CN-121996322-A - Task branching control method and system for large language model pre-analysis preprocessing drive

CN121996322ACN 121996322 ACN121996322 ACN 121996322ACN-121996322-A

Abstract

The application relates to the technical field of electric digital data processing and discloses a task branching control method and a task branching control system driven by large language model pre-analysis pretreatment, wherein the method comprises the steps of analyzing a document set to be processed, generating semantic data blocks, and distributing index identifiers with hierarchical relations for each semantic data block; the method comprises the steps of adopting a pre-constructed feature extraction model to perform feature extraction on semantic data blocks to obtain a multi-dimensional feature set, constructing a structured index table, dynamically searching a candidate data block list from the structured index table according to a received user query task, generating a data block loading strategy according to the multi-dimensional feature set associated with each candidate data block in the candidate data block list, and carrying out branching control on the selected quantized large language model to process the candidate data blocks according to the data block loading strategy to obtain a task processing result. The method provided by the application improves the processing capacity of the quantized large language model on complex and long document tasks and reduces the calculation power consumption.

Inventors

XIE XIN
ZHANG SHIZHE
LV YUNYI
ZHAO ZITONG
CHEN HENG
Wu Gangchao
ZHANG QI
LI KANGPING

Assignees

华景乐游(深圳)智慧科技有限公司
郑州曙光云科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260117

Claims (10)

1. A task branching control method driven by pre-analysis pretreatment of a large language model is characterized by comprising the following steps: Analyzing a document set to be processed by adopting a multi-granularity self-adaptive semantic segmentation algorithm, generating semantic data blocks, and distributing index identifiers with hierarchical relations for each semantic data block; Carrying out feature extraction on each semantic data block by adopting a pre-constructed feature extraction model to obtain a multi-dimensional feature set; constructing a structured index table according to the semantic data blocks, the index identifiers and the multi-dimensional feature set, and dynamically searching from the structured index table according to a received user query task to obtain a candidate data block list; generating a data block loading strategy according to the multi-dimensional feature set associated with each candidate data block in the candidate data block list, wherein the data block loading strategy is set to be determined based on task adaptation degree and topic correlation degree between the multi-dimensional feature set and the user query task and the hierarchical relation between the index identifiers; And according to the data block loading strategy, the branching control is carried out on the selected quantized large language model to process the candidate data blocks, an intermediate processing result corresponding to each candidate data block is generated, and the intermediate processing results are integrated according to the corresponding index identifiers to obtain a task processing result.
2. The task branching control method driven by large language model pre-analysis preprocessing as claimed in claim 1, wherein the analyzing the document set to be processed by adopting the multi-granularity self-adaptive semantic segmentation algorithm to generate the semantic data block comprises the following steps: Carrying out multi-granularity scanning on the documents to be processed in the set of the documents to be processed, and identifying candidate segmentation points under different granularity levels to obtain a candidate segmentation point sequence; Calculating the topic consistency score of text units at two sides of each candidate segmentation point, selecting an optimal segmentation point sequence from the candidate segmentation point sequences according to the topic consistency score and a preset self-adaptive threshold, and dynamically calculating the self-adaptive threshold according to at least one of document type, text length and domain keyword density; And executing semantic segmentation on the document to be processed based on the optimal segmentation point sequence, and constructing hierarchical relations among segmentation blocks according to granularity nesting relations in the semantic segmentation process to obtain semantic data blocks with hierarchical relations.
3. The large language model pre-analysis preprocessing driven task branching control method as claimed in claim 2, wherein the dynamic calculation method comprises: Fitting the document type, the text length, the domain keyword density and the topic consistency scores of the segmentation points with different granularity of the manually marked segmented document to obtain a self-adaptive threshold calculation formula; And according to the document type, the paragraph length and the domain keyword density corresponding to the document to be processed, adopting the self-adaptive threshold calculation formula to obtain the self-adaptive threshold corresponding to the document to be processed.
4. The task branching control method driven by pre-analysis preprocessing of a large language model according to claim 1, wherein the feature extraction is performed on each semantic data block by adopting a pre-built feature extraction model to obtain a multi-dimensional feature set, and the method comprises the following steps: Processing each semantic data block by adopting a pre-constructed feature extraction model to obtain a theme vector, a task adaptation type label and a cross-block association degree matrix of the semantic data block, wherein the feature extraction model comprises a semantic encoder and a multi-task output head; adopting a pre-constructed risk assessment engine to perform data conflict detection on the topic vector to obtain the data conflict risk level of the semantic data block; And constructing a multi-dimensional feature set of the semantic data block according to the topic vector, the task adaptation type label, the cross-block association degree matrix and the data conflict risk level.
5. The method for task branching control driven by preanalysis preprocessing of large language model as in claim 4, wherein said structured index table is a hierarchical cascade architecture, at least comprising: the task adaptation sub-index layer is constructed based on the index identifiers corresponding to the semantic data blocks and the task adaptation type labels and is used for realizing rapid screening based on task types; the semantic adaptation sub-index layer is constructed based on the index identifiers corresponding to the semantic data blocks and the topic vectors and is used for realizing retrieval based on topic similarity; and a context association sub-index layer which is constructed based on the index identifiers corresponding to the semantic data blocks and the cross-block association degree matrix and is used for supporting the retrieval of association contexts.
6. The method for task branching control driven by large language model pre-analysis preprocessing as set forth in claim 5, wherein said dynamically retrieving from said structured index table based on a received user query task, obtaining a candidate data block list, comprises: analyzing the received user query task to obtain the query task type, the semantic query vector and the context association requirement; Searching at the task adaptation sub-index layer based on the query task type to obtain a first candidate set; Based on the semantic query vector, performing similarity search on the first candidate set in the semantic adaptation sub-index layer to obtain a second candidate set; Performing association expansion on the elements in the second candidate set in the context association sub-index layer based on the context association requirement to obtain a third candidate set; and obtaining a candidate data block list according to the semantic data blocks associated with the second candidate set and the semantic data blocks associated with the third candidate set.
7. The large language model pre-analysis preprocessing driven task branching control method as set forth in claim 6, wherein said generating a data block loading strategy according to said multi-dimensional feature set associated with each candidate data block in said candidate data block list comprises: Obtaining task adaptation degree and topic correlation degree between each multi-dimensional feature set and the user query task according to the multi-dimensional feature set associated with each candidate data block in the candidate data block list; obtaining a comprehensive priority score according to the task adaptation degree, the topic relevance degree and the risk penalty coefficient; And obtaining an initial loading queue according to the hierarchical relation between each index identifier, and carrying out queue adjustment on the candidate data blocks of the same hierarchy of the initial loading queue according to the comprehensive priority score to obtain a data block loading strategy.
8. The method for task branching control of a large language model pre-analysis preprocessing driver according to claim 7, wherein said branching control of processing of said candidate data blocks by a selected quantized large language model according to said data block loading strategy, generating intermediate processing results corresponding to each of said candidate data blocks, comprises: Setting a fixed capacity of the sliding window controller according to the context window length limit of the selected quantized large language model; Filling the content of the corresponding candidate data block into the current sliding window of the sliding window controller according to the data block loading strategy, and performing data block length check in the content filling process to ensure that the total length of the current sliding window is smaller than or equal to the fixed capacity; Performing secondary semantic splitting on single candidate data blocks with lengths exceeding the fixed capacity, and distributing extension index identifiers for the split sub-semantic data blocks; Inputting the content of the current sliding window into the quantized large language model to generate an intermediate processing result associated with the index identifier or the extended index identifier.
9. The method for task branching control driven by large language model pre-analysis preprocessing as set forth in claim 4, wherein said integrating said intermediate processing results according to the corresponding index identifiers to obtain task processing results includes: integrating the intermediate processing results according to the hierarchical relationship of the index identifiers or the extension index identifiers associated with the intermediate processing results to obtain initial task processing results; And carrying out semantic conflict correction on the initial task processing result according to the data conflict risk level of the semantic data block associated with the intermediate processing result to obtain a task processing result.
10. The task branching control system for the large language model pre-analysis pretreatment drive is used for realizing the task branching control method for the large language model pre-analysis pretreatment drive according to any one of claims 1-9, and is characterized by comprising a document analysis module, a feature extraction module, a data block retrieval module, a loading strategy generation module and a task processing module; The document analysis module is used for analyzing a document set to be processed by adopting a multi-granularity self-adaptive semantic segmentation algorithm, generating semantic data blocks, and distributing index identifiers with hierarchical relations for each semantic data block; The feature extraction module is used for carrying out feature extraction on each semantic data block by adopting a pre-constructed feature extraction model to obtain a multi-dimensional feature set; the data block retrieval module is used for constructing a structured index table according to the semantic data block, the index identifier and the multi-dimensional feature set, and dynamically retrieving from the structured index table according to a received user query task to obtain a candidate data block list; The loading strategy generation module is used for generating a data block loading strategy according to the multi-dimensional feature set associated with each candidate data block in the candidate data block list, wherein the data block loading strategy is set to be determined based on task adaptation degree and topic correlation degree between the multi-dimensional feature set and the user query task and the hierarchical relation between the index identifiers; and the task processing module is used for controlling the processing of the selected quantized large language model on the candidate data blocks by branching according to the data block loading strategy, generating an intermediate processing result corresponding to each candidate data block, and integrating the intermediate processing results according to the corresponding index identifiers to obtain a task processing result.

Description

Task branching control method and system for large language model pre-analysis preprocessing drive Technical Field The invention relates to the technical field of electric digital data processing, in particular to a task branching control method and system driven by large language model pre-analysis preprocessing. Background With the breakthrough development of large language model technology, the method is widely applied to the fields of document understanding, information retrieval, intelligent question and answer and the like. However, the operation of the large language model depends on the support of high-performance computing hardware, the computing card of an enterprise cannot meet the requirements of the large language model, and the prior art reduces the occupation of the storage space and the consumption of computing resources of the model on the premise of ensuring that the core capability of the model is basically not damaged by performing weight quantization, pruning, knowledge distillation and other operations on the original large language model, so that the large language model can be deployed under the support of a small number of computing cards. However, when the quantized large language model is applied to the processing of unstructured documents with huge mass and complex structures of enterprises, a series of severe technical bottlenecks are faced, wherein the complex task processing capacity is insufficient due to model performance loss, the context window of the quantized large language model is limited and cannot adapt to a long text office scene, and the multi-task concurrent processing capacity is poor due to insufficient computing power. It can be seen that how to improve the processing capability of the quantized large language model on complex tasks and long document tasks and how to reduce the computational effort consumption has become a technical problem to be solved by those skilled in the art. Disclosure of Invention The invention provides a task branching control method and a task branching control system for a large language model pre-analysis preprocessing driver, which are used for solving the technical problems of improving the processing capacity of a quantized large language model on complex tasks and long-document tasks and reducing the calculation power consumption. In a first aspect, the present invention provides a task branching control method driven by pre-analysis and preprocessing of a large language model, where the method includes: Analyzing a document set to be processed by adopting a multi-granularity self-adaptive semantic segmentation algorithm, generating semantic data blocks, and distributing index identifiers with hierarchical relations for each semantic data block; Carrying out feature extraction on each semantic data block by adopting a pre-constructed feature extraction model to obtain a multi-dimensional feature set; constructing a structured index table according to the semantic data blocks, the index identifiers and the multi-dimensional feature set, and dynamically searching from the structured index table according to a received user query task to obtain a candidate data block list; generating a data block loading strategy according to the multi-dimensional feature set associated with each candidate data block in the candidate data block list, wherein the data block loading strategy is set to be determined based on task adaptation degree and topic correlation degree between the multi-dimensional feature set and the user query task and the hierarchical relation between the index identifiers; And according to the data block loading strategy, the branching control is carried out on the selected quantized large language model to process the candidate data blocks, an intermediate processing result corresponding to each candidate data block is generated, and the intermediate processing results are integrated according to the corresponding index identifiers to obtain a task processing result. Preferably, the analyzing the document set to be processed by adopting the multi-granularity adaptive semantic segmentation algorithm to generate the semantic data block includes: Carrying out multi-granularity scanning on the documents to be processed in the set of the documents to be processed, and identifying candidate segmentation points under different granularity levels to obtain a candidate segmentation point sequence; Calculating the topic consistency score of text units at two sides of each candidate segmentation point, selecting an optimal segmentation point sequence from the candidate segmentation point sequences according to the topic consistency score and a preset self-adaptive threshold, and dynamically calculating the self-adaptive threshold according to at least one of document type, text length and domain keyword density; And executing semantic segmentation on the document to be processed based on the optimal segmentation point sequence, and c