A multi-strategy retrieval intelligent question answering system based on a knowledge graph

Abstract

This invention discloses a multi-strategy retrieval intelligent question-answering system based on knowledge graphs. Through the collaboration of question parsing, knowledge retrieval, answer generation, and evaluation modules, intelligent question answering is achieved: The system receives user questions, parses the question intent using a large language model, and extracts entities, relation types, and query types; it then performs a multi-strategy progressive retrieval based on a graph database knowledge graph, sequentially performing batch exact matching of entities, batch querying of relations, full-text index search, fuzzy matching, and source file keyword search; it scores and filters the retrieval results based on relevance; it generates answers based on the filtered knowledge using a large language model and calculates the confidence level of the answers by integrating multiple factors; and it implements result caching. This invention, through the synergy of multi-strategy progressive retrieval and intelligent filtering, effectively solves the problems of inaccurate retrieval, poor answer quality, and slow response speed in traditional question-answering systems, significantly improving the accuracy, recall, and response speed of question answering.

Description

Technical Field

[0001] This invention relates to the technical field of artificial intelligence and natural language processing, and in particular to the technical field of a multi-strategy retrieval intelligent question answering system based on knowledge graphs. Background Technology

[0002] Traditional knowledge-based question-answering systems have limitations in their retrieval strategies, often relying on a single method, making it difficult to balance retrieval precision, recall, and response speed. One type of approach focuses on exact matching or structured queries, which, while ensuring retrieval speed, cannot effectively handle the diversity and semantic relevance of natural language expressions, such as abbreviations of technical terms, synonyms, and colloquial expressions, resulting in low recall and the omission of a large amount of relevant knowledge. Another type of approach tends to use vector similarity retrieval, which can improve semantic recall, but it is insufficient in handling exact matching of proper entity names and ensuring factual accuracy, and it usually has a high computational cost, affecting response speed.

[0003] Furthermore, due to a weak foundation in the retrieval phase, the quality of answers generated by existing systems fluctuates significantly. More critically, current technologies generally lack a quantitative and transparent confidence assessment system. The system cannot inform users of the reliability of the answers, nor can it dynamically adjust the certainty and level of detail in the responses based on the confidence level. This makes it difficult for users to judge the reliability of answers in professional scenarios, severely limiting the practical application value of question-answering systems. Therefore, while achieving efficient retrieval, existing technological systems struggle to systematically guarantee the accuracy and reliability of answers, failing to meet high-standard application requirements. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the prior art and propose a multi-strategy retrieval intelligent question answering system based on knowledge graphs, which can achieve intelligent retrieval with high accuracy and high recall, and provide reliable answer quality assessment.

[0005] To achieve the above objectives, this invention proposes a multi-strategy retrieval intelligent question-answering system based on knowledge graphs, comprising: The question parsing module receives natural language questions submitted by users and generates a question intent object containing the query type and entities. The knowledge retrieval module, connected to the question parsing module, is used to perform multi-strategy progressive retrieval from the graph database knowledge graph based on the question intent object; The answer generation module, connected to the knowledge retrieval module, is used to generate answers based on the retrieval results and calculate confidence levels. The knowledge retrieval module is configured to execute the following retrieval strategies sequentially, from fastest to slowest: first, perform batch exact matching of entities, followed by batch query of relationships; if the sum of the results obtained by executing the aforementioned two strategies reaches or exceeds the dynamic threshold corresponding to the query type, the retrieval process is terminated early; if not, then full-text index search, fuzzy matching, and source file keyword search are executed sequentially. Furthermore, the knowledge retrieval module is also configured to perform relevance scoring and filtering on the entities and relationships retrieved through the aforementioned strategy; The answer generation module is configured to calculate the confidence level by comprehensively evaluating six factors: knowledge completeness, knowledge quantity, question type matching degree, entity coverage, source file relevance, and knowledge relevance.

[0006] Preferably, the adjustment rule for the dynamic threshold is as follows: for entity queries or relationship queries, the threshold is a first numerical range; for aggregated statistical queries, the threshold is a second numerical range higher than the first numerical range; the first and second numerical ranges can be configured according to the application scenario. This preferred solution, by distinguishing query types, makes the early exit mechanism trigger more accurate and reasonable, further optimizing retrieval efficiency.

[0007] Preferably, the knowledge retrieval module is further configured to: first, perform batch exact matching of entities; then, perform batch query of relationships; if the sum of the results obtained by executing the aforementioned two strategies does not reach the dynamic threshold, then perform full-text index search; if the number of full-text index search results still does not meet the requirements, then perform fuzzy matching; if the number of entities and relationships retrieved by the aforementioned strategies is extremely small, then perform source file keyword search. This preferred scheme clarifies the detailed execution order and triggering conditions of multiple strategies, forming an efficient and controllable progressive funnel of "fast and accurate retrieval → semantic expansion search," which is the core process of this invention.

[0008] Preferably, when performing relevance scoring, the knowledge retrieval module uses scoring factors including exact matching of entity / relationship names with question keywords, matching of keywords contained in entity / relationship descriptions with complete phrases in the question, and, for entities, the weight of exact name matching is higher than the weight of keyword inclusion in the description; for relations, the weight of subject / object name matching is higher than the weight of relation description matching. This preferred scheme establishes a refined set of quantitative scoring rules, which can effectively distinguish the importance of knowledge fragments with different matching degrees, providing a reliable basis for subsequent filtering.

[0009] Preferably, in the relevance scoring, the weight of complete phrase matching is greater than the weight of exact name matching. This preferred approach emphasizes a higher regard for semantic completeness, giving higher priority to knowledge that fully expresses the question's intent, thus improving the relevance of the answer.

[0010] Preferably, when calculating confidence level, the answer generation module uses a weighted average of the six factors: knowledge completeness (20%-30%), knowledge quantity (15%-25%), question type matching (10%-20%), entity coverage (10%-20%), source file relevance (5%-15%), and knowledge relevance (10%-20%). Furthermore, the knowledge quantity score is calculated using a non-linear function to avoid unreasonably high scores due to an excessive amount of knowledge. This preferred solution achieves a scientific quantification of answer credibility through a multi-dimensional, weighted comprehensive evaluation model. The non-linear processing avoids an overemphasis on quantity, making the evaluation results more objective.

[0011] Preferably, the problem parsing module is configured to: first, parse based on preset rules; if this fails, then invoke a large language model for parsing; and perform post-processing on the parsing results, including entity synonym expansion and entity type inference. This preferred solution adopts a "rule-first" two-layer parsing strategy, which significantly reduces the response time and computational cost for handling common problems while ensuring parsing accuracy.

[0012] Preferably, the system also includes a cache management module for three-level caching of question parsing results, knowledge retrieval results, and answer generation results. The cache keys are the hash value of the natural language question, the hash value of the question intent object, and the combined hash value of the question, intent, and retrieval results, respectively. This preferred solution significantly improves the system's response speed to repeated or similar requests and reduces backend load by designing differentiated caching strategies for different computation stages.

[0013] Preferably, the system achieves intelligent question answering by performing the following steps: receiving natural language questions submitted by users; parsing the questions and generating a question intent object containing query type and entities; performing a multi-strategy progressive retrieval from a graph database knowledge graph based on the question intent object; the execution includes: first performing batch exact matching of entities, followed by batch querying of relationships; if the sum of the obtained results does not reach the corresponding dynamic threshold, then performing full-text index search and fuzzy matching in sequence; if the number of entities and relationships is still very small, then performing source file keyword search; scoring the relevance of the retrieval results and filtering them; generating answers based on the filtered knowledge, and comprehensively evaluating the six factors to calculate the confidence level of the answer; returning a result containing the answer and the confidence level. This preferred solution, from the perspective of the overall system operation process, fully elaborates on the technical path from question input to answer output, and clearly demonstrates the collaborative working method of each core module.

[0014] The beneficial effects of this invention: By deeply integrating multi-strategy progressive retrieval, relevance scoring filtering, confidence assessment, and intelligent caching optimization, this invention constructs a complete, efficient, and accurate intelligent question-answering closed loop. It effectively solves the problems of inaccurate retrieval, poor answer quality, and slow response speed in traditional question-answering systems. It realizes the transformation of question-answering systems from "single strategy" to "multi-strategy collaboration," from "blind retrieval" to "intelligent filtering," and from "no confidence" to "comprehensive evaluation," significantly improving the accuracy, efficiency, and user experience of knowledge-based question answering.

[0015] While some of the technical elements of this invention are known technologies, their specific combination produces a significant synergistic effect: First, the synergy between the early exit mechanism and the multi-strategy progressive retrieval allows the system to quickly obtain sufficient results in the precise matching stage, skipping subsequent slow queries and reducing response time from several seconds to hundreds of milliseconds, achieving a doubling of response speed. Second, the synergy between relevance scoring and multi-strategy retrieval enables the system to filter the most relevant knowledge from the retrieval results of different strategies, increasing accuracy from 62% to 89% and recall from 58% to 92%, achieving a fundamental improvement in retrieval accuracy. Third, the synergy between multi-factor confidence assessment and answer generation allows the system to dynamically adjust its answer style based on knowledge quality, providing detailed answers at high confidence and cautious prompts at low confidence, achieving a fundamental improvement in answer reliability. Finally, the synergy between the three-level caching mechanism and multi-strategy retrieval results in a cache hit rate of 68%, further reducing response time from hundreds of milliseconds to tens of milliseconds. These synergistic effects make the technical effects of this invention far exceed the effects of simply adding up the technical elements, demonstrating significant inventiveness and practicality.

[0016] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the workflow and system modules of a multi-strategy retrieval intelligent question answering system based on knowledge graphs according to the present invention.

[0018] In the diagram: 1- Question Parsing Module: Responsible for receiving user questions and extracting the question intent, including entity, relation type, and query type, through rule parsing or large language model parsing; 2- Knowledge Retrieval Module: Responsible for performing multi-strategy progressive retrieval from the graph database knowledge graph, including batch exact matching of entities, batch querying of relations, full-text index search, fuzzy matching, and source file search, and performing relevance scoring and filtering on the retrieval results; 3- Answer Generation Module: Responsible for generating answers based on the filtered knowledge through the large language model, calculating confidence, and supporting streaming output; 4- Cache Management Module: Responsible for caching the parsing, retrieval, and generation results, supporting both memory and Redis backends; 5- Monitoring Service Module: Responsible for recording metrics, performance analysis, and alerts. Detailed Implementation

[0019] The following will be combined with the appendix Figure 1 The technical solutions in the embodiments of the present invention will be clearly and completely described.

[0020] The physical infrastructure of the system includes a cluster of servers interconnected via a network, graph databases supporting Cypher (a graph database query language) (such as Neo4j, JanusGraph, Nebula Graph, etc.), storage devices, and user terminals. The servers run the program modules of this invention. The graph database stores entities and relationships in the knowledge graph. The storage devices store cached data, log files, etc. The user terminals include PCs and mobile devices for submitting questions and viewing answers.

[0021] Question Parsing Phase: Question parsing module 1 receives natural language questions submitted by users on the front end, such as "What is NCR?" or "What HSE work is Zhang San responsible for?". This module then first attempts to quickly parse based on rules, identifying common problem patterns: definition-based questions (matching patterns like "What is X?", "What is X?"), aggregation / statistical questions (matching keywords like "all", "which", "how many", etc.), process-based questions (matching keywords like "how", "how", "process", etc.), and relationship-based questions (matching patterns like "X belongs to Y", "X is responsible for Y", etc.). If a rule match is successful, the parsing result is returned directly, avoiding calls to the large language model and significantly improving parsing speed (from hundreds of milliseconds to several milliseconds). If rule parsing fails, the large language model is called for parsing: the question is input into the large language model (e.g., Qwen-8B (Tongyi Qianwen - 8 billion parameter version)), and the output is required to include a list of entities, a list of entity types, relationship types, query types (entity / relation / path), and a summary of the question in JSON (JavaScript Object Notation, a data exchange format). Post-processing of the parsed results includes: entity synonym expansion (e.g., "HSE" expanded to "safety", "environmental protection", "HSE management"), entity type inference (based on entity name patterns, e.g., "quality department" inferred to "department"), keyword extraction (using jieba word segmentation and part-of-speech tagging to extract important words such as nouns and proper nouns), deduplication, and stop word filtering. Parsing results are cached for 1 hour; cached results can be used directly for the same problem.

[0022] Knowledge Retrieval Phase: Knowledge Retrieval Module 2 performs a multi-strategy progressive retrieval from the graph database knowledge graph based on the question intent object. First, it performs batch exact matching (Phase 1, response time <10ms): using the graph database index to batch query entity names, such as querying entities like "Zhang San" and "HSE". If an exact match finds more than a preset number (specifically 3) of entities, it exits early, skipping subsequent slower queries, significantly improving retrieval speed. If exact matching results are insufficient, it performs full-text index search (Phase 2, response time <200ms): using the graph database full-text index to perform keyword search on entity descriptions, supporting relevance scoring and multi-word search. For example, searching for "target assessment" can find entities like "safety production target assessment system," even if the entity names do not match exactly. If the full-text index search results are still insufficient, it performs fuzzy matching query (Phase 3, response time >500ms): using string inclusion operators (such as Cypher's CONTAINS) to search for keywords in entity names and descriptions, merging multiple keywords into a single query to reduce database round trips. If there are few entities and relations, perform a source file keyword search: retrieve knowledge from entities and relations whose source file names contain the keywords; this is a last resort strategy. During the search, each entity and relation is scored for relevance: extract keywords from the question, calculate the matching degree between entity names and keywords, with the following scoring rules: exact match of entity name has a weight of 3.0, entity description containing keywords has a weight of 1.0, complete two-word phrase match has a weight of 5.0, and complete three-word phrase match has a weight of 8.0; calculate the matching degree between relation subject / object and keywords, with the following scoring rules: relation subject / object name match has a weight of 3.0, relation description match has a weight of 1.5, and complete phrase match has a weight of 5.0; based on the score filtering and sorting, retain knowledge with a score higher than a preset threshold (specifically 1.0), and sort them in descending order of score. Search results are cached for 30 minutes; cached results can be used directly for the same intent.

[0023] Answer Generation Stage: Module 3 generates answers based on the filtered knowledge using a large language model. First, it calculates confidence and formatted knowledge context in parallel. Confidence calculation considers six factors with the following weights: knowledge completeness (25%), knowledge quantity (20%), question type matching (15%), entity coverage (15%), source document relevance (10%), and knowledge relevance (15%). The knowledge completeness score is calculated as: completeness_score = complete_items / total_items, where complete_items is the number of entities and relations with descriptions or source documents, and total_items is the total number of items. The knowledge quantity score uses a logarithmic function for non-linear scoring. The specific formula is: entity_score = min(1.0, log10(max(1, entity_count)) / log10(10)), relation_score = min(1.0, log10(max(1, relation_count)) / log10(10)), quantity_score = entity_score × 0.3 + relation_score × 0.5 + path_score × 0.2, to avoid excessively high scores when there are too many knowledge items. Question type matching score: If the query type is entity and an entity is found, the matching score is 1.0. If a relation or path is found, the matching score decreases to 0.5-0.6. Entity coverage score: coverage = matched_entities / total_query_entities. If the coverage is less than 0.5, partial matching is checked and a score of 0.3-0.7 is given. Source document relevance score: linear increase for 1-5 source documents (0.2-1.0), optimal range for 5-15 (score 1.0), slightly decrease for more than 15. Knowledge relevance score: severe penalty for relevance ratio below 30% (score = ratio × 0.5), moderate penalty for relevance ratio between 30%-50% (score = ratio × 0.8), and direct use of ratio as score for relevance ratio ≥ 50%. Overall confidence calculation formula: confidence = completeness_score × 0.25 + quantity_score × 0.20 + type_match_score × 0.15 + entity_coverage_score × 0.15 + source_score × 0.10 + relevance_score × 0.15, yielding a confidence score between 0.0 and 1.0.The knowledge context formatting includes entity information (name, type, description, source file) and relation information (subject, relation type, object, description, source file). The response style is then adjusted based on confidence level: at high confidence (≥0.8), the model is prompted to provide a "relatively confident and well-developed" response; at medium confidence (0.5-0.8), the model is prompted to be "cautious," providing the known parts first, then explaining the uncertain parts; at low confidence (<0.5), the model is prompted to "prioritize explaining known content and unknown parts," and the response should be "concise and conservative." The question, intent, knowledge context, and confidence level are input into a large language model (such as DeepSeek-R1-14B, a version with 14 billion parameters) to generate the answer. Streaming output is supported: first, metadata (confidence, source, number of entities and relations) is returned; then, the answer text is returned chunk by chunk; finally, a completion marker is returned, supporting real-time display of the answer generation process. The generated results are cached for 30 minutes; the same question, intent, and knowledge can directly use the cached results.

[0024] Cache optimization phase: Cache management module 4 implements a three-level cache for the results of question parsing, knowledge retrieval, and answer generation. The question parsing cache uses the MD5 hash value of the question (Message-Digest Algorithm 5) as the cache key, with a cache time of 1 hour, avoiding duplicate parsing of the same question. The knowledge retrieval cache uses the MD5 hash value of the question intent object as the cache key, with a cache time of 30 minutes. The answer generation cache uses the combined MD5 hash value of the question, intent, and knowledge as the cache key, with a cache time of 30 minutes. It supports two backends: memory caching and Redis (Remote DictionaryServer). The memory cache uses an LRU (Least Recently Used) eviction policy, deleting the oldest item when the cache is full. The Redis cache supports TTL (Time To Live) automatic expiration, suitable for distributed deployment. Cache hits can significantly improve response time, reducing it from several seconds to tens of milliseconds.

[0025] Monitoring and Alert Phase: Monitoring service module 5 records all query metrics, including query count, success rate, response time (segmented recording of parsing time, retrieval time, and generation time), and error type. Performance analysis is performed: P50, P95, and P99 response times are calculated to analyze performance differences between different query types. Alert checks are conducted: a warning is triggered when the error rate exceeds 10%, when the average response time exceeds 10 seconds, and when the P95 response time exceeds 20 seconds. A comprehensive report is generated, including system statistics, performance metrics, alarm information, and health status, and supports exporting reports in JSON format.

[0026] Example 1: In the field of enterprise knowledge management, a user asks "What is NCR?". The system identifies the question as a definition-type question through rule parsing, with the query type being "entity" and the entity being ["NCR"]. The knowledge retrieval module performs batch exact matching, finding an entity named "NCR" that includes the description "Non-Conformance Report" and source file information. Relevance score calculation: Exact match of entity name, score 3.0; description contains keywords, score 1.0; total score 4.0, exceeding the threshold of 1.0, so the entity is retained. The answer generation module calculates confidence: knowledge completeness 1.0 (with description and source file), knowledge quantity 1.0 (one relevant entity found), type matching degree 1.0 (entity query finds entity), entity coverage 1.0 (entity in the question is found), source file relevance 1.0 (with source file), knowledge relevance 1.0 (highly relevant), overall confidence 0.95. Based on the high confidence, the model generates a confident and detailed answer: "NCR is an abbreviation for Non-Conformance Report, used to record and track non-conformances to quality management requirements...". The system returns the answer, knowledge, confidence level of 0.95, and a list of source files, with a response time of 1.2 seconds.

[0027] Example 2: In the field of enterprise organization management, a user asks, "What are all the departments?". The system identifies this as an aggregated statistical question through rule parsing, with the query type being "entity," the entity type being ["department"], and the entity list being empty (mainly retrieved using entity_types). The knowledge retrieval module identifies this as an aggregated statistical query, retrieves all department entities by type, finds 15 department entities (such as "Quality Department," "Safety Department," "Engineering Department," etc.), and retrieves the relationships between these departments. Relevance score: All department entities contain the keyword "department," and all scores are above the threshold, so all are retained. The answer generation module calculates the confidence score: Knowledge completeness 0.9 (most have descriptions), Knowledge quantity 1.0 (15 entities found), Type matching 1.0 (entity query found entities), Entity coverage 1.0 (high coverage when retrieved by type), Source file relevance 1.0 (multiple source files exist), Knowledge relevance 1.0 (highly relevant), and overall confidence score 0.98. Based on the high confidence level, the model generates a detailed list-style answer: "According to the knowledge graph, there are 15 departments, including: 1. Quality Department... 2. Safety Department... 3. Engineering Department...". The system returns the answer, the 15 department entities, a confidence level of 0.98, and a list of source files, with a response time of 2.5 seconds (slightly longer due to retrieving more entities).

[0028] Example 3 (Comparative Experiment): To verify the technical effect of the present invention, a comparative experiment was designed, using the same set of 50 test questions and employing three retrieval strategies: Strategy A (single exact match strategy), Strategy B (single vector retrieval strategy), and Strategy C (the multi-strategy progressive retrieval of the present invention). Test environment: Neo4j 5.25 database, containing 10,000 entities and 15,000 relations; server configuration: 8-core CPU, 32GB memory. The test results are shown in the table below: Analysis: Strategy A (single exact match) has high accuracy when entity names are exactly matched, but it cannot handle cases where names are not exactly matched, resulting in low recall. Strategy B (single vector retrieval) can handle semantic similarity, but its response time is long and its recognition of technical terms is not accurate enough. Strategy C (the multi-strategy progressive retrieval of this invention) significantly improves response speed while maintaining high accuracy through multi-strategy collaboration and an early exit mechanism. It filters irrelevant knowledge through relevance scoring, significantly improving recall and F1 score. In addition, the three-level caching mechanism of this invention achieves a cache hit rate of 68%, further improving response speed. Experimental data fully demonstrates the technical effectiveness and superiority of this invention.

[0029] The above embodiments are illustrative of the present invention and are not intended to limit the present invention. Any simple modifications to the present invention are within the scope of protection of the present invention.

Claims

1. A knowledge graph-based multi-strategy retrieval intelligent question-answering system, characterized in that... ,include: The question parsing module receives natural language questions submitted by users and generates a question intent object containing the query type and entities. The knowledge retrieval module, connected to the question parsing module, is used to perform multi-strategy progressive retrieval from the graph database knowledge graph based on the question intent object; The answer generation module, connected to the knowledge retrieval module, is used to generate answers based on the retrieval results and calculate confidence levels. The knowledge retrieval module is configured to execute the following retrieval strategies sequentially, from fastest to slowest: first, perform batch exact matching of entities, followed by batch query of relationships; if the sum of the results obtained by executing the aforementioned two strategies reaches or exceeds the dynamic threshold corresponding to the query type, the retrieval process is terminated early; if not, then full-text index search, fuzzy matching, and source file keyword search are executed sequentially. Furthermore, the knowledge retrieval module is also configured to perform relevance scoring and filtering on the entities and relationships retrieved through the aforementioned strategy; The answer generation module is configured to calculate the confidence level by comprehensively evaluating six factors: knowledge completeness, knowledge quantity, question type matching degree, entity coverage, source file relevance, and knowledge relevance.

2. The system according to claim 1, characterized in that... The adjustment rules for the dynamic threshold are as follows: for entity queries or relationship queries, the threshold is a first numerical range; for aggregate statistical queries, the threshold is a second numerical range higher than the first numerical range; the first numerical range and the second numerical range can be configured according to the application scenario.

3. The system according to claim 1, characterized in that... The knowledge retrieval module is further configured as follows: First, perform a batch exact match of entities; Then perform a batch query of relationships; If the sum of the results obtained by executing the aforementioned two strategies does not reach the dynamic threshold, then a full-text index search is performed; If the number of full-text index search results still does not meet the requirement, then fuzzy matching will be performed; If the number of entities and relationships retrieved using the aforementioned strategy is extremely small, then perform a keyword search of the source file.

4. The system according to claim 1 or 3, characterized in that... After performing the search, the knowledge retrieval module is further configured to: score the relevance of the retrieved entities and relationships. The scoring factors include the exact match between the entity / relationship name and the question keywords, the matching of keywords contained in the entity / relationship description and the matching of complete phrases in the question. For entities, the weight of exact name matching is higher than the weight of keyword inclusion in the description; for relationships, the weight of subject / object name matching is higher than the weight of relationship description matching.

5. The system according to claim 4, characterized in that... In the relevance scoring, the weight of complete phrase matching is greater than the weight of exact name matching.

6. The system according to claim 1, characterized in that... When calculating the confidence level, the answer generation module performs a weighted average of the six factors: the weight of knowledge completeness is 20%-30%, the weight of knowledge quantity is 15%-25%, the weight of question type matching degree is 10%-20%, the weight of entity coverage is 10%-20%, the weight of source file relevance is 5%-15%, and the weight of knowledge relevance is 10%-20%. Furthermore, the knowledge quantity score is calculated using a non-linear function to avoid unreasonable score increases due to an excessive amount of knowledge.

7. The system according to claim 1, characterized in that... The problem parsing module is configured to: firstly, parse based on preset rules; if it fails, call the large language model for parsing; and then perform post-processing on the parsing results, including entity synonym expansion and entity type inference.

8. The system according to claim 1, characterized in that... The system also includes a cache management module for three-level caching of question parsing results, knowledge retrieval results, and answer generation results. The cache keys are the hash value of the natural language question, the hash value of the question intent object, and the combined hash value of the question, intent, and retrieval results.

9. The system according to any one of claims 1 to 8, characterized in that... The system achieves intelligent question answering by performing the following steps: S1: Receives natural language questions submitted by users; S2: Parse the question and generate a question intent object containing the query type and entities; S3: Based on the question intent object, perform a multi-strategy progressive retrieval from the graph database knowledge graph; the execution includes: first performing batch exact matching of entities, then performing batch query of relationships; if the sum of the number of results obtained does not reach the corresponding dynamic threshold, then perform full-text index search and fuzzy matching in sequence; if the number of entities and relationships is still very small, then perform source file keyword search; S4: Score the relevance of the search results and filter them; S5: Generate an answer based on the filtered knowledge, and calculate the confidence level of the answer by comprehensively evaluating the six factors; S6: Return the result containing the stated answer and confidence level.

Images