CN-121998159-A - Large-model-based frequency track declaration coordination question-answering system and method

CN121998159ACN 121998159 ACN121998159 ACN 121998159ACN-121998159-A

Abstract

The application relates to the field of satellite frequency orbit resources, in particular to a frequency orbit declaration coordination question-answering system and method based on a large model. The system comprises a data processing module, a model training optimization module, a model fine tuning deployment module and a frequency track reporting and answering coordination function, wherein the data processing module is used for acquiring original data and preprocessing to construct a pre-training data set and an answer-to-question data set, the model training optimization module is used for taking the pre-training data set as the original data set, taking the answer-to-question data set as a target data set, generating a training data set through topic distribution modeling and text importance evaluation and combining a sampling screening strategy, optimizing a training process of a large model based on data screening and dynamic batch updating strategies of training loss, further completing the pre-training of the large model, and the model fine tuning deployment module is used for fine tuning the pre-trained large model by adopting the answer-to-question data set and deploying the fine-tuned large model to realize the frequency track reporting and answering-question function. The application can improve the efficiency and accuracy of the reporting coordination of the frequency track.

Inventors

ZHONG LEI
CHEN ZHIMIN
YAO XIUJUAN
GAO XIANG
LI ZHEN
WANG JING
LI XUE

Assignees

中国科学院国家空间科学中心

Dates

Publication Date: 20260508
Application Date: 20251217

Claims (9)

1. A frequency track declaration coordination question-answering system based on a large model is characterized by comprising: the data processing module is used for acquiring original data and preprocessing the original data to construct a pre-training data set and a question-answer data set; The model training optimization module is used for taking the pre-training data set as an original data set, taking the question-answer data set as a target data set, generating a training data set by combining topic distribution modeling and text importance evaluation and a sampling screening strategy, optimizing the training process of the large model based on the data screening and dynamic batch updating strategy of training loss so as to finish the pre-training of the large model, and The model fine tuning deployment module is used for fine tuning the pre-trained large model by adopting the question-answer pair data set and deploying the fine-tuned large model to realize the function of reporting, coordination and question-answer of the frequency track.
2. The large model based frequency track declaration coordination question-answering system according to claim 1, wherein the raw data includes a radio rule series file.
3. The large model-based frequency track declaration coordination question-answering system according to claim 1, wherein the data processing module includes a data acquisition and format conversion unit, a data structuring processing unit, a data cleaning optimizing unit, and a question-answering pair generating unit, wherein, The data acquisition and format conversion unit adopts text data and table data for extracting the original data and stores the extracted text data and table data; The data structuring processing unit is used for respectively converting the text data and the form data of the original data into independent data; The data cleaning and optimizing unit is used for cleaning the converted data through the steps of rule filtering, data deduplication and text division to obtain the pre-training data set; The question-answer pair generating unit is used for extracting corresponding answers from the pre-training data set through an automatic script based on the frequency track declaration coordination high-frequency query question types to generate the question-answer data set.
4. The big model-based frequency track declaration coordination question-answering system according to claim 3, wherein the data structuring processing unit is used for dividing text data based on paragraphs, converting each piece of text data into independent data respectively, splicing header information of a row content table in table data with specific contents of a row to which the header information belongs, further converting the row content table into independent data, splicing header information of a column content table in the table data with contents of a column to which the header information belongs, and further converting the column content table into independent data.
5. The large model based frequency track declaration coordination question answering system according to claim 3, wherein the frequency track declaration coordination question types include business consultation, limit value query and footnote retrieval.
6. The large model based frequency track declaration coordination question-answering system according to claim 1, wherein the model training optimization module includes a data selection sub-module and a training optimization sub-module, wherein, A data selection sub-module for taking the pre-training data set as the original data set The question-answer pair data set is taken as a target data set Through topic distribution modeling and text importance evaluation, a topic model is adopted to perform initial data set With the target data set Subject mining is carried out to respectively obtain original data sets And a target data set For computing the original data set Similarity of the medium text t and subject distribution of the target data set as the original data set Importance weight of middle text t For calculating importance scores in combination with noise factors Selecting a preset number of samples according to the descending order of the scores to form the training data set; the training optimization sub-module is used for splitting the training data set into a plurality of batches, screening high-value data based on training loss of data of each batch, dynamically updating the training batches, and improving the convergence speed of the large model.
7. The large model based frequency track declaration coordination question-answering system according to claim 6, wherein, The topic model is an LDA topic model, and the topic mining adopts a Gibbs sampling algorithm; The original data set Importance weight of middle text t The method comprises the following steps: ; Wherein, the Representing the original dataset The topic distribution of the medium text t, For the number of topics in the topic model, Representative text t belonging to the first Probability of individual topics; Representing a target dataset Subject distribution of (a); The importance score Calculation based on Gumbel sampling strategy: ; Wherein, the Is standard gummel noise.
8. The large model based frequency track declaration coordination question-answering system according to claim 6, wherein the training optimization sub-module is configured to, for each batch of data Forward propagation to obtain data Training loss of (2) Based on training loss Calculation data Selection probability of (a) According to the probability of selection Screening the high-value data, wherein, ; Wherein, the Is the number of losses to be incurred, Indicating the latest A cumulative distribution function of the loss components of the individual training data; Is a selectivity factor; And when the screened high-value data reaches the preset batch size, carrying out back propagation and parameter updating on the batch data.
9. A big model-based frequency track declaration coordination question-answering method, implemented based on the big model-based frequency track declaration coordination question-answering system of any one of claims 1-8, comprising: step1, acquiring original data through a data processing module and preprocessing the original data to construct a pre-training data set and a question-answer data set; Step 2, taking the pre-training data set as an original data set, taking the question-answer data set as a target data set, carrying out topic distribution modeling and text importance evaluation through a model training optimization module, and generating a training data set by combining a sampling screening strategy; and 3, the model fine tuning deployment module adopts question and answer pair data sets to carry out fine tuning on the pre-trained large model, and deploys the fine-tuned large model to realize a frequency track declaration coordination question and answer function.

Description

Large-model-based frequency track declaration coordination question-answering system and method Technical Field The application relates to the field of satellite frequency orbit resources, in particular to a frequency orbit declaration coordination question-answering system and method based on a large model. Background The artificial earth satellite refers to an unmanned spacecraft manufactured by human beings and launched onto the earth orbit, and radio frequency and satellite orbit required by satellite launching are non-renewable natural resources, so that the artificial earth satellite has great economic and use values. In recent years, the system has the advantages that the system is in a state of accelerating the preemption of the frequency track resources with limited total amount, the satellite constellation in China lacks the frequency track resources with interception, high priority and convenient use, and is in a disadvantageous position in the international rule formulation and cooperation, and the system is in urgent need of actively reporting and reserving the frequency track resources. The existing method mainly accelerates the reporting coordination flow from an auxiliary angle, and each application needs to rely on the traditional information retrieval mode, so that the efficiency is low, and the accuracy cannot be guaranteed. The question-answering system based on the large model is used as an emerging artificial intelligence technology, can understand the information requirement of users, which is put forward by using natural language, and is an effective method for solving the problems. High quality questioning and answering data is required for large models in specific fields, but similar research work is lacking in the satellite communication field at present. In addition, the training of the large model requires a large amount of operations on massive data for a long time, the existing large model training optimization algorithm only focuses on screening training data or optimizing the training process, and only optimizes the training data from a single direction, so that the research of a method capable of efficiently training the large model is also very important. Disclosure of Invention The application aims to overcome the defects of the prior art, and provides a frequency track reporting, coordination and question-answering system and method based on a large model. In order to solve the technical problems, the frequency track declaration coordination question-answering system based on the large model provided by the technical scheme of the application comprises the following components: the data processing module is used for acquiring original data and preprocessing the original data to construct a pre-training data set and a question-answer data set; The model training optimization module is used for taking the pre-training data set as an original data set, taking the question-answer data set as a target data set, generating a training data set by combining topic distribution modeling and text importance evaluation and a sampling screening strategy, optimizing the training process of the large model based on the data screening and dynamic batch updating strategy of training loss so as to finish the pre-training of the large model, and The model fine tuning deployment module is used for fine tuning the pre-trained large model by adopting the question-answer pair data set and deploying the fine-tuned large model to realize the function of reporting, coordination and question-answer of the frequency track. As an improvement of the above system, the raw data includes a radio rule series file. The system comprises a data acquisition and format conversion unit, a data structuring processing unit, a data cleaning and optimizing unit and a question-answer pair generating unit, wherein the data acquisition and format conversion unit adopts text data and form data for extracting original data and stores the text data and the form data of the original data, the data structuring processing unit is used for respectively converting the text data and the form data of the original data into independent data, the data cleaning and optimizing unit is used for cleaning the converted data through rule filtering, data deduplication and text dividing steps to obtain a pre-training data set, and the question-answer pair generating unit is used for extracting corresponding answers from a pre-training data set through an automatic script based on a frequency track reporting cooperative high-frequency query problem type to generate the question-answer pair data set. As an improvement of the system, the data structuring processing unit is used for dividing text data based on paragraphs, converting each segment of text data into independent data respectively, splicing the header information of a row of content tables in table data with the specific content of the row to which the header information belongs