Search enhancement generation method and apparatus

Abstract

The application provides a retrieval enhancement generation method and device, the method comprising: performing semantic analysis on an original query to obtain a structured query intention containing a core entity and reasoning requirements; generating a target query variant based on the structured query intention and performing filtering to obtain a query expansion set; inputting the query expansion set and the structured query intention into a hybrid retrieval module to obtain an initial retrieval set; performing multi-dimensional confidence evaluation on retrieval results in the initial retrieval set, and rearranging based on the evaluation results to obtain target retrieval results. The application performs query expansion based on the structured query intention, thereby expanding the semantic coverage of retrieval; performs path reasoning by introducing a dynamic knowledge graph in hybrid retrieval, thereby connecting scattered deep entity relationships and bridging the semantic gap between retrieval results and user real query intentions; and rearranges the initial retrieval set based on multi-dimensional confidence evaluation, thereby guaranteeing the high reliability of the target retrieval results.

Description

Technical Field

[0001] This invention relates to the field of enhanced search generation technology, and in particular to a search enhancement generation method and apparatus. Background Technology

[0002] Retrieval-Augmented Generation (RAG) technology combines retrieved external knowledge with a large language model to improve the accuracy and reliability of the model's output, and has been widely used in scenarios such as intelligent question answering and knowledge base querying. In these applications, achieving accurate and efficient retrieval of external knowledge is a key prerequisite for meeting users' complex information acquisition needs.

[0003] To meet the aforementioned retrieval needs, existing retrieval enhancement solutions typically employ vector retrieval technology. This primarily involves converting user queries into vector representations through embedding models, and then matching and retrieving single-round related information from the database based on semantic similarity. Furthermore, in practical applications, some existing solutions also attempt to utilize manually preset rules to perform simple variations on the query statements, or combine pre-built static knowledge graphs as auxiliary data sources for knowledge retrieval.

[0004] However, existing models have a weak ability to understand technical terms and fuzzy queries, which can easily lead to an overly narrow search scope and an inability to comprehensively cover potential relevant information. When faced with scenarios requiring traversing multiple layers of entity relationships to obtain target information, existing solutions often only match scattered surface-level information, resulting in a severe disconnect between search results and the user's true query intent, creating a semantic gap. Furthermore, existing search processes are prone to mixing a large amount of irrelevant or low-quality interference information into the final results, thus seriously affecting the accuracy of subsequent generation stages. Summary of the Invention

[0005] This invention provides a retrieval enhancement generation method and apparatus to address the shortcomings of existing retrieval enhancement generation schemes, which are limited by the semantic matching depth of vector retrieval and the timeliness of static knowledge bases, making it difficult to accurately parse professional terms and fuzzy queries, and easily falling into the dilemma of surface information matching and noise interference when handling multi-hop association reasoning, resulting in retrieval results being disconnected from the user's deep intent.

[0006] This invention provides a retrieval enhancement generation method, comprising the following steps.

[0007] Obtain the original query input by the user, and perform semantic parsing on the original query to obtain a structured query intent; the structured query intent includes core entities and reasoning requirements; Based on the structured query intent, target query variants are generated, and the target query variants are filtered to obtain a set of query extensions; The query expansion set and the structured query intent are input into the hybrid retrieval module to obtain the initial retrieval set output by the hybrid retrieval module. The hybrid retrieval module includes a vector retrieval module, a knowledge graph retrieval module, and a result merging module. The vector retrieval module is used to perform semantic retrieval in a vector database based on the query expansion set to obtain candidate text blocks. The knowledge graph retrieval module is used to perform path reasoning in a dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is associated with the core entity in multiple hops. The result merging module is used to determine the initial retrieval set based on the candidate text blocks and the associated knowledge information. The search results in the initial search set are evaluated using a multi-dimensional confidence level to obtain the evaluation results. Based on the evaluation results, the initial search set is rearranged to obtain the target search results.

[0008] According to the retrieval enhancement generation method provided by the present invention, the reasoning requirement includes target attributes and reasoning type; The semantic parsing of the original query to obtain the structured query intent includes: The original query is semantically parsed to obtain the core entity, the target attribute, and the reasoning type; Based on the core entity, the target attribute, and the reasoning type, the structured query intent is constructed; The reasoning type includes single-hop reasoning or multi-hop reasoning.

[0009] According to a retrieval enhancement generation method provided by the present invention, the target query variant includes a first query variant, a second query variant, and a third query variant; The generation of the target query variant based on the structured query intent includes: Based on the core entities in the structured query intent, attribute retrieval is performed in the dynamic knowledge graph to obtain the associated attribute information of the core entities; Based on the associated attribute information, the original query is expanded with constraints to obtain the first query variant; Based on the core entity and target attribute in the structured query intent, synonym replacement processing is performed to obtain the second query variant; Based on the reasoning type in the structured query intent, the original query is processed by sentence transformation to obtain the third query variant.

[0010] According to a retrieval enhancement generation method provided by the present invention, the step of performing a multi-dimensional confidence evaluation on the retrieval results in the initial retrieval set to obtain the evaluation result includes: Based on the vector similarity between the first vector of the retrieval result and the second vector of the original query, a semantic matching score is determined. Based on the consistency between the entity relationships in the search results and the records stored in the dynamic knowledge graph, a consistency evaluation score is determined; Based on the degree of coverage of the search results to the core entities and the target attributes, an information integrity score is determined; The evaluation result is determined based on the semantic matching score, the consistency assessment score, and the information integrity score.

[0011] According to a retrieval enhancement generation method provided by the present invention, the step of filtering the target query variants to obtain a query expansion set includes: Determine the semantic similarity between the target query variant and the original query; Filter out the target query variants whose semantic similarity is lower than a preset threshold to obtain the retained target query variants; Based on the reserved target query variant, construct the query extension set.

[0012] According to a retrieval enhancement generation method provided by the present invention, the step of performing path reasoning in a dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is multi-hop related to the core entity includes: Starting with the core entity, retrieve the entity relationship path that matches the reasoning requirement in the dynamic knowledge graph; Perform a multi-step traversal along the entity relationship path to obtain the associated nodes that satisfy the reasoning requirements, as well as the associated description information corresponding to the associated nodes; Based on the entity relationship path and the association description information, the association knowledge information is determined.

[0013] According to a retrieval enhancement generation method provided by the present invention, the step of determining the initial retrieval set based on the candidate text blocks and the associated knowledge information includes: The candidate text blocks, the entity relationship paths in the associated knowledge information, and the associated description information are aggregated to obtain a preliminary retrieval set; The initial search set is obtained by deduplicating the preliminary search set.

[0014] According to a retrieval enhancement generation method provided by the present invention, the construction steps of the dynamic knowledge graph include: Obtain a pre-constructed original knowledge graph; the original knowledge graph stores entity nodes, relation edges between entity nodes, and node attributes; A multimodal extraction method is used to extract results from a domain corpus; the extraction results include entity, relation, and attribute information; the domain corpus includes at least one of text format, table format, and chart format; The entity is matched with the entity nodes in the original knowledge graph to obtain the matching result; When the matching result indicates that there is an attribute conflict or a relationship conflict, the original knowledge graph is incrementally updated based on the extraction result to obtain the dynamic knowledge graph; When the matching result indicates that there is no conflict, the original knowledge graph is updated based on the extraction result to obtain the dynamic knowledge graph.

[0015] According to a retrieval enhancement generation method provided by the present invention, the step of incrementally updating the original knowledge graph based on the extraction result to obtain the dynamic knowledge graph includes: When new corpus content is detected in the domain corpus, or when an entity relationship error instruction is received from a user, the original knowledge graph is incrementally updated based on the extraction results to obtain the dynamic knowledge graph.

[0016] The present invention also provides a retrieval enhancement generation device, comprising the following units: The acquisition unit is used to acquire the original query input by the user and perform semantic parsing on the original query to obtain a structured query intent; the structured query intent includes core entities and reasoning requirements; A filtering unit is used to generate target query variants based on the structured query intent, and to filter the target query variants to obtain a set of query extensions; An input unit is used to input the query expansion set and the structured query intent into the hybrid retrieval module to obtain an initial retrieval set output by the hybrid retrieval module. The hybrid retrieval module includes a vector retrieval module, a knowledge graph retrieval module, and a result merging module. The vector retrieval module is used to perform semantic retrieval in a vector database based on the query expansion set to obtain candidate text blocks. The knowledge graph retrieval module is used to perform path reasoning in a dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is multi-hop related to the core entity. The result merging module is used to determine the initial retrieval set based on the candidate text blocks and the associated knowledge information. An evaluation unit is used to perform multi-dimensional confidence evaluation on the search results in the initial search set, obtain evaluation results, and rearrange the initial search set based on the evaluation results to obtain the target search results.

[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the retrieval enhancement generation method as described above.

[0018] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the retrieval enhancement generation method as described above.

[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the retrieval enhancement generation method as described above.

[0020] The retrieval enhancement generation method and apparatus provided by this invention expands the query based on structured query intent, thereby broadening the semantic coverage of the retrieval. This avoids, to some extent, the problem of existing models having a narrow retrieval scope due to weak understanding of technical terms and fuzzy queries, and comprehensively covers potential relevant information. By introducing dynamic knowledge graphs into hybrid retrieval to perform path reasoning to obtain multi-hop association information of core entities, it achieves in-depth mining and association of logical paths between entities, overcoming the problem of only matching scattered surface information when facing multi-layer entity association scenarios, and bridging the semantic gap between retrieval results and users' true query intent. Finally, the initial retrieval set is rearranged based on multi-dimensional confidence assessment, thereby effectively quantifying and filtering irrelevant or low-quality interference information, ensuring the high reliability of the target retrieval results. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is one of the flowcharts of the retrieval enhancement generation method provided by the present invention.

[0023] Figure 2 This is the second flowchart of the retrieval enhancement generation method provided by the present invention.

[0024] Figure 3 This is the internal hierarchical architecture of the dynamic knowledge graph module provided by this invention.

[0025] Figure 4 This is a schematic diagram of the hybrid retrieval execution process provided by the present invention.

[0026] Figure 5This is an architecture diagram of the retrieval enhancement generation system provided by the present invention.

[0027] Figure 6 This is a schematic diagram of the retrieval enhancement generation device provided by the present invention.

[0028] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] This invention provides a retrieval enhancement generation method, which can be executed by a computing device or server of an enhanced retrieval system. Figure 1 This is one of the flowcharts illustrating the retrieval enhancement generation method provided by the present invention, such as... Figure 1 As shown, the method includes the following: Step 110: Obtain the original query input by the user and perform semantic parsing on the original query to obtain a structured query intent; the structured query intent includes core entities and reasoning requirements.

[0031] Specifically, firstly, the user's original query can be obtained, and semantic parsing can be performed on the original query to obtain a structured query intent.

[0032] Here, the original query refers to the natural language question or search text information entered by the user in retrieval scenarios such as intelligent question answering and knowledge base query.

[0033] Semantic parsing of the original query can typically be performed using a pre-trained natural language processing model or an intent recognition algorithm as the parsing engine. Specifically, after obtaining the original query, it is used as input to the intent recognition model. The intent recognition model performs word segmentation, part-of-speech tagging, and syntactic analysis on the original query to output a structured query intent.

[0034] Among them, the structured query intent is used to reflect the set of key information obtained after abstracting and structuring the deep semantics of the original query.

[0035] Here, the structured query intent includes at least core entities and reasoning requirements. Core entities reflect the main objects, names of people, names of organizations, or concept nodes involved in the original query; reasoning requirements reflect the user's expectation of obtaining logical relationships, deep attributes, or target information that requires traversing multiple layers of entity relationships related to the core entities.

[0036] Step 120: Generate target query variants based on the structured query intent, and filter the target query variants to obtain a query extension set.

[0037] Specifically, after obtaining the structured query intent, target query variants can be generated based on the structured query intent, and the target query variants can be filtered to obtain a set of query extensions.

[0038] Considering that the original expressions of user queries are often limited or ambiguous, this step enriches the semantic coverage of the retrieval through a dynamic expansion mechanism.

[0039] Here, target query variants refer to query expressions with similar or related retrieval intents generated through multi-dimensional language transformation or logical derivation based on the core semantics of the original query. In practice, the system will automatically generate multiple sets of differentiated target query variants based on the structured query intent, employing a combination of expansion strategies such as model-based synonym generalization, rule-based condition supplementation, and syntax-based sentence transformation. This aims to effectively broaden the retrieval boundaries and avoid the risks of insufficient recall and missing key information caused by a single query paradigm.

[0040] Because the generated target query variants may contain content that deviates from the original semantics or is redundant and invalid, it is necessary to filter these variants. Filtering methods may include calculating the relevance or similarity between each target query variant and the original query, and removing variants with low relevance. The high-quality, highly relevant target query variants retained after filtering constitute the query expansion set. The query expansion set refers to the expanded query statements that, after filtering, can be directly used for parallel retrieval.

[0041] Step 130: Input the query expansion set and the structured query intent into the hybrid retrieval module to obtain the initial retrieval set output by the hybrid retrieval module; the hybrid retrieval module includes a vector retrieval module, a knowledge graph retrieval module, and a result merging module; the vector retrieval module is used to perform semantic retrieval in the vector database based on the query expansion set to obtain candidate text blocks; the knowledge graph retrieval module is used to perform path reasoning in the dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is multi-hop associated with the core entity; the result merging module is used to determine the initial retrieval set based on the candidate text blocks and the associated knowledge information.

[0042] Specifically, after obtaining the query expansion set, the query expansion set and the structured query intent can be input into the hybrid retrieval module to obtain the initial retrieval set output by the hybrid retrieval module.

[0043] Here, the hybrid retrieval module refers to a composite retrieval architecture that integrates different underlying retrieval mechanisms. After the query extension set and structured query intent are input into the hybrid retrieval module simultaneously, the sub-modules within the hybrid retrieval module will execute retrieval tasks in parallel.

[0044] The hybrid retrieval module can include a vector retrieval module, a knowledge graph retrieval module, and a result merging module. The vector retrieval module is used to perform semantic retrieval based on the query expansion set to obtain candidate text blocks, while the knowledge graph retrieval module is used to perform path reasoning in the dynamic knowledge graph based on reasoning needs and core entities to obtain associated knowledge information that is linked to the core entities in multiple hops.

[0045] Here, the vector retrieval module primarily utilizes deep learning embedding techniques to convert each query variant in the query expansion set into a high-dimensional vector representation. This high-dimensional vector representation is then matched against data in a pre-built vector database to retrieve semantically similar fragments, i.e., candidate text blocks. Candidate text blocks refer to unstructured or semi-structured document fragments retrieved based on semantic similarity matching.

[0046] Here, the knowledge graph retrieval module primarily relies on an underlying dynamic knowledge graph for logical retrieval. The dynamic knowledge graph refers to a structured knowledge network extracted from a domain corpus and capable of real-time maintenance and updates. This structured knowledge network contains a large number of entity nodes, relational edges, and attribute information, and supports real-time addition, modification, and deletion of entity relationships in response to constantly changing data within the domain, ensuring the timeliness of knowledge. Based on this, the knowledge graph retrieval module triggers and executes path reasoning algorithms, such as graph traversal and graph neural network reasoning, within the dynamic knowledge graph according to the reasoning requirements and core entities in the structured query intent. By connecting dispersed knowledge nodes, such as product-manufacturer-raw material supplier, or raw material supplier-location-regional policy, it obtains related knowledge information that cannot be directly matched by literal semantics alone and has multi-hop logical connections with the core entities.

[0047] Among them, the associated knowledge information is used to reflect the multi-hop factual data and associated descriptive information obtained through deep relational reasoning of the knowledge graph.

[0048] Here, the result merging module is used to determine the initial retrieval set based on candidate text blocks and related knowledge information. The specific integration process can employ set deduplication algorithms, entity alignment mechanisms, etc., to eliminate overlapping or redundant information in the two types of retrieval results, organically combining candidate text blocks and related knowledge information to determine the initial retrieval set. The initial retrieval set refers to the candidate information set that initially summarizes the results of semantic retrieval and graph reasoning retrieval.

[0049] Step 140: Perform multi-dimensional confidence assessment on the search results in the initial search set to obtain the assessment results, and rearrange the initial search set based on the assessment results to obtain the target search results.

[0050] Specifically, after obtaining the initial search set, the search results in the initial search set can be evaluated in multiple dimensions to obtain the evaluation results. Based on the evaluation results, the initial search set can be rearranged to obtain the target search results.

[0051] To prevent irrelevant or erroneous information from interfering with subsequent generation or display processes, quality control of the retrieved results is necessary. Multi-dimensional confidence assessment refers to a process that goes beyond a single semantic score, combining semantic, logical, and factual aspects of the search results to quantify their reliability.

[0052] After the above evaluation, each search result in the initial search set is assigned a comprehensive confidence score or reliability index, thus obtaining the evaluation result. Here, the evaluation result reflects the multi-dimensional reliability quantification information of the initial search set. Finally, based on the ranking of the scores in the evaluation result, all search results in the initial search set are reordered, and the header results with high confidence are extracted as the final target search results. Here, the target search results refer to the highly accurate and reliable final search information sources selected after rigorous evaluation and reordering.

[0053] The method provided in this invention expands the query based on structured query intent, thereby broadening the semantic coverage of the retrieval. This avoids, to some extent, the problem of existing models having a narrow retrieval scope due to their weak understanding of technical terms and fuzzy queries, and comprehensively covers potential relevant information. By introducing dynamic knowledge graphs into hybrid retrieval to perform path reasoning and obtain multi-hop association information of core entities, it achieves in-depth mining and association of logical paths between entities, overcoming the problem of only matching scattered surface information when facing multi-layered entity association scenarios, and bridging the semantic gap between retrieval results and the user's true query intent. Finally, the initial retrieval set is rearranged based on multi-dimensional confidence assessment, thereby effectively quantifying and filtering irrelevant or low-quality interference information, ensuring the high reliability of the target retrieval results.

[0054] Based on the above embodiments, the reasoning requirements include target attributes and reasoning types; Step 110, which involves semantic parsing of the original query to obtain a structured query intent, includes: Step 1101: Perform semantic parsing on the original query to obtain the core entity, the target attribute, and the reasoning type; Step 1102: Construct the structured query intent based on the core entity, the target attribute, and the reasoning type; The reasoning type includes single-hop reasoning or multi-hop reasoning.

[0055] Specifically, reasoning requirements include target attributes and reasoning types.

[0056] Accordingly, semantic parsing can be performed on the original query to obtain the core entity, target attribute, and reasoning type.

[0057] Here, the target attribute refers to the specific characteristics, multi-dimensional characterization, or target state that the user expects to obtain regarding the core entity. For example, when the original query is "the recent scientific research output of the alma mater of Li, the CEO of a new energy company", the target attribute would be "scientific research output in the energy field".

[0058] The reasoning type reflects the information in the hierarchy of knowledge nodes or the depth of logical deduction required to answer the original query. Here, the reasoning type can include single-hop reasoning or multi-hop reasoning. Single-hop reasoning refers to a shallow logical reasoning process that obtains the answer through a single entity and its direct attributes; multi-hop reasoning refers to a deep logical reasoning process that requires bridging at least two entities and the multiple layers of relationships between them to obtain the answer.

[0059] In practical implementation, the system can use a fine-tuned BERT (Bidirectional Encoder Representations from Transformers) model to perform deep semantic parsing of the original query. The BERT model identifies and outputs a structured representation—the core entity, target attribute, and inference type—by understanding syntax and contextual semantics. For example, parsing the original query using the BERT model can identify the core entities as "Li" and "a university," the target attribute as research output in the energy field, and, because it involves traversing multiple nodes such as the person, the company they work for, and the university they graduated from, the inference type is identified as multi-hop inference.

[0060] The method provided in this invention constructs a structured query intent by refining the identification of core entities, target attributes, and single / multi-hop reasoning types. This enables a more granular and accurate structured decomposition of natural language queries, thereby avoiding deviations in the retrieval direction due to incomplete intent identification and ensuring the accuracy of mixed retrieval under different reasoning paths.

[0061] Based on the above embodiments, the target query variants include a first query variant, a second query variant, and a third query variant; Step 120, which involves generating a target query variant based on the structured query intent, includes: Step 1201: Based on the core entity in the structured query intent, perform attribute retrieval in the dynamic knowledge graph to obtain the associated attribute information of the core entity; Step 1202: Expand the constraints of the original query based on the associated attribute information to obtain the first query variant; Step 1203: Based on the core entity and target attribute in the structured query intent, perform synonym replacement processing to obtain the second query variant; Step 1204: Based on the reasoning type in the structured query intent, perform sentence transformation processing on the original query to obtain the third query variant.

[0062] Specifically, the target query variants may include a first query variant, a second query variant, and a third query variant.

[0063] Accordingly, attribute retrieval can be performed on the dynamic knowledge graph based on the core entity in the structured query intent to obtain the associated attribute information of the core entity. Then, the original query is expanded with constraints based on the associated attribute information to obtain the first query variant. Here, the core entity is used as the query node, and the entity node corresponding to the core entity is accurately located in the dynamic knowledge graph using graph database query language. Subsequently, the relation edges directly connected to the entity node are traversed to extract the corresponding attribute key-value pairs. For example, the "subject field" attribute value of the entity node "a certain university" is obtained as "energy science", thus obtaining the associated attribute information. Further, the extracted associated attribute information is transformed into natural language phrases or keyword fragments. Through preset string concatenation rules or prompt word templates based on a large language model, the attribute fragments are inserted into the limiting context of the original query, or directly concatenated as appended limiting conditions at the end of the original query, thereby generating a more specific first query variant.

[0064] Here, the first query variant refers to the extended query obtained after applying contextual logical constraints based on the objective association attributes of the graph.

[0065] Then, based on the core entities and target attributes in the structured query intent, synonym replacement processing can be performed to obtain a second query variant. Specifically, a pre-built domain-specific thesaurus or domain-specific word vector model is pre-loaded, such as a Word2Vec or BERT model fine-tuned based on a specific corpus; key terms corresponding to the core entities and target attributes in the original query are extracted, and hard matching is performed in the thesaurus, or the core entities and target attributes are input into the word vector model to calculate the cosine similarity between word vectors, and the top K candidate words with the highest cosine similarity are extracted. Then, the original terms are replaced with the extracted candidate words according to natural language grammar, thereby generating the second query variant.

[0066] Finally, based on the reasoning type in the structured query intent, the original query can be transformed to obtain a third query variant. After the system identifies the reasoning type as multi-hop reasoning, it calls the internally integrated syntactic dependency parser or a finely tuned large language model to decompose and reconstruct the original interrogative or complex nested sentence structure into a declarative sentence structure or a multi-keyword combination sentence structure oriented towards the target attribute, such as "a certain university-recent-energy-paper-patent". The reconstruction rules focus on stripping away intermediate modifiers, directly reducing the syntactic distance between the cross-hop entity and the target attribute, thus obtaining the third query variant.

[0067] Here, the second query variant refers to the expanded query obtained by replacing the core element with synonyms or near-synonyms; the third query variant refers to the expanded query obtained by changing the grammatical structure or key word order of the original sentence according to the requirements of reasoning depth.

[0068] In specific implementation, for the first query variant, attribute retrieval can be performed on the dynamic knowledge graph based on the core entity "a certain university" to obtain its associated attribute information "subject field - energy science", which is then used as a query constraint to generate the first query variant "Li Mou - alma mater - a certain university - energy - scientific research achievements"; for the second query variant, a domain word vector model is used to obtain synonyms of core terms, and the words in the original query are replaced with synonyms to generate the second query variant "a certain new energy company - CEO - alma mater - scientific research output - energy field"; for the third query variant, based on the multi-hop reasoning type, the syntactic structure of the original query is deconstructed and transformed to generate the third query variant "a certain university - recent years - energy - papers - patents" which is easier to match directly.

[0069] The method provided in this invention generates various query variants through synonym replacement, sentence transformation, and constraint condition expansion based on dynamic knowledge graphs. This greatly enriches the semantic expression of query intent and, to a certain extent, overcomes the limitations of the embedding model in understanding a single fuzzy query statement, significantly improving the retrieval coverage of potentially relevant information.

[0070] Based on the above embodiments, step 140, which involves performing a multi-dimensional confidence assessment on the search results in the initial search set to obtain the assessment results, includes: Step 1401: Determine the semantic matching score based on the vector similarity between the first vector of the retrieval result and the second vector of the original query; Step 1402: Determine the consistency evaluation score based on the consistency between the entity relationships in the search results and the records stored in the dynamic knowledge graph; Step 1403: Determine the information integrity score based on the degree of coverage of the core entity and the target attribute by the search results; Step 1404: Determine the evaluation result based on the semantic matching score, the consistency evaluation score, and the information integrity score.

[0071] Specifically, firstly, the semantic matching score can be determined based on the vector similarity between the first vector of the retrieval result and the second vector of the original query.

[0072] Here, the first vector refers to the dense feature representation obtained by encoding each search result in the initial search set through the embedding model; the second vector refers to the feature representation obtained by encoding the user's original query in the same way.

[0073] The semantic matching score is used to reflect the semantic relevance of the text between the first vector and the second vector.

[0074] Then, a consistency assessment score can be determined based on the consistency between the entity relationships in the search results and the records stored in the dynamic knowledge graph.

[0075] Here, entity relations are used to reflect the connection or interaction status between the objective entities actually presented in the search results text. Records stored in a dynamic knowledge graph refer to authoritative entity association data that is maintained and verified in real time in the system's underlying knowledge base.

[0076] The consistency assessment score reflects the degree of consistency or logical contradiction between the factual descriptions in the search results and the authoritative facts in the system map.

[0077] Furthermore, the information integrity score can be determined based on the degree of coverage of core entities and target attributes by the search results.

[0078] Here, coverage reflects the proportion or range of core entities and target attributes that are matched in a single search result, as required by the original query. The information completeness score refers to the comprehensiveness of the search results in presenting key information elements.

[0079] Finally, the evaluation results can be determined based on semantic matching scores, consistency assessment scores, and information integrity scores.

[0080] In practical implementation, the system can set corresponding weights for the three evaluation dimensions mentioned above: semantic matching score is set to 0.4, entity relationship consistency score is set to 0.3, and information integrity score is set to 0.3. For example, when scoring a retrieved "abstract of an energy paper from a university in 2024": the system calculates the vector similarity between the first and second vectors, converts it to a 0-10 scale, and obtains a semantic matching score of 9; it verifies that the entity associations in the search results are completely consistent with the records stored in the dynamic knowledge graph, and obtains a consistency evaluation score of 10, where partial contradictions earn 5 points and complete contradictions earn 0 points; the evaluation finds that the result completely covers all core entities and target attributes of the query intent, and obtains an information integrity score of 10, where partial coverage earns 3-7 points and no coverage earns 0 points. Finally, based on a weighted sum (9×0.4 + 10×0.3 + 10×0.3 = 9.6 points), the evaluation result (confidence score) of the search result is obtained. The system then reorders the initial search set according to the confidence score from high to low.

[0081] Based on the above embodiments, the step 120 of filtering the target query variants to obtain the query extension set includes: Step 121: Determine the semantic similarity between the target query variant and the original query; Step 122: Filter out the target query variants whose semantic similarity is lower than a preset threshold to obtain the retained target query variants; Step 123: Based on the reserved target query variant, construct the query expansion set.

[0082] Specifically, firstly, the semantic similarity between the target query variant and the original query can be determined. Semantic similarity reflects the degree of approximation or correlation between the target query variant and the original query in a multidimensional semantic feature space.

[0083] Then, target query variants with semantic similarity below a preset threshold can be filtered out to obtain retained target query variants. The preset threshold refers to the lowest numerical benchmark used to distinguish whether a variant is of acceptable quality, either manually set by the system or determined based on historical experience.

[0084] Here, retaining target query variants refers to valid query variants with semantic similarity higher than or equal to a preset threshold that have not been filtered out by the system.

[0085] Finally, a query expansion set can be constructed based on the retained target query variants. In practice, the system calculates the semantic similarity scores between the first, second, and third query variants and the original query. For example, the system sets a preset threshold of 0.7. Calculations show that the semantic similarity scores of the first, second, and third query variants with the original query are all higher than 0.85, meeting the requirement of being higher than 0.7. Therefore, no variants are filtered, and these three retained target query variants are combined to form the final query expansion set for parallel retrieval.

[0086] The method provided in this invention filters target query variants by calculating semantic similarity and setting a preset threshold, thereby eliminating invalid variants that deviate from the original query intent or are unclear in meaning, thus avoiding the inclusion of retrieval noise caused by excessive divergence, and ensuring that the query expansion set maintains extremely high quality and semantic accuracy while expanding its scope.

[0087] Based on the above embodiments, the step of performing path reasoning in the dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is multi-hop related to the core entity includes: Step 21: Using the core entity as the starting node, retrieve the entity relationship path that matches the reasoning requirement in the dynamic knowledge graph; Step 22: Perform a multi-step traversal along the entity relationship path to obtain the associated nodes that meet the reasoning requirements, as well as the associated description information corresponding to the associated nodes; Step 23: Determine the associated knowledge information based on the entity relationship path and the associated description information.

[0088] Specifically, firstly, starting with the core entity, one can retrieve the entity relationship path that matches the reasoning requirements in the dynamic knowledge graph.

[0089] Here, entity relationship path refers to the logical transmission link in a dynamic knowledge graph, which consists of multiple entity nodes and interconnected relationship edges connected end to end.

[0090] Then, multiple traversals can be performed along the entity relationship path to obtain the associated nodes that meet the reasoning requirements, as well as the associated description information corresponding to the associated nodes.

[0091] Here, an associated node refers to an entity object that meets the original query reasoning requirements and is finally reached after traversing multiple paths in the knowledge graph. The associated description information reflects the text content attached to the associated node or relation edge, providing specific attribute descriptions or factual explanations of the related entity.

[0092] Finally, associated knowledge information can be determined based on entity relationship paths and associated description information.

[0093] In practical implementation, for multi-hop reasoning queries, the knowledge graph retrieval module uses core entities in the intent recognition results, such as "Li Mou," as the starting node. It then executes path reasoning algorithms, such as graph traversal algorithms, within the dynamic knowledge graph. By continuously cascading downwards, it traverses the entity relationship path "E-R1-E1-R2-E2-...-An," for example, traversing the path "Li Mou-Graduated from-a certain university-Discipline-Energy Science-Research Achievements-2023-2024 Papers / Patents." This process obtains associated nodes that satisfy the target attribute ("2023-2024 Papers"), along with corresponding descriptive information such as paper titles, authors, and abstracts. This information is ultimately integrated into the associated knowledge information.

[0094] The method provided in this invention connects originally scattered and isolated knowledge nodes by performing multi-step traversal of entity relationship paths starting from the core entity in a dynamic knowledge graph. This overcomes the shortcomings of traditional vector retrieval in handling complex logical relationships to a certain extent, effectively solves the reasoning discontinuity problem in multi-hop retrieval scenarios, and significantly improves the logical depth of retrieval results.

[0095] Based on the above embodiments, determining the initial retrieval set based on the candidate text blocks and the associated knowledge information includes: Step 31: The candidate text blocks, the entity relationship paths in the associated knowledge information, and the associated description information are aggregated to obtain a preliminary retrieval set; Step 32: Perform deduplication on the preliminary search set to obtain the initial search set.

[0096] Specifically, firstly, candidate text blocks, entity relationship paths in associated knowledge information, and associated description information can be aggregated to obtain a preliminary retrieval set.

[0097] Here, the preliminary retrieval set refers to the original mixed data set that directly combines the unstructured candidate text blocks obtained from vector retrieval with the structured paths and descriptions obtained from knowledge graph logical reasoning, without undergoing any redundancy removal processing.

[0098] In practice, since vector retrieval and knowledge graph retrieval are performed in parallel, information overlap is inevitable in the results of the two retrieval paths. The system first combines the obtained Top-5 candidate text blocks with three entity relationship paths and their associated descriptive information to form a preliminary retrieval set. Then, the preliminary retrieval set is deduplicated to obtain an initial retrieval set. Here, the system uses set deduplication algorithms, such as text similarity-based deduplication or entity referencing resolution algorithms, to remove parts of the candidate text blocks that are semantically highly repetitive with the associated descriptive information, such as duplicate paper information, while retaining complementary candidate text blocks, entity relationship paths, and associated descriptive information to form the final initial retrieval set.

[0099] The method provided in this invention achieves seamless integration of structured logical graph data and unstructured semantic text data by summarizing and deduplicating the two search results. This effectively eliminates redundant and overlapping data while preserving breadth and depth, ensuring the purity and richness of information in the initial search set.

[0100] Based on the above embodiments, the steps for constructing the dynamic knowledge graph include: Step 41: Obtain the pre-constructed original knowledge graph; the original knowledge graph stores entity nodes, relation edges between entity nodes, and node attributes; Step 42: Extract the extraction results from the domain corpus using a multimodal extraction method; the extraction results include entity, relation, and attribute information; the domain corpus includes at least one of text format, table format, and chart format; Step 43: Match the entity with the entity nodes in the original knowledge graph to obtain the matching result; Step 44: When the matching result indicates that there is an attribute conflict or a relationship conflict, perform an incremental update on the original knowledge graph based on the extraction result to obtain the dynamic knowledge graph; Step 45: When the matching result indicates that there is no conflict, update the original knowledge graph based on the extraction result to obtain the dynamic knowledge graph.

[0101] Specifically, firstly, a pre-constructed original knowledge graph can be obtained, which stores entity nodes, relational edges between entity nodes, and node attributes.

[0102] Here, the original knowledge graph refers to pre-constructed structured graph data in a baseline state. Entity nodes are independent basic units of knowledge in the original knowledge graph, representing specific names, organizations, concepts, etc. Relationship edges are structural lines connecting different entity nodes, representing specific semantic associations or logical interactions between entities, such as semantic relationships or logical reasoning relationships. Node attributes describe the inherent characteristics, category status, or specific numerical values ​​of entity nodes, providing additional dimensions.

[0103] Then, multimodal extraction can be used to extract results from the domain corpus, wherein the extraction results include entity, relation and attribute information; the domain corpus includes at least one of text format, table format and chart format.

[0104] Domain corpora are used to reflect the raw data sets collected from specific industries or scenarios as sources of knowledge extraction. Text format, table format, and chart format refer to the different multimodal data carrier representations presented by the domain corpora.

[0105] Furthermore, the entity can be matched with the entity nodes in the original knowledge graph to obtain the matching results.

[0106] Here, the matching result is used to reflect the alignment status determined after comparing the newly extracted entities with existing knowledge graph nodes through entity linking technology. The alignment status includes consistency, addition, or contradiction.

[0107] When the matching result indicates that there is an attribute conflict or a relationship conflict, the original knowledge graph is incrementally updated based on the extraction result to obtain a dynamic knowledge graph; when the matching result indicates that there is no conflict, the original knowledge graph is updated based on the extraction result to obtain a dynamic knowledge graph.

[0108] In its implementation, the system employs a multimodal information extraction model to extract new entities, relationships, and attributes from domain corpora containing text, tables, or charts, such as corporate annual reports and academic papers. Subsequently, entity linking technology is used to match the newly extracted entities with nodes in the original knowledge graph. If the matching results reveal attribute conflicts—for example, old records show someone works for company A, while new corpora show they work for company B—the system uses conflict detection algorithms to resolve the factual discrepancies, such as based on timestamp overlay or confidence level judgments, and performs incremental updates. If no conflict exists, the new knowledge is directly incorporated, thereby generating and maintaining a highly timely dynamic knowledge graph.

[0109] The method provided in this invention achieves dynamic iteration and contradiction resolution of knowledge graphs through multimodal extraction combined with an incremental update mechanism of conflict detection and matching. This overcomes, to some extent, the pain point that static graphs cannot adapt to the rapid evolution of knowledge, and ensures the high timeliness and accuracy of the underlying supporting data.

[0110] Based on the above embodiments, step 44 includes: Step 44-1: When new corpus content is detected in the domain corpus, or when an entity relationship error instruction is received from the user, the original knowledge graph is incrementally updated based on the extraction results to obtain the dynamic knowledge graph.

[0111] Specifically, when new corpus content is detected in the domain corpus, or when an entity relationship error instruction is received from a user, the original knowledge graph is incrementally updated based on the extraction results to obtain a dynamic knowledge graph.

[0112] In its implementation, the system employs an automated knowledge update triggering mechanism. On one hand, the system monitors the domain corpus in real time. When it detects new data in the academic database, such as "energy field papers from a certain university in 2024," it automatically triggers the extraction and update process. On the other hand, when the terminal system receives entity relationship correction instructions from manual review or user feedback, it immediately activates the incremental update logic to correct and overwrite erroneous records in the original knowledge graph.

[0113] The method provided in this invention establishes a dual update triggering mechanism based on new corpus monitoring and user feedback, thereby endowing the system with agile knowledge self-healing and self-growth capabilities. This effectively avoids long-term contamination of search results by outdated and erroneous knowledge, further consolidating the timeliness and reliability of the entire enhanced search system in practical application scenarios.

[0114] Based on any of the above embodiments Figure 2 This is the second flowchart of the retrieval enhancement generation method provided by the present invention, as shown below. Figure 2As shown, the system first executes the dynamic knowledge graph construction module, which includes three parallel core sub-processes: knowledge extraction, knowledge fusion, and dynamic maintenance. Next, it proceeds to the query intent parsing and dynamic expansion stage. Specifically, it identifies the user's core needs through intent recognition, generates rich semantic expressions using variants, and ensures expansion quality through variant filtering. Following this, a hybrid retrieval process is performed. In this stage, the system simultaneously initiates vector retrieval and knowledge graph retrieval to balance semantic relevance and logical deduction depth, and merges the results to eliminate data redundancy. Further, confidence assessment and ranking are conducted. Multi-dimensional scoring is used to quantify the recall results, and priority is determined based on the result ranking. Finally, the system outputs the target retrieval results (including confidence scores) and reaches the termination node, thus completing a closed-loop augmented retrieval process.

[0115] Based on any of the above embodiments Figure 3 This is the internal hierarchical architecture of the dynamic knowledge graph module provided by this invention, such as... Figure 3 As shown, a domain corpus containing multi-dimensional data such as text, tables, and charts is used as the input source and fed into the knowledge extraction unit. The knowledge extraction unit extracts features from massive amounts of unstructured data by calling a multimodal information extraction model. The extracted raw knowledge then enters the knowledge fusion unit, where entity linking technology is used to align knowledge, and a conflict detection algorithm is used to resolve factual inconsistencies to ensure accuracy. The fused high-fidelity knowledge is output to the dynamic maintenance unit for unified management. The dynamic maintenance unit also receives external error correction information from user feedback (incorrect entity relationships) and achieves automatic knowledge growth and self-healing through an update trigger mechanism. Under the unified scheduling of the dynamic maintenance unit, the system updates and maintains the entity node set, relationship edge set, and attribute information set in real time. Finally, this well-managed structured data is output through a multi-hop inference data interface, providing underlying data support with strong logical deduction capabilities for the aforementioned knowledge graph retrieval submodule.

[0116] Based on any of the above embodiments Figure 4 This is a schematic diagram of the hybrid retrieval execution process provided by the present invention, such as... Figure 4As shown, the system takes the extended query set as input and then splits into two parallel retrieval branches for execution: the left branch uses a vector retrieval module to vectorize query variants and match them with a vector database to obtain Top-N semantically similar text blocks, ensuring comprehensive semantic coverage; the right branch uses a knowledge graph retrieval module to first extract the core entity E, then execute a path reasoning algorithm and traverse the entity relationship path E-R1-E1-...-An, ultimately obtaining multi-hop association information to solve deep logic derivation problems. The retrieval products from both branches converge in a result merging module, where a set deduplication algorithm is called to remove redundancy, thus retaining text blocks and entity relationship paths. Finally, the system integrates the heterogeneous knowledge and outputs the initial retrieval set.

[0117] Based on any of the above embodiments Figure 5 This is an architecture diagram of the retrieval enhancement generation system provided by the present invention, such as... Figure 5 As shown, the retrieval enhancement generation system consists of a data layer, a service layer, and an application layer from bottom to top. The bottom data layer serves as the foundation, deploying a domain corpus, a knowledge graph database (Neo4j), a vector database (Milvus), and a retrieval log database, responsible for the persistent storage of all heterogeneous data. The middle service layer is the core logic layer of the system, containing four collaborative service modules: query processing service, dynamic knowledge graph service, hybrid retrieval service, and confidence assessment service. The query processing service comprises an intent recognition unit C1, a variant generation unit C2, and a variant filtering unit C3. The dynamic knowledge graph service provides entity relationship data. The hybrid retrieval service integrates a vector retrieval submodule D1 and a knowledge graph retrieval submodule D2. The initial retrieval results produced by the hybrid retrieval service are directly input into the confidence assessment service for reliability quantification. The top application layer connects with external large models through the RAG system's application programming interface (API), completing the final delivery of retrieval enhancement capabilities. This architecture ensures efficient collaboration across the entire process of knowledge extraction, query processing, hybrid retrieval, and evaluation and ranking through logical connections between different levels.

[0118] The retrieval enhancement generation apparatus provided by the present invention is described below. The retrieval enhancement generation apparatus described below can be referred to in correspondence with the retrieval enhancement generation method described above.

[0119] Based on any of the above embodiments, the present invention provides a retrieval enhancement generation device. Figure 6 This is a schematic diagram of the retrieval enhancement generation device provided by the present invention, as shown below. Figure 6 As shown, the device includes: The acquisition unit 610 is used to acquire the original query input by the user and perform semantic parsing on the original query to obtain a structured query intent; the structured query intent includes core entities and reasoning requirements; The filtering unit 620 is used to generate target query variants based on the structured query intent, and filter the target query variants to obtain a set of query extensions; Input unit 630 is used to input the query expansion set and the structured query intent into the hybrid retrieval module to obtain the initial retrieval set output by the hybrid retrieval module; the hybrid retrieval module includes a vector retrieval module, a knowledge graph retrieval module, and a result merging module; the vector retrieval module is used to perform semantic retrieval in a vector database based on the query expansion set to obtain candidate text blocks; the knowledge graph retrieval module is used to perform path reasoning in a dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is multi-hop associated with the core entity; the result merging module is used to determine the initial retrieval set based on the candidate text blocks and the associated knowledge information; The evaluation unit 640 is used to perform multi-dimensional confidence evaluation on the search results in the initial search set, obtain the evaluation result, and rearrange the initial search set based on the evaluation result to obtain the target search result.

[0120] The apparatus provided in this invention expands the query based on structured query intent, thereby broadening the semantic coverage of the retrieval. This avoids, to some extent, the problem of existing models having a narrow retrieval scope due to weak understanding of technical terms and fuzzy queries, and comprehensively covers potential relevant information. By introducing dynamic knowledge graphs into hybrid retrieval to perform path reasoning to obtain multi-hop association information of core entities, it achieves in-depth mining and association of logical paths between entities, overcoming the problem of only matching scattered surface information when facing multi-layered entity association scenarios, and bridging the semantic gap between retrieval results and the user's true query intent. Finally, the initial retrieval set is rearranged based on multi-dimensional confidence assessment, thereby effectively quantifying and filtering irrelevant or low-quality interference information, ensuring the high reliability of the target retrieval results.

[0121] Based on any of the above embodiments, the reasoning requirement includes target attributes and reasoning type; The acquisition unit 610 is specifically used for: The original query is semantically parsed to obtain the core entity, the target attribute, and the reasoning type; Based on the core entity, the target attribute, and the reasoning type, the structured query intent is constructed; The reasoning type includes single-hop reasoning or multi-hop reasoning.

[0122] Based on any of the above embodiments, the target query variant includes a first query variant, a second query variant, and a third query variant; The filter unit 620 is specifically used for: Based on the core entities in the structured query intent, attribute retrieval is performed in the dynamic knowledge graph to obtain the associated attribute information of the core entities; Based on the associated attribute information, the original query is expanded with constraints to obtain the first query variant; Based on the core entity and target attribute in the structured query intent, synonym replacement processing is performed to obtain the second query variant; Based on the reasoning type in the structured query intent, the original query is processed by sentence transformation to obtain the third query variant.

[0123] Based on any of the above embodiments, the evaluation unit 640 is specifically used for: Based on the vector similarity between the first vector of the retrieval result and the second vector of the original query, a semantic matching score is determined. Based on the consistency between the entity relationships in the search results and the records stored in the dynamic knowledge graph, a consistency evaluation score is determined; Based on the degree of coverage of the search results to the core entities and the target attributes, an information integrity score is determined; The evaluation result is determined based on the semantic matching score, the consistency assessment score, and the information integrity score.

[0124] Based on any of the above embodiments, the filtering unit 620 is specifically used for: Determine the semantic similarity between the target query variant and the original query; Filter out the target query variants whose semantic similarity is lower than a preset threshold to obtain the retained target query variants; Based on the reserved target query variant, construct the query extension set.

[0125] Based on any of the above embodiments, the input unit 630 is specifically used for: Starting with the core entity, retrieve the entity relationship path that matches the reasoning requirement in the dynamic knowledge graph; Perform a multi-step traversal along the entity relationship path to obtain the associated nodes that satisfy the reasoning requirements, as well as the associated description information corresponding to the associated nodes; Based on the entity relationship path and the association description information, the association knowledge information is determined.

[0126] Based on any of the above embodiments, the input unit 630 is specifically used for: The candidate text blocks, the entity relationship paths in the associated knowledge information, and the associated description information are aggregated to obtain a preliminary retrieval set; The initial search set is obtained by deduplicating the preliminary search set.

[0127] Based on any of the above embodiments, a map construction unit is further included, wherein the map construction unit specifically includes: Obtain the original knowledge graph unit, which is used to obtain the pre-constructed original knowledge graph; the original knowledge graph stores entity nodes, relationship edges between the entity nodes, and node attributes; An extraction unit is used to extract extraction results from a domain corpus using a multimodal extraction method; the extraction results include entity, relation, and attribute information; the domain corpus includes at least one of text format, table format, and chart format; A matching unit is used to match the entity with entity nodes in the original knowledge graph to obtain a matching result; An incremental update unit is used to perform an incremental update on the original knowledge graph based on the extraction result when the matching result indicates that there is an attribute conflict or a relationship conflict, so as to obtain the dynamic knowledge graph. An update unit is used to update the original knowledge graph based on the extraction result when the matching result indicates that there is no conflict, so as to obtain the dynamic knowledge graph.

[0128] Based on any of the above embodiments, the incremental update unit is specifically used for: When new corpus content is detected in the domain corpus, or when an entity relationship error instruction is received from a user, the original knowledge graph is incrementally updated based on the extraction results to obtain the dynamic knowledge graph.

[0129] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7As shown, the electronic device may include a processor 710, a communications interface 720, a memory 730, and a communication bus 740, wherein the processor 710, communications interface 720, and memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a retrieval enhancement generation method. This method includes: acquiring a user-inputted original query and performing semantic parsing on the original query to obtain a structured query intent; the structured query intent includes core entities and reasoning requirements; generating target query variants based on the structured query intent and filtering the target query variants to obtain a query expansion set; inputting the query expansion set and the structured query intent into a hybrid retrieval module to obtain an initial retrieval set output by the hybrid retrieval module; the hybrid retrieval module includes a vector retrieval module and a knowledge graph retrieval module. The system includes a result merging module; a vector retrieval module for performing semantic retrieval in a vector database based on the query expansion set to obtain candidate text blocks; a knowledge graph retrieval module for performing path reasoning in a dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is associated with the core entity in multiple hops; a result merging module for determining the initial retrieval set based on the candidate text blocks and the associated knowledge information; a multi-dimensional confidence evaluation of the retrieval results in the initial retrieval set to obtain the evaluation results; and a rearrangement of the initial retrieval set based on the evaluation results to obtain the target retrieval results.

[0130] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0131] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the retrieval enhancement generation method provided by the above methods. The method includes: obtaining a user-inputted original query and performing semantic parsing on the original query to obtain a structured query intent; the structured query intent includes core entities and reasoning requirements; generating target query variants based on the structured query intent and filtering the target query variants to obtain a query expansion set; inputting the query expansion set and the structured query intent into a hybrid retrieval module to obtain the output of the hybrid retrieval module. The initial retrieval set is determined by the following steps: the hybrid retrieval module includes a vector retrieval module, a knowledge graph retrieval module, and a result merging module; the vector retrieval module performs semantic retrieval in a vector database based on the query expansion set to obtain candidate text blocks; the knowledge graph retrieval module performs path reasoning in a dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is multi-hop related to the core entity; the result merging module determines the initial retrieval set based on the candidate text blocks and the associated knowledge information; the retrieval results in the initial retrieval set are evaluated using multi-dimensional confidence to obtain evaluation results, and the initial retrieval set is rearranged based on the evaluation results to obtain the target retrieval results.

[0132] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the retrieval enhancement generation method provided by the above methods. This method includes: acquiring a user-inputted original query and performing semantic parsing on the original query to obtain a structured query intent; the structured query intent includes core entities and reasoning requirements; generating target query variants based on the structured query intent and filtering the target query variants to obtain a query expansion set; inputting the query expansion set and the structured query intent into a hybrid retrieval module to obtain an initial retrieval set output by the hybrid retrieval module; the hybrid retrieval module... The module includes a vector retrieval module, a knowledge graph retrieval module, and a result merging module. The vector retrieval module is used to perform semantic retrieval in the vector database based on the query expansion set to obtain candidate text blocks. The knowledge graph retrieval module is used to perform path reasoning in the dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is multi-hop related to the core entity. The result merging module is used to determine the initial retrieval set based on the candidate text blocks and the associated knowledge information. The retrieval results in the initial retrieval set are evaluated with multi-dimensional confidence to obtain the evaluation results. Based on the evaluation results, the initial retrieval set is rearranged to obtain the target retrieval results.

[0133] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0134] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A retrieval enhancement generation method, characterized in that, include: Obtain the original query input by the user, and perform semantic parsing on the original query to obtain the structured query intent; The structured query intent includes core entities and reasoning requirements; Based on the structured query intent, target query variants are generated, and the target query variants are filtered to obtain a set of query extensions; The query expansion set and the structured query intent are input into the hybrid retrieval module to obtain the initial retrieval set output by the hybrid retrieval module. The hybrid retrieval module includes a vector retrieval module, a knowledge graph retrieval module, and a result merging module. The vector retrieval module is used to perform semantic retrieval in the vector database based on the query expansion set to obtain candidate text blocks. The knowledge graph retrieval module is used to perform path reasoning in the dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is associated with the core entity in multiple hops. The result merging module is used to determine the initial retrieval set based on the candidate text blocks and the associated knowledge information; The search results in the initial search set are evaluated using a multi-dimensional confidence level to obtain the evaluation results. Based on the evaluation results, the initial search set is rearranged to obtain the target search results.

2. The retrieval enhancement generation method according to claim 1, characterized in that, The reasoning requirements include target attributes and reasoning types; The semantic parsing of the original query to obtain the structured query intent includes: The original query is semantically parsed to obtain the core entity, the target attribute, and the reasoning type; Based on the core entity, the target attribute, and the reasoning type, the structured query intent is constructed; The reasoning types include single-hop reasoning or multi-hop reasoning.

3. The retrieval enhancement generation method according to claim 2, characterized in that, The target query variants include a first query variant, a second query variant, and a third query variant; The generation of the target query variant based on the structured query intent includes: Based on the core entities in the structured query intent, attribute retrieval is performed in the dynamic knowledge graph to obtain the associated attribute information of the core entities; Based on the associated attribute information, the original query is expanded with constraints to obtain the first query variant; Based on the core entity and target attribute in the structured query intent, synonym replacement processing is performed to obtain the second query variant; Based on the reasoning type in the structured query intent, the original query is processed by sentence transformation to obtain the third query variant.

4. The retrieval enhancement generation method according to claim 2, characterized in that, The step of performing a multi-dimensional confidence assessment on the search results in the initial search set to obtain the assessment results includes: Based on the vector similarity between the first vector of the retrieval result and the second vector of the original query, a semantic matching score is determined. Based on the consistency between the entity relationships in the search results and the records stored in the dynamic knowledge graph, a consistency evaluation score is determined; Based on the degree of coverage of the search results to the core entities and the target attributes, an information integrity score is determined; The evaluation result is determined based on the semantic matching score, the consistency assessment score, and the information integrity score.

5. The retrieval enhancement generation method according to any one of claims 1 to 4, characterized in that, The filtering of the target query variants to obtain a query extension set includes: Determine the semantic similarity between the target query variant and the original query; Filter out the target query variants whose semantic similarity is lower than a preset threshold to obtain the retained target query variants; Based on the reserved target query variant, construct the query extension set.

6. The retrieval enhancement generation method according to any one of claims 1 to 4, characterized in that, The process of performing path reasoning in a dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is multi-hop related to the core entity includes: Starting with the core entity, retrieve the entity relationship path that matches the reasoning requirement in the dynamic knowledge graph; Perform a multi-step traversal along the entity relationship path to obtain the associated nodes that meet the reasoning requirements, as well as the associated description information corresponding to the associated nodes; Based on the entity relationship path and the association description information, the association knowledge information is determined.

7. The retrieval enhancement generation method according to claim 6, characterized in that, Determining the initial retrieval set based on the candidate text blocks and the associated knowledge information includes: The candidate text blocks, the entity relationship paths in the associated knowledge information, and the associated description information are aggregated to obtain a preliminary retrieval set; The initial search set is obtained by deduplicating the preliminary search set.

8. The retrieval enhancement generation method according to any one of claims 1 to 4, characterized in that, The construction steps of the dynamic knowledge graph include: Obtain a pre-constructed original knowledge graph; the original knowledge graph stores entity nodes, relation edges between entity nodes, and node attributes; A multimodal extraction method is used to extract results from a domain corpus; the extraction results include entity, relation, and attribute information; the domain corpus includes at least one of text format, table format, and chart format; The entity is matched with the entity nodes in the original knowledge graph to obtain the matching result; When the matching result indicates that there is an attribute conflict or a relationship conflict, the original knowledge graph is incrementally updated based on the extraction result to obtain the dynamic knowledge graph; When the matching result indicates that there is no conflict, the original knowledge graph is updated based on the extraction result to obtain the dynamic knowledge graph.

9. The retrieval enhancement generation method according to claim 8, characterized in that, The step of incrementally updating the original knowledge graph based on the extraction results to obtain the dynamic knowledge graph includes: When new corpus content is detected in the domain corpus, or when an entity relationship error instruction is received from a user, the original knowledge graph is incrementally updated based on the extraction results to obtain the dynamic knowledge graph.

10. A retrieval enhancement generation device, characterized in that, include: The acquisition unit is used to acquire the original query input by the user and perform semantic parsing on the original query to obtain a structured query intent; The structured query intent includes core entities and reasoning requirements; A filtering unit is used to generate target query variants based on the structured query intent, and to filter the target query variants to obtain a set of query extensions; An input unit is used to input the query expansion set and the structured query intent into the hybrid retrieval module to obtain an initial retrieval set output by the hybrid retrieval module; the hybrid retrieval module includes a vector retrieval module, a knowledge graph retrieval module, and a result merging module; the vector retrieval module is used to perform semantic retrieval in a vector database based on the query expansion set to obtain candidate text blocks, and the knowledge graph retrieval module is used to perform path reasoning in a dynamic knowledge graph based on the reasoning requirements and the core entity to obtain the associated knowledge information that is associated with the core entity in multiple hops; The result merging module is used to determine the initial retrieval set based on the candidate text blocks and the associated knowledge information; An evaluation unit is used to perform multi-dimensional confidence evaluation on the search results in the initial search set, obtain evaluation results, and rearrange the initial search set based on the evaluation results to obtain the target search results.

Images