CN-121980069-A - Large-model-based data visualization narrative method and system

CN121980069ACN 121980069 ACN121980069 ACN 121980069ACN-121980069-A

Abstract

The application relates to the technical field of large model application, and discloses a data visualization narrative method and system based on a large model. The method comprises the steps of obtaining structured original data in an ERP system, a CRM system and an Internet of things (IoT) device, generating a unified event set, executing synchronous identification of cross-source events, constructing a heterogeneous event diagram, outputting a causal chain set based on a graph annotation path decoding model, and outputting a visual narrative template and a display instruction set. Compared with the prior art which mainly relies on a single-mode data stream or static business rule to carry out causal modeling, the method and the system have the advantages that under the conditions that enterprises exist in multi-business system coordination and semantic description is inconsistent, the technical problems of high-reliability causal chain reconstruction and structured narrative output cannot be achieved.

Inventors

DING HUAFENG
ZHANG JUN
DUAN JIA
ZHANG BO

Assignees

深圳思特顺科技有限公司

Dates

Publication Date: 20260505
Application Date: 20251205

Claims (10)

1. A method of data visualization narrative based on a large model, the method comprising: Step 10, three types of structured original data are obtained, wherein the three types of structured original data comprise ERP structured original data, CRM structured original data and IoT structured original data; Step S20, executing a cross-source event synchronous identification task by adopting a weighted time consistency driven mode collaborative mapping mechanism based on a unified event set E, and outputting a synchronous mapping table Stab; step S30, executing a cross-source event heterogeneous graph construction task by adopting a hierarchical causal driven direction perception edge generation mechanism based on a synchronous mapping table Stab, and outputting a heterogeneous event graph G; step S40, inputting the heterogeneous event graph G into a preset graph meaning path decoding model, and outputting a causal chain set C by the graph meaning path decoding model; And S50, executing a structured narrative chain generation task by adopting a hierarchical fusion mechanism based on time causal dependency and semantic causal based on a causal chain set, and outputting a visual narrative template and a display instruction set.
2. The method of claim 1, wherein in step S10, three types of structured raw data are obtained, the three types of structured raw data comprise ERP structured raw data, CRM structured raw data and IoT structured raw data, the step of performing a time sequence and semantic fusion task by using a window-driven cross-modal feature alignment mechanism based on the three types of structured raw data, and the step of generating a unified event set E specifically comprises: Step S101, presetting a uniform time granularity delta t, and constructing a sliding time window sequence W according to the uniform time granularity delta t, wherein ERP structured original data is extracted from an ERP system based on the sliding time window sequence W, the ERP structured original data comprises yield log records, shutdown times log records and inventory change log records, CRM structured original data is extracted from a CRM system based on the sliding time window sequence W, the CRM structured original data comprises customer order data, service response data and complaint record data, and the IoT structured original data is extracted from an IoT system based on the sliding time window sequence W, and comprises original sensor sampling data; step S102, carrying out window compression processing on ERP structured original data by adopting a multi-index sliding difference method to generate a first modal event vector group Performing emotion label extraction and semantic embedding processing on the CRM structured original data by adopting python pandas library and transformers library to generate a second modal event vector group Window compression processing is carried out on the IoT structured raw data by adopting a multi-polar encoding method, and a third modal event vector set is generated ; Step S103, based on the first modal event vector group Second modal event vector set And a third modal event vector group And carrying out semantic fusion processing by adopting a cross-modal nested mapping mechanism, and outputting a unified event set E.
3. The large model based data visualization narrative method as set forth in claim 2, wherein in step S103, based on the first modal event vector set Second modal event vector set And a third modal event vector group The method comprises the steps of carrying out semantic fusion processing by adopting a cross-modal nested mapping mechanism and outputting a unified event set E, and specifically comprises the following steps: Step S1031, based on the first modal event vector group And a second modal event vector group Calculating a first similarity index by adopting a cosine similarity mode, and obtaining a vector group based on a first modal event And third modal event vector group Calculating a second similarity index by adopting a standardized Euclidean distance reciprocal mode, and obtaining a second modal event vector group And a third modal event vector group Calculating a third similarity index by using a KL divergence approximation method, and respectively constructing a first mode alignment scoring vector based on the first similarity index, the second similarity index and the third similarity index Second modality alignment scoring vector And a third modality alignment scoring vector ; Step S1032, introducing a nested attention network, aligning scoring vectors with a first modality Alignment of scoring vectors in a second modality as the dominant Query of a nested attention network And a third modality alignment scoring vector The nested attention network carries out weighting treatment on the KeyValue by adopting a modal residual error weighting fusion method according to the dominant Query and the key value, and outputs a modal fusion vector sequence S; Step S1032, a unified event set E is constructed and output based on the modal fusion vector sequence S and the unified time granularity delta t.
4. The method for visualizing a narrative based on data in a large model as set forth in claim 1, wherein in step S20, a cross-source event synchronization identification task is executed based on a unified event set E by using a weighted time consistency driven modality collaborative mapping mechanism, and the step of outputting a synchronization mapping table Stab specifically includes: Step S201, extracting the ith event node from the unified event set E And the jth event node Computing event nodes And event node The three-dimensional degree Score (i, j) comprises a time consistency Score, a main mode weight Score and a semantic vector cosine similarity Score, wherein the time consistency Score is used for indicating whether a time stamp difference value is in a cross-system time drift tolerance range; Step S202, when an event node Derived from ERP structured raw data, and the three-dimensional Score (i, j) is higher than a preset modal adaptive fusion threshold When marking event nodes Event synchronization pairs for "ERP system and IoT system"; when an event node Derived from ERP structured raw data, and the three-dimensional Score (i, j) is higher than a preset cross-department business association threshold When marking event nodes Event synchronization pairs for an ERP system and a CRM system; step S203, a synchronous mapping table Stab is constructed based on all event synchronous pairs in the unified event set E by adopting a weighted bipartite graph matching principle.
5. The method for visualizing a narrative on the basis of data in a large model as set forth in claim 4, wherein in step S30, a cross-source event heterogeneous graph construction task is performed by using a hierarchical causal driven direction-aware edge generation mechanism based on a synchronization map Stab, and a heterogeneous event graph G is output, which specifically includes: Step S301, carrying out semantic label analysis and clustering on event synchronization pairs by adopting an event semantic nested label clustering method based on a synchronization mapping table Stab, and outputting event semantic subgraph clustering ; Step S302, constructing a direction perception weight for an event synchronization pair by adopting a multi-factor direction perception weighting method based on a synchronization mapping table Stab; Step S303, clustering according to event semantics And constructing a heterogeneous event graph G by adopting an edge density constraint driving composition method by the direction perception weight.
6. The method for visualizing a narrative based on data on a large model as set forth in claim 1, wherein in step S40, the step of inputting the heterogeneous event graph G into a preset annotation path decoding model, the annotation path decoding model outputting a causal link set C, specifically comprises: Step S401, presetting a graph attention path decoding model, wherein the graph attention path decoding model comprises an input layer, a causal edge attention aggregation layer, a path decoding layer, a causal chain evaluation layer, an output layer and a causal chain set, wherein the input layer is used for receiving a heterogeneous event graph G; Step S402, acquiring a historical heterogeneous event diagram and a corresponding causal chain template set, taking the historical heterogeneous event diagram as input of a graph annotation path decoding model, taking the corresponding causal chain template set as output of the graph annotation path decoding model, and executing pre-training on the graph annotation path decoding model by combining Yolov confidence cross entropy loss functions; Step S403, inputting the heterogeneous event graph G into a pre-trained graph annotation path decoding model, and outputting a causal chain set C.
7. The method of claim 1, wherein in step S50, a causal chain set is used to perform a structured narrative chain generation task based on a temporal causal dependency and semantic causal hierarchical fusion mechanism, and the step of outputting a visual narrative template and a presentation instruction set specifically comprises: step S501, extracting a plurality of causal path sequence sets with consistent time by adopting a depth-first traversal mode based on occurrence time stamps of all events in a causal link set C; Step S502, building a four-tuple representation structure of an event name, event time, event reason and event influence structure aiming at each causal path sequence in a causal path sequence set, and extracting by adopting a causal attention mechanism according to the four-tuple representation structure to obtain a structural semantic embedded sequence; step S503, inputting a structured semantic embedding sequence into a preset large narrative generation model, and outputting a visual narrative template and a display instruction set by the large narrative generation model, wherein the large narrative generation model is obtained by fine adjustment of a general purpose model generated based on GPT languages by adopting a parameter efficient LoRA optimization mode guided by a causal role prompt template.
8. A large model based data visualization narrative system, for use in a large model based data visualization narrative method of any one of claims 1 to 7, comprising: The data fusion acquisition module is used for acquiring three types of structured raw data, wherein the three types of structured raw data comprise ERP structured raw data, CRM structured raw data and IoT structured raw data; performing time sequence and semantic fusion tasks by adopting a window-driven cross-modal feature alignment mechanism based on three types of structured raw data to generate a unified event set E; the mode collaborative recognition module is used for executing a cross-source event synchronous recognition task by adopting a mode collaborative mapping mechanism driven by weighted time consistency based on the unified event set E and outputting a synchronous mapping table Stab; the heterogram construction module is used for executing a cross-source event heterogram construction task by adopting a hierarchical causal driven direction perception edge generation mechanism based on the synchronous mapping table Stab and outputting a heterogram G; The causal chain decoding module is used for inputting the heterogeneous event graph G into a preset graph meaning path decoding model, and outputting a causal chain set C by the graph meaning path decoding model; And the narrative template generation module is used for executing a structured narrative chain generation task based on a causal chain set by adopting a causal dependency and semantic causal hierarchical fusion mechanism and outputting a visual narrative template and a display instruction set.
9. A large model based data visualization narrative apparatus, comprising a memory, a processor, and a large model based data visualization narrative program stored on the memory and executable on the processor, the large model based data visualization narrative program when executed by the processor implementing a large model based data visualization narrative method of any one of claims 1 to 7.
10. A computer program product comprising a large model based data visualization narrative program, which when executed by a processor implements a large model based data visualization narrative method of any one of claims 1 to 7.

Description

Large-model-based data visualization narrative method and system Technical Field The invention relates to the technical field of large model application, in particular to a large model-based data visualization narrative method and system. Background Currently, with the continuous acceleration of enterprise digital transformation, a large amount of structured business data is recorded in heterogeneous platforms distributed in different systems, including Enterprise Resource Planning (ERP) systems, customer Relationship Management (CRM) systems, and internet of things (IoT) systems. The ERP system often comprises flow events such as order processing, inventory management, financial data and the like, the CRM system records event information such as customer interaction records, sales behaviors, service feedback and the like, and the IoT system acquires monitoring data such as equipment running states, sensor alarms, maintenance logs and the like in real time. However, due to heterogeneous sources, non-uniform semantic tags, obvious time stamp format differences and inconsistent event granularity spans of the data, in the actual enterprise operation insight, root cause analysis and strategy backtracking process, a complete time closed loop and a complete semantic closed loop are difficult to construct, and explanatory visual narrative content facing a manager is formed. Existing data visualization tools (such as BI systems or simple chart plugins) focus on the graphical presentation of statistical indicators, lacking automatic event extraction, causal chain analysis and narrative mainline generation capabilities. In some applications, although a process mining or event sequence analysis tool is introduced to perform auxiliary analysis, the modeling basis of the process mining or event sequence analysis tool often depends on static rules, preset maps or shallow event matching logic, and is difficult to deal with the actual business scenes of cross-modal, weak labels and multi-hop causality. For example, the system may not automatically identify whether a customer churn is associated with a cross-domain event such as an equipment anomaly alert, inventory shortage, order delay, etc., nor generate a report or graphic mixing narrative content with logical inference chains and language narrative capabilities. Therefore, a technical method for automatically identifying a causal chain and visually generating a structured narrative under the conditions of complex cross-system data sources and inconsistent event granularity is still needed, so as to improve the intelligentization level of data-driven business understanding and auxiliary decision making. Disclosure of Invention Aiming at the technical defects, the invention aims to provide a data visualization narrative method based on a large model, and aims to solve the technical problems that in the prior art, causal modeling is mainly carried out by relying on single-mode data flow or static business rules, and especially under the condition that enterprises have multi-business system coordination and inconsistent semantic description, causal chain reconstruction and structured narrative output with high credibility cannot be realized. In order to solve the technical problems, the invention adopts the following technical proposal that the invention provides a data visualization narrative method based on a large model, The large model-based data visualization narrative method comprises the following steps: Step 10, three types of structured original data are obtained, wherein the three types of structured original data comprise ERP structured original data, CRM structured original data and IoT structured original data; Step S20, executing a cross-source event synchronous identification task by adopting a weighted time consistency driven mode collaborative mapping mechanism based on a unified event set E, and outputting a synchronous mapping table Stab; step S30, executing a cross-source event heterogeneous graph construction task by adopting a hierarchical causal driven direction perception edge generation mechanism based on a synchronous mapping table Stab, and outputting a heterogeneous event graph G; step S40, inputting the heterogeneous event graph G into a preset graph meaning path decoding model, and outputting a causal chain set C by the graph meaning path decoding model; And S50, executing a structured narrative chain generation task by adopting a hierarchical fusion mechanism based on time causal dependency and semantic causal based on a causal chain set, and outputting a visual narrative template and a display instruction set. Preferably, in step S10, three types of structured raw data are obtained, wherein the three types of structured raw data comprise ERP structured raw data, CRM structured raw data and IoT structured raw data, and the step of generating a unified event set E is specifically included: Step S101, presetting a uniform time granularity delta