Structured knowledge base-oriented search-enhanced intelligent question-answering method and system

Abstract

The application provides a structured knowledge base-oriented retrieval enhanced intelligent question and answer method and system, relates to the natural language processing and information retrieval technical field, and the method comprises the following steps: after receiving a natural language scientific research question input by a user, performing structured semantic analysis under the constraint of a pre-defined domain knowledge mode, and outputting a structured query statement; performing a database level retrieval operation on the structured query statement, and outputting a structured retrieval result set; performing semantic clustering compression processing on the structured retrieval result set, screening to obtain a representative evidence subset; performing constraint content generation on the representative evidence subset, and outputting a verifiable scientific research question and answer result. The technical problems that the prior art has inaccurate scientific research field knowledge retrieval, redundant evidence and low reliability of the question and answer result are solved. The technical effects that the retrieval accuracy in the scientific research question and answer scene is improved and the generated reliability is improved under the premise of ensuring the answer explainability and credibility are achieved.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and information retrieval technology, specifically to a retrieval-enhanced intelligent question-answering method and system for structured knowledge bases. Background Technology

[0002] In recent years, the rapid development of large-scale language models has significantly promoted the progress of natural language processing technology. Among them, retrieval-enhanced generation technology, by combining external knowledge retrieval mechanisms with generation models, not only improves the fluency of generated text but also significantly enhances the accuracy, traceability, and factual consistency of answers.

[0003] However, existing research on enhanced retrieval generation techniques mainly focuses on general open domains such as news and encyclopedias. Their core relies on large-scale, weakly structured general corpora, and their application in knowledge-intensive scientific research remains relatively limited. This is primarily due to the characteristics and complex requirements of scientific literature itself. Scientific literature is typically lengthy, with rich hierarchical organization, containing a large amount of technical terminology, conditional descriptions, experimental or technological process descriptions, and conclusive information, making it difficult to achieve high-quality knowledge retrieval through simple text segmentation or keyword matching. Furthermore, scientific questions often involve multi-dimensional constraints such as object attributes, process characteristics, conditional ranges, temporal order, and contextual dependencies. Traditional retrieval methods based on vector similarity or keyword filtering struggle to accurately express and effectively combine complex query intents. In addition, to obtain complete semantic support, scientific question answering usually requires combining multiple documents or paragraphs as input. This direct splicing method easily exceeds the input window limit of the generation model and may lead to the weakening or forgetting of key information, thus affecting the generation quality. Scientific literature often contains descriptions with highly similar content or different expressions but consistent semantics. If there is a lack of effective result aggregation and compression mechanisms, directly returning search results will make it difficult to form a structured, high-level knowledge summary, which will easily lead to information redundancy and semantic repetition.

[0004] In summary, existing technologies suffer from technical problems such as inaccurate knowledge retrieval in the scientific research field, redundant evidence, and low reliability of question-and-answer results. Summary of the Invention

[0005] The purpose of this application is to provide a retrieval-enhanced intelligent question-answering method and system for structured knowledge bases, which addresses the technical problems of inaccurate knowledge retrieval in the scientific research field, redundant evidence, and low reliability of question-answering results in existing technologies.

[0006] In view of the above problems, this application provides a retrieval-enhanced intelligent question answering method and system for structured knowledge bases.

[0007] The first aspect of this application provides a retrieval-enhanced intelligent question-answering method for structured knowledge bases. The method includes: receiving a natural language research question input by a user; performing structured semantic parsing under predefined domain knowledge model constraints to output a structured query statement; performing a database-level retrieval operation on the structured query statement to output a set of structured retrieval results; performing semantic clustering compression processing on the set of structured retrieval results to obtain a representative evidence subset; and generating constrained content from the representative evidence subset to output verifiable research question-answering results.

[0008] Optionally, the natural language research question is subjected to question element identification to obtain key elements; terminology normalization and structured element extraction operations are performed on the key elements to obtain a set of structured elements, wherein the set of structured elements includes, but is not limited to, target entities, retrieval target fields, constraint sets, and return field sets; under the constraints of the domain knowledge model, the set of structured elements is compiled into an executable structured query statement, wherein the structured query statement includes retrieval object limits, filtering conditions, and return field projections.

[0009] Optionally, the target retrieval object is determined based on the retrieval object being limited to a pre-built domain knowledge database; the filtering conditions are applied to perform multi-field joint filtering of the target retrieval object, and the original database query results are output; multiple query records in the original database query results are converted into multiple domain knowledge object records, wherein the fields of the domain knowledge object records include entities, processes, conditions, and attributes; after the multiple domain knowledge object records are encapsulated in a structured data manner, the data content of the returned field projection is retained, and the structured retrieval result set is output.

[0010] Optionally, when the intent of the natural language research question requires a statistical answer, the structured query statement may also include statistical summary expressions.

[0011] Optionally, when applying the filtering conditions to perform multi-field joint filtering of the target retrieval object, the statistical summary expression is simultaneously applied to perform aggregation calculations and output the original database query results.

[0012] Optionally, the structured retrieval result set is converted into a semantic modeling text representation; and the semantic modeling text representation is automatically clustered based on semantic similarity to obtain a cluster set; representative records are selected from each cluster of the cluster set according to clustering frequency, representativeness or diversity strategies to form the representative evidence subset, wherein the representative evidence subset satisfies coverage constraints, redundancy constraints and size constraints.

[0013] Optionally, multiple structured domain knowledge objects in the representative evidence subset are converted into natural language text sequences; the natural language text sequences are integrated and encapsulated into a single context input using a preset template; constraint generation rules are executed on the single context input, and the verifiable scientific question-and-answer results are output.

[0014] Optionally, during the generation of the verifiable research question-and-answer results, only content based on the knowledge records in the context input is allowed to be generated, and the introduction of external unverified information is prohibited; when the representative evidence subset is empty, a predefined no-knowledge response is output; the verifiable research question-and-answer results are prohibited from using enumerative output, subjective inference, or extended reasoning.

[0015] A second aspect of this application provides a retrieval-enhanced intelligent question-answering system for structured knowledge bases. The system includes: a query rewriting module, used to perform structured semantic parsing under predefined domain knowledge model constraints after receiving a natural language research question input by a user, and output a structured query statement; a structured knowledge retrieval module, used to perform database-level retrieval operations on the structured query statement and output a set of structured retrieval results; a semantic clustering and result compression module, used to perform semantic clustering compression processing on the set of structured retrieval results to obtain a representative subset of evidence; and a constrained generation module, used to generate constrained content from the representative subset of evidence and output verifiable research question-answering results.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0017] The method provided in this application receives a natural language research question input by a user, performs structured semantic parsing under the constraints of a predefined domain knowledge model, and outputs a structured query statement; performs a database-level retrieval operation on the structured query statement, and outputs a set of structured retrieval results; performs semantic clustering and compression processing on the structured retrieval result set to obtain a representative evidence subset; and generates constrained content from the representative evidence subset to output verifiable research question-and-answer results. This achieves the technical effect of automatically mapping natural language research questions to executable queries in a structured database, and generating answers under constrained conditions after semantic aggregation and compression of the retrieval results, thereby improving the retrieval accuracy and generation reliability in research question-and-answer scenarios while ensuring the interpretability and credibility of the answers.

[0018] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application 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 merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an intelligent question-answering method for retrieval enhancement based on a structured knowledge base, as provided in this application.

[0021] Figure 2 This application provides a schematic diagram of the structure of an intelligent question-answering system for retrieval enhancement based on a structured knowledge base.

[0022] Figure labeling: Query rewriting module 11, structured knowledge retrieval module 12, semantic clustering and result compression module 13, constrained generation module 14. Detailed Implementation

[0023] This application provides a retrieval-enhanced intelligent question-answering method and system for structured knowledge bases, addressing the technical problems of inaccurate knowledge retrieval, redundant evidence, and low reliability of question-answering results in existing technologies. It achieves the technical effect of automatically mapping natural language research questions into executable queries in a structured database, and generating answers under constrained conditions after semantic aggregation and compression of the retrieval results. This improves the accuracy and reliability of retrieval in scientific research question-answering scenarios while ensuring the interpretability and credibility of the answers.

[0024] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0025] Example 1, as Figure 1 As shown, this application provides a retrieval-enhanced intelligent question-answering method for structured knowledge bases, which includes:

[0026] After receiving natural language research questions input by users, the system performs structured semantic parsing under the constraints of predefined domain knowledge patterns and outputs structured query statements.

[0027] Furthermore, after receiving a natural language research question input by a user, structured semantic parsing is performed under the constraints of a predefined domain knowledge model to output a structured query statement. The method includes: identifying question elements of the natural language research question to obtain key elements; performing terminology normalization and structured element extraction operations on the key elements to obtain a set of structured elements, wherein the set of structured elements includes, but is not limited to, target entities, retrieval target fields, constraint sets, and return field sets; and compiling the set of structured elements into an executable structured query statement under the constraints of the domain knowledge model, wherein the structured query statement includes retrieval object limitations, filtering conditions, and return field projections.

[0028] Specifically, after receiving a natural language research question input by the user, structured semantic parsing is performed under the constraints of a predefined domain knowledge pattern. The domain knowledge pattern constraint refers to a predefined research domain knowledge framework, including table structure, field dictionary, entity type and relation constraints, and alias mapping rules for material names, such as the mapping between Chinese names, chemical formulas and standard English names. In the process of structured semantic parsing, the received natural language research questions are first identified in terms of question elements. The text of the natural language research questions undergoes standardized preprocessing, including removing irrelevant symbols, standardizing unit expressions, and standardizing numerical representations. Based on a pre-built domain terminology dictionary, material name database, chemical formula rule database, and process method list, the text is segmented into words and phrases. Candidate entities are identified through a combination of dictionary matching, pattern matching, and rule matching. For material names or chemical formulas, identification and verification are performed using chemical formula composition rules, such as element symbols, numerical structural features, and material standard name tables. For process methods, they are determined by matching verb phrase structures with a predefined process type mapping table. For experimental conditions and their constraints, extraction is performed using numerical and unit combination rules, range expression rules (e.g., greater than, not exceeding, between), and conditional trigger words (e.g., under certain conditions, when using…). The identified numerical values ​​are then mapped to the corresponding physical quantity fields. The above identification results are categorized according to entity type, and ambiguous and duplicate items are removed. The output is a set of key elements that can be directly mapped to the knowledge base field structure, including but not limited to core information such as material entities, process methods, experimental conditions and their constraint expressions, so as to generate executable structured search requests under the constraints of domain knowledge patterns.

[0029] Then, terminology normalization and structured element extraction operations are performed on the key elements, including at least: Entity normalization: mapping material names, chemical formulas, process names, equipment names, etc., to standard entity identifiers in the knowledge base; Field alignment: mapping natural language intents to fields corresponding to the knowledge base field dictionary, such as materials, methods, condition parameters, result indicators, etc.; Constraint parsing: parsing comparison relationships and range expressions into executable filtering predicates, including greater than, less than, interval, contain, etc.; Unit normalization: optionally normalizing the units and numerical ranges of parameters such as temperature and time; Semi-structured parsing: generating corresponding nested condition expressions when knowledge base fields contain JSON or nested metadata. After the above processing, a structured element set is formed. The structured element set includes at least the target entity, the retrieval target field, the constraint set, and the return field set. Among them, the target entity is the core of the retrieval object in the research question, the retrieval target field is the corresponding field in the knowledge base, the constraint set is the explicit conditions or default range conditions in the user question, and the return field set is the final output result field.

[0030] Under the constraints of a predefined domain knowledge model, a set of structured elements is compiled into a single read-only, executable structured query statement. This structured query statement can be SQL or a query statement for a document-oriented database. The structured query statement includes at least: search object limitations, filtering conditions, and return field projections. Search object limitations are used to determine the query scope; filtering conditions are such as conditional constraints; and return field projections only return fields of interest to the user, such as process parameters and experimental yields.

[0031] For example, receiving a natural language research question Q from a user: "What are the three methods for synthesizing MoS2?", the intent of the question is parsed to retrieve and provide representative records related to MoS2 synthesis methods, and output the names of the three methods and their core principles / process descriptions. Under the constraints of the domain knowledge model, question Q undergoes structured parsing and terminology normalization, including: normalizing and mapping MoS2 to the standard entity MoS2 in the knowledge base; mapping synthesis methods to knowledge base fields / relationships, such as synthesis_method related fields or method relationships; selecting the set of returned fields based on the question intent, such as method name, principle / process description, and condition-related fields, and then generating a single, read-only, executable structured query statement for performing precise retrieval in the structured knowledge base.

[0032] By performing structured semantic parsing under the constraints of predefined domain knowledge patterns, natural language questions are transformed into precise structured query statements. This enables accurate understanding of user intent and precise retrieval of relevant information from the database, avoiding the inaccuracies and ambiguities of natural language understanding, thereby improving the accuracy and reliability of question-answering results.

[0033] Perform a database-level retrieval operation on the structured query statement and output a set of structured retrieval results.

[0034] Furthermore, a database-level retrieval operation is performed on the structured query statement to output a set of structured retrieval results. The method includes: determining the target retrieval object based on the retrieval object being limited to a pre-constructed domain knowledge database; applying the filtering conditions to perform multi-field joint filtering on the target retrieval object and outputting the original database query results; converting multiple query records in the original database query results into multiple domain knowledge object records, wherein the fields of the domain knowledge object records include entities, processes, conditions, and attributes; after performing structured data encapsulation on the multiple domain knowledge object records, retaining the data content projected by the returned fields, and outputting the set of structured retrieval results.

[0035] Specifically, a domain knowledge database is pre-constructed, with knowledge objects as the basic storage unit, including but not limited to fields such as material name, method name, principle / process description, and experimental / process conditions. Based on the search object definition, the target search object is determined in the pre-constructed domain knowledge database using a data dictionary or schema mapping table. The search object definition refers to the data range or data table category explicitly indicated in the query statement, such as polystyrene synthesis experimental records or performance test data of a certain type of material. This is used to locate the corresponding data type in the database metadata mapping table, thereby avoiding cross-domain erroneous searches.

[0036] Multi-field joint filtering is performed on the target retrieval object based on the filtering conditions. This multi-field joint filtering refers to combining multiple conditions into a composite filtering expression at the database level using logical operators such as AND, OR, and BETWEEN, and then calling the database query optimizer to match the index fields to improve query efficiency. Through multi-field joint filtering, the original database query results are output, which are multiple structured data records that meet the conditions. These original database query records are then transformed into multiple domain knowledge object records. These domain knowledge object records are structured expressions of the database physical records after semantic reconstruction. Their fields include: entity, process, condition, and attribute. Entities refer to the core research object, such as material names or chemical substances; processes refer to experimental or technological steps, such as catalytic polymerization or heat treatment; conditions refer to experimental parameters or constraints, such as temperature, pressure, and time; and attributes refer to results or performance indicators, such as yield, purity, and strength. The specific transformation process relies on a predefined field-semantic mapping rule base to map database fields to a unified knowledge object structure, ensuring semantic consistency and readability for subsequent processing.

[0037] The multiple domain knowledge object records are then encapsulated in a structured data format. These knowledge objects are organized using a unified data structure, such as JSON, and the fields required by the user's question are retained based on the returned field projection. This returned field projection means outputting only the fields specified in the query statement, rather than all database fields, thereby reducing redundant information and improving subsequent processing efficiency. After encapsulation, a final structured search result set is formed. This set uses domain knowledge objects as basic units and includes at least the entity, process, condition, and attribute information fields related to the question's intent. For example, executing a structured query in the domain knowledge base yields a set of structured search results that meet the conditions. This set contains multiple knowledge object records related to the MoS2 synthesis method.

[0038] By performing precise database-level retrieval operations on structured query statements and generating a set of structured search results, it is possible to ensure that the question-and-answer results come directly from real records in structured domain knowledge databases, achieve a strict correspondence between query conditions and data entries, avoid semantic drift and fuzzy search problems that occur during natural language matching, and thus improve the accuracy and reliability of search results.

[0039] Furthermore, when the intent of the natural language research question requires a statistical answer, the structured query statement also includes statistical summary expressions.

[0040] Specifically, when the intent of a natural language research question requires a statistical answer, after completing the identification of question elements and extraction of structured elements, the statistical attributes of the natural language research question intent are further determined. This involves identifying whether the question contains statistical semantic features such as quantity statistics, proportion calculation, mean analysis, maximum / minimum value filtering, or distribution comparison. The statistical answer refers to the final output of the natural language research question not being a single record description, but rather a result formed after aggregation calculations based on multiple structured records, such as average yield, maximum efficiency, sample size, and qualified rate. First, the question text is matched using a statistical trigger word rule base, identifying keywords such as average, highest, lowest, total, percentage, distribution, statistics, and number of types. Simultaneously, sentence structure analysis is used to determine whether it involves summarizing multiple records rather than querying a single instance. After confirming the statistical intent, a statistical summary expression field is added to the existing structured element set. The statistical summary expression refers to structured instructions used to instruct the database to perform aggregation operations, including aggregation function types such as AVG for average, MAX for maximum, COUNT for count, SUM for summation, grouping fields, and necessary sorting or filtering conditions.

[0041] When the intent of a natural language research question requires a statistical answer, structured query statements contain statistical summary expressions, which can perform aggregation calculations simultaneously during the retrieval process and directly output statistical results. This meets the needs of data statistics and analysis in the research field and improves the functionality and application scope of question-and-answer results.

[0042] Furthermore, when applying the filtering conditions to perform multi-field joint filtering of the target retrieval object, the statistical summary expression is simultaneously applied to perform aggregation calculations and output the original database query results.

[0043] Specifically, after determining the target retrieval object, when the natural language research question contains statistical intent, while applying the aforementioned filtering conditions for multi-field joint filtering, the aforementioned statistical summary expression is simultaneously introduced for database-level aggregation calculations. Field type validation rules are used to confirm whether the target field is numerical or statistically valid, ensuring the legality of the aggregation operation. During the query compilation phase, the filtering conditions and statistical summary expression are structurally merged to generate a database query statement containing aggregation functions. When executing database-level retrieval operations, condition filtering is first performed based on the index mechanism to form a data set that meets the constraints. Then, aggregation calculations are performed directly on the data set that meets the constraints within the domain knowledge database, rather than performing secondary statistics at the application layer, thus ensuring the accuracy of the statistical results and computational efficiency. After execution, the output original database query results are no longer individual experimental records, but rather structured statistical results that have already undergone aggregation calculations, or a set of statistical results generated by grouping dimensions.

[0044] By moving statistical calculations to the database execution layer and performing them synchronously with conditional filtering, it is ensured that the original database query results are strictly generated based on data that meets the conditions, avoiding data omissions or duplicate statistics, while improving query efficiency and result consistency.

[0045] The structured retrieval result set is subjected to semantic clustering compression processing of homogeneous records to obtain a representative subset of evidence.

[0046] Furthermore, the structured retrieval result set is subjected to semantic clustering compression processing of homogeneous records to obtain a representative evidence subset. The method includes: converting the structured retrieval result set into a semantic modeling text representation; automatically clustering the semantic modeling text representation based on semantic similarity to obtain a cluster set; and selecting representative records from each cluster of the cluster set according to clustering frequency, representativeness, or diversity strategies to form the representative evidence subset, wherein the representative evidence subset satisfies coverage constraints, redundancy constraints, and size constraints.

[0047] Specifically, the structured search results set is subjected to semantic clustering compression processing of homogeneous records in order to select a representative subset of evidence. Homogeneous records refer to multiple records that are highly similar in terms of entities, process paths, key experimental conditions or result trends, and differ only in individual parameters or expression forms. If compression processing is not performed, it is easy to cause information redundancy and interfere with the generation of subsequent content. Specifically, the structured search results are first transformed into a semantic modeling text representation. Each structured domain knowledge object record undergoes field standardization, which involves reorganizing entity, process, condition, and attribute fields according to a predefined semantic expression order, and standardizing unit formats, terminology, and numerical precision. Based on a fixed semantic description template, the structured fields are mapped into readable natural language expressions, such as the sentence structure "Under [condition], [process] is used to process [entity], resulting in [attribute result]". During the mapping process, numerical fields are bound to corresponding physical quantity labels to ensure semantic clarity, while retaining the original field identifiers as implicit index information for later backtracking. If multiple conditions exist, they are connected according to logical relation words such as "and," "with," and "under...conditions," forming a semantically complete and structurally consistent text description. Finally, each original structured record is transformed into a semantically consistent, informationally complete, and similarity-calculating semantic modeling text representation. Through field order standardization, unit unification, and terminology standardization rules, the consistency of text expression is ensured, providing a comparable basis for subsequent similarity calculations.

[0048] Each semantically modeled text representation is transformed into a measurable vector form, for example, through text feature encoding or text embedding, mapping the entity, process, condition, and attribute information of the text into a high-dimensional semantic space, forming a high-dimensional numerical vector, enabling the semantic information of the text to be quantified and compared in the vector space. Cosine similarity or distance metrics, such as Euclidean distance, are used to calculate the semantic similarity between semantically modeled text representations. Based on semantic similarity, unsupervised clustering algorithms, such as hierarchical clustering, density clustering, or K-means, are used to automatically cluster the semantically modeled text representations, grouping texts with highly similar semantics into the same cluster to form a cluster set. During the clustering process, a similarity threshold or maximum intra-cluster distance constraint can be set to control the cluster density while ensuring significant semantic differences between different clusters, thus obtaining a set of clusters that are internally homogeneous and externally distinct. Based on clustering frequency, representativeness, or diversity strategies, representative records with significant meaning are selected from each cluster set to form a representative subset of evidence. Representativeness can be determined based on the cluster center record, i.e., the record with the smallest average distance from other records in the cluster, the record with the highest statistical frequency, or the record with the most typical parameters. The diversity strategy refers to maintaining differential coverage between different clusters to avoid selecting only a single type of experimental results.

[0049] Meanwhile, the representative evidence subset satisfies coverage, redundancy, and size constraints. Coverage constraint means the selected records should cover the main experimental types or result ranges. Redundancy constraint means the semantic similarity between records in the representative evidence subset should not exceed a preset threshold, reducing synonymous / near-synonymous duplicate records. Size constraint means the number of evidence records is controlled within a reasonable range to meet the input length limitations of subsequent generation and mitigate context overflow. For example, by performing semantic clustering compression processing on the aforementioned structured retrieval result set to obtain a representative evidence subset that can include records of the following three representative methods: mechanical stripping, chemical vapor deposition (CVD), and hydrothermal / solvothermal methods.

[0050] By performing semantic clustering compression on the structured search results set to homogeneous records, a representative subset of evidence is selected. This process removes redundant information while ensuring the coverage of key information, thereby improving the quality and usability of search results and enabling users to obtain key information more quickly.

[0051] Constraint content is generated from the representative subset of evidence, and verifiable research question-and-answer results are output.

[0052] Furthermore, constraint content generation is performed on the representative subset of evidence to output verifiable research question-and-answer results. The method includes: converting multiple structured domain knowledge objects in the representative subset of evidence into natural language text sequences; integrating and encapsulating the natural language text sequences into a single context input using a preset template; executing constraint generation rules on the single context input; and outputting the verifiable research question-and-answer results.

[0053] Furthermore, the constraint generation rules include: during the generation of verifiable scientific question-and-answer results, only content based on knowledge records in the context input is allowed to be generated, and the introduction of external unverified information is prohibited; when the representative evidence subset is empty, a predefined no-knowledge response is output; and the verifiable scientific question-and-answer results are prohibited from using enumerative output, subjective inference, or extended reasoning.

[0054] Specifically, the entity, process, condition, and attribute fields of each structured domain knowledge object record in the representative evidence subset are standardized, including uniform naming, unit format, and numerical precision. Then, each field is converted into a natural language phrase according to field mapping rules, and concatenated into a complete sentence according to a predefined logical order to form a natural language text sequence. The above steps are repeated for all records in the representative evidence subset to generate multiple standardized and readable natural language text sequences, each sequence corresponding to one piece of structured evidence.

[0055] Define a pre-defined template to clarify the contextual organization and logical order. For example, the pre-defined template includes introductory text, numbered lists, or separators to identify the location and content of each experimental record. Insert each natural language text sequence generated from the representative evidence subset into the pre-defined template in sequence, while consistently using line breaks, semicolons, or numbered separators to ensure the independence and logical integrity of each record. Add contextual prompts to the pre-defined template, such as "Please answer the research question based on the following experimental record," to limit the output to verifiable research questions and answers that only refer to the input content. This unifies and encapsulates multiple texts into a single contextual input, avoiding ambiguity caused by fragmented information or disordered record order, and achieving a complete, structured, and semantically coherent contextual representation.

[0056] Constraint generation rules are applied to a single contextual input. These rules include: during the generation of verifiable research question-and-answer results, only content based on knowledge records from the contextual input is allowed, and the introduction of unverified external information is prohibited to ensure the verifiability of the results; when no knowledge related to the question is retrieved, i.e., the representative evidence subset is empty, a predefined no-knowledge response is output, such as "no relevant experimental data available"; the output format is restricted to provide the method name and its core principles / process in a concise and objective academic expression, prohibiting enumerative output, subjective inference, or extended reasoning to avoid generating conclusions irrelevant to the evidence or unreliable ones. By using a single contextual input as the sole source of information and strictly enforcing the constraint generation rules, complete, natural, and fluent verifiable research question-and-answer results are generated and output. These verifiable research question-and-answer results are supported by evidence, traceable, and logically consistent research answers. Their content strictly corresponds to the experimental records and data in the representative evidence subset, such as describing and summarizing experimental conditions, response results, or statistical indicators, thereby ensuring both the readability of the research question-and-answer and the verifiability and reliability of the answer. For example, for a user-input natural language research question: "What are the three methods for synthesizing MoS2?", the output verifiable research question-and-answer results should include at least the following three synthesis methods and brief descriptions: Mechanical exfoliation: using external force to peel off bulk MoS2 into single-layer or few-layer thin films and transfer them to a substrate; Chemical vapor deposition (CVD): molybdenum source and sulfur source react under high temperature conditions and deposit to form a MoS2 thin film on the substrate surface; Hydrothermal / solvothermal method: in a closed reactor under high temperature and high pressure conditions, molybdenum salt reacts with sulfur source to generate MoS2 nanomaterials.

[0057] By transforming a selected set of representative evidence into directly understandable and verifiable research question-and-answer results, a closed-loop generation from data to knowledge is achieved. By controlling the output range through constraint rules, the retrieval accuracy and generation reliability in research question-and-answer scenarios are improved while ensuring the interpretability and credibility of the answers, thereby meeting the high requirements of research scenarios for the credibility of results.

[0058] In summary, the retrieval enhancement intelligent question answering method for structured knowledge bases provided in this application has the following technical effects:

[0059] Unlike existing retrieval enhancement methods based on vector similarity or coarse-grained text matching, this application receives a natural language research question input by a user, performs structured semantic parsing under the constraints of a predefined domain knowledge pattern, and outputs a structured query statement. It then performs a database-level retrieval operation on the structured query statement, outputting a set of structured retrieval results. Next, it performs semantic clustering and compression processing on the structured retrieval result set to obtain a representative evidence subset. Finally, it generates constrained content from the representative evidence subset, outputting verifiable research question-and-answer results. This achieves the technical effect of automatically mapping natural language research questions to executable queries in a structured database, and generating answers under constrained conditions after semantic aggregation and compression of the retrieval results. This improves the retrieval accuracy and generation reliability in research question-and-answer scenarios while ensuring the interpretability and credibility of the answers.

[0060] Furthermore, the retrieval-enhanced intelligent question-answering method for structured knowledge bases provided in this application is applicable to scientific research and professional application fields with structured or semi-structured knowledge as the core carrier, including but not limited to materials science, chemical processes, and scientific literature analysis, and is used to achieve accurate retrieval of domain knowledge, semantic aggregation analysis, and verifiable generative question answering.

[0061] Example 2, based on the same inventive concept as the retrieval-enhanced intelligent question-answering method for structured knowledge bases in the preceding examples, such as... Figure 2 As shown, this application provides a retrieval-enhanced intelligent question-answering system for structured knowledge bases, wherein the retrieval-enhanced intelligent question-answering system for structured knowledge bases includes:

[0062] The query rewriting module 11 is used to perform structured semantic parsing under the constraints of a predefined domain knowledge model after receiving a natural language research question input by the user, and output a structured query statement.

[0063] Specifically, the query rewriting module 11 is used to receive natural language research questions input by users and automatically generate query statements that can be directly executed by the structured database under the constraints of a predetermined domain knowledge model. This achieves automatic alignment between natural language intent and precise retrieval requests, thereby solving the problems in existing RAG technology in the research field where natural language questions are difficult to accurately convert into executable retrieval requests and where inconsistent cross-language terminology makes it difficult for structured retrieval to express precisely.

[0064] The query rewriting module 11 is characterized by: using domain knowledge patterns, including table structure, field dictionary, entity type, and relational constraints, as the query generation boundary to guide the large language model to output a single, read-only, and executable structured query statement, avoiding the generation of non-executable or inconsistent queries with the knowledge base; supporting terminology standardization and cross-language mapping for key elements in scientific research problems such as material names, chemical formulas, process methods, and experimental conditions, uniformly mapping Chinese expressions and aliases to standard English fields and standardized values ​​in the knowledge base, thereby reducing ambiguity and missed detections; supporting structured expression of multiple conditional constraints, parsing comparative relationships in natural language, such as greater than, less than, range, and containment, into database-level filtering predicates, and optionally normalizing the units and numerical ranges of parameters such as temperature and time; supporting nested conditional retrieval generation for semi-structured fields such as JSON, enabling process / experimental metadata such as temperature, time, and equipment information to be written into query statements in an executable nested filtering manner; and supporting automatic selection of query granularity and return field set according to the problem intent, such as returning complete records, key field projections, or statistical summaries, reducing redundant output caused by irrelevant fields while ensuring retrieval effectiveness.

[0065] The structured knowledge retrieval module 12 is used to perform database-level retrieval operations on the structured query statement and output a set of structured retrieval results.

[0066] Specifically, the structured knowledge retrieval module 12 executes the structured query generated by the query rewriting module 11 based on a pre-built domain knowledge database, and returns the domain knowledge records that meet the conditions.

[0067] The structured knowledge retrieval module 12 is characterized by: using domain knowledge objects as the basic storage and retrieval units, wherein the knowledge objects are used to describe the entities, processes, conditions and their attribute information involved in scientific research activities; supporting multi-field joint filtering and multi-condition combination retrieval, which can accurately express the complex constraints implicit in scientific research problems; and outputting the retrieval results in the form of structured data, providing stable and controllable data input for subsequent semantic analysis, clustering processing and generation stages.

[0068] The structured knowledge retrieval module 12 enables high-precision retrieval of domain knowledge, providing a reliable data foundation for subsequent semantic aggregation and generation processes.

[0069] The semantic clustering and result compression module 13 is used to perform semantic clustering compression processing on the structured retrieval result set to obtain a representative evidence subset.

[0070] Specifically, to address the issue of a large number of structured search results with highly similar semantic content, a semantic clustering and result compression module 13 is introduced to deduplicate and refine the information of the large number of homogeneous records retrieved, thereby selecting the most representative subset of records with the highest information density. The semantic clustering and result compression module 13 includes the following processes: converting structured knowledge records into textual representations that can be used for semantic modeling; automatically clustering the search results based on semantic similarity to identify sets of records that are semantically consistent or highly similar; selecting representative records from each cluster according to clustering frequency, representativeness, or diversity strategies; and significantly reducing the context size of the input generation model while ensuring coverage of key information.

[0071] The semantic clustering and result compression module 13 enables the transformation from multiple similar fact descriptions to high-level, de-redundant knowledge summaries, effectively alleviating the problems of contextual redundancy and information interference in scientific research question-and-answer scenarios.

[0072] The constrained generation module 14 is used to generate constrained content for the representative subset of evidence and output verifiable scientific research question-and-answer results.

[0073] Specifically, the constrained generation module 14 is used to generate the final answer based on representative knowledge records selected by the semantic clustering and result compression module 13 under strictly controlled context conditions. The constraint mechanisms of the constrained generation module 14 include: allowing content generation only based on retrieved knowledge records and prohibiting the introduction of unverified external information; outputting a predefined no-knowledge response when no knowledge related to the question is retrieved to avoid the model fabricating answers; and restricting the form of the generated content to be expressed in a concise and objective academic style, avoiding enumerative output, subjective inference, or extended reasoning.

[0074] The constrained generation module 14 ensures that the output results are authentic, interpretable, and usable for scientific research while maintaining smooth generation, thus meeting the high credibility requirements of intelligent question-answering systems in scientific research scenarios.

[0075] Furthermore, the query rewriting module 11 is also used to: identify problem elements of the natural language research question to obtain key elements; perform terminology normalization and structured element extraction operations on the key elements to obtain a structured element set, wherein the structured element set includes, but is not limited to, target entities, retrieval target fields, constraint sets, and return field sets; and compile the structured element set into an executable structured query statement under the constraints of the domain knowledge model, wherein the structured query statement includes retrieval object limits, filtering conditions, and return field projections.

[0076] Furthermore, the structured knowledge retrieval module 12 is also used to: determine the target retrieval object based on the retrieval object being limited to a pre-built domain knowledge database; apply the filtering conditions to perform multi-field joint filtering of the target retrieval object, and output the original database query results; convert multiple query records in the original database query results into multiple domain knowledge object records, wherein the fields of the domain knowledge object records include entities, processes, conditions, and attributes; after performing structured data encapsulation on the multiple domain knowledge object records, retain the data content of the returned field projection, and output the structured retrieval result set.

[0077] Furthermore, the query rewriting module 11 is also used to: when the intent of the natural language research question requires a statistical answer, the structured query statement also includes a statistical summary expression.

[0078] Furthermore, the structured knowledge retrieval module 12 is also used to: simultaneously apply the statistical summary expression to perform aggregation calculations when applying the filtering conditions to perform multi-field joint filtering of the target retrieval object, and output the original database query results.

[0079] Furthermore, the semantic clustering and result compression module 13 is also used to: convert the structured retrieval result set into a semantic modeling text representation; and automatically cluster the semantic modeling text representation based on semantic similarity to obtain a cluster set; and select representative records from each cluster of the cluster set according to clustering frequency, representativeness or diversity strategies to form the representative evidence subset, wherein the representative evidence subset satisfies coverage constraints, redundancy constraints and size constraints.

[0080] Furthermore, the constrained generation module 14 is also used to: convert multiple structured domain knowledge objects in the representative evidence subset into natural language text sequences; integrate and encapsulate the natural language text sequences into a single context input using a preset template; execute constraint generation rules on the single context input, and output the verifiable scientific question-and-answer results.

[0081] Furthermore, the constrained generation module 14 is also used to: allow only the generation of content based on the knowledge records in the context input during the generation of the verifiable scientific question-and-answer results, and prohibit the introduction of external unverified information; output a predefined no-knowledge response when the representative evidence subset is empty; and prohibit the verifiable scientific question-and-answer results from using enumerative output, subjective inference, or extended reasoning.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The retrieval enhancement intelligent question answering method and specific examples for structured knowledge bases described in the foregoing embodiment one are also applicable to the retrieval enhancement intelligent question answering system for structured knowledge bases described in this embodiment. Through the foregoing detailed description of the retrieval enhancement intelligent question answering method for structured knowledge bases, those skilled in the art can clearly understand the retrieval enhancement intelligent question answering system for structured knowledge bases described in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0083] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0084] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A retrieval-enhanced intelligent question-answering method for structured knowledge bases, characterized in that, The method includes: After receiving natural language research questions input by users, the system performs structured semantic parsing under the constraints of predefined domain knowledge patterns and outputs structured query statements. Perform a database-level retrieval operation on the structured query statement and output a set of structured retrieval results; The structured retrieval result set is subjected to semantic clustering compression processing of homogeneous records to obtain a representative subset of evidence; Constraint content is generated from the aforementioned representative subset of evidence, and verifiable research question-and-answer results are output. After receiving a natural language research question input by a user, the method performs structured semantic parsing under the constraints of a predefined domain knowledge model and outputs a structured query statement. The method includes: The key elements of the natural language research problem are identified through problem element identification. The key elements are subjected to terminology normalization and structured element extraction operations to obtain a set of structured elements, wherein the set of structured elements includes, but is not limited to, target entity, retrieval target field, constraint set and return field set; Under the constraints of the domain knowledge model, the structured element set is compiled into an executable structured query statement, wherein the structured query statement includes retrieval object limits, filtering conditions, and return field projections; The method includes performing a database-level retrieval operation on the structured query statement and outputting a set of structured retrieval results. The target search object is determined based on the search object being limited to a pre-constructed domain knowledge database; Apply the filtering conditions to perform multi-field joint filtering on the target retrieval object, and output the original database query results; The original database query results are transformed into multiple domain knowledge object records, wherein the domain knowledge object records are composed of entities, processes, conditions and attributes. After encapsulating the multiple domain knowledge object records in a structured data manner, the data content of the returned field projection is retained, and the structured search result set is output. The method involves generating binding content from the representative subset of evidence and outputting verifiable research question-and-answer results, including: The representative subset of evidence is converted into a sequence of natural language text. The natural language text sequence is integrated and encapsulated into a single context input using a preset template; The constraint generation rule is executed on the single context input, and the verifiable scientific question-and-answer result is output.

2. The retrieval-enhanced intelligent question-answering method for structured knowledge bases as described in claim 1, characterized in that, When the intent of the natural language research question requires a statistical answer, the structured query statement also includes statistical summary expressions.

3. The retrieval-enhanced intelligent question-answering method for structured knowledge bases as described in claim 2, characterized in that, When applying the filtering conditions to perform multi-field joint filtering of the target retrieval object, the statistical summary expression is simultaneously applied to perform aggregation calculations and output the original database query results.

4. The retrieval-enhanced intelligent question-answering method for structured knowledge bases as described in claim 1, characterized in that, The method involves performing semantic clustering compression on the structured search result set to obtain a representative subset of evidence. The structured search result set is then transformed into a semantically modeled text representation; The semantically modeled text representations are automatically clustered based on semantic similarity to obtain a cluster set; Based on clustering frequency, representativeness, or diversity strategies, representative records are selected from each cluster of the cluster set to form the representative evidence subset, wherein the representative evidence subset satisfies coverage constraints, redundancy constraints, and size constraints.

5. The retrieval-enhanced intelligent question-answering method for structured knowledge bases as described in claim 1, characterized in that, The constraint generation rules include: During the generation of verifiable scientific research question-and-answer results, only content based on the knowledge records in the context input is allowed to be generated, and the introduction of external unverified information is prohibited. When the representative subset of evidence is empty, output a predefined no-knowledge response; Verifiable research Q&A results must not be presented in an enumerative manner, through subjective inference, or through extended reasoning.

6. A retrieval-enhanced intelligent question-answering system for structured knowledge bases, characterized in that, The system is used to implement the retrieval enhancement intelligent question answering method for structured knowledge bases as described in any one of claims 1 to 5, wherein the system comprises: The query rewriting module is used to perform structured semantic parsing under the constraints of predefined domain knowledge patterns after receiving natural language research questions input by users, and output structured query statements. The structured knowledge retrieval module is used to perform database-level retrieval operations on the structured query statement and output a set of structured retrieval results. The semantic clustering and result compression module is used to perform semantic clustering and compression processing on the structured retrieval result set to obtain a representative evidence subset; The constrained generation module is used to generate constrained content for the representative subset of evidence and output verifiable scientific research question-and-answer results.

Images