046] The interactive path exploration module is configured to maintain a set of candidate reasoning paths based on a bundle search strategy, calculate the information gain of missing conditions, generate clarification questions, and prune or expand the conditional decision graph based on user feedback until the optimal reasoning path is locked.

[0047] The dynamic retrieval and synthesis module is configured to generate the final query based on the optimal reasoning path, perform a hybrid retrieval of vectors and keywords, and use a large language model to generate the final answer based on the retrieval results.

[0048] A third aspect of this invention discloses an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the efficient knowledge retrieval method based on user intent according to any one of the first aspects of this disclosure.

[0049] A fourth aspect of this invention discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of an efficient knowledge retrieval method based on user intent, as described in any of the first aspects of this disclosure.

[0050] The beneficial effects of this invention are as follows:

[0051] 1. Structured Intent Modeling: By using conditional decision graphs, unstructured queries are transformed into a computable search space, solving the problem that complex logic cannot be captured by vectors.

[0052] 2. Proactive interactive clarification: Utilizing an information gain-driven questioning strategy, it can proactively identify ambiguities and guide users to clarify them, just like an expert, significantly improving the resolution rate of fuzzy queries.

[0053] 3. Efficient reasoning: By introducing a bundle search mechanism, the most promising K paths are always retained in multiple rounds of interaction, avoiding the combination explosion of local optima in single-path greedy algorithms and full-graph search, thus achieving a balance between efficiency and accuracy.

[0054] 4. Dynamic Adaptation: Supports dynamic updates of graph structure (soft pruning and expansion) and edge weight learning based on user feedback, and the system has the ability to continuously optimize. Attached Figure Description

[0055] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0056] Figure 1 A flowchart illustrating an efficient knowledge retrieval method based on user intent according to an embodiment of the present invention;

[0057] Figure 2 This is a flowchart illustrating the overall architecture and data processing of the system of the present invention.

[0058] Figure 3 This is a flowchart of the algorithm for constructing a Conditional Decision Graph (CDG).

[0059] Figure 4This is a flowchart of the active interaction algorithm based on beam search.

[0060] Figure 5 This is a structural diagram of an efficient knowledge retrieval system based on user intent according to an embodiment of the present invention;

[0061] Figure 6 This is a structural diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] The core of this invention lies in simulating the thinking process of human experts: when faced with an incomplete problem, several possible problem-solving paths (reasoning paths) are first formed, and then some paths are verified or eliminated by actively asking questions, and finally the accurate answer is found by following the most likely path.

[0064] Example 1:

[0065] The first aspect of this invention discloses an efficient knowledge retrieval method based on user intent. Figure 1 This is a flowchart of an efficient knowledge retrieval method based on user intent according to an embodiment of the present invention, such as... Figure 1 and 2 As shown, the method includes:

[0066] Step S1: Receive the user's initial natural language query, perform semantic parsing on the initial natural language query, and extract the initial set of known conditions;

[0067] In step S1, the semantic parsing adopts a sequence labeling and relation extraction model based on a pre-trained language model; the conditions in the initial known condition set include entity, attribute, or relation triplet.

[0068] This application employs a method that combines pattern-based text mining with statistical learning.

[0069] Pattern matching: Defines a set of syntactic patterns to capture explicit dependencies. For example:

[0070]

[0071] For text where no explicit pattern is matched, analyze the co-occurrence of conditions across a large number of documents. Use Point Mutual Information (PMI) or causal discovery algorithms (such as PC) to determine dependency direction. The basic idea is: if the occurrence of condition A significantly increases the probability of condition B appearing in subsequent paragraphs or related chapters, rather than the other way around, then we tend to establish an edge A→B. In the initial stage, edges without a clear direction can be set as bidirectional or pre-defined based on domain knowledge.

[0072] Step S2: Based on the initial set of known conditions and combined with a preset domain knowledge base, construct or expand a conditional decision graph; the conditional decision graph is a directed graph that contains nodes representing business conditions and edges representing dependencies between conditions;

[0073] In step S2, the construction process of the conditional decision graph includes:

[0074] The initial set of known conditions is used as the seed node;

[0075] Breadth-first search is used to match relevant conditions from the domain knowledge base as expansion nodes;

[0076] The semantic similarity between nodes is calculated to perform deduplication and fusion, and directed edges and initial edge weights are established based on logical dependencies or statistical co-occurrence probabilities in domain knowledge.

[0077] This invention employs a multi-level strategy hybrid approach, balancing accuracy and efficiency, rather than relying on a single rule. See below for details:

[0078] Level 1: Rule base matching: Predefine a thesaurus / near-synonyms (which can be extracted from domain terminology standards), for example, {"within the red line", "within the ecological protection red line", "ecological red line range"} are mapped to the same standardized expression "located within the ecological protection red line".

[0079] Level 2: Semantic Vector Matching: For new expressions not covered by the rule base, the Sentence-BERT model with domain fine-tuning is used to calculate their sentence vectors. If the cosine similarity of the vectors of two conditions is higher than the preset threshold θ (e.g., 0.85), they are considered to be the same condition.

[0080] Level 3: Manual Review Queue: For ambiguous cases with similarity in the range [θ-Δ,θ] (e.g., [0.75, 0.85]), the system adds them to the review queue, where domain experts ("document administrator" role) make the final decision. The result of the decision is fed back to update the rule base.

[0081] This hierarchical strategy ensures the consistency of core concepts while allowing the system to learn and expand during operation, which aligns with the "dynamically updatable" feature described in the patent.

[0082] This application adopts a "semi-automatic expert-guided" construction method, rather than a completely unsupervised one, to ensure the accuracy of the initial graph, which is more feasible in practical systems.

[0083] 1. Seed Node Extraction: Extract a batch of seed condition nodes (such as "Land Type is Cultivated Land" and "Area is Greater Than the Minimum Map Patch Standard") from the table of contents, chapter titles and core clauses of core domain documents (such as "Technical Regulations for Land Change Survey" and "National Standard for Land Use Classification") through rules and key phrases.

[0084] 2. Relationship labeling (expert participation): Invite 3-5 domain experts to use graphical labeling tools to label the relationships between these seed nodes, clarifying the direction of the edges and the initial weights (experts can provide three levels of weights based on experience: high, medium, and low, corresponding to quantization of 0.9, 0.6, and 0.3).

[0085] 3. Graph Expansion (Automated): Using this expert-verified seed graph as the core, the BuildCDG algorithm is run. Starting from these seed nodes, the algorithm automatically scans the entire document library, using the aforementioned "node deduplication" and "edge identification" methods to expand new nodes and edges. For each newly expanded edge, its initial weight is... It can be set to a default value (such as 0.5), or assigned a value based on the confidence level of the associated textual evidence.

[0086] Step S3: Generate multiple candidate reasoning paths in the conditional decision graph, and perform iterative interactive optimization of the candidate reasoning paths based on the BeamSearch strategy;

[0087] The iterative interactive optimization includes: calculating the information gain of missing conditions in the current candidate inference path, selecting conditions with high information gain to generate clarification questions and interact with the user, updating the comprehensive score of the condition decision graph and candidate inference path based on user feedback, until a preset termination condition is met, and determining the optimal inference path.

[0088] In step S3, the formula for calculating the comprehensive score Total(p,t) is as follows:

[0089]

[0090] Where p represents the path and t represents the current round;

[0091] S prior (p) represents the path prior score, calculated as the weights λ of all edges on the path. ij The product;

[0092] S coverage (p,t) is the coverage score, representing the proportion of verified conditions out of the total conditions for the path;

[0093] S specificity (p,t) is the specificity score, calculated based on the inverse document frequency (IDF) of the path conditions;

[0094] |C missing (p,t)∣ represents the number of missing conditions on the path;

[0095] α is the missing condition penalty factor, with a value range of (0,1], and a preferred value of 0.8. β is the balance coefficient, used to balance the coverage score and the specific score, with a value range of [0,1], and a preferred value of 0.7.

[0096] In step S3, the information gain is calculated based on the expected change in path scores; for path p and missing condition c, the information gain IG(p,c) is calculated as follows:

[0097] IG(p,c)=∑ v∈{True,False} P(c=v)·∣Stotal(p∣c=v)-Stotal(p)∣;

[0098] Where P(c=v) is the prior probability that condition c takes the value v, and Total(p|c=v) is the estimated comprehensive score of path p after assuming condition c takes the value v;

[0099] The system selects the condition with the highest weighted information gain to generate a natural language question to ask the user.

[0100] In step S3, updating the conditional decision graph based on user feedback includes:

[0101] If the user's feedback is positive, mark the corresponding condition node as verified;

[0102] If the user feedback is negative, perform soft pruning: retain the condition node, but multiply the weight of all edges connected to the condition node by a preset penalty factor γ, thereby reducing the prior score of the path passing through the node.

[0103] If the user provides a specific value, create a new condition node and associate it with the current node.

[0104] Wherein, the edge weight λ ij Dynamic updates are performed based on Bayesian smoothing, and the formula is as follows:

[0105]

[0106] Where Count(c i ∧c j ) represents condition c in historical interactions. i and c j The number of times the numbers co-occur on the successful path, Count(c i) is condition c i The total number of occurrences, μ is the smoothing hyperparameter, λ prior This represents the prior probability.

[0107] The initial weight of each edge during a system cold start with no historical interaction data. It is set to 0.5. This is a neutral prior, meaning "the strength of the dependency cannot be determined". This weight will be updated according to Bayesian principles in subsequent interactions.

[0108] The statistical object: Count(ci∧cj) is defined as the number of times that conditions ci and cj are both true on the ultimately determined optimal path pbest in all question-and-answer sessions that are judged as "successfully answered" by the system. Count(ci) is the number of times condition ci appears on the optimal path in all these successful sessions.

[0109] Bayesian Update Explanation: Weight λ ij Essentially, it is the conditional probability P(c) j =True|c i =True) estimate. We use Bayesian estimation (which is Laplace smoothed) to update it because it can smooth out small sample data and avoid extreme probabilities of 0 or 1.

[0110]

[0111] Molecular Count(c i ∧c j )+Γ: can be understood as observing c i Caused c j The virtual number of times it is true. Γ (e.g., set to 1) is the "prior success count", representing the number of times c is true when there is no data. i It may lead to c j Established.

[0112] Denominator Count(c i +Δ: This can be understood as the virtual total number of times ci is observed. Δ (usually Δ = 2Γ, if set to 2) is the total number of prior attempts, which determines the strength of the prior.

[0113] This formula is equivalent to our initial "prior belief" of Γ successes and (Γ-Δ) failures. As real interaction data (Count) accumulates, the weight estimates gradually converge to the actual observed frequencies. The introduction of Laplace smoothing ensures that even if a condition ci occurs infrequently in the early stages, its corresponding λ value will not become extreme (0 or 1) due to a single accidental co-occurrence or non-co-occurrence, thus improving the robustness of the system.

[0114] Furthermore, the pruning logic for "BeamSearch" is as follows:

[0115] 1. The main strategy of fault tolerance and backtracking mechanism: path score penalty rather than hard deletion. When the user answers "No" to condition c, we do not immediately physically delete all relevant edges in the graph, but instead adopt a soft pruning strategy:

[0116] 1) Significantly reduce the weights of relevant edges: Multiply the weights λ of all edges connected to c as the truth by a penalty factor γ (e.g., 0.1), so that the prior scores s of paths passing through these edges are reduced. prior A sharp decline.

[0117] Here, γ specifically represents the soft pruning penalty factor, which is used to multiply the weight of the related edge when the user negates a certain condition. The value range is (0,1), and the preferred value is 0.1.

[0118] 2) Path marking instead of deletion: Paths containing condition c, where c is rejected by the user, are marked as "questionable". This is reflected in their overall score S. total An additional, huge penalty is introduced, causing its ranking to plummet to the bottom of Beam, and it will be naturally eliminated in the next iteration (it will not be selected in the Top-K selection).

[0119] 3) "Undo" Mechanism: The system retains complete CDG and historical interaction logs. If subsequent dialogues indicate that a previous negative answer may have been incorrect, the system administrator (or advanced user) can revoke the acceptance of that negative answer through background operations. The system will then recalculate the weights and path scores of the affected edges based on the corrected information, and the "soft-pruned" paths may re-enter the candidate set. This demonstrates the system's fault tolerance and learning capabilities.

[0120] 2. Complete termination conditions:

[0121] The beam search interaction loop terminates when any of the following conditions are met:

[0122] 1) Success condition: There exists a path p in Beamt, whose... (All conditions are met).

[0123] 2) Resource constraints: The number of interaction rounds t reaches the preset maximum number of rounds T (for example, T = 5 to avoid infinite loops), where T represents the maximum number of interaction rounds, a positive integer, with typical values ​​of 3 ≤ T ≤ 8.

[0124] 3) Search space exhaustion condition: Beamt is an empty set (all paths are pruned or have too low a score).

[0125] 4) User actively exits: The user clicks "Stop asking questions" or enters "Skip" or other commands.

[0126] 5) Confidence threshold (optional enhancement): The path ranked first in Beamt has a composite score S. total If the information gain of the missing information is above a certain absolute threshold (e.g., 0.8) and the information gain of the missing information is below a certain threshold, the system can "confidently" use the current information for retrieval without asking any further questions.

[0127] Step S4: Generate the final retrieval query based on the verified condition set in the optimal reasoning path, perform the retrieval in the domain knowledge base, and synthesize the final answer to return to the user.

[0128] In step S4, the retrieval process employs a hybrid retrieval and fusion ranking strategy:

[0129] The optimal reasoning path is reconstructed into a structured query statement using a large language model;

[0130] Semantic retrieval based on vector embedding and inverted index retrieval based on keyword matching are performed respectively to obtain two sets of candidate documents;

[0131] The Reciprocal Ranking (RRF) algorithm is used to fuse the two sets of candidate documents.

[0132] The fused documents are finely sorted using the Cross-Encoder model, and the Top-N documents are selected as reference materials to be input into the large language model to generate the answer.

[0133] Example 2:

[0134] A second aspect of this invention provides an efficient knowledge retrieval system based on user intent, such as... Figure 5 As shown, the system adopts the method in Embodiment 1. The system 100 of the present invention mainly includes the following core modules:

[0135] The intent parsing module 101 is configured to receive the user's initial natural language query, perform entity recognition and relation extraction using a pre-trained language model, and output an initial set of known conditions; specifically, it is responsible for converting the user's query into structured initial conditions and initializing the search space.

[0136] The graph construction and maintenance module 102 is configured to store and manage conditional decision graphs, and dynamically expand the graph structure based on the initial set of known conditions and the domain knowledge base; specifically, it uses a BERT-BiLSTM-CRF model chain for entity and relation extraction.

[0137] The interactive path exploration module 103 is configured to maintain a set of candidate reasoning paths based on a beam search strategy, calculate the information gain of missing conditions, generate clarification questions, and prune or expand the conditional decision graph based on user feedback until the optimal reasoning path is locked; the core control unit maintains a path set (Beam) of size K, and determines the questioning strategy for the next round by calculating the score (Stotal) and information gain (IG).

[0138] The dynamic retrieval and synthesis module 104 is configured to generate the final query based on the optimal inference path, perform a hybrid retrieval of vectors and keywords, and generate the final answer based on the search results using a large language model. It is responsible for the final document retrieval and answer generation. It integrates the Milvus vector database, the Elasticsearch search engine, and the GPT series of large language models.

[0139] The calculation of information gain IG(p,c) is key to achieving efficient interaction in this method. Its core idea is to evaluate how much clarifying a missing condition c contributes to reducing the uncertainty of path p.

[0140] Calculation formula and steps:

[0141] For a given path p and its missing condition c, instead of calculating traditional entropy, we define information gain based on the expected change in path scores, which is more consistent with our scoring model:

[0142]

[0143] Among them, S′ total (p,t|c=v) is the expected score of path p when the condition c takes the value v.

[0144] Specific calculation example:

[0145] Assume that at the current time t, there is a path p with a current overall score S. total (p,t) = 0.6. It has two missing conditions: cA (whether it is inside the red line) and cB (whether the area is greater than 1 hectare).

[0146] Estimate based on prior knowledge (or simple statistics) in CDG:

[0147] P(c A =True) = 0.2, P(c A =False) = 0.8;

[0148] P(c B =True)=0.5, P(c B =False) = 0.5;

[0149] Simultaneously, the simulation evaluated the path scores under different conditions:

[0150] If cA = True, the path score may jump to S′. total (p|c A =T) = 0.9;

[0151] If cA = False, the path score may plummet to S′. total (p|c A =F) = 0.1;

[0152] If cB = True, the path score becomes S′. total (p|c b =T) = 0.7;

[0153] If cB = False, the path score becomes S′. total (p|c B =F) = 0.5.

[0154] Calculate the information gain of cA:

[0155] IG(p,c A )=P(c A =T)×|0.6-0.9|+P(c A =F)×|0.6-0.1|=0.2×0.3+0.8×0.5

[0156] =0.06 + 0.40 = 0.46

[0157] Calculate the information gain of cB:

[0158] IG(p,c B )=P(c B =T)×|0.6-0.7|+P(c B =F)×|0.6-0.5|=0.5×0.1+0.5×0.1

[0159] =0.05 + 0.05 = 0.10

[0160] Analysis and Conclusion: Calculations show that IG(p,cA) = 0.46 is significantly greater than IG(p,cB) = 0.10. This means that the clarification condition cA (whether it is within the red line) plays a decisive role in determining the final fate of this path: a "yes" answer greatly increases the path's probability, while a "no" answer almost negates it. The clarification condition cB has little impact on the score. Therefore, the system prioritizes asking questions related to cA, such as: "Is the plot you inquired about located within the ecological protection red line?" This example clearly demonstrates how IG calculation quantifies the "value" of a question, thereby intelligently and efficiently guiding the dialogue, obtaining the maximum amount of information with the fewest questions—a concentrated manifestation of the algorithm's ingenious design.

[0161] Example 3: Key Algorithms and Processes

[0162] 1. Data Structure and Construction of Conditional Decision Graph (CDG)

[0163] CDG is defined as G = (V, E, Λ), where V is a condition node (e.g., "located within the ecological red line"), E is a directed edge, and Λ is the set of edge weights.

[0164] 11) Construction logic: Using the parsed initial conditions as seeds, expand relevant nodes from the domain knowledge base through breadth-first search (BFS).

[0165] 12) Node deduplication: A hierarchical strategy is adopted.

[0166] Level 1 is rule base matching (thesaurus); Level 2 is Sentence-BERT semantic vector similarity matching (threshold θ = 0.85); θ specifically represents the node semantic similarity threshold, used to determine whether two conditions are equivalent, with a value range of [0.7, 0.95], and a preferred value of 0.85. Level 3 is manual review.

[0167] 13) Weight calculation: Edge weight λ ij = P(target|source), with an initial value of 0.5 or based on expert annotations (0.9 / 0.6 / 0.3). As the system runs, the weights are updated using the Bayesian smoothing formula:

[0168]

[0169] Where μ = 2 is the smoothing parameter.

[0170] 2. Path Comprehensive Scoring Model

[0171] In each round of interaction t, for path p, its total score Statotal(p,t) is calculated as follows:

[0172]

[0173] in,

[0174] S prior (p)=Πλ ij Represents the prior probability of the path;

[0175] λ ij The weights of directed edges in the conditional decision graph represent conditional probabilities, with values ​​ranging from [0,1].

[0176] S coverage : Number of verified conditions / Total number of path conditions.

[0177] S specificity The average IDF value of all conditions on the path, encouraging the selection of paths that contain more specific information.

[0178] α = 0.8: Missing condition penalty factor; the more missing conditions, the lower the score.

[0179] β = 0.7: Balance factor.

[0180] 3. Problem selection based on information gain

[0181] To obtain the maximum amount of information with the fewest questions, the system calculates the information gain IG for the missing condition c:

[0182] IG(p,c)=∑ v∈{T，F} P(c=v)·S total (p|c=v)-S total (p);

[0183] The system selects the condition with the highest weighted IG to generate a natural language question. For example, if it is confirmed that "whether it is within the red line" can cause a drastic change in the path score (either extremely high or extremely low), then that question will be asked first.

[0184] 4. Soft pruning and dynamic renewal

[0185] When a user answers "No" to a condition, the system does not directly delete the node, but instead performs soft pruning: the weight of the edge connected to the condition is multiplied by γ = 0.1. This causes the score of the path containing the condition to drop sharply, thus naturally eliminating it in the next round of bundle search, but retains the possibility of "regret" (i.e., if it is found later that the user answered incorrectly, the weight can be restored).

[0186] Example 4: Real-world application scenario

[0187] According to this method, in some specific embodiments, the following is true:

[0188] I. Data Processing and Initial Condition Decision Graph Construction

[0189] Step 10: Semantic parsing to obtain the initial set of known conditions

[0190] Input: The initial query Q (natural language string) entered by the user.

[0191] Method: A finely tuned joint model for named entity recognition and relation extraction is invoked. This model uses a pre-trained language model BERT as the encoder, followed by a BiLSTM-CRF layer for entity sequence labeling, and outputs the relationships between entities through a relation classification head.

[0192] Output: The initial set of known conditions Each condition is either a structured triple (entity 1, relation, entity 2) or a unary assertion. For example, for the query "How should farmland within the ecological red line be changed?", the parsing yields:

[0193] (The land parcel is located within the ecological protection red line) → Condition c1

[0194] (Current land type: arable land) → Condition c2

[0195] (Intent: Change of land use category) → Condition c3

[0196] Specifically, the model was fine-tuned on a dataset containing approximately 5,000 domain sentences, 30,000 labeled entities, and 8,000 sets of relations. The entity labeling system includes B-GEO-FEATURE (geographic features) and B-LAND-CLASS (land categories); the relation labels include LOCATED_IN (located in) and IS_A (is).

[0197] Step 20: Constructing and expanding the conditional decision graph

[0198] Input: Initial set of known conditions Domain knowledge base.

[0199] Method: Execute a conditional decision graph construction algorithm.

[0200] Data Structure: The Conditional Decision Graph (CDG) is stored in memory as an adjacency list. Each ConditionNode contains a unique identifier (id), a natural language description (desc), a type, and a verification status (is_verified).

[0201] Edge weight: Each directed edge ConditionalEdge contains source / target node IDs and weight λ∈[0,1], representing the conditional probability P(target=True|source=True).

[0202] Build process:

[0203] 1) The conditions in the graph are used as initial nodes to add to graph G.

[0204] 2) Initialize a frontier queue to be expanded, containing these initial nodes.

[0205] 3) When Frontier is not empty, pop node c.

[0206] 4) Query all RelatedConditions related to c from the domain knowledge base.

[0207] 5) For each relevant condition cnew:

[0208] a. Node deduplication: If the semantic similarity between the description of cnew and the existing nodes in the graph (cosine similarity calculated by the domain-fine-tuned Sentence-BERT model) is lower than the threshold θ = 0.85, then it is considered a new node and added to graph G and Frontier.

[0209] b. Add an edge: Add a directed edge from c to cnew. The initial weight of the edge λinit is set to 0.5. If a clear logical dependency can be parsed from the knowledge base (such as "if A then B"), then it is set to 0.9.

[0210] Output: A preliminary expanded conditional decision graph G.

[0211] II. Interactive Path Exploration Driven by Beam Search

[0212] Step 30: Generate the initial inference path set and initialize the bundle.

[0213] Input: Conditional decision graph G, known set of conditions

[0214] Method: Execute the path generation algorithm. From Starting from each condition node in the algorithm, perform a depth-first search on G until a "leaf node" with no outgoing edges (representing a possible result or final operation) is encountered, or the path length exceeds the limit L. max =5. Each path p is an ordered list of nodes.

[0215] Path scoring: Calculate the overall score s for each path p. total (p):

[0216]

[0217] S prior (p)=∏ e∈Edges(p) λ e The path prior is the product of edge weights.

[0218] Coverage score.

[0219] Specificity. N is the total number of nodes, and df(c) is the document frequency of condition c.

[0220] C missing (p): The set of unverified conditions on path p.

[0221] α = 0.8: Penalty coefficient for missing conditions.

[0222] β = 0.7: Balance coefficient between coverage and specificity scores.

[0223] Output: The initial path set P0 and its scores. Select the K highest-scoring paths from P0 (bundle width K = 3) to form the initial beam Beam0, where K represents the beam width, i.e., the number of candidate paths retained in each round, a positive integer, typically 2 ≤ K ≤ 5.

[0224] Step 40: Multi-round interaction, path evaluation and pruning

[0225] Input: Initial Beam t-1 The current set of known conditions Conditional decision graph G.

[0226] Method: This is a cyclical process that continues until the termination condition is met. The steps for each round are as follows:

[0227] Step 41: Candidate Problem Generation: For each path p in Beamt-1:

[0228] like Skip to step 5.

[0229] Otherwise, calculate the information gain IG(p,c) for each missing condition c∈Cmissing(p).

[0230]

[0231] Where P(c=T) and P(c=F) are the prior probabilities of the condition being true / false (initially calculated from the knowledge base or both set to 0.5), S′ total (p|c=v) is the estimated score of path p assuming c takes the value v.

[0232] The condition for selecting the maximum value of IG(p,c) is c. * .

[0233] Calling a large language model (such as GPT-5) to generate C++ * The natural language question q (e.g., given the condition "located within permanent basic farmland", generate "Is this plot of land within the scope of permanent basic farmland?").

[0234] (q,c) * ,p) Add to the candidate problem set.

[0235] Step 42: Question Selection and Questioning: Select candidate questions according to condition c * Grouping. Calculate the total weighted information gain for each group: GroupScore(c) = ∑ (q,c,p)∈Group(c) IG(p,c)×S total (p). Ask the user a question corresponding to the condition with the highest GroupScore.

[0236] Step 43: Process user feedback: Obtain user response At.

[0237] If At is "Yes" / "No": Update the validation status of condition c* and add...

[0238] If "No", perform soft pruning: multiply the weight λ of all edges in G that originate from c* by the penalty factor γ = 0.1.

[0239] If At is a "specific value" (such as "paddy field"): Create a new condition node, associate it with c*, and mark it as verified.

[0240] If At is "Don't know" or "Uncertain": The system temporarily sets the confidence level of condition c* to zero, and it will no longer be used as a key decision-making basis in this session, but it will remain in the graph. The system will then select the condition with the second highest information gain from the other missing conditions of this path to ask a question.

[0241] Step 44: Update the bundle: based on the updated G and Recalculate the Cmissing and Total scores for all paths. Select the K highest-scoring paths from all active paths to form a new Beamt bundle.

[0242] Termination conditions (if any one of these conditions is met, the loop will exit):

[0243] 1) A certain path in Beamt

[0244] 2) The maximum number of interaction rounds T = 5 is reached.

[0245] 3) Beamt is empty.

[0246] 4) The user terminates the service voluntarily.

[0247] Output: pbest optimal path (or the path with the highest score in the last round).

[0248] III. Precise Search and Answer Generation

[0249] Step 50: Generate the final query and hybrid search

[0250] Input: optimal path pbest.

[0251] Method: Input all validated condition descriptions from pbest into a large language model (such as ChatGLM) for query reconstruction, generating a final query statement Qfinal with a clear structure and rich keywords. For example: "Special provisions for the approval process of changes in the current cultivated land category within the ecological protection red line".

[0252] Hybrid Search:

[0253] 1) Vector retrieval: Use the bge-large-zh model to vectorize Qfinal, retrieve it from the Milvus vector database, and return the top-100 relevant document fragments.

[0254] 2) Keyword search: Use Elasticsearch (BM25 algorithm) to search Qfinal and return the top 100 relevant document fragments.

[0255] 3) Result fusion: The two result lists are merged using the inverse fusion ranking method, and the Top-50 are taken as the candidate set Dcandidates.

[0256]

[0257] Step 60: Fine sorting and answer synthesis

[0258] Input: Candidate document set Dcandidates, final query Qfinal, original user question Q.

[0259] Method: Fine sorting was performed. A cross-encoder model (such as cross-encoder / ms-marco-MiniLM-L-12-v2) was used to score the relevance of each document d in Qfinal and Dcandidates, and the documents were sorted in descending order of score. The top-5 documents were selected as the final reference segments.

[0260] Answer generation: Input the original question Q, the final query Qfinal, and the Top-5 reference fragments into a large language model with strong instruction compliance capabilities (such as GPT-5), and ask it to generate a professional, accurate, and source-marked final answer. The prompt explicitly indicates that citations should use the format "[Reference 1]".

[0261] In some natural resource sectors, the above methods are implemented in the following scenario: A user (natural resource surveyor) enters a fuzzy query: "Can aquaculture be carried out on land within the red line zone where fruit trees are planted?"

[0262] Step 110: Semantic parsing: The system parses the initial conditions:

[0263] c1: (Plot of land, located within the ecological protection red line)

[0264] c2: (Current land use: planting fruit trees)

[0265] c3: (User intent: Yes, to build aquaculture facilities)

[0266] Step 210: Graph Construction and Path Generation: The system expands the CDG from the domain knowledge base and generates initial paths. Assume two main paths are generated:

[0267] Path P1: [c1, c4 (belonging to forest land / orchard land), c5 (policy on building a breeding farm in forest land), c6 (requires approval from forestry department)].

[0268] Path P2: [c1, c7 (belonging to arable land), c8 (restriction on "non-grain conversion" of arable land), c9 (prohibition of damaging the topsoil), c10 (aquaculture facilities are construction activities)].

[0269] Initial scores: P1(0.7), P2(0.6). Beam width K=2, therefore Beam0={P1,P2}.

[0270] Step 310: First round of interaction:

[0271] Information gain calculation: For P1, calculate the information gain (IG) of missing conditions c4 (whether it belongs to woodland / orchard) and c5, and find that c4 has a higher IG. For P2, calculate the IG of c7 (whether it belongs to cultivated land).

[0272] Question Selection: Compare the GroupScores of c4 and c7, assuming c4 is higher. The system asks: "Regarding the 'land parcel with fruit trees' you mentioned, is the current land use classification in the land survey 'orchard' (belonging to garden land) or some other land use category?"

[0273] User feedback: The user replied, "It is an orchard identified by the Third National Land Survey."

[0274] System update: c4 is confirmed to be true (belonging to the orchard). Edge weights related to "orchard" in CDG are retained. P1's score increases to 0.85, while P2's score decreases to 0.4 due to its weaker association with "orchard". After the update, Beam1 = {P1} (P2 is eliminated).

[0275] Step 410: Second round of interaction:

[0276] The system follows path P1 to calculate the IG of the next critical missing condition c5 (specific policies for building livestock farms in woodlands / orchards) and asks: "Is the livestock facility you plan to build a large-scale farm or a family-run farm?" (This question is used to refine the application of the policy).

[0277] The user replied: "Small-scale, like a family raising dozens of sheep."

[0278] System update: Based on the answer, the relevant policy node c11 for "small-scale family farming" has been located. The update path is P1'. At this point, all key conditions are complete (c1, c4, c11). The loop terminates, pbest = P1′.

[0279] Steps 510 and 610: Retrieval and Generation:

[0280] The system combines the conditions in pbest into a final query: "Approval policy for the construction of family-run free-range livestock and poultry facilities in orchards within the ecological protection red line".

[0281] Perform a hybrid search and fine sorting to find relevant legal provisions (such as a clause in the "Notice of the Ministry of Natural Resources on Regulating the Management of Aquaculture Facilities in Forest Land").

[0282] The answer generator synthesizes this information and generates the final answer: "According to the provisions of the 'Notice', the construction of family-run free-range breeding facilities in orchards within the ecological protection red line must first apply to the county-level natural resources and forestry authorities to complete the registration procedures for facility agriculture land, and ensure that the orchard planting layer and ecological environment are not damaged...[Document 1]".

[0283] In addition, regarding the hardware environment, the system can be deployed on a server cluster configured with Intel Xeon processors and NVIDIA A100 GPUs. Graph data is stored in Neo4j, vector data in Milvus, and text indexing uses Elasticsearch.

[0284] Furthermore, to demonstrate the practical effectiveness of this method, a comparative experiment was designed in a "fuzzy intent" scenario. This experiment focuses on high-frequency, fuzzy queries in the field of natural resource survey and monitoring.

[0285] Experimental setup:

[0286] Test set: A test set containing 200 typical fuzzy queries was collected and constructed from real-world business scenarios. These queries share the common characteristic of incomplete information, requiring further clarification. Examples include: "How do I get approval for this land?" and "Is it okay to plant this crop?".

[0287] Baseline scheme:

[0288] A: BM25: Based on the standard BM25 algorithm of Elasticsearch, it retrieves data directly using the raw query.

[0289] B: Traditional RAG (Vector Retrieval + Generation): Uses the bge-large-zh model for vector retrieval, without any interactive clarification mechanism.

[0290] C (Our current solution): CDG + beam search interactive retrieval: Using this solution, the maximum number of interactive rounds T = 5, and the beam width K = 3.

[0291] Evaluation indicators:

[0292] 1. Intent recognition accuracy: The degree of matching (F1 score) between the "user's true intent" (i.e. the set of conditions in pbest) finally determined by the system after the interaction and the standard answer labeled by the expert.

[0293] 2. Final answer satisfaction (human rating): The final answer generated by the system will be blind-rated by 3 domain experts, using a 5-point scale (1 point: completely irrelevant / incorrect; 5 points: completely accurate and detailed).

[0294] 3. Average number of interaction rounds: This refers only to the solution of this invention, and is the average number of questions required to successfully answer a fuzzy query.

[0295] The experimental results are shown in Table 1:

[0296] Table 1

[0297] Evaluation indicators Option A (BM25) Option B (Traditional RAG) Option C (this option) Intent recognition accuracy (F1) 0.32 0.41 0.89 Final answer satisfaction (average score) 1.8 2.5 4.3 Average number of interaction rounds N / A N / A 2.7

[0298] Data Analysis and Conclusions: The method of this invention improves the accuracy of intent recognition by 117% ((0.89-0.41) / 0.41) compared to the optimal baseline scheme (traditional RAG), and improves the answer satisfaction by 72%. This directly proves that this scheme fundamentally solves the persistent problem of poor retrieval performance under fuzzy queries through two core mechanisms: "structured intent parsing" and "proactive interactive clarification." This invention can accurately locate user intent with an average of only 2.7 rounds of interaction, proving the efficiency of bundled search guidance, avoiding lengthy and rambling dialogues, and improving user experience. This set of data strongly proves that the technical effect of this scheme is not a simple superposition of existing technologies (BM25+RAG+dialogue), but rather produces a synergistic gain, achieving superior technical results in solving the specific technical problem of "accurate identification of fuzzy intents."

[0299] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

[0300] A third aspect of this invention discloses an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the efficient knowledge retrieval method based on user intent, as disclosed in any of the first aspects of this invention.

[0301] Figure 6 This is a structural diagram of an electronic device according to an embodiment of the present invention, such as... Figure 6 As shown, the electronic device includes a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, Near Field Communication (NFC), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0302] Those skilled in the art will understand that Figure 6 The structure shown is only a structural diagram of the part related to the technical solution of this disclosure, and does not constitute a limitation on the electronic device to which the solution of this application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0303] A fourth aspect of this invention discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the efficient knowledge retrieval method based on user intent according to any one of the first aspects of this invention.

[0304] Please note that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A highly efficient knowledge retrieval method based on user intent, characterized in that, Includes the following steps: Step S1: Receive the user's initial natural language query, perform semantic parsing on the initial natural language query, and extract the initial set of known conditions; Step S2: Based on the initial set of known conditions and combined with a preset domain knowledge base, construct or expand a conditional decision graph; the conditional decision graph is a directed graph that contains nodes representing business conditions and edges representing dependencies between conditions; Step S3: Generate multiple candidate reasoning paths in the conditional decision graph, and perform iterative interactive optimization of the candidate reasoning paths based on the bundle search strategy; The iterative interactive optimization includes: calculating the information gain of missing conditions in the current candidate inference path, selecting conditions with high information gain to generate clarification questions and interact with the user, updating the condition decision graph and the comprehensive score of the candidate inference path based on user feedback, until a preset termination condition is met, and determining the optimal inference path. Step S4: Generate the final retrieval query based on the verified condition set in the optimal reasoning path, perform the retrieval in the domain knowledge base, and synthesize the final answer to return to the user.

2. The method according to claim 1, characterized in that, In step S1, the semantic parsing adopts a sequence labeling and relation extraction model based on a pre-trained language model; the conditions in the initial known condition set include entity, attribute, or relation triplet. In step S2, the construction process of the conditional decision graph includes: The initial set of known conditions is used as the seed node; Breadth-first search is used to match relevant conditions from the domain knowledge base as expansion nodes; The semantic similarity between nodes is calculated to perform deduplication and fusion, and directed edges and initial edge weights are established based on logical dependencies or statistical co-occurrence probabilities in domain knowledge.

3. The method according to claim 1, characterized in that, In step S3, the formula for calculating the comprehensive score Total(p,t) is as follows: Where p represents the path and t represents the current round; S prior (p) represents the path prior score, calculated as the weights λ of all edges on the path. ij The product; S coverage (p,t) is the coverage score, representing the proportion of verified conditions out of the total conditions for the path; S specificity (p,t) is the specificity score, calculated based on the inverse document frequency (IDF) of the path conditions; |Cmissing(p,t)| represents the number of missing conditions on the path, i.e., the power of the exponent; α is the missing condition penalty factor, and β is the balance coefficient.

4. The method according to claim 1, characterized in that, In step S3, the information gain is calculated based on the expected change in path scores; for path p and missing condition c, the information gain IG(p,c) is calculated as follows: IG(p,c)=∑ v∈{True,False} P(c=v)·∣Stotal(p∣c=v)-Stotal(p)∣; Where P(c=v) is the prior probability that condition c takes the value v, and Total(p|c=v) is the estimated comprehensive score of path p after assuming condition c takes the value v; The system selects the condition with the highest weighted information gain to generate a natural language question to ask the user.

5. The method according to claim 1, characterized in that, In step S3, updating the conditional decision graph based on user feedback includes: If the user's feedback is positive, mark the corresponding condition node as verified; If the user feedback is negative, perform soft pruning: retain the condition node, but multiply the weight of all edges connected to the condition node by a preset penalty factor γ, thereby reducing the prior score of the path passing through the node. If the user provides a specific value, create a new condition node and associate it with the current node.

6. The method according to claim 1, characterized in that, In step S4, the retrieval process employs a hybrid retrieval and fusion ranking strategy: The optimal reasoning path is reconstructed into a structured query statement using a large language model; Semantic retrieval based on vector embedding and inverted index retrieval based on keyword matching are performed respectively to obtain two sets of candidate documents; The Reciprocal Ranking (RRF) algorithm is used to fuse the two sets of candidate documents. The fused documents are finely sorted using the Cross-Encoder model, and the Top-N documents are selected as reference materials to be input into the large language model to generate the answer.

7. The method according to claim 2, characterized in that, The edge weight λ ij Dynamic updates are performed based on Bayesian smoothing, and the formula is as follows: Where Count(c i ∧c j ) represents condition c in historical interactions. i and c j The number of times the numbers co-occur on the successful path, Count(c i ) is condition c i The total number of occurrences, μ is the smoothing hyperparameter, λ prior This represents the prior probability.

8. A high-efficiency knowledge retrieval system based on user intent, characterized in that, include: The intent parsing module is configured to receive the user's initial natural language query, use a pre-trained language model to perform entity recognition and relation extraction, and output an initial set of known conditions. The graph construction and maintenance module is configured to store and manage conditional decision graphs, and dynamically expand the graph structure based on the initial set of known conditions and the domain knowledge base. The interactive path exploration module is configured to maintain a set of candidate reasoning paths based on a bundle search strategy, calculate the information gain of missing conditions, generate clarification questions, and prune or expand the conditional decision graph based on user feedback until the optimal reasoning path is locked. The dynamic retrieval and synthesis module is configured to generate the final query based on the optimal reasoning path, perform a hybrid retrieval of vectors and keywords, and use a large language model to generate the final answer based on the retrieval results.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the efficient knowledge retrieval method based on user intent according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the efficient knowledge retrieval method based on user intent according to any one of claims 1 to 7.

Images