Multi-source heterogeneous database-oriented spatio-temporal data natural language query method and system

Abstract

The invention provides a spatio-temporal data natural language query method and system for a multi-source heterogeneous database, and the method comprises the steps: at least inputting a natural language query demand inputted by a user into a demand conversion model, and obtaining a structured spatio-temporal query instance outputted by the demand conversion model, the structured spatio-temporal query instance conforms to a predefined structured spatio-temporal query specification; performing grammar mapping on the structured spatio-temporal query instance to obtain respective original query statements of a plurality of target databases required by the query; and executing query in the corresponding target database based on the obtained original query statement, and outputting a query result required by the user based on a result returned by each target database. By structuring the spatio-temporal query specification and integrating the large language model, a common user can also retrieve cross-platform and multi-source heterogeneous spatio-temporal data based on a natural language, the retrieval difficulty is reduced, and the retrieval efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of spatiotemporal data query technology, and in particular to a method and system for spatiotemporal data natural language query for multi-source heterogeneous databases. Background Technology

[0002] Currently, spatiotemporal data exhibits characteristics of multi-source heterogeneity and diverse types. To adapt to differences in data types and applications, the industry generally adopts a hybrid storage model combining relational databases, non-relational databases, and spatial data engines to accommodate different types of spatiotemporal data. Query requirements have expanded from simple conditional filtering to determining spatial relationships and performing joint statistics across time periods and regions.

[0003] However, different types of database interfaces and query syntax are not uniform, requiring multiple sets of statements to be written for the same retrieval. Cross-source queries are inefficient, making it difficult for ordinary business personnel who have not received training or learning to query such spatiotemporal data. Summary of the Invention

[0004] This invention provides a spatiotemporal data natural language query method and system for multi-source heterogeneous databases, which solves the defects of existing technologies that require writing multiple sets of statements for the same retrieval and have low efficiency in cross-source queries, and realizes a spatiotemporal data query method that can be implemented based on natural language description.

[0005] This invention provides a natural language query method for spatiotemporal data in multi-source heterogeneous databases, including: The user's natural language query request is input into the request transformation model to obtain a structured spatiotemporal query instance output by the request transformation model. The structured spatiotemporal query instance conforms to a predefined structured spatiotemporal query specification. Syntax mapping is performed on the structured spatiotemporal query instance to obtain the native query statements of several target databases required for this query; Based on the obtained native query statement, execute the query in the corresponding target database, and output the query results required by the user based on the results returned by each target database; The demand transformation model is obtained by fine-tuning a large language model; the structured spatiotemporal query specification is obtained by extending spatiotemporal operators on the basis of the general query language CQL.

[0006] According to the present invention, a spatiotemporal data natural language query method for multi-source heterogeneous databases is provided, wherein the spatiotemporal operators include spatial predicates, temporal predicates, spatial functions, and temporal functions.

[0007] According to the present invention, a method for spatiotemporal data natural language querying for multi-source heterogeneous databases further includes, before the step of inputting the user-inputted natural language query request into a request transformation model: Construct a training corpus, which is a paired dataset consisting of natural language query requirements and their corresponding structured spatiotemporal query instances; Define a loss function, which includes sequence generation loss, type consistency loss, topological validity loss, and semantic mapping consistency loss. The type consistency loss is used to penalize outputs where the type of the spatiotemporal operator parameter field predicted by the model is inconsistent with the expected type of the operator. The topological validity loss is used to penalize spatial operators that do not conform to spatial topological rules. The semantic mapping consistency loss is used to penalize outputs that do not match the spatiotemporal operators triggered by semantics in natural language with the spatiotemporal operators selected by the model. On the training corpus, the large language model is fine-tuned based on the defined loss function to obtain the demand conversion model.

[0008] According to the present invention, a method for spatiotemporal data natural language querying for multi-source heterogeneous databases includes the step of inputting the user's natural language query request into a request transformation model and obtaining a structured spatiotemporal query instance output by the request transformation model, specifically comprising: The demand transformation model encodes the natural language query demand input by the user at the encoding end, generates a dynamic query semantic vector, and then calls the rule prompt vector obtained by pre-encoding the rule prompt text. After concatenating it with the dynamic query semantic vector, a fusion vector containing spatiotemporal constraint features is obtained. The demand transformation model filters the predicted probability distribution for each generated time step at the decoding end using a mask vector. The mask vector is dynamically constructed based on field type consistency constraints, topological legality constraints, and semantic mapping consistency constraints, and is used to shield spatiotemporal operators or parameter field candidates that do not meet the constraints. Based on the filtered probability distribution, the structured spatiotemporal query instance is generated by sampling. The rule prompt text is pre-built based on metadata from multiple target databases, including type consistency constraints, topology legality constraints, and semantic mapping consistency constraints.

[0009] This invention provides a method for spatiotemporal data natural language querying for multi-source heterogeneous databases. The step of performing syntax mapping on the structured spatiotemporal query instance to obtain the native query statements for each of the multiple target databases specifically includes: The data table attributes in the structured spatiotemporal query instance are parsed to determine the target database type to which each data table belongs; An abstract syntax tree is constructed based on the structured instance, and the nodes of the abstract syntax tree are traversed to bind data source properties; Traverse the abstract syntax tree from bottom to top, and perform flow splitting on the tree nodes according to the preset database dialect mapping rules: If the expressions corresponding to the current node and its child nodes all belong to the same data source, the node is converted into the native expression string of the target database and synthesized upwards to generate a local native query statement; If the expression corresponding to the current node involves different data sources, or performs cross-database join and aggregation operations, the conversion to a native query statement is aborted, and the node is marked as an in-memory task to be executed in the middleware. The output includes a hybrid execution plan of the local native query and the in-memory task; The memory tasks include memory computation tasks generated by node expression transformation, memory connection tasks generated by cross-library connection, and memory aggregation tasks generated by aggregation operations.

[0010] According to the present invention, a natural language query method for spatiotemporal data in multi-source heterogeneous databases includes the following steps: executing a query in the corresponding target database based on the obtained native query statement, and outputting the query results required by the user based on the results returned by each target database. The local native query statements in the hybrid execution plan are executed concurrently to obtain each single-source basic result set from the corresponding target database, and the single-source basic result sets are uniformly converted into a standard data format. Based on the memory join task in the hybrid execution plan, a federated join is performed on the single-source basic result set in the middleware memory to generate a wide table data stream; Traverse the wide table data stream, dynamically bind multi-source field values ​​to the abstract syntax tree operator context reserved by the memory computing task, perform algebraic operations and logical judgments, and complete cross-source filtering and computation of data. Based on the memory aggregation task in the hybrid execution plan, memory bucketing and reduction calculations are performed on the computed data stream; Perform global sorting and field projection to output the query results required by the user.

[0011] This invention also provides a spatiotemporal data natural language query system for multi-source heterogeneous databases, comprising: The conversion module is used to input the user's natural language query requirements into the demand conversion model to obtain a structured spatiotemporal query instance output by the demand conversion model. The structured spatiotemporal query instance conforms to a predefined structured spatiotemporal query specification. The mapping module is used to perform syntax mapping on the structured spatiotemporal query instance to obtain the original query statements of several target databases required for this query. The output module is used to execute queries in the corresponding target databases based on the obtained raw query statements, and output the query results required by the user based on the results returned by each target database; The demand transformation model is obtained by fine-tuning a large language model; the structured spatiotemporal query specification is obtained by extending spatiotemporal operators on the basis of the general query language CQL.

[0012] 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 program to implement the spatiotemporal data natural language query method for multi-source heterogeneous databases as described above.

[0013] 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 spatiotemporal data natural language query method for multi-source heterogeneous databases as described above.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the spatiotemporal data natural language query method for multi-source heterogeneous databases as described above.

[0015] The present invention provides a natural language query method and system for spatiotemporal data in multi-source heterogeneous databases. By extending spatiotemporal operators on the basis of existing general query languages, a set of structured spatiotemporal query specifications is defined. After integrating a large language model, ordinary users can also perform cross-platform, multi-source heterogeneous spatiotemporal data retrieval based on natural language, reducing the difficulty of retrieval and improving the efficiency of retrieval. This allows users to focus on the core business logic of their needs without having to worry about the physical structure of the data, providing key technical support for the integrated management and efficient utilization of cross-domain data resources. Attached Figure Description

[0016] 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.

[0017] Figure 1 This is one of the flowcharts illustrating the spatiotemporal data natural language query method for multi-source heterogeneous databases provided by the present invention; Figure 2 This is a general spatiotemporal retrieval language class diagram of the spatiotemporal data natural language query method for multi-source heterogeneous databases provided by the present invention; Figure 3This is the second flowchart of the spatiotemporal data natural language query method for multi-source heterogeneous databases provided by the present invention; Figure 4 This is a schematic diagram of the spatiotemporal data natural language query system for multi-source heterogeneous databases provided by the present invention; Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0018] 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.

[0019] The following is combined Figures 1 to 3 This invention introduces a spatiotemporal data natural language query method for multi-source heterogeneous databases, such as... Figure 1 As shown, it includes: Step 101: Input the natural language query requirements input by the user into the demand transformation model to obtain a structured spatiotemporal query instance output by the demand transformation model. The structured spatiotemporal query instance conforms to a predefined structured spatiotemporal query specification. The demand transformation model is obtained by fine-tuning a large language model; the structured spatiotemporal query specification is obtained by extending spatiotemporal operators on the basis of the general query language CQL.

[0020] User-inputted natural language queries include queries involving time and / or spatial relationships. For example, querying the number of newly added shops within 1 kilometer of subway station A in the past year contains the time query "past year" and the spatial query "within 1 kilometer of subway station A".

[0021] Typically, the data required for such spatiotemporal queries is stored in different databases, and the interfaces and query syntax of the multiple different types of databases involved in such queries are not consistent, making it difficult for ordinary business personnel to independently write query statements for each database to perform the queries.

[0022] Although in some feasible implementations, a large language model can be trained to generate query statements for different databases based on the user's natural language query requirements, allowing the user to query each database using the query statements output by the large language model and obtain the desired query results, this approach requires training a large language model for each database separately to ensure that the large language model can accurately output the corresponding query statements, reduce the probability of hallucinations, has high training costs, and the effectiveness is difficult to guarantee.

[0023] Therefore, in this invention, the spatiotemporal operators are first extended on the basis of the Common Query Language (CQL) to define a new structured spatiotemporal query specification, which is called the Geospatial Common Query Language (GCQL) in this embodiment. Based on this, by fine-tuning the large language model, the large language model can output the corresponding structured spatiotemporal query instance, that is, the GCQL statement corresponding to the natural language, based on the natural language query requirements input by the user, so as to perform subsequent query steps through syntax mapping. The large language model only needs to learn the relationship between the natural language containing the spatiotemporal query requirements and GCQL, which effectively reduces the training difficulty and training cost.

[0024] Optionally, a training set is pre-built, and each piece of data in the training set is a natural language query statement containing spatiotemporal query requirements and its corresponding GCQL statement. Using the GCQL statement as a label, the large language model is fine-tuned on the training set to obtain the requirement transformation model.

[0025] Alternatively, the large language model can be any inference model such as DeepSeek.

[0026] Based on this, the demand transformation model receives the natural language query requirements input by the user and can output the corresponding structured spatiotemporal query instance, which is the GCQL statement corresponding to the user's query requirements.

[0027] Step 102: Perform syntax mapping on the structured spatiotemporal query instance to obtain the native query statements of several target databases required for this query; Since the structured spatiotemporal query specification defined in this invention is a spatiotemporal retrieval language with strict organizational format and semantic rules, it can map the structured spatiotemporal query instance output by the demand transformation model into the native query statement of any database through syntax mapping.

[0028] Specifically, the first step is to determine all the target databases that may correspond to the natural language query request input by the user. For example, the demand transformation model obtained by fine-tuning in this embodiment is used for the retrieval application of natural resource data in a certain region. The target databases involved in the query request of the user using the demand transformation model can be determined in advance. Therefore, the mapping relationship between GCQL and the native query statement of each possible target database can also be pre-built. Thus, when applying the application, the pre-configured mapping relationship can be directly called to map the structured spatiotemporal query instance to the native query statement of each of the corresponding target databases.

[0029] Step 103: Execute the query in the corresponding target database based on the obtained original query statement, and output the query results required by the user based on the results returned by each target database; Based on the native query statements corresponding to each database, query operations can be performed in the corresponding target database.

[0030] Optionally, when the query requirements corresponding to the structured spatiotemporal query instance all belong to the same source target database, the original query statement of the corresponding target database can be obtained directly based on the structured spatiotemporal query instance and the query can be executed. The query results required by the user can be output based on the returned results.

[0031] When the query requirements corresponding to the structured spatiotemporal query instance belong to target databases from different sources, the subquery corresponding to each target database is determined, and the corresponding native query statement is generated. The query is then retrieved from each target database, and the returned results are subjected to cross-source calculation, association, and aggregation to serve as the query results required by the user.

[0032] This invention defines a set of structured spatiotemporal query specifications by extending spatiotemporal operators on the basis of existing general query languages. After integrating a large language model, it enables ordinary users to retrieve cross-platform, multi-source heterogeneous spatiotemporal data based on natural language, reducing the difficulty of retrieval and improving retrieval efficiency. This allows users to focus on the core business logic of their needs without having to worry about the physical structure of the data, providing key technical support for the integrated management and efficient utilization of cross-domain data resources.

[0033] In the spatiotemporal data natural language query method for multi-source heterogeneous databases of the present invention, the spatiotemporal operators include spatial predicates, temporal predicates, spatial functions, and temporal functions.

[0034] Specifically, CQL is a language created by the Open Geospatial Information Consortium for Web service specifications. It extracts concepts shared by General Query Language and Structured Query Language for matching and defines abstract classes in the query language.

[0035] In this implementation, GCQL extends CQL with spatiotemporal predicates and spatiotemporal functions, and the structural relationships between classes are as follows: Figure 2 As shown.

[0036] Specifically, a spatial predicate is introduced to determine whether geometric values ​​conform to a certain relationship. Optionally, it includes commonly used spatial predicates as shown in Table 1 below: Table 1

[0037] The spatial predicates mentioned above are all binary spatial relation decision operators, which are called in GCQL in the form of {"<spatial predicate>":[Expression A, Expression B]}.

[0038] A time predicate is introduced to determine whether time values ​​conform to a certain relationship. Optionally, it includes commonly used time predicates as shown in Table 2 below: Table 2

[0039] The time predicates mentioned above are all binary time relation decision operators, and in GCQL, time expressions are called in the form of {"<time predicate>":[Expression A, Expression B]}.

[0040] Spatial functions are introduced for the creation, measurement, and spatial geometric calculation of geometric objects. Optionally, commonly used spatial functions as shown in Table 3 below are included: Table 3

[0041] The aforementioned spatial functions are called in GCQL in the form of {"function": {"name":"<function name>", "arguments":[<parameter values>,...]}}, and can be used for spatial property calculations, conditional expression construction, and as operands for spatial predicates.

[0042] Time functions are introduced for creating time objects, extracting time elements, and calculating time intervals. Optionally, commonly used time functions as shown in Table 4 below are included: Table 4

[0043] The aforementioned time functions are called in GCQL in a uniform form: {"function": {"name":"<function name>", "arguments":[<parameter value>,...]}}. They can be used for calculating time attributes, constructing conditional logic, and as operands for time predicates.

[0044] By using the above method, the structured spatiotemporal query specification of the present invention can be constructed, which can fully express the spatiotemporal relationship of query requirements and facilitate mapping with the original query languages ​​of various databases.

[0045] In the spatiotemporal data natural language query method for multi-source heterogeneous databases of the present invention, before the step of inputting the user-inputted natural language query request into the request transformation model, the method further includes: Construct a training corpus, which is a paired dataset consisting of natural language query requirements and their corresponding structured spatiotemporal query instances; To fine-tune the large language model and obtain the required requirement transformation model, it is first necessary to identify the heterogeneous target database systems that may be involved in the user requirements, extract the directory information of each system, obtain the name, column name, data type, coordinate reference system (CRS) of geometric fields of all data tables, and the connection relationship between tables, and build a metadata table.

[0046] Based on this, a training corpus is constructed, which consists of natural language (NL) query requirements and their corresponding GCQL structured statements (i.e., structured spatiotemporal query instances), covering various semantic types such as basic retrieval, multi-source joint query, and aggregation statistics.

[0047] Define a loss function, which includes sequence generation loss, type consistency loss, topological validity loss, and semantic mapping consistency loss. The type consistency loss is used to penalize outputs where the type of the spatiotemporal operator parameter field predicted by the model is inconsistent with the expected type of the operator. The topological validity loss is used to penalize spatial operators that do not conform to spatial topological rules. The semantic mapping consistency loss is used to penalize outputs that do not match the spatiotemporal operators triggered by semantics in natural language with the spatiotemporal operators selected by the model. To reduce the illusion of large language models and improve the model's ability to understand spatiotemporal data, the total loss function used for fine-tuning the large language model in this implementation is as follows: ; in, For the total loss, The loss for sequence generation of a conventional language model is calculated based on paired samples of "natural language query - GCQL target sequence" in the training corpus; , and These are type consistency loss, topological legitimacy loss, and semantic mapping consistency loss, respectively.

[0048] Specifically, in order to enhance the model's understanding of spatiotemporal semantics, this implementation first constructs type consistency constraints based on the parameter definitions of spatiotemporal operators, constructs topological legality constraints based on geographic spatiotemporal specifications, establishes semantic mapping consistency constraints by combining extracted database metadata tables with natural language expression features, and summarizes them to form a structured constraint library.

[0049] In one feasible implementation, each constraint can be organized into a constraint table.

[0050] For example, a type consistency constraint table can be constructed where each data entry includes a spatiotemporal operator identifier, a parameter position index, and a set of allowed field data types. For instance, the time predicate `BEFORE` determines whether one time object precedes another; its parameter must be a time data type. Spatial or numeric fields cannot be used for time order determination and will not be recorded in the allowed type set for that operator parameter in the type consistency constraint table.

[0051] Construct a topological legality constraint table. Each entry in the table includes a spatial operator identifier, a parameter geometry type, and a topological relationship legality flag, used to define compatibility rules between geometric dimensions. For example, the spatial function CROSSES calculates intersection relationships, while point-to-point and point-to-line relationships do not occur and are not recorded in the topological legality constraint table.

[0052] A single data entry in the semantic mapping consistency constraint table includes common query semantics and their corresponding feasible spatiotemporal operators. For example, the query semantic "find within 1 kilometer of point A" can be achieved by both the Buffer and Distance spatial operators. Therefore, the record "nearby" represents the query semantics, and the records "Buffer" and "Distance" represent the corresponding feasible spatiotemporal operators.

[0053] Based on this, let the input natural query language be x, and the GCQL generated by the model contain K spatiotemporal loss operators. This represents the spatiotemporal operator at the k-th position. The number of parameters is Operator The first j The parameters are Total number of parameters .

[0054] Type consistency loss can be determined as follows: Define a type indicator: ; In the formula, Indicates the type of parameter field used in the model prediction. This represents the set of field types allowed at the j-th parameter position of the k-th spatiotemporal operator given by the constraint library. The formula for calculating type consistency loss is as follows: .

[0055] Topology legitimacy loss can be determined as follows: Define the topology indicator: ; In the formula, This represents a spatial operator and its geometric dimension combination currently generated by the model. This represents the set of topologically valid combinations (i.e., the pre-constructed topological validity constraint table). The formula for calculating the topological validity loss is as follows: .

[0056] The semantic mapping legitimacy loss can be determined as follows: Candidate consistency metric calculation: ; In the formula, For the first k The set of spatiotemporal operator candidates corresponding to each semantic trigger. For the model to input natural language queries Under the condition of decoding the first k The prediction probability of the candidate spatiotemporal operator o is selected based on semantic time. The formula for calculating the semantic mapping consistency loss is as follows: .

[0057] On the training corpus, the large language model is fine-tuned based on the defined loss function to obtain the demand conversion model.

[0058] The training corpus, constraint library, and loss function constructed in the above manner can be used to fine-tune the large language model on the training corpus until the model converges, resulting in a demand transformation model that can accurately understand spatiotemporal data.

[0059] In the spatiotemporal data natural language query method for multi-source heterogeneous databases of the present invention, the step of inputting the user's natural language query request into a request transformation model and obtaining a structured spatiotemporal query instance output by the request transformation model specifically includes: The demand transformation model encodes the natural language query demand input by the user at the encoding end, generates a dynamic query semantic vector, and then calls the rule prompt vector obtained by pre-encoding the rule prompt text. After concatenating it with the dynamic query semantic vector, a fusion vector containing spatiotemporal constraint features is obtained. The demand transformation model filters the predicted probability distribution for each generated time step at the decoding end using a mask vector. The mask vector is dynamically constructed based on field type consistency constraints, topological legality constraints, and semantic mapping consistency constraints, and is used to shield spatiotemporal operators or parameter field candidates that do not meet the constraints. Based on the filtered probability distribution, the structured spatiotemporal query instance is generated by sampling. The rule prompt text is pre-built based on metadata from multiple target databases, including type consistency constraints, topology legality constraints, and semantic mapping consistency constraints.

[0060] Based on this, the demand transformation model vectorizes the natural language query demand input by the user into a dynamic query semantic vector at the encoding end, and calls the pre-stored rule suggestion vector to concatenate with the dynamic query semantic vector so that the resulting fusion vector contains both the user's query semantics and the spatiotemporal constraint features in the rule suggestion vector.

[0061] Furthermore, a mask vector is dynamically constructed based on the three constraint tables stored in the constraint library and the natural language query requirements input by the user, so as to block out spatiotemporal operators or parameter field candidates that do not meet the constraint conditions during the decoding stage.

[0062] Specifically, at the decoding end, the demand transformation model uses a mask vector to filter the predicted probability distribution for each generated time step, and generates a structured spatiotemporal query instance based on the filtered probability distribution.

[0063] It is understandable that, in the training phase, the demand conversion model in this implementation also needs to inject the rule hint vector and mask vector into the model in each round of training, based on the above. As an optimization criterion, the model is trained with spatiotemporal consistency constraints until the model converges.

[0064] Through the above methods, such as Figure 3 As shown in the blue section, during the fine-tuning of the large language model, the loss function is comprehensively optimized. The constructed training corpus (NL-GCQL corpus) and the spatiotemporal semantic enhancement mechanism based on the constraint library enable the demand transformation model to more accurately understand the spatiotemporal information of the input natural language query and output more accurate structured spatiotemporal query instances.

[0065] In the spatiotemporal data natural language query method for multi-source heterogeneous databases of the present invention, the step of performing syntactic mapping on the structured spatiotemporal query instance to obtain the native query statements of each of the multiple target databases specifically includes: The data table attributes in the structured spatiotemporal query instance are parsed to determine the target database type to which each data table belongs; An abstract syntax tree is constructed based on the structured instance, and the nodes of the abstract syntax tree are traversed to bind data source properties; Traverse the abstract syntax tree from bottom to top, and perform flow splitting on the tree nodes according to the preset database dialect mapping rules: If the expressions corresponding to the current node and its child nodes all belong to the same data source, the node is converted into the native expression string of the target database and synthesized upwards to generate a local native query statement; If the expression corresponding to the current node involves different data sources, or performs cross-database join and aggregation operations, the conversion to a native query statement is aborted, and the node is marked as an in-memory task to be executed in the middleware. The output includes a hybrid execution plan of the local native query and the in-memory computing task; The memory tasks include memory computation tasks generated by node expression transformation, memory connection tasks generated by cross-library connection, and memory aggregation tasks generated by aggregation operations.

[0066] During the syntax mapping stage, the tables attribute of the structured spatiotemporal query instance output by the model is first parsed, and the metadata directory is queried to determine the target database type to which each data table belongs.

[0067] Then, extract the spatiotemporal predicates, spatiotemporal functions, and logical combination relationships from the WHERE conditions and SELECT computed items of GCQL, convert the expressions into syntax tree nodes, construct the complete AST structure of the structured spatiotemporal query instance from top to bottom, and label each field node in the AST with its corresponding data source based on the metadata table.

[0068] Then, the abstract syntax tree is traversed from bottom to top, and the tree nodes are mapped using a single-source composition and multi-source pushdown splitting strategy according to the preset database dialect mapping rules: Single-source transformation: If the current node and its child nodes all belong to the same data source, the corresponding node is converted into the native expression string of the target database, and then synthesized upwards to generate a local native query expression.

[0069] Multi-source pushdown: If the current node involves different data sources, or performs cross-database join and aggregation operations, the conversion to a native query statement is stopped, its original AST operator structure is preserved, and the node is marked as an "in-memory computing task" to be executed in the middleware.

[0070] Then, the join clauses are parsed, and the join tasks are divided.

[0071] For single-source operations, perform unit joins: if the left and right tables belong to the same database instance, directly convert to the target database's native join syntax.

[0072] For multi-source operations, perform federated joins: if the left and right tables belong to different database instances, or the target database does not support join operations, no join statement is generated, the node is marked as an "in-memory join task" and pushed down for processing.

[0073] Furthermore, the aggregation clauses are parsed to divide the aggregation tasks.

[0074] For single-source operations, perform single-source aggregation: if the aggregation key belongs to the same data source, directly convert it to the target database's native aggregation syntax.

[0075] For multi-source operations, perform federated aggregation: if the aggregation involves multiple source fields, do not generate an aggregation statement for the target database, and mark the node as an "in-memory aggregation task" for pushdown processing.

[0076] Finally, based on the generated transformation fragments and pushdown strategy, for the parts determined to be "single-source operations", they are assembled according to the syntax specifications of the target database to generate local executable query statements corresponding to each data source; for multi-source operations, a hybrid execution plan including in-memory tasks is generated.

[0077] The above process is as follows Figure 3 The yellow part is shown in the image.

[0078] The target database includes relational databases and non-relational databases, specifically including PostgreSQL (PostGIS), MySQL, DM, Kingbase, Oracle, SpatiaLite, MongoDB, ElasticSearch, and Redis; in other feasible implementations, the target database can also be determined according to the specific application field and user needs.

[0079] The preset database dialect mapping rules represent the mapping relationship between GCQL and the native query languages ​​of each target database. In this embodiment, these rules are stored and retrieved in the form of a mapping table. While existing mapping tables can be used for the CQL mapping, this embodiment additionally constructs a mapping table between the spatiotemporal operators and each target database.

[0080] As an example, the mapping relationship between the spatiotemporal operators defined in this implementation and each target database can be shown in Table 5-8 below: Table 5 Spatial Predicate Mapping Table

[0081] Table 6 Time Predicate Mapping Table

[0082] Table 7 Spatial Function Mapping Table

[0083] Table 8 Time Function Mapping Table

[0084] The above mapping table only illustrates the mapping relationship between some spatiotemporal operators and the SQL of the target database. Those skilled in the art can define their own mappings based on the actual target database used.

[0085] In the spatiotemporal data natural language query method for multi-source heterogeneous databases of this invention, the step of executing the query in the corresponding target database based on the obtained native query statement and outputting the query result required by the user based on the results returned by each target database specifically includes: The local native query statements in the hybrid execution plan are executed concurrently to obtain each single-source basic result set from the corresponding target database, and the single-source basic result sets are uniformly converted into a standard data format. Based on the memory join task in the hybrid execution plan, a federated join is performed on the single-source basic result set in the middleware memory to generate a wide table data stream; Traverse the wide table data stream, dynamically bind multi-source field values ​​to the abstract syntax tree operator context reserved by the memory computing task, perform algebraic operations and logical judgments, and complete cross-source filtering and computation of data. Based on the memory aggregation task in the hybrid execution plan, memory bucketing and reduction calculations are performed on the calculated data stream, and global sorting and field projection are performed to output the final query results.

[0086] In the data extraction and merging stage, such as Figure 3 As shown in the green section, the system first executes the generated set of native query statements for each target database in the corresponding target database to obtain the result set of each unit. Then, all returned data is uniformly converted into JSON type, the spatial field is uniformly parsed and the sequence number is uniformly converted into GeoJSON type, and the time field is uniformly converted into standard timestamp type.

[0087] Based on the memory-connection task in the generated hybrid execution plan, the heterogeneous dataset is obtained. In memory, a hash connection or nested loop strategy is used to complete the logical pairing and attribute alignment of the data, and a wide table data stream is generated.

[0088] Load the marked in-memory computation task, traverse the wide table data stream, dynamically bind multi-source field values ​​to variable contexts, execute algebraic operations and logical judgments defined by nodes, and complete data filtering or attribute expansion to perform cross-source computation.

[0089] Based on the marked memory aggregation task, initialize the grouping container, traverse the data stream, map the data to logical buckets according to the aggregation key, perform reduction calculations within the buckets, and generate the final result to complete the federated aggregation.

[0090] Finally, the result set is sorted and the fields are projected to output the query results required by the user.

[0091] In a specific example, the user's natural language query request is: to query the spatial distribution of forestry sample plots newly added in the East and West districts of City A in the past year, and to count the number of sample plots in each district.

[0092] Inputting this natural language query requirement into the requirement transformation model, the resulting structured spatiotemporal query example output by the model is as follows: ; Based on this, the above structured spatiotemporal query instance is syntactically mapped to obtain the corresponding native query statement for the target database. The steps are as follows: Parse the data table attributes to determine the corresponding target database. In this example, the forestry data table forest_plot is in a PostgreSQL database, and the administrative division data table admin_district is in an Oracle database.

[0093] Then, perform single-source synthesis A (Oracle): It is found that both EQUALS under the OR logic involve admin_district (Oracle), and they are merged and converted into the native SQL IN clause: "district_name" IN ('A City East District', 'A City West District').

[0094] Single-source synthesis B (PostgreSQL): Recognizes the AFTER time condition and converts it to native PG SQL: "create_time">= (NOW() - INTERVAL '1 year').

[0095] Multi-source pushdown: INTERSECTS involves cross-source fields, aborts the conversion, and marks it as an "in-memory computation task".

[0096] Perform a federated join: Cross-source heterogeneous data with no common key, marked as an "in-memory join task".

[0097] Performing federated aggregations: Aggregation keys (Oracle) and statistical fields (PG) are heterogeneous, marked as "in-memory aggregation task".

[0098] Then, the native local query statement is generated, as shown below: Generate Oracle query statement (SQL_A): SELECT district_id, district_name, geom FROM admin_district WHERE district_name IN ('East District of City A', 'West District of City A') Generate PostgreSQL query statements (SQL_B): SELECT plot_id, location, create_time FROM forest_plot WHERE create_time>= (NOW() - INTERVAL '1 year').

[0099] Finally, the query operation is performed based on the local native query statement, and the query results are output: Perform a federated join: Execute SQL_A, returning 2 records (face object P1 from the East region and face object P2 from the West region). Execute SQL_B, returning N records (all sample location objects from the past year). Execute a nested loop to pair {P1, P2} with {N points} using Cartesian products, generating a 2N-row wide table data stream to be verified.

[0100] Perform cross-source computation: Traverse the data stream, bind the plot coordinates of each row to the corresponding administrative region polygon (East or West) to the context, and perform geometric determination: Determine if the plot is within the East region? -> If yes, keep it; otherwise, discard it. Determine if the plot is within the West region? -> If yes, keep it; otherwise, discard it.

[0101] Handling federated aggregation: Initialize two grouping buckets in memory based on district_name, distribute the valid records retained from cross-source computation into the buckets, and perform COUNT reduction on each bucket.

[0102] The final output is a JSON list, including two sets of statistical results: JSON[ { "district_name": "East District of City A", "district_id": "1001", "plot_cnt": 8 }, { "district_name": "West District of City A", "district_id": "1002", "plot_cnt": 12 }] The complete processing flow obtained through the above method can be executed entirely by computer, enabling ordinary business personnel to perform cross-platform, multi-source heterogeneous spatiotemporal data retrieval based on natural language, reducing retrieval difficulty and improving retrieval efficiency.

[0103] The spatiotemporal data natural language query system for multi-source heterogeneous databases provided by the present invention will be described below. The spatiotemporal data natural language query system for multi-source heterogeneous databases described below can be referred to in correspondence with the spatiotemporal data natural language query method for multi-source heterogeneous databases described above.

[0104] like Figure 4 As shown, the spatiotemporal data natural language query system for multi-source heterogeneous databases includes a conversion module 401, a mapping module 402, and an output module 403.

[0105] The conversion module 401 is used to input the natural language query requirements input by the user into the demand conversion model to obtain a structured spatiotemporal query instance output by the demand conversion model, wherein the structured spatiotemporal query instance conforms to a predefined structured spatiotemporal query specification. The demand transformation model is obtained by fine-tuning a large language model; the structured spatiotemporal query specification is obtained by extending spatiotemporal operators on the basis of the general query language CQL.

[0106] User-inputted natural language queries include queries involving time and / or spatial relationships. For example, querying the number of newly added shops within 1 kilometer of subway station A in the past year contains the time query "past year" and the spatial query "within 1 kilometer of subway station A".

[0107] Typically, the data required for such spatiotemporal queries is stored in different databases, and the interfaces and query syntax of the multiple different types of databases involved in such queries are not consistent, making it difficult for ordinary business personnel to independently write query statements for each database to perform the queries.

[0108] Although in some feasible implementations, a large language model can be trained to generate query statements for different databases based on the user's natural language query requirements, allowing the user to query each database using the query statements output by the large language model and obtain the desired query results, this approach requires training a large language model for each database separately to ensure that the large language model can accurately output the corresponding query statements, reduce the probability of hallucinations, has high training costs, and the effectiveness is difficult to guarantee.

[0109] Therefore, in this invention, the spatiotemporal operators are first extended on the basis of the Common Query Language (CQL) to define a new structured spatiotemporal query specification, which is called the Geospatial Common Query Language (GCQL) in this embodiment. Based on this, by fine-tuning the large language model, the large language model can output the corresponding structured spatiotemporal query instance, that is, the GCQL statement corresponding to the natural language, based on the natural language query requirements input by the user, so as to perform subsequent query steps through syntax mapping. The large language model only needs to learn the relationship between the natural language containing the spatiotemporal query requirements and GCQL, which effectively reduces the training difficulty and training cost.

[0110] Optionally, a training set is pre-built, and each piece of data in the training set is a natural language query statement containing spatiotemporal query requirements and its corresponding GCQL statement. Using the GCQL statement as a label, the large language model is fine-tuned on the training set to obtain the requirement transformation model.

[0111] Alternatively, the large language model can be any inference model such as DeepSeek.

[0112] Based on this, the demand transformation model receives the natural language query requirements input by the user and can output the corresponding structured spatiotemporal query instance, which is the GCQL statement corresponding to the user's query requirements.

[0113] Mapping module 402 is used to perform syntax mapping on the structured spatiotemporal query instance to obtain the original query statements of several target databases required for this query. Since the structured spatiotemporal query specification defined in this invention is a spatiotemporal retrieval language with strict organizational format and semantic rules, it can map the structured spatiotemporal query instance output by the demand transformation model into the native query statement of any database through syntax mapping.

[0114] Specifically, the first step is to determine all the target databases that may correspond to the natural language query request input by the user. For example, the demand transformation model obtained by fine-tuning in this embodiment is used for the retrieval application of natural resource data in a certain region. The target databases involved in the query request of the user using the demand transformation model can be determined in advance. Therefore, the mapping relationship between GCQL and the native query statement of each possible target database can also be pre-built. Thus, when applying the application, the pre-configured mapping relationship can be directly called to map the structured spatiotemporal query instance to the native query statement of each of the corresponding target databases.

[0115] The output module 403 is used to execute queries in the corresponding target databases based on the obtained native query statements, and output the query results required by the user based on the results returned by each target database; Based on the native query statements corresponding to each database, query operations can be performed in the corresponding target database.

[0116] Optionally, when the query requirements corresponding to the structured spatiotemporal query instance all belong to the same source target database, the original query statement of the corresponding target database can be obtained directly based on the structured spatiotemporal query instance and the query can be executed. The query results required by the user can be output based on the returned results.

[0117] When the query requirements corresponding to the structured spatiotemporal query instance belong to target databases from different sources, the subquery corresponding to each target database is determined, and the corresponding native query statement is generated. The query is then retrieved from each target database, and the returned results are subjected to cross-source calculation, association, and aggregation to serve as the query results required by the user.

[0118] This invention defines a set of structured spatiotemporal query specifications by extending spatiotemporal operators on the basis of existing general query languages. After integrating a large language model, it enables ordinary users to retrieve cross-platform, multi-source heterogeneous spatiotemporal data based on natural language, reducing the difficulty of retrieval and improving retrieval efficiency. This allows users to focus on the core business logic of their needs without having to worry about the physical structure of the data, providing key technical support for the integrated management and efficient utilization of cross-domain data resources.

[0119] Figure 5 Example: A schematic diagram of the physical structure of an electronic device, such as... Figure 5As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a spatiotemporal data natural language query method for multi-source heterogeneous databases. The method includes: inputting the user's natural language query requirements into a requirement transformation model to obtain a structured spatiotemporal query instance output by the requirement transformation model, wherein the structured spatiotemporal query instance conforms to a predefined structured spatiotemporal query specification; performing syntactic mapping on the structured spatiotemporal query instance to obtain the native query statements of several target databases required for this query; executing the query in the corresponding target database based on the obtained native query statements, and outputting the query results required by the user based on the results returned by each target database; wherein the requirement transformation model is obtained by fine-tuning a large language model; and the structured spatiotemporal query specification is obtained by extending spatiotemporal operators on the basis of the general query language CQL.

[0120] Furthermore, the logical instructions in the aforementioned memory 530 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, 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.

[0121] 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 spatiotemporal data natural language query method for multi-source heterogeneous databases provided by the above methods. The method includes: inputting the natural language query requirements input by the user into a requirement transformation model to obtain a structured spatiotemporal query instance output by the requirement transformation model, wherein the structured spatiotemporal query instance conforms to a predefined structured spatiotemporal query specification; performing syntactic mapping on the structured spatiotemporal query instance to obtain the native query statements of several target databases required for this query; executing the query in the corresponding target database based on the obtained native query statements, and outputting the query results required by the user based on the results returned by each target database; wherein the requirement transformation model is obtained by fine-tuning a large language model; and the structured spatiotemporal query specification is obtained by extending spatiotemporal operators on the basis of the general query language CQL.

[0122] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a spatiotemporal data natural language query method for multi-source heterogeneous databases provided by the above methods. The method includes: inputting a user-inputted natural language query requirement into a requirement transformation model to obtain a structured spatiotemporal query instance output by the requirement transformation model, wherein the structured spatiotemporal query instance conforms to a predefined structured spatiotemporal query specification; performing syntactic mapping on the structured spatiotemporal query instance to obtain the native query statements of several target databases required for this query; executing the query in the corresponding target database based on the obtained native query statements, and outputting the query results required by the user based on the results returned by each target database; wherein the requirement transformation model is obtained by fine-tuning a large language model; and the structured spatiotemporal query specification is obtained by extending spatiotemporal operators on the basis of the general query language CQL.

[0123] 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.

[0124] 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.

[0125] 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 method for spatiotemporal data natural language querying for multi-source heterogeneous databases, characterized in that, include: The user's natural language query request is input into the request transformation model to obtain a structured spatiotemporal query instance output by the request transformation model. The structured spatiotemporal query instance conforms to a predefined structured spatiotemporal query specification. Syntax mapping is performed on the structured spatiotemporal query instance to obtain the native query statements of several target databases required for this query; Based on the obtained native query statement, execute the query in the corresponding target database, and output the query results required by the user based on the results returned by each target database; The demand transformation model is obtained by fine-tuning a large language model; the structured spatiotemporal query specification is obtained by extending spatiotemporal operators on the basis of the general query language CQL.

2. The spatiotemporal data natural language query method for multi-source heterogeneous databases according to claim 1, characterized in that, The spatiotemporal operators include spatial predicates, temporal predicates, spatial functions, and temporal functions.

3. The spatiotemporal data natural language query method for multi-source heterogeneous databases according to claim 1, characterized in that, Before the step of inputting the user's natural language query request into the request transformation model, the method further includes: Construct a training corpus, which is a paired dataset consisting of natural language query requirements and their corresponding structured spatiotemporal query instances; Define a loss function, which includes sequence generation loss, type consistency loss, topological validity loss, and semantic mapping consistency loss. The type consistency loss is used to penalize outputs where the type of the spatiotemporal operator parameter field predicted by the model is inconsistent with the expected type of the operator. The topological validity loss is used to penalize spatial operators that do not conform to spatial topological rules. The semantic mapping consistency loss is used to penalize outputs that do not match the spatiotemporal operators triggered by semantics in natural language with the spatiotemporal operators selected by the model. On the training corpus, the large language model is fine-tuned based on the defined loss function to obtain the demand conversion model.

4. The spatiotemporal data natural language query method for multi-source heterogeneous databases according to claim 1, characterized in that, The step of inputting the user's natural language query request into the request transformation model and obtaining the structured spatiotemporal query instance output by the request transformation model specifically includes: The demand transformation model encodes the natural language query demand input by the user at the encoding end, generates a dynamic query semantic vector, and then calls the rule prompt vector obtained by pre-encoding the rule prompt text. After concatenating it with the dynamic query semantic vector, a fusion vector containing spatiotemporal constraint features is obtained. The demand transformation model filters the predicted probability distribution for each generated time step at the decoding end using a mask vector. The mask vector is dynamically constructed based on field type consistency constraints, topological legality constraints, and semantic mapping consistency constraints, and is used to shield spatiotemporal operators or parameter field candidates that do not meet the constraints. Based on the filtered probability distribution, the structured spatiotemporal query instance is generated by sampling. The rule prompt text is pre-built based on metadata from multiple target databases, including type consistency constraints, topology legality constraints, and semantic mapping consistency constraints.

5. The spatiotemporal data natural language query method for multi-source heterogeneous databases according to any one of claims 1-4, characterized in that, The step of performing syntax mapping on the structured spatiotemporal query instance to obtain the native query statements for each of the multiple target databases specifically includes: The data table attributes in the structured spatiotemporal query instance are parsed to determine the target database type to which each data table belongs; An abstract syntax tree is constructed based on the structured instance, and the nodes of the abstract syntax tree are traversed to bind data source properties; Traverse the abstract syntax tree from bottom to top, and perform flow splitting on the tree nodes according to the preset database dialect mapping rules: If the expressions corresponding to the current node and its child nodes all belong to the same data source, the node is converted into the native expression string of the target database and synthesized upwards to generate a local native query statement; If the expression corresponding to the current node involves different data sources, or performs cross-database join and aggregation operations, the conversion to a native query statement is aborted, and the node is marked as an in-memory task to be executed in the middleware. The output includes a hybrid execution plan of the local native query and the in-memory task; The memory tasks include memory computation tasks generated by node expression transformation, memory connection tasks generated by cross-library connection, and memory aggregation tasks generated by aggregation operations.

6. The spatiotemporal data natural language query method for multi-source heterogeneous databases according to claim 5, characterized in that, The steps of executing queries in the corresponding target databases based on the obtained native query statements and outputting the query results required by the user based on the results returned by each target database specifically include: The local native query statements in the hybrid execution plan are executed concurrently to obtain each single-source basic result set from the corresponding target database, and the single-source basic result sets are uniformly converted into a standard data format. Based on the memory join task in the hybrid execution plan, a federated join is performed on the single-source basic result set in the middleware memory to generate a wide table data stream; Traverse the wide table data stream, dynamically bind multi-source field values ​​to the abstract syntax tree operator context reserved by the memory computing task, perform algebraic operations and logical judgments, and complete cross-source filtering and computation of data. Based on the memory aggregation task in the hybrid execution plan, memory bucketing and reduction calculations are performed on the computed data stream; Perform global sorting and field projection to output the query results required by the user.

7. A spatiotemporal data natural language query system for multi-source heterogeneous databases, characterized in that, include: The conversion module is used to input the user's natural language query requirements into the demand conversion model to obtain a structured spatiotemporal query instance output by the demand conversion model. The structured spatiotemporal query instance conforms to a predefined structured spatiotemporal query specification. The mapping module is used to perform syntax mapping on the structured spatiotemporal query instance to obtain the original query statements of several target databases required for this query. The output module is used to execute queries in the corresponding target databases based on the obtained raw query statements, and output the query results required by the user based on the results returned by each target database; The demand transformation model is obtained by fine-tuning a large language model; the structured spatiotemporal query specification is obtained by extending spatiotemporal operators on the basis of the general query language CQL.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the spatiotemporal data natural language query method for multi-source heterogeneous databases as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the spatiotemporal data natural language query method for multi-source heterogeneous databases as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the spatiotemporal data natural language query method for multi-source heterogeneous databases as described in any one of claims 1 to 6.

Images