Data processing method and system based on table large model

Abstract

The application discloses a data processing method and system based on a table large model, semantic indexing is performed on a knowledge subgraph in a table knowledge graph, a target knowledge subgraph is searched from the table knowledge graph, and finally the searched target knowledge subgraph is input into the table large model together with prompt information to generate a recommended result. Since the screened and optimized table knowledge subgraph is input into the table large model together with the prompt information, the output of the table large model is more accurate, the understanding ability of the model for complex data is enhanced, and the model has stronger generalization ability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to a data processing method and system based on a table large model. BACKGROUND

[0002] As an advanced information processing method, large model recommendation technology has been widely applied and developed in recent years. The basic principle of this technology is to understand and predict user behavior or preferences by constructing and training large-scale machine learning models. These models are usually based on deep learning frameworks such as neural networks, which can automatically extract features from large amounts of data and discover patterns hidden in the data. In the field of electric power, the processing and analysis of table data is a key step to ensure the stable operation of the power grid, optimize resource allocation, and improve system efficiency. The core of table analysis technology is to convert unstructured or semi-structured data into a structured format that can be understood and analyzed by machines.

[0003] In traditional table data processing, structured data is often converted into linear serialized text for processing, and table data is input into a large model in the form of serialized text. This method ignores the rich structural features of the original data. As a highly organized information carrier, the inherent row-column relationship, cell dependency, and cross-row and cross-column data patterns are important for understanding and analyzing data. However, when these complex relationships are simply flattened into one-dimensional text, valuable information is lost, limiting the large model's ability to understand the deeper meaning of the data, resulting in inaccurate output from the large model. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a data processing method and system based on a table large model, which inputs the filtered and optimized table knowledge subgraph together with the prompt information into the table large model, making the output of the table large model more accurate, not only enhancing the model's understanding of complex data, but also making it have stronger generalization ability.

[0005] To achieve the above-mentioned purpose, the embodiments of the present application provide a data processing method based on a table large model, comprising:

[0006] extracting features from the table knowledge graph related to electric power to obtain first semantic information of the table knowledge graph, and constructing a table knowledge subgraph according to the first semantic information;

[0007] associating the nodes in the table knowledge subgraph according to the first semantic information to optimize the table knowledge vector database;

[0008] At least two second semantic information of the to-be-processed table data is acquired, and a corresponding target knowledge subgraph is retrieved from the optimized table knowledge vector database according to the second semantic information.

[0009] All target knowledge subgraphs are aggregated to obtain a target knowledge subgraph set.

[0010] The target knowledge subgraph set and the predefined prompt information are input into the table large model, so that the table large model outputs a recommendation result.

[0011] As an improvement of the above scheme, the to-be-processed table data includes a title and a description of the table, wherein the title is information in the first row or the first column of the table, and the description is detailed power-related data recorded in the table.

[0012] As an improvement of the above scheme, the feature extraction of the power-related table knowledge graph obtains first semantic information of the table knowledge graph, including:

[0013] The nodes and edges in the table knowledge graph are extracted by using a multi-modal large model to obtain the first semantic information.

[0014] As an improvement of the above scheme, the construction of the table knowledge subgraph according to the first semantic information includes:

[0015] According to the first semantic information, a first graph neural network model is used to aggregate information of each node and its adjacent nodes to obtain a table knowledge subgraph corresponding to each node.

[0016] As an improvement of the above scheme, the association of the nodes in the table knowledge subgraph according to the first semantic information includes:

[0017] The association relationship of the nodes in the table knowledge subgraph is determined according to the first semantic information, and a graph model is constructed according to the association relationship.

[0018] The information of the graph model is iterated, so that the state vector of each node contains global relationship information through information iteration.

[0019] The state vectors of all nodes and the remaining nodes having an association relationship are integrated to optimize the table knowledge vector database.

[0020] As an improvement of the above scheme, the acquisition of at least two second semantic information of the to-be-processed table data includes:

[0021] The to-be-processed table data is extracted by using a multi-modal large model to obtain at least two second semantic information.

[0022] As an improvement of the above scheme, after the target knowledge sub-graph set is obtained, the method further comprises:

[0023] Reference semantic information of the original prompt information is obtained.

[0024] Each target knowledge sub-graph in the target knowledge sub-graph set is sorted according to the reference semantic information, and an updated knowledge sub-graph set is obtained.

[0025] As an improvement of the above scheme, the target knowledge sub-graph set and the predefined prompt information are input into the table large model, so that the table large model outputs the recommended result, which comprises:

[0026] The second graph neural network model is used to process information of each target knowledge sub-graph in the target knowledge sub-graph set, and structure information of each target knowledge sub-graph is obtained.

[0027] The structure information of the target knowledge sub-graph is mapped into the embedding space of the table large model, and soft prompt embedding information is obtained.

[0028] The soft prompt embedding information and the predefined prompt information are input into the table large model, so that the table large model outputs the recommended result.

[0029] To achieve the above object, the embodiment of the present application further provides a data processing system based on a table large model, comprising:

[0030] A first semantic information extraction module is configured to extract features of a table knowledge graph related to electric power to obtain first semantic information of the table knowledge graph.

[0031] A table knowledge sub-graph construction module is configured to construct a table knowledge sub-graph according to the first semantic information.

[0032] A table knowledge vector database optimization module is configured to associate nodes in the table knowledge sub-graph according to the first semantic information to optimize a table knowledge vector database.

[0033] A second semantic information extraction module is configured to obtain at least two second semantic information of to-be-processed table data.

[0034] A target knowledge sub-graph acquisition module is configured to acquire corresponding target knowledge sub-graphs from the optimized table knowledge vector database according to the second semantic information.

[0035] A target knowledge sub-graph aggregation module is configured to aggregate all target knowledge sub-graphs to obtain a target knowledge sub-graph set.

[0036] The recommendation result acquisition module is configured to input the target knowledge subgraph set and the predefined prompt information into the table large model, so that the table large model outputs a recommendation result.

[0037] As an improvement of the above-mentioned scheme, the system further comprises:

[0038] The reference semantic information extraction module is configured to acquire reference semantic information of the original prompt information.

[0039] The knowledge subgraph set updating module is configured to sort each target knowledge subgraph in the target knowledge subgraph set according to the reference semantic information, to obtain an updated knowledge subgraph set.

[0040] Compared with the prior art, the data processing method and system based on the table large model disclosed by the present application have the following advantages: the knowledge subgraphs in the table knowledge graph are semantically indexed, then the target knowledge subgraphs are retrieved from the table knowledge graph, and finally the retrieved target knowledge subgraphs are input into the table large model together with the prompt information to generate a recommendation result. Since the screened and optimized table knowledge subgraphs are input into the table large model together with the prompt information, the output of the table large model is more accurate, the model's understanding ability for complex data is enhanced, and the model has stronger generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a flowchart of the data processing method based on the table large model provided by the embodiment of the present application;

[0042] Figure 2 is a flowchart of the optimization of the table knowledge vector database provided by the embodiment of the present application;

[0043] Figure 3 is a flowchart of the table large model outputting a recommendation result provided by the embodiment of the present application;

[0044] Figure 4 is a structural block diagram of the data processing system based on the table large model provided by the embodiment of the present application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0046] In the field of electric power, the processing and analysis of tabular data is a key step in ensuring the stable operation of the power grid, optimizing resource allocation, and improving system efficiency. The core of tabular analysis technology lies in the conversion of unstructured or semi-structured data into structured formats that can be understood and analyzed by machines. Given the increasingly stringent requirements for data security and privacy protection, how to efficiently perform tabular data analysis while ensuring information security has become a pressing problem.

[0047] In the embodiments of the present application, the Self-Consistency Chain-of-Thought (CoT-SC) technology is used to optimize the retrieval efficiency of tabular knowledge graphs. The Self-Consistency Chain-of-Thought technology is an advanced technology aimed at improving the accuracy and consistency of model reasoning, especially when dealing with complex structured data such as tabular data. The core idea of the Self-Consistency Chain-of-Thought technology is to simulate the human thinking process, i.e., analyzing and verifying problems from multiple angles and levels, to ensure the rationality and accuracy of the final results. This technology first builds a multi-level knowledge graph to capture the potential relationships and patterns in the data. These knowledge graphs not only contain explicit relationships between entities but also cover implicit associations, enabling machines to better understand the meaning behind the information. Then, based on the rich knowledge background, the model generates a series of possible reasoning paths, each representing a way or perspective to solve a specific problem. This approach emphasizes exploring diverse solutions rather than relying solely on a single best answer. The key to this technology lies in how to evaluate and select the most appropriate reasoning path, where a self-consistency checking mechanism is introduced. It judges the rationality of each possible answer by checking the internal logic consistency of each answer.

[0048] In the embodiments of the present application, a table large model is used to complete the recommendation. The table large model technology is a new method for processing structured data that has emerged in recent years with the development of artificial intelligence and machine learning. It aims to automatically analyze, understand and generate structured table data by using generative large model technology, especially deep neural network models. The core of this technology is to capture the complex relationships between data in the table and make predictions or recommendations based on these relationships. The table large model usually includes an input layer, a hidden layer and an output layer, where the input layer is responsible for receiving the original table data, the hidden layer extracts feature representations through multiple nonlinear transformations, and finally the output layer gives the corresponding results according to the task requirements, such as classification, regression, etc. In order to enhance the expression ability of the model, the table large model often uses self-attention mechanism (Self-Attention Mechanism) and other advanced architectural elements, so that the model can perform well when processing large and complex table data. In specific implementation, the table large model can be designed based on different deep neural networks, including but not limited to graph neural networks, recurrent neural networks, convolutional neural networks, and variants of the Transformer architecture, etc. These methods have different focuses, for example, the graph neural network model is good at capturing the dependency relationships between cells in the table; the recurrent neural network model is suitable for processing sequential data, although it has the problem of gradient disappearance in long sequence processing; the convolutional neural network model is mainly used to detect local patterns in table data; and the Transformer architecture with its powerful parallel computing capability and effective modeling of long-distance dependencies has become one of the most popular solutions. In addition, to improve the generalization ability and stability of the model, regularization techniques, dropout, batch normalization (Batch Normalization) and other strategies are also used.

[0049] Referring to Figure 1 , Figure 1 is a flowchart of a data processing method based on a table large model provided by the embodiments of the present application, the method comprising:

[0050] S1, feature extraction is performed on the table knowledge graph related to power to obtain first semantic information of the table knowledge graph, and a table knowledge subgraph is constructed according to the first semantic information;

[0051] S2, the nodes in the table knowledge subgraph are associated according to the first semantic information to optimize the table knowledge vector database;

[0052] S3, at least two second semantic information of the table data to be processed are obtained, and according to the second semantic information, the corresponding target knowledge subgraph is retrieved from the optimized table knowledge vector database;

[0053] S4, aggregate all target knowledge sub-graphs to obtain a target knowledge sub-graph set;

[0054] S5, input the target knowledge sub-graph set and the predefined prompt information into the table large model to make the table large model output a recommendation result.

[0055] For example, by performing semantic indexing on the knowledge sub-graph in the table knowledge graph, then retrieving the target knowledge sub-graph from the table knowledge graph, and finally inputting the retrieved target knowledge sub-graph and the prompt information into the table large model to generate the recommendation result, since the screened and optimized table knowledge sub-graph and the prompt information are input into the table large model, the output of the table large model is more accurate, not only enhancing the model's understanding ability of complex data, but also making it have stronger generalization ability.

[0056] Specifically, in step S1, the feature extraction is performed on the table knowledge graph related to power to obtain first semantic information of the table knowledge graph, including: performing feature extraction on nodes and edges in the table knowledge graph by using a multi-modal large model to obtain first semantic information.

[0057] For example, the present application aims to enhance the recommendation performance of the table large model in the power field by using external table knowledge database. The table knowledge database in the power field contains rich structured table data and knowledge (such as summarizing the load conditions of various electrical equipment, the current-carrying capacity of different specifications of cables, the power generation capacity of photovoltaic power stations at different times, etc.), which presents these information in the form of triples, meeting: , both represent entities (nodes), represent the relationship between the two entities, where N and E represent the entity set and the relationship set in the table data, respectively. Through the above process, the table data is formalized into a table knowledge graph (TKG). The formalization process is a database system that converts table data (such as relational databases, Excel files, etc.) into structured knowledge graphs. Its core is to map rows, columns, and cells in the table to entities, attributes, and relationships in the knowledge graph through semantic modeling. For example, a table containing electrical equipment parameters can be converted into a "transformer" entity (one entity represents one node) in the knowledge graph, with its rated voltage, winding number, and "installed in" relationship with the "substation" entity. The table knowledge graph TKG is composed of a series of triples of "entity-relationship-entity". These triples are extracted from table data (such as mapping table rows / columns to entities , mapping the logical association between rows and columns to relationships ), and finally forming a structured knowledge network to support subsequent semantic indexing, reasoning, and retrieval operations.

[0058] Furthermore, graph neural network models are a key technology in graph machine learning, widely applied to various graph tasks. By iteratively updating node representations to aggregate information from neighboring nodes, this neural network model can effectively capture the latent topology and relational structure of a graph. Formally, typical graph neural network operations can be described as follows:

[0059] (1);

[0060] in, Represents a node In the Features of the layer; Represents a node In the Features of +1 layer; This indicates a feature combination operation (such as concatenation or addition), which combines the node's own features with the aggregated features of its neighbors; It is a node The neighbor set, that is, the set of nodes The set of other directly connected nodes; Indicates the first The aggregation function of layer +1 is used to aggregate the features of neighboring nodes. AGG is an aggregation function used to aggregate the features of neighbors and combine the information of neighbors with the node itself. Equation (1) describes the update process of node features in a graph neural network. By aggregating the features of neighboring nodes and combining them with its own features, node features of the next layer are generated to capture the local association information in the graph structure.

[0061] It should be noted that retrieving knowledge from traditional knowledge graphs decomposes them into nodes or triples and performs local retrieval only around the target entity within the knowledge graph. However, these methods simply retrieve the first-order neighbors of an entity, making it difficult to capture higher-order neighborhood effects between entities / items during the recommendation process, and they are not suitable for application scenarios involving tabular data. Therefore, to effectively retrieve knowledge from tabular knowledge graphs, this invention proposes performing semantic indexing on Hop-Field-based tabular knowledge subgraphs. This allows for flexible decomposition of the tabular knowledge graph and provides a comprehensive view of nodes within it. Therefore, this invention first introduces a pre-trained multimodal large model (PMLM), such as CogAgent, to capture the node semantic information and edge semantic information of nodes, obtaining the first semantic information.

[0062] The process of obtaining node semantic information satisfies the following formula:

[0063] (2);

[0064] in, is the node semantic information obtained after processing by the multi-modal large model; is the node ; is the dimension of the output vector , is the dimension of the output representation.

[0065] Similarly, the present application also captures the edge semantic information of the edge / relationship in the table knowledge graph, satisfying the following formula:

[0066] (3) ;

[0067] wherein, is the edge semantic information obtained after processing by the multi-modal large model; is the text attribute of the edge / relationship .

[0068] Through the PMLM model, the text attributes of the nodes and the edges / relationships are converted into dimensional semantic vectors and , to provide a semantic basis for subsequent knowledge graph analysis (such as semantic indexing, relationship reasoning) with the semantic vectors and as the first semantic information.

[0069] Specifically, in step S1, the table knowledge subgraph is constructed according to the first semantic information, including: according to the first semantic information, using a first graph neural network model to aggregate information of each node and its adjacent nodes to obtain a table knowledge subgraph corresponding to each node.

[0070] Illustratively, in order to retrieve the implicit knowledge in the fine-grained graph structure from the table knowledge graph TKG, the present application introduces a first graph neural network model, i.e. to aggregate the information of entity neighbors, is a learnable parameter in the first graph neural network model, such as usually containing a weight matrix, a bias term, etc., which is optimized through a backpropagation algorithm in the model training process to minimize the loss function (such as the difference between the predicted value and the true value). Therefore, the order embedding of the center entity (node) can be defined by the following formula:

[0071] (4) ; ​

[0072] in, Central Entity of Hierarchical embedding, representing the central entity In the Feature representation after aggregating neighbor information at each layer; For neighboring nodes exist The embedding of the order represents the feature representation of the neighbor node in the previous layer; For the edge exist The relational embedding of order describes the semantic information of edges. It is a node and neighboring nodes The edge between; It is a node The neighbor set. This formula describes the aggregation of central entities through the first graph neural network model. neighboring nodes and its connecting edges Information to generate central entity In the Layered embedding representations are used to capture the relationships between entities in a tabular knowledge graph.

[0073] It should be noted that for each entity (node), its A jump can be viewed as containing itself. The table knowledge subgraph representation of skipping neighbors. Therefore, the present invention can represent the table knowledge subgraph. , represented as The collection of all knowledge subgraphs. For each subgraph, the present invention stores its representation in a tabular knowledge vector database.

[0074] For example, each entity is accessed through " The way of "jumping" puts oneself and The jump range is a combination of neighbors (other entities that are directly or indirectly related) to form a group that includes itself and... A table-based knowledge subgraph that skips neighbors. For example, a device entity, when... When the value is 2, the subgraph not only includes the device itself, but also its directly connected (1-hop) and indirectly connected (2-hop) entities, constructing a locally related knowledge structure. All such subgraphs generated from entities (denoted as...) Summarize and form a set This means that the set Extracted from the tabular knowledge graph (TKG), consisting of various entities A jump represents the sum of the subgraphs formed. For a set... Each subgraph in The table knowledge graph is stored in a table knowledge vector database in a specific representation form (such as a vector). This storage manner facilitates subsequent retrieval, analysis and application of the subgraph, and provides data support for reasoning and recommendation based on table knowledge.

[0075] Specifically, in step S2, referring to Figure 2 , Figure 2 is a flowchart for optimizing a table knowledge graph, and the nodes in the table knowledge subgraph are associated according to the first semantic information to optimize a table knowledge vector database, comprising:

[0076] S21, determining the association relationship of the nodes in the table knowledge subgraph according to the first semantic information, and constructing a graph model according to the association relationship;

[0077] S22, information iteration is performed on the graph model, so that the state vector of each node contains global relationship information through information iteration;

[0078] S23, integrating the state vectors of all nodes and the remaining nodes having an association relationship to optimize the table knowledge vector database.

[0079] It should be noted that although many existing retrieval enhancement generation technologies can model user (electric power equipment) preferences by retrieving table knowledge through large models, a large amount of retrieval time is still consumed for each retrieval project, which can seriously reduce user experience and cause user loss. At the same time, due to the rich interaction between users and popular projects, most models tend to retain rich knowledge of popular projects, thereby performing poorly on cold-start projects. Therefore, the present application designs a table knowledge reasoning method based on self-consistent thinking chain technology to optimize the retrieval efficiency of the table knowledge vector database. Specifically, by converting the structured data in the table into a form suitable for model processing, the present application constructs a graph model that can reflect the internal relationship of the data. In this graph model, each node represents an entity (such as a row or column in a table), and the edge represents the relationship strength between entities (nodes). It is defined as follows:

[0080] (5) ;

[0081] wherein, represents the relationship strength between entities and , and are the state vectors of entities and , respectively, and the state vector is a mathematical representation of these entities in the graph neural network, which continuously integrates global information through iteration, is a function for computing the similarity or correlation between two entities. This process emphasizes the capture of high-order neighbor effects, not just limited to directly adjacent entities, but reflects deeper relationships through multiple iteration updates.

[0082] The text attributes of nodes (such as devices, attributes) and edges (such as connection relationships between devices) in the table knowledge subgraph are converted into semantic vectors by pre-training a multi-modal large model (PMLM), giving semantic information to the knowledge subgraph and forming the initial representation of the graph model. The graph model is an abstract mathematical framework, unlike the knowledge subgraph, which is the composition element and input basis of the graph model. The graph model realizes the mining and optimization of global relationships through the integration and reasoning of multiple knowledge subgraphs (such as aggregating the neighbor information of entities in the knowledge subgraph). The graph model is used to represent and analyze entities (nodes) and their relationships (edges). The core function is to optimize the state vector of entities through iteration, to mine high-order and global relationships between entities. For example, in the power equipment scenario, the graph model can discover the indirect relationship between "transformer failure" and "regional power grid load fluctuation" and "other associated device states" through multiple iterations, realizing local to global relationship reasoning. Self-consistent thinking chain reasoning is based on the semantic representation of the above knowledge subgraph. The nodes in the graph model are entities, and the edges are relationship strengths . Through iterative update rules (formula 6 below), the relationship estimation between entities is continuously optimized. For example, in the power equipment scenario, semantic indexing first captures the semantics of the device "transformer" and its attributes. Self-consistent reasoning further mines the high-order association relationship between the transformer and other devices (such as circuit breakers and lines) through the power grid topology, realizing the transition from simple semantic representation to complex relationship reasoning.

[0083] Using self-consistent thinking chain technology for reasoning means that on the basis of the above graph model, the relationship estimation between entities is continuously optimized using a recursive method. In this process, the algorithm adjusts the state of each node according to the current understanding of all relationships in order to achieve global consistency. Specifically, each iteration uses the following update rule:

[0084] (6) ;

[0085] where , represent the state vector of entity and entity in the th iteration, respectively, containing the comprehensive information of the entity in the current round (such as node information and neighbor node information); is the set of all neighbors directly connected to entity , representing the relationship between the entity and its neighbors Other entities that have a direct relationship; It is a connection and The weights reflect the strength or importance of the relationship between the two, while It is a non-linear transformation function responsible for integrating information from its neighbors and updating its own state.

[0086] This approach ensures the effective propagation and integration of information across the entire network, enhancing the model's ability to recognize complex patterns. In each iteration, each entity collects information from its neighbors, weighted according to their respective properties. Aggregate neighbor states And in combination with one's current state ,pass The function updates its own state. After multiple iterations, information is fully propagated and integrated throughout the graph model. Eventually, the state vector of each entity will contain comprehensive relationship information between itself and all other entities, enabling the graph model to reach a globally consistent stable state.

[0087] It should be noted that self-consistent thought chain reasoning iteratively updates the state vector of each node (entity), ensuring that it not only contains its own semantics (such as the model information of "transformer T1001"), but also aggregates relational information from neighboring nodes (such as the load information of "substation S001" associated through the "installed at" relation). Through multiple iterations, the state vector of each node eventually contains global relational information of the entire graph (such as high-order dependencies between devices, synergistic effects across substations, etc.).

[0088] Finally, after several iterations, the graph model will reach a stable state (the model achieves smooth convergence; the number of iterations depends on the actual training process). At this point, the state vector of each entity contains comprehensive relationship information between it and all other entities, thus completing the optimization of the tabular knowledge vector database. Based on these optimized state vectors, the final tabular knowledge reasoning task can be performed.

[0089] Furthermore, for each node, the matching degree between the current node and the remaining nodes with which it has an association is calculated, and the state vectors of the remaining nodes are sorted according to the matching degree. For example, to predict other entities that an entity might be interested in, the following formula can be used for scoring:

[0090] (7);

[0091] in, Refers to the entity To the entity Recommended score, Entities To the entity The state vector, It is a scoring function that combines the state vectors of the source and target entities to evaluate the degree of matching between them. The scoring is based on the iteratively optimized entity state vectors, and the process of evaluating the degree of matching between entities provides a quantitative basis for knowledge reasoning and recommendation. The higher the score, the better the matching of the entity. and entity The higher the matching degree, the more potentially relevant entities can be sorted based on these scores, and the most relevant content to the target entity can be filtered out. This further optimizes the tabular knowledge vector database for accurate knowledge reasoning or recommendation tasks.

[0092] Specifically, in step S3, the table data to be processed includes a table title and a description, wherein the title is information in the first row or first column of the table, and the description is detailed power-related data recorded in the table.

[0093] For example, the objective of this invention is to make recommendations based on existing text and tabular data using a large tabular model. The first row or column can be headings such as transformer model, rated power, manufacturer, and operating status, while the description refers to the detailed data recorded in the table. Specifically, let... and Let n and m represent the user set and item set, respectively, where n and m are the number of users and the number of items, respectively. The number of users refers to the total number of individual users served by the recommendation system. In the power industry, users can correspond to business entities (such as substations, maintenance teams, dispatchers, etc.). Each entity is an independent user, and its interactive behaviors (such as querying equipment maintenance plans and processing fault work orders) constitute the association between users and items. For example, n=200 means the system serves 200 substation users. The number of items refers to the total amount of interactive knowledge content in the recommendation system. In the power industry, items include equipment-related knowledge (such as maintenance plans, fault cases, and technical parameter tables) and business-related knowledge (such as dispatching strategies, analysis reports, and procurement lists). For example, m=5000 means the system stores 5000 recommendable equipment maintenance plans.

[0094] The goal of recommendation technology based on large tabular models is to understand user preferences by modeling users' past item interactions. It means that among them User The length of the interaction sequence. It is worth noting that the project... The side information mainly refers to the table's title and description. This invention considers allowing the large table model to be selected from the candidate set. Select User Most preferred items Where M is the number of candidate items, and the candidate set C typically contains one positive sample and M-1 negative samples. Specifically, for a large table model with fixed parameters... In this invention, the input-output sequence pair is denoted as ,in It is a recommended query / suggestion, which consists of a task description and the user's past items, and the output is... It is a prediction from a large table model.

[0095] Specifically, obtaining at least two second semantic information pieces of the table data to be processed includes: using a multimodal large model to extract features from the table data to be processed to obtain at least two second semantic information pieces.

[0096] For example, for queries retrieving table knowledge subgraphs, this invention utilizes the same PMLM as the first indexing step to ensure that the query and the representation of the table knowledge subgraph are in a consistent embedding space. To this end, this invention defines the items to be retrieved... Text attributes (e.g., project title and description). Defined as a query, and its second semantic information is obtained. It satisfies the following formula:

[0097] (8);

[0098] Next, from Search results for related The K most similar subgraphs satisfy the following formula:

[0099] (9);

[0100] in, Used to measure second semantic information and any knowledge subgraph in the tabular knowledge vector database The representation of Similarity metrics between them; This is a collection of subgraph sets in a tabular knowledge vector database; The operator indicates that from the set The top K elements with the highest similarity are selected as the target knowledge subgraph. .

[0101] Specifically, in step S4, all target knowledge subgraphs are aggregated to obtain a target knowledge subgraph set. .

[0102] For example, from the user Historical Interaction All the required items are retrieved, and all the corresponding target knowledge sub-graphs are aggregated to form a target knowledge sub-graph set The aggregation methods mainly include average aggregation, maximum pool aggregation and hierarchical aggregation, and the maximum pool aggregation is mainly used in the application, because the information of the neighbor nodes can be aggregated by selecting the maximum value in each dimension, and the above method can highlight the main features of the nodes.

[0103] Further, after obtaining the target knowledge sub-graph set, the method further includes: obtaining reference semantic information of the original prompt information; and sorting each target knowledge sub-graph in the target knowledge sub-graph set according to the reference semantic information to obtain an updated knowledge sub-graph set.

[0104] For example, in order to improve the recommendation performance of the table large model, the next key step of the application is to rearrange the target knowledge sub-graphs, if the user has a long list of historical interactions on the item, directly inputting all the retrieved target knowledge sub-graphs into the multi-modal large model may cause information overload. Therefore, re-ranking is performed to shorten the retrieved target knowledge sub-graphs and ensure that the most relevant target knowledge sub-graphs are displayed first in the prompt. Specifically, the original prompt information is used as a query to re-rank. To maintain consistency, the application uses the same PMLM to capture the reference semantic information of the original prompt information and re-rank the target knowledge sub-graphs in G to obtain a Top-N knowledge sub-graph set that satisfies the following formula:

[0105] (10) ;

[0106] wherein, is used to measure the similarity measure index between the reference semantic information and the representation of the target knowledge sub-graph , is an operator, indicating that the first N elements with the highest similarity are further selected from the set , and at this time = .

[0107] Specifically, in step S5, referring to Figure 3 , Figure 3 is a flowchart of the table large model output recommendation result provided by the embodiment of the application, and the target knowledge sub-graph set and the predefined prompt information are input into the table large model to make the table large model output a recommendation result, which includes:

[0108] S51. Use the second graph neural network model to process information on each target knowledge subgraph in the target knowledge subgraph set to obtain the structural information of each target knowledge subgraph.

[0109] S52. Map the structural information of the target knowledge subgraph to the embedding space of the large table model to obtain soft hint embedding information;

[0110] S53. Input the soft hint embedding information and the predefined hint information into the table large model so that the table large model outputs recommendation results.

[0111] For example, to facilitate a larger model's better understanding of the structure of the retrieved target knowledge subgraph and to avoid long contexts, this invention further integrates a second graph neural network model. ,parameter (and It is the same as other adjustable parameters, but the specific value is different. Different methods are used to enhance the representation learning of structural information, and the processing procedure satisfies the following formula:

[0112] (11);

[0113] Formula (11) is derived by reordering the knowledge subgraph set. Structural information is extracted to generate a structural representation of each target knowledge subgraph. It can capture the topological structure of entities (such as equipment and substations) and relationships (such as the relationship "the equipment is installed in the substation") in the subgraph, avoiding the loss of structural information caused by directly flattening the subgraph into text.

[0114] Through parameters Multilayer perceptron , knowledge subgraph set The structural information of all target knowledge subgraphs in the dataset is concatenated and represented as follows: ,Will Mapping this to the embedding space of the large table model yields soft-tooltip embedding information. It satisfies the following formula:

[0115] (12);

[0116] Formula (12) converts graph structure features into continuous vectors that the large table model can understand, and integrates them with text input (such as prompt templates) in the same space. The extracted subgraph embeddings... As a "soft prompt", it is directly spliced into the word segmentation representation of the input table model. The essence of the soft prompt is a continuous vector optimized by training, which is more flexible than discrete text prompts and can implicitly convey the structural information of the subgraph. The pre-defined prompt information includes system prompts, user prompts, roles, etc. The pre-defined prompt information is different from the original prompt information in subgraph reordering. The pre-defined prompt information provides a structured framework for a task, a task director, and a knowledge subgraph to form the input of the table model. The structure of the recommended task is explicitly recommended, for example, the task instruction is: recommend maintenance strategy according to device operation data. After the combination of the soft prompt and the prompt information, the composite input of "structured knowledge + task instruction" is formed, which guides the table model to generate recommendations that conform to the business logic of electric power.

[0117] Further, the model training process designed by the present application can be roughly regarded as soft prompt optimization, wherein the retrieved knowledge subgraph is a series of soft graph prompts. In form, its generation process can be represented as follows:

[0118] (13) ;

[0119] wherein, is the main parameter of the table model; is the multi-layer perception (MLP) parameter; is the input prompt information; is the true label; represents the probability of the sequence r positions are multiplied, which embodies the characteristics of sequence generation; is the result of the kth position in the predicted sequence, represents the result generated before the kth position (i.e. ).

[0120] The formula describes the probability of generating the predicted sequence when the rearranged subgraph set and the query are given. It decomposes the sequence generation process into the multiplication of the conditional probability of each position k, and the prediction of each position k depends on the previously generated result , the soft prompt embedding information and the prompt information . This embodies a sequence generation process that gradually depends on historical information, knowledge subgraphs and user needs when generating recommended results (such as electric power equipment maintenance strategy sequences). By jointly optimizing and other parameters, the probability of generating correct recommended sequences is maximized to ensure that the recommended results meet the user's needs and the logic of the table knowledge graph. Therefore, the present application only needs to learn two graph neural network parameters (i.e. and and ), while the backbone parameters of the tabular large model are frozen. Therefore, the present application updates , , and parameters through cross-entropy loss , where is the true label, and is the final prediction value of the tabular large model.

[0121] The data processing method based on the tabular large model provided by the embodiment of the present application has the following beneficial effects:

[0122] 1. The traditional method is often limited to retrieving triplets or first-order neighbor information of entities when processing tabular data in the power field, ignoring the importance of higher-order neighbor effects, resulting in a decline in recommendation performance. The present application effectively captures more extensive associated information by introducing Hop-Field technology to semantically index the knowledge sub-graphs in the tabular knowledge base. This method not only improves the accuracy of recommendations, but also discovers implicit relationship patterns in complex data networks, providing more accurate and personalized recommendation results for users. In addition, by using self-consistent thinking chain reasoning technology, the present application can adaptively determine which items need to be retrieved or enhanced, avoiding the low efficiency problem caused by indiscriminate information retrieval, and significantly improving the running efficiency of the model.

[0123] 2. Usually, tabular data in the power field is input into the large model in the form of serialized text, which cannot fully exploit the rich structural features hidden in the data. The present application designs a method for retrieving specific sub-graphs from the tabular knowledge vector database, so that the model can directly process data with rich structural information, greatly improving data utilization efficiency. At the same time, by reordering and optimizing the retrieved knowledge sub-graphs, the quality of recommendations is further enhanced. This process ensures that only the most relevant and valuable information is used to generate the final recommendation results, greatly improving the relevance of recommendations and user satisfaction. Therefore, the present application not only solves the problem of tabular data representation, but also fundamentally improves the performance and user experience of large model recommendation technology.

[0124] 3、The present application combines the advantages of advanced self-consistency thinking chain technology and table large model, realizes the great improvement of recommendation technology in accuracy and efficiency. By feeding the screened and optimized table knowledge subgraph and prompt template into the table large model, the present application can generate more accurate and context-related recommendation content. This combination not only enhances the model's understanding ability of complex data, but also makes it have stronger generalization ability, which can maintain stable performance in different application scenarios. In summary, the present application provides a comprehensive and efficient table large model recommendation technology solution, aiming to overcome the main challenges in current recommendation methods and provide new ideas and directions for future related research.

[0125] Referring to Figure 4 , Figure 4 is a structural block diagram of a data processing system 100 based on a table large model provided by an embodiment of the present application, the data processing system 100 based on the table large model comprises:

[0126] A first semantic information extraction module 11 is configured to perform feature extraction on the power-related table knowledge graph to obtain first semantic information of the table knowledge graph.

[0127] A table knowledge subgraph construction module 12 is configured to construct a table knowledge subgraph according to the first semantic information.

[0128] A table knowledge vector database optimization module 13 is configured to associate nodes in the table knowledge subgraph according to the first semantic information to optimize a table knowledge vector database.

[0129] A second semantic information extraction module 14 is configured to obtain at least two second semantic information of to-be-processed table data.

[0130] A target knowledge subgraph acquisition module 15 is configured to retrieve corresponding target knowledge subgraphs from the optimized table knowledge vector database according to the second semantic information.

[0131] A target knowledge subgraph aggregation module 16 is configured to aggregate all target knowledge subgraphs to obtain a target knowledge subgraph set.

[0132] A recommendation result acquisition module 17 is configured to input the target knowledge subgraph set and predefined prompt information into a table large model to make the table large model output a recommendation result.

[0133] Specifically, the system further comprises:

[0134] A reference semantic information extraction module is configured to obtain reference semantic information of original prompt information.

[0135] The knowledge subgraph set updating module is configured to sort each target knowledge subgraph in the target knowledge subgraph set according to the reference semantic information, and obtain an updated knowledge subgraph set.

[0136] Specifically, the table data to be processed includes a title and a description of the table, wherein the title is information in a first row or a first column of the table, and the description is detailed power-related data recorded in the table.

[0137] Specifically, the first semantic information extraction module 11 is specifically configured to: extract features of nodes and edges in the table knowledge graph by using a multi-modal large model to obtain first semantic information.

[0138] Specifically, the table knowledge subgraph construction module 12 is specifically configured to: aggregate information of each node and its adjacent nodes by using a first graph neural network model according to the first semantic information to obtain a table knowledge subgraph corresponding to each node.

[0139] Specifically, the table knowledge vector database optimization module 13 is specifically configured to: determine an association relationship of nodes in the table knowledge subgraph according to the first semantic information, and construct a graph model according to the association relationship; perform information iteration on the graph model to make a state vector of each node contain global relationship information through the information iteration; and integrate the state vectors of all nodes and the remaining nodes having the association relationship to optimize the table knowledge vector database.

[0140] Specifically, the second semantic information extraction module 14 is specifically configured to: extract features of the table data to be processed by using a multi-modal large model to obtain at least two second semantic information.

[0141] Specifically, the recommendation result acquisition module 17 is specifically configured to: perform information processing on each target knowledge subgraph in the target knowledge subgraph set by using a second graph neural network model to obtain structure information of each target knowledge subgraph; map the structure information of the target knowledge subgraph to an embedding space of a table large model to obtain soft prompt embedding information; and input the soft prompt embedding information and predefined prompt information into the table large model to make the table large model output a recommendation result.

[0142] It is worth noting that the working processes of the various modules in the table large model-based data processing system 100 described in the embodiments of the present application can refer to the working processes of the table large model-based data processing method described in the above embodiments, which will not be repeated here.

[0143] The above is the preferred embodiment of the present application, it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, can also make a number of improvements and refinements, these improvements and refinements are also considered to be within the scope of the present application.

Claims

1. A data processing method based on a large table model, characterized in that, include: Feature extraction is performed on the table-related knowledge graph to obtain the first semantic information of the table-related knowledge graph, and a table-related knowledge subgraph is constructed based on the first semantic information; The nodes in the table knowledge subgraph are associated based on the first semantic information to optimize the table knowledge vector database; Obtain at least two second semantic information pieces of the table data to be processed, and retrieve the corresponding target knowledge subgraph from the optimized table knowledge vector database based on the second semantic information; Aggregate all target knowledge subgraphs to obtain a target knowledge subgraph set; The target knowledge subgraph and predefined prompts are input into the large table model so that the large table model can output recommendation results.

2. The data processing method based on a large table model as described in claim 1, characterized in that, The table data to be processed includes a table title and a description, wherein the title is the information in the first row or first column of the table, and the description is the detailed power-related data recorded in the table.

3. The data processing method based on a large table model as described in claim 1, characterized in that, The feature extraction of the electricity-related tabular knowledge graph yields the first semantic information of the tabular knowledge graph, including: The first semantic information is obtained by extracting features from the nodes and edges in the table knowledge graph using a multimodal large model.

4. The data processing method based on a large table model as described in claim 1, characterized in that, The step of constructing a tabular knowledge subgraph based on the first semantic information includes: Based on the first semantic information, the information of each node and its neighboring nodes is aggregated using the first graph neural network model to obtain a table knowledge subgraph corresponding to each node.

5. The data processing method based on a large table model as described in claim 1, characterized in that, The step of associating nodes in the table knowledge subgraph based on the first semantic information to optimize the table knowledge vector database includes: The association relationships of nodes in the table knowledge subgraph are determined based on the first semantic information, and a graph model is constructed based on the association relationships; Information iteration is performed on the graph model so that the state vector of each node contains global relational information. Integrate the state vectors of all nodes and the remaining nodes with which they have relationships to optimize the tabular knowledge vector database.

6. The data processing method based on a large table model as described in claim 1, characterized in that, The acquisition of at least two second semantic information pieces of the table data to be processed includes: The table data to be processed is subjected to feature extraction using a multimodal large model to obtain at least two second semantic information.

7. The data processing method based on a large table model as described in claim 1, characterized in that, After obtaining the target knowledge subgraph, the method further includes: Obtain reference semantic information from the original prompt message; The target knowledge subgraph in the target knowledge subgraph set is sorted according to the reference semantic information to obtain the updated knowledge subgraph set.

8. The data processing method based on a large table model as described in claim 1, characterized in that, The step of inputting the target knowledge subgraph and predefined prompt information into the table-based large model so that the table-based large model outputs recommendation results includes: The second graph neural network model is used to process information on each target knowledge subgraph in the target knowledge subgraph set to obtain the structural information of each target knowledge subgraph. The structural information of the target knowledge subgraph is mapped to the embedding space of the large table model to obtain soft hint embedding information; The soft hint embedding information and predefined hint information are input into the table model so that the table model can output recommendation results.

9. A data processing system based on a large table model, characterized in that, include: The first semantic information extraction module is used to extract features from the electricity-related tabular knowledge graph to obtain the first semantic information of the tabular knowledge graph. A table knowledge subgraph construction module is used to construct a table knowledge subgraph based on the first semantic information; The table knowledge vector database optimization module is used to associate nodes in the table knowledge subgraph based on the first semantic information in order to optimize the table knowledge vector database. The second semantic information extraction module is used to obtain at least two second semantic information pieces of the table data to be processed. The target knowledge subgraph acquisition module is used to retrieve the corresponding target knowledge subgraph from the optimized tabular knowledge vector database based on the second semantic information. The target knowledge subgraph aggregation module is used to aggregate all target knowledge subgraphs to obtain a target knowledge subgraph set; The recommendation result acquisition module is used to input the target knowledge subgraph and predefined prompt information into the table model so that the table model can output recommendation results.

10. The data processing system based on a large table model as described in claim 9, characterized in that, The system also includes: The reference semantic information extraction module is used to obtain the reference semantic information of the original prompt information; The knowledge subgraph update module is used to sort each target knowledge subgraph in the target knowledge subgraph set according to the reference semantic information to obtain the updated knowledge subgraph set.

Images