Knowledge graph retrieval method based on multi-channel recall and relationship enhancement

Abstract

The invention provides a knowledge graph retrieval method based on multi-channel recall and relation enhancement, and belongs to the technical field of information retrieval. The invention provides a knowledge graph retrieval method and system integrating three-channel concurrent recall, unified scoring and quantified cutting, extraction type relation enhancement and double-threshold legality verification and path-level reasoning and unified sorting. Cross-channel score alignment and tracing are realized through a unified score registry; denoising is stabilized through a quantile threshold; the relation is reliably enhanced through a prompt chain, label mapping and a structure template; the node confidence, the edge weight, the structure regularity, the time recency, the service / regulation priority and the context relevancy are converged into a single path score through PathScore; and the front end presents in a'forest-path 'double view, and outputs an auditable evidence list. The system adopts a pluggable design of an event bus and an SPI (Serial Peripheral Interface), so that the mobility and the SLA under different compliance environments are ensured.

Description

Technical Field

[0001] This specification relates to the field of information retrieval technology, and in particular to a knowledge graph retrieval method based on multi-channel recall and relationship enhancement. Background Technology

[0002] Currently, when using large language models with knowledge graph-supported retrieval enhancement generation systems for natural language question-answering reasoning tasks, the retrieved graph information may contain noise or irrelevant paths, reducing the accuracy of answer generation in question-answering reasoning tasks. In addition, large language models rely excessively on external knowledge, neglecting the reasoning and knowledge capabilities they already possess, and may even be misled in scenarios where the model could have answered correctly, thus reducing the stability and performance of knowledge graph-based large language models.

[0003] A Chinese patent, publication number CN120950652A, discloses a video retrieval method for matching text information. The question-answering method includes: responding to a received natural language question, obtaining a corresponding candidate path set using a knowledge graph retrieval system; performing coarse and fine filtering on the candidate path set using a large language model to obtain a coarsely filtered path set and a target path set, constructing a structured prompt instruction, obtaining a first answer result output by the knowledge graph retrieval system and a second answer result output by the large language model, calculating their respective confidence scores, and filtering and fusing the first and second answer results according to their confidence scores to generate the target answer result corresponding to the natural language question information.

[0004] Traditional "inverted index + vector fine sorting" has problems with heterogeneous and weakly connected data, such as inconsistent dimensions across channels, fragmented results, lack of path-level interpretation, and difficulty in verification under compliance requirements. In high-concurrency and multi-compliance scenarios, it is difficult to balance coverage and accuracy, interpretability and latency. Summary of the Invention

[0005] This invention provides a knowledge graph retrieval method based on multi-channel recall and relation enhancement. It proposes a three-channel (inverted index + node vector + relation vector) concurrent recall and achieves cross-channel score calibration (channel calibration function), domain adaptive weighting (source weight) and posterior robustness processing through a unified score registry (NSR), unifying the units and solving the problem of instability in rearrangement. At the same time, it retains the source of scores and evidence pointers, laying the foundation for subsequent interpretability.

[0006] In some embodiments, the following steps are included: S1: Preprocessing and multi-vector embedding, specifically to complete hierarchical word segmentation, synonym mapping and syntax preservation, and generate multi-carrier vectors; S2: Three-channel concurrent recall, specifically including parallel recall of candidates through inverted index, nodes, and relationships; S3: Unified scoring and registration, specifically including calibrating and weighting the scores of each channel, combining them into a unified dimensional score, and writing it into the unified scoring registry; S4: Quantile pruning, which specifically includes calculating quantile thresholds for nodes and relationships respectively, and deleting low-quality long-tail candidates; S5: Relationship enhancement, specifically including outputting enhanced relationship information of the cue chain after hierarchical cue chain classification; S6: Structure and deduplication verification, specifically including verification of direction, depth and loop according to the structure template, deduplication within UID-Pair by leaf priority strategy, and setting thresholds for root / trunk / leaf layer; S7: Graph construction and path reasoning, specifically including constructing a graph with valid edges, constructing a forest with BFS units, enumerating paths with DFS units, and unifying the path ranking within the domain with unified path scores; S8: Output and Audit, specifically including generating roadmaps and evidence lists, and recording the evolution from channel scores to path scores; S9: Deployment and Operations, specifically including stateless microservices driven by event bus; hot switching of inverted index, vector, graph, and LLM encapsulated in SPI.

[0007] Through the aforementioned technical means, traditional retrieval only returns a "list of results matching keywords", while this technology returns "path-based results with logical relationships, evidence chains, and auditable characteristics" (such as the risk path of "Company A → Guarantee → Company B → Debt Default"). The core solution of this application is to accurately and interpretably find related information and logical paths from massive amounts of heterogeneous knowledge (text, structured data, etc.).

[0008] Furthermore, in step S3, the unified scoring registry contains records of the candidate's unified dimensional score, source channel, weight, timestamp, evidence pointer, domain identifier, and context gating factor, and provides quantile statistical processing.

[0009] Using the aforementioned technical methods, the scores from different channels are converted into scores between 0 and 1.

[0010] Furthermore, step S3 also includes normalizing or equivalently calibrating the scores of different retrieval channels so that they all fall on a unified number axis, and updating the parameters in combination with historical statistics or validation sets.

[0011] Furthermore, in step S4, the quantile threshold includes quantile thresholds calculated separately for each node and relationship.

[0012] By using the above-mentioned technical means, a score threshold is set, only candidates with high scores are retained, and obviously irrelevant low-value candidates are removed. Based on the statistical distribution of candidate scores, high-quality candidates are screened, low-quality long tails and noise are eliminated, and the computing power consumption of subsequent stages is reduced. The core is accurate noise reduction based on quantiles.

[0013] Furthermore, in step S6, the structural template includes a set of rules for relational direction, maximum depth, loop constraints, and hierarchical compatibility matrix.

[0014] Furthermore, the unified path score in step S7 also includes synthesizing factors into a single scalar for unified ranking of paths of different lengths and components. The factors include node confidence, edge weight, structural regularity, temporal recency, regulatory priority, and contextual relevance.

[0015] Using the above technical means, a domain knowledge graph is constructed with effective relationships as edges and entities as nodes. BFS is used to construct the root node as the target entity, and DFS is used to enumerate all possible logical paths. Node confidence (average of entity unified score), edge weight (relation score × compromise coefficient + average endpoint confidence × compromise coefficient), structure regularization (path length compliance), and time recentity (the more recent the evidence time, the higher the weight).

[0016] Furthermore, step S5 also includes, L1 master node alignment; L2 name endpoint extensions include aliases, abbreviations, and synonyms; L3 vector nearest neighbor comparison; L4 evidence constraint extraction includes extracting triples and slots from evidence fragments; L5 conflict resolution includes label mapping, name disambiguation, and hierarchical merging.

[0017] Furthermore, in step S5, the enhanced chain relationship output includes triples, confidence, evidence pointers, and timestamps.

[0018] Using the above technical means, based on the LLM suggestion chain, the candidate entities and relationships after pruning are deeply optimized, and hierarchical relationship completion and calibration are performed.

[0019] The beneficial effects of this invention are: 1. Output the results in the form of "visual path + evidence list" for easy viewing and auditing by users; 2. Deploy the entire process into the actual business system to ensure high concurrency and high availability, while supporting flexible adjustments. Attached Figure Description

[0020] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is a diagram of the concurrent recall and unified scoring module according to Embodiment 3 of this specification. Detailed Implementation

[0021] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0022] It should be understood that the terms “system,” “device,” “unit,” and / or “module” used herein are one way to distinguish different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0023] As indicated in this specification and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0024] Flowcharts are used in this specification to illustrate the operations performed by the system according to embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0025] Implementation: 1: Deployment on enterprise knowledge and risk control data: Inverted index uses ES / OpenSearch, vector index uses Milvus / Faiss, graph engine uses Neo4j / TigerGraph or a self-developed KV graph, LLM is called via SPI and can smoothly switch between private and public clouds; degradation is triggered by distributed tracking and indicator thresholds (shortening the alert chain or switching the rule engine), and automatic reverting occurs after indicator recovery. Specifically, 1. Implementation environment: Linux / x86_64; Docker / K8s; gRPC + Protobuf; ES≥8.x, Milvus≥2.x, Neo4j≥5.x (or equivalent alternatives), LLM is called via SPI.

[0026] 2. Minimum feasible process: S1→S9 is coherent and reproducible; the output JSON (roadMap / resultList) includes evidence offset, timestamps and score evolution.

[0027] 3. Replaceable components: Any SPI-compatible inverted index / vector / graph / LLM engine is interchangeable without changing the core mechanism and claim points.

[0028] 4. Reproducible experiment: Sample 100,000 entities / edges, recall each channel K≈100, quantile threshold P90, length regularization and the median of the recommended interval for time recentity, with NDCG@K and audit consistency as the main indicators.

[0029] Please refer to Table 1 for the mechanisms and effects of each step.

[0030] Table 1. Process Step Comparison Table S1 - Three-channel fusion and unified dimensions include: S11: Map the raw scores of each channel to a unified number line using a channel calibration function; S12: Weighted synthesis based on domain adaptive source weights; S13: Perform posterior robustness processing on the synthesis results; S14: Register the uniform-dimensional score along with its source and evidence pointers to the NSR, such as Figure 1 .

[0031] S2 - Partial Pruning (Node / Relationship Two-Level) includes: S21: Calculate the quantile thresholds for nodes and relationships respectively; S22: Retain candidates with a threshold value or higher, remove long-tail noise, and maintain a balance between coverage and computing power; S3-relationship enhancement includes: S31: Relationships may only be extracted from registered evidence fragments; S32: The output must include a triplet, confidence, evidence pointers, and a timestamp; S33: Align with authoritative tag library; if alignment fails, mark it as "requires manual review".

[0032] S4-Double Threshold Legality includes: S41: The semantic strength has reached the preset threshold; S42: Structural template verification passed (direction, depth, loop, and hierarchy compatibility); S43: Only when the relationship between the two conditions is satisfied can it be included in the diagram; S5 - Path Score and Unified Ranking: S51: The path score is determined by: node confidence, edge weight, structural regularity, temporal recency, business / regulatory priority, and contextual relevance. S52: Edge weights integrate channel scores and endpoint node confidence, with configurable or learnable coefficients; S53: To avoid numerical instability, it is preferable to use logarithmic domain summation for calculation; S54: The order of priority is as follows: shorter path first, closer evidence update first, endpoint node confidence and higher priority.

[0033] S6-Path Control (Merge and Prune): S61: Perform fusion on repeated path segments with the same starting point; S62: Perform overall pruning on weak paths below the threshold to avoid path explosion; S7 - Context Gating and Online Learning: S71: Introduce a gating factor to participate in the ranking based on the relevance of the overall path to the query context; S72: Minimize the sorting loss by using online feedback (clicks / reviews) and update the channel weights, thresholds, and regularization parameters; S73: Data drift triggers channel calibration function and threshold recalibration.

[0034] In summary, this embodiment features three-channel concurrent recall; unified NSR dimensionality; quantile pruning; extractive relation enhancement and dual-threshold legality; and path-level unified sorting and evidence binding based on PathScore.

[0035] Example 2: This embodiment takes the scenario of personalized learning path recommendation for students in the field of education as an example. It processes heterogeneous student learning data (including mastery of subject knowledge points, records of homework errors, classroom interaction data, learning behavior trajectories, exam scores, etc.). The core requirement is to accurately locate the correlation path of "student → knowledge gap → suitable learning resources" from massive educational data (such as "student A → not mastering the solution of quadratic equations → suitable specialized explanation videos + tiered practice questions"). At the same time, it meets the requirements of personalization, scientificity and traceability of education and teaching, and solves the pain points of existing technologies such as "chaotic cross-channel learning data dimensions, fragmented knowledge point associations, and lack of targeted learning path recommendations".

[0036] Deploy an event bus (using Kafka), a stateless microservice cluster (based on Spring Cloud), and a unified storage layer (MySQL for NSR tables, Elasticsearch for inverted indexes, Neo4j for knowledge graphs, and Redis for vector caches), and encapsulate three core modules via an SPI interface: 1. Vector Model: Employs a domain-adjusted version of BERT - Education Edition (supports vector generation of student names, knowledge point names, and learning behavior descriptions); 2. LLM Module: Uses a finely tuned LLaMA-2-7B (for enhancing cue chain relationships, adapted for knowledge reasoning in the education field). 3. Path reasoning engine: Integrates custom BFS / DFS units (supports dynamic weight calculation based on learning ability).

[0037] Step S1: Preprocessing and Multi-Vector Embedding – A Mechanism for Preserving Cognitive-Behavioral Dual Dimensions Existing technologies mostly generate vectors based solely on "semantic similarity," ignoring the "influence of cognitive level on the association of knowledge points" (e.g., "solution method for a quadratic equation" and "discriminant of the roots of a quadratic equation" are semantically similar but have a sequential relationship in cognitive level). This step proposes a three-in-one multi-vector generation scheme of **"layered word segmentation + cognitive hierarchy tree + learning behavior mapping"**, with the specific operation as follows: Layered word segmentation: First layer (basic word segmentation): Perform basic word segmentation on the input query "What are the weak knowledge points of student A in mathematics, and recommend suitable learning resources", to obtain "student A / mathematics / weak knowledge points / learning resources"; The second layer (domain-specific word segmentation): Combined with educational dictionaries, word segmentation is supplemented to break down "weak math knowledge points" into "weak algebra / weak geometry / weak statistics," and "learning resources" are associated with "explanation videos / practice questions / mind maps / micro-lessons." The third layer (cognitive word segmentation): Based on Bloom's Taxonomy of Cognitive Objectives, a cognitive hierarchy tree is constructed, marking the cognitive relationship between "Student A" (subject) → "weak math knowledge points" (object) → "learning resources" (solution), generating the cognitive tag sequence "SUBJ-OBJ-SOL." Learning behavior mapping is based on a K12 education domain learning behavior dictionary, constructing a mapping table. For example: Student identifier: "Student A" → "Student A," "Student ID 001" (school identifier), "A_001" (platform identifier); Knowledge point name: "Quadratic equation" → "Quadratic equation" (abbreviation), "Solution and application of quadratic equation" (full name); Learning behavior: "Repeatedly making mistakes" → "Multiple errors," "Accuracy rate below 60%" (synonymous behavior), "Knowledge confusion" (associated behavior).

[0038] Multi-carrier vector generation Three types of vectors are generated, while maintaining cognitive and behavioral dual-dimensional information: Node vector: Input the text of "Student A" and its identifier + cognitive label "SUBJ" into BERT to generate a 768-dimensional node vector V_node; Relation vector: Input the text of "weak knowledge points", "learning resources" and their related phrases + cognitive label "REL" into BERT to generate a 768-dimensional relation vector V_rel; Context vector: Encode the complete query sentence + cognitive hierarchy tree structure into a context vector V_ctx, which is used for context gating of the NSR table later.

[0039] The innovation of this step lies in the fact that it integrates the encoding of cognitive hierarchy into vector generation for the first time, solving the problem of "confusion in the cognitive order of knowledge points" in existing technologies (such as avoiding recommending "the discriminant of the roots of a quadratic equation" to students who have not mastered "the basic concept of a quadratic equation"), and laying the foundation for subsequent accurate recall.

[0040] Step S2: Three-channel concurrent recall - inter-channel learning data association preloading mechanism The existing "three-channel parallelism" technology is only "physical parallelism," with no data interaction between channels, which easily leads to fragmented recall results (e.g., the inverted index channel recalls "student A's math mistakes," and the node channel recalls "knowledge points on quadratic equations," which are unrelated). This step proposes a concurrent recall scheme of "learning data association preloading + shared cache scheduling," with the specific operations as follows: channel initialization and preloading. When starting the inverted index, node, and relationship channels, "learning data association preloading" is triggered via the event bus: Inverted Index Channel: Build an inverted index in Elasticsearch that follows the structure of "student identifier - knowledge point tag - learning behavior". Preload the list of document IDs corresponding to all identifiers of "student A" in the query, and associate it with the inverted chain of learning tags such as "weak in math" and "repeatedly making mistakes". Node Channel: The top 50 nearest neighbor nodes of the node vector V_node of "Student A" are preloaded in Redis (such as "Quadratic Equation Knowledge Point" and "Function Knowledge Point", which are weak or related knowledge points of Student A). The nearest neighbor calculation uses cosine similarity (threshold 0.75). Relationship Channel: Preload the Top 30 nearest neighbor relationships of the relationship vector V_rel for "weak knowledge points" and "learning resources" (such as "knowledge gaps", "adaptive explanations", and "tiered exercises"), and associate the corresponding "student node - knowledge point node" pairs (such as the weak relationship "student A → quadratic equation").

[0041] Concurrent recall execution The three channels interact with data in real time through a shared cache (Redis Cluster) to avoid duplicate recalls: The inverted index channel recalls documents identified as "Student A" and tagged with "weak in math," yielding 150 candidates (including Student A's incorrect answers to quadratic equations and function-related questions); the node channel further filters nodes related to "weak in math" based on pre-loaded Top 50 nearest neighbor nodes, resulting in 42 candidates (including quadratic equations and Pythagorean theorem knowledge points); and the relationship channel filters relationship pairs where "student A is the subject," yielding 65 candidates (e.g., "Student A → quadratic equations → knowledge gaps" and "Student A → functions → misunderstandings").

[0042] The innovation of this step lies in achieving data collaboration between channels through "learning data association preloading," avoiding the fragmentation of results caused by the "parallel but isolated" nature of existing technologies. This improves recall efficiency by 45% (compared to traditional three-channel recall without preloading) and candidate relevance by 70%. Step S3: Unified scoring and registration - learning ability adaptive calibration mechanism. Existing technologies often use fixed normalization functions (such as Min-Max) for "score calibration," which cannot adapt to the differences in learning abilities among different students (e.g., the difficulty weight of the same knowledge point is different for high-achieving students and struggling students). This step proposes a unified scoring scheme of "learning ability adaptive calibration function + mastery level factor," achieving cross-channel dimensional unification based on the unified scoring registry (NSR). The specific operation is as follows: The NSR table is stored in MySQL partitioned tables, with the following core fields: Table 2 NSR Table Core Fields The raw scores for each channel are processed using a "learning ability adaptive calibration function", as shown in the following formula: S_{calibrated} = \frac{S_{raw} - \mu_s}{\sigma_s} \times w_s \times g\times l in: At the same time, a "mastery level factor" f_m = 1 - \frac{mastery\_level}{10} (where mastery\_level is the student's score for mastering the knowledge point) is introduced, ultimately unifying the score formula: S_{unified} = S_{calibrated} \times f_m Example calculation (candidate "Student A → Quadratic Equation → Knowledge Gap"): 1. S_{raw}: Raw channel scores (e.g., inverted channel BM25 score = 20.2, node channel cosine similarity = 0.85, relation channel similarity = 0.93); 2. \mu_s / \sigma_s: Historical mean / standard deviation of this channel under the learning ability of student s (based on the student's learning data statistics over the past 3 months, such as node channel \mu=0.78, \sigma=0.10); 3. w_s: Channel weight corresponding to student s' learning ability (dynamically adjusted based on the student's academic ranking, e.g., the channel weight for struggling students is set to 0.45). 4.g: Context gating factors (e.g., the match between the candidate "student A's weakness in quadratic equations" and the query is 0.95); 5.l: Learning progress factor (based on the student's current grade level's knowledge point teaching progress, such as l=1 if the quadratic equation has been learned, and l=0.5 if it has not been learned).

[0043] 6. The raw scores of the relation channel are S_{raw}=0.93, mu=0.78, sigma=0.10, w_s=0.45, g=0.95, l=1; 7.S_{calibrated}=(0.93-0.78) / 0.10 \times 0.45 \times 0.95 \times 1 ≈0.641; 8. Mastery level = 3, f_m = 1 - 3 / 10 = 0.7; 9. Finally, S_{unified} = 0.641 × 0.7 ≈ 0.449 (written into the NSR table).

[0044] Based on "student learning effect feedback", update \mu_s, \sigma_s, w_s: If the learning resources recommended by a certain channel (such as the node channel) are effectively learned by students and the mastery of knowledge points increases by ≥30%, then its weight is increased by 0.05, and \mu_s and \sigma_s are adjusted (incorporating the learning data of the latest day).

[0045] The innovation of this step lies in the fact that it integrates "student learning ability weight", "learning progress" and "mastery level" into score calibration for the first time, solving the problems of "inconsistent dimensions across channels and low matching degree between scores and students' actual situation" in existing technologies. It also improves the Pearson correlation coefficient between scores and students' knowledge weaknesses to 0.92 (compared to 0.65 of traditional Min-Max normalization).

[0046] It is worth noting that in step S4: quantile pruning; the existing "quantile pruning" technology for knowledge points and resources often adopts "global unified quantile" (e.g., taking the 90th quantile for all candidates), which can easily lead to "over-pruning of knowledge point candidates (missing low-scoring but crucial weak knowledge points)" or "under-pruning of resource candidates (retaining a large number of unsuitable learning resources)". This step proposes a "knowledge point-resource classification quantile threshold" mechanism, which accurately removes noise based on the quantile statistics of the NSR table. The specific operation is as follows: quantile statistics calculation classifies candidates in the NSR table according to "knowledge point type" and "resource suitability", and calculates the 95th quantile value (\theta_{95}) as the pruning threshold: knowledge point candidate classification and threshold: Core weak knowledge points (knowledge points corresponding to students' recent frequently missed questions, such as quadratic equations): α_{95} = 0.30 (lenient threshold, to avoid missing key weak points); Secondary weak knowledge points (knowledge points with occasional errors, such as the Pythagorean theorem): α_{95} = 0.40 (strict threshold, to eliminate weakly related knowledge points); Unlearned knowledge points (knowledge points not covered in the teaching progress): α_{95} = 0.25 (lenient threshold, to leave room for pre-study recommendations); Candidate resource categories and thresholds: Highly applicable resources (resources directly matching students' weak points, such as videos specifically explaining quadratic equations): α_{95} = 0.35 (lenient threshold, to retain core applicable resources); Lowly applicable resources (generalized subject resources, such as comprehensive middle school math practice problems): α_{95} = 0.48 (strict threshold, to eliminate unsuitable resources).

[0047] If the number of candidates for a certain category exceeds a preset threshold (e.g., more than 40 candidates for core weak knowledge points), its quantile threshold is increased by 5% (e.g., from 0.30 to 0.32); if the number of candidates is less than 8, the threshold is decreased by 5% (e.g., from 0.35 to 0.33) to avoid "too many candidates leading to wasted computing power" or "too few candidates leading to a broken learning path".

[0048] Iterate through the NSR table of candidates and remove candidates whose uniform scores are lower than the corresponding category threshold: 1. Candidate knowledge points: 32 minor weak knowledge points (score < 0.40) and 4 unlearned knowledge points (score < 0.25) were removed, leaving 38 core weak knowledge points; 2. Resource Candidates: Remove 25 low-fit resources (score < 0.48) and retain 40 high-fit resources.

[0049] The innovation of this step lies in dynamically adjusting the quantile threshold based on the candidate type. Compared with the existing globally unified threshold, the noise candidate elimination rate is increased by 60%, while the key candidate retention rate is increased to 99% (85% in traditional methods), which greatly reduces the computational cost of subsequent graph construction and inference.

[0050] Step S5 specifically includes, L1 (master node alignment): matching the student ID of "Student A" with the node candidates in the NSR table to confirm the core node "Quadratic Equation Knowledge Point" (Knowledge Point ID: M090105). L2 (Name Endpoint Expansion): Expand the aliases of "quadratic equation in one variable" ("quadratic equation", "solution of quadratic equation in one variable"), related knowledge points ("linear equation in one variable", "quadratic function"), and simultaneously expand the synonyms of "knowledge gaps" ("insufficient mastery", "misunderstanding", "application error"). L3 (Vector Nearest Neighbor Comparison): The node vector of "quadratic equation" is compared with the vectors of all "student A's weak knowledge points" in the knowledge graph. It is found that the vector similarity between "quadratic equation" and "discriminant of roots" is 0.83 (Supplementary association: Student A made a mistake in solving the quadratic equation because he did not master the discriminant of roots). L4 (Evidence Constraint Extraction): An alias based on L2, it extracts triples from students' error notebooks and homework correction records.

[0051] L5 (Conflict Resolution): Two "equation" knowledge points were identified (simple equations for elementary school students and quadratic equations for junior high school students), and the "segment-priority strategy" was activated.

[0052] It is worth noting that step S7, graph construction and path reasoning—the learning ability dynamic weight path reasoning algorithm—also includes constructing a ninth-grade math learning knowledge graph in Neo4j using valid relationships after structural verification as "edges" and candidate nodes as "nodes." The algorithm performs DFS enumeration on the core learning tree, prioritizing the output of paths with "basic knowledge points first, length ≤ 4, and the highest weak factor." The unified path score (S_{path}) is composed of 6 weighted factors, with the weights dynamically adjusted according to the student's learning ability.

[0053] In step S9, dynamic scheduling of microservice resources is implemented based on Kubernetes, with pre-defined resource strategies according to the "education business scenario": High-concurrency scenarios (such as end-of-semester learning assessments, QPS=1500+): automatically increase the number of microservice instances for the inverted index channel (Elasticsearch) and the vector channel (Redis) (expanding from 4 to 10), increase the CPU core allocation from 2 cores to 6 cores, and increase the memory allocation from 4GB to 12GB; Low-concurrency scenarios (such as holiday learning periods, QPS=80-): reduce the number of instances (reducing the inverted index channel instances to 2), and release resources for model updates (such as fine-tuning of LLM in the education field and iteration of vector models); Sudden scenarios (such as large-scale online learning activities, with a sudden increase in QPS): trigger the "emergency resource pool" (reserving 25% of resources) to prioritize the resource supply of core channels (relationship channels, path inference engine).

[0054] In summary, this technology overcomes the limitations of existing technologies, such as "isolated channels, fixed weights, and difficulty in tracking teaching effectiveness." Through innovative mechanisms such as cognitive-behavioral dual-dimensional embedding, adaptive calibration of learning ability, completion of the evidence chain of learning levels, and full tracing of the evolution of learning path scores, it achieves knowledge graph retrieval that is "precise in retrieval, personalized in recommendation, traceable in teaching effectiveness, and highly available." It can be widely applied to educational scenarios such as K-12 education, vocational education, and higher education that require personalized learning path recommendations.

Claims

1. A knowledge graph retrieval method based on multi-channel recall and relationship enhancement, characterized in that, Includes the following steps: S1: Preprocessing and multi-vector embedding, specifically to complete hierarchical word segmentation, synonym mapping and syntax preservation, and generate multi-carrier vectors; S2: Three-channel concurrent recall, specifically including parallel recall of candidates through inverted index, nodes, and relationships; S3: Unified scoring and registration, specifically including calibrating and weighting the scores of each channel, combining them into a unified dimensional score, and writing it into the unified scoring registry; S4: Quantile pruning, which specifically includes calculating quantile thresholds for nodes and relationships respectively, and deleting low-quality long-tail candidates; S5: Relationship enhancement, specifically including enhancing the output of the cue chain relationship after hierarchical cue chain; S6: Structure and deduplication verification, specifically including verification of direction, depth and loop according to the structure template, deduplication within UID-Pair by leaf priority strategy, and setting thresholds for root / trunk / leaf layer; S7: Graph construction and path reasoning, specifically including constructing a graph with valid edges, constructing a forest with BFS units, enumerating paths with DFS units, and unifying the path ranking within the domain with unified path scores; S8: Output and Audit, specifically including generating roadmaps and evidence lists, and recording the evolution from channel scores to path scores; S9: Deployment and operation, specifically including stateless microservices driven by event bus; Inverted indexes, vectors, graphs, and LLMs are hot-swapped in SPI packages.

2. The knowledge graph retrieval method based on multi-channel recall and relationship enhancement as described in claim 1, characterized in that, In step S3, the unified scoring registry contains records of the candidate's unified dimensional score, source channel, weight, timestamp, evidence pointer, domain identifier, and context gating factor, and provides quantile statistical processing.

3. The knowledge graph retrieval method based on multi-channel recall and relationship enhancement as described in claim 2, characterized in that, Step S3 further includes normalizing or equivalently calibrating the scores of different retrieval channels so that they all fall on a unified number axis, and updating the parameters in combination with historical statistics or validation sets.

4. The knowledge graph retrieval method based on multi-channel recall and relationship enhancement as described in claim 3, characterized in that, In step S4, the quantile threshold includes quantile thresholds calculated separately for each node and relationship.

5. The knowledge graph retrieval method based on multi-channel recall and relationship enhancement as described in claim 4, characterized in that, In step S6, the structural template includes relational direction, maximum depth, loop constraints, and a set of hierarchical compatibility matrix rules.

6. The knowledge graph retrieval method based on multi-channel recall and relationship enhancement as described in claim 5, characterized in that, The unified path score in step S7 also includes synthesizing factors into a single scalar for unified sorting of paths of different lengths and components. The factors include node confidence, edge weight, structural regularity, temporal recency, regulatory priority, and contextual relevance.

7. The knowledge graph retrieval method based on multi-channel recall and relationship enhancement as described in claim 6, characterized in that, Step S5 also includes, L1 master node alignment; L2 name endpoint extensions include aliases, abbreviations, and synonyms; L3 vector nearest neighbor comparison; L4 evidence constraint extraction includes extracting triples and slots from evidence fragments; L5 conflict resolution includes label mapping, name disambiguation, and hierarchical merging.

8. The knowledge graph retrieval method based on multi-channel recall and relationship enhancement as described in claim 7, characterized in that, In step S5, the output of the chain relationship enhancement includes triples, confidence, evidence pointers, and timestamps.

Images